CSI-HC: A WiFi-Based Indoor Complex Human Motion

Research ArticleCSI-HC A WiFi-Based Indoor Complex Human MotionRecognition Method

Zhanjun Hao 12 Yu Duan 1 Xiaochao Dang 12 and Tong Zhang1

1College of Computer Science and Engineering Northwest Normal University Lanzhou Gansu 730070 China2Gansu Province Internet of ings Engineering Research Center Northwest Normal University Lanzhou Gansu 730070 China

Correspondence should be addressed to Zhanjun Hao zhanjunhao126com

Received 28 May 2019 Revised 9 December 2019 Accepted 22 January 2020 Published 26 February 2020

Academic Editor Filippo Gandino

Copyright copy 2020 Zhanjun Hao et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

WiFi indoor personnel behavior recognition has become the core technology of wireless network perception However theexisting human behavior recognition methods have great challenges in terms of detection accuracy intrusion and complexity ofoperations In this paper we firstly analyze and summarize the existing human motion recognition schemes and due to theexistence of the problems in them we propose a noninvasive highly robust complex human motion recognition scheme based onChannel State Information (CSI) that is CSI-HC and the traditional Chinese martial art XingYiQuan is verified as a complexmotion background CSI-HC is divided into two phases offline and online In the offline phase the human motion data arecollected on the commercial Atheros NIC and a powerful denoising method is constructed by using the Butterworth low-passfilter and wavelet function to filter the outliers in the motion data en through Restricted Boltzmann Machine (RBM) trainingand classification we establish offline fingerprint information In the online phase SoftMax regression is used to correct the RBMclassification to process the motion data collected in real time and the processed real-time data are matched with the offlinefingerprint information On this basis the recognition of a complex human motion is realized Finally through repeated ex-periments in three classical indoor scenes the parameter setting and user diversity affecting the accuracy of motion recognition areanalyzed and the robustness of CSI-HC is detected In addition the performance of the proposedmethod is compared with that ofthe existing motion recognition methods e experimental results show that the average motion recognition rate of CSI-HC inthree classic indoor scenes reaches 854 in terms of motion complexity and indoor recognition accuracy Compared with otheralgorithms it has higher stability and robustness

1 Introduction

Benefiting from the widespread deployment of the wirelesscommunication infrastructure human behavior recognitionbased on wireless communication network technology hasbecome a core technology to promote various applications[1 2] Traditionally in order to recognize human behaviorphysical sensing devices (eg ultra-wideband (UWB) radiofrequency identification (RFID) and acceleration sensors)are first required to be worn on the human body or deployedin the environment On this basis the information collectedby these physical devices is read to facilitate the identifi-cation of the personrsquos behavioral state Although this tra-ditional behavior recognition method has been widely used

and achieved good results most of them require specificsensor equipment In addition WiFi-based behavior rec-ognition overcomes the shortcomings of traditionalmethods that is it can automatically recognize humanbehavior without the user wearing a sensor or device and hasbeen widely used in real life including smart home remotehealth care campus security severe illness patient care andelderly activity detection [3]

Currently WiFi-based human motion sensing tech-nology has broad development prospects Compared withthe human behavior recognition of the motion sensor andoptical camera the advantage of WiFi-based human motionrecognition is that the coverage of the WiFi signal is widethere is no sense of dead angle and there is no sense of light

HindawiMobile Information SystemsVolume 2020 Article ID 3185416 20 pageshttpsdoiorg10115520203185416

requirement and obscurity Not only can it run on the cheapcommercial WiFi device but also it has lower hardware costand maintenance cost than the motion sensor and opticalcamera e traditional indoor human behavior detectionmainly depends on the received signal strength (RSS) methodbut the experimental results show that the RSS method haspoor motion recognition and low stability Compared withRSS CSI is a finer measure of the physical layer which de-scribes the amplitude attenuation and phase shift of a wirelesssignal based on which it can effectively identify a variety ofbehaviors from vital signs to basic behavior as well ascomplex activities [4ndash7] In reference [4] Liu developed asystem to track vital signs such as the heart rate and respi-ration rate during sleep by analyzing CSI signals in real timeeWi-Sleep system proposed in reference [5] can extract therespiratory information of users in various sleep positions byidentifying the rhythm patterns associated with breathing Inreference [6] WiDraw is a hand motion tracking systemwhich uses the angle of arrival (AOA) value of theWiFi signalon the mobile device to track the hand motion trajectory eWiWho framework proposed in reference [7] can use CSI forhuman gait recognition

According to the different recognition algorithms andapplication scenes a WiFi-based human motion recognitionsystem can be divided into two categories one is the model-based recognition system and the other is the fingerprintrecognition system e main difference between them iswhether a priori learning is required [8] e model-basedrecognition system can recognize human motion withouttraining For example the Fresnel model of WiFi humanrecognition is established in reference [9] However themodel-based recognition system needs to access a largenumber of access points (APs) to accurately identify humanbehavior which leads to a significant increase in hardwarecost andmaintenance cost Based on the fingerprint databaserecognition scheme through offline phase training andonline phase recognition for pattern matching humanmotion recognition can be realized

However the existing WiFi-based human motion rec-ognition scheme recognizes the action is relatively simplethe actual scene availability is not strong or it is a dailybehavior such as walking opening the door sleeping or asingle human body standing picking up and sitting downSome complex motions cannot be accurately identified andthe application scenarios are relatively simple and theexisting application scenarios are diversified (special scenedetection such as prison hospitals wildlife behavior detec-tion indoor elderly activity detection and construction sitesafety detection) erefore it is imperative to find acomplex motion recognition scheme Considering the ap-plication scene in real life we use WiFi to identify theChinese traditional martial art XingYiQuan motion in theindoor environment We guide users to carry out correctfitness motions so as to maintain human health

e main contributions of our work are as follows

(1) In this paper a complex human motion recognitionscheme CSI-HC based on WiFi is proposed andverified with the background of the Chinese

traditional martial art XingYiQuan Its advantage isthat the detection motion is relatively complex anddoes not need human wearing equipment It canwork effectively on cheap commercial devices and isof great value in guiding and monitoring humanhealth campaigns

(2) CSI-HC uses the amplitude of the signal received bythe receiver array antenna to establish a mappingrelationship with the different actions of the humanbody It uses the Butterworth low-pass filter andSym8 wavelet function to construct a powerfuldenoising method to filter outliers in motion datae offline fingerprint construction is completedthrough RBM training In the online phase SoftMaxregression is used to modify the RBM classificationto accurately sense the complex motions of differenthuman bodies

(3) We analyze the key factors that affect the perceptualeffect such as the distance between transceiver de-vices and transmitter contracting rate find the ap-propriate parameter settings and explore the impactof user diversity on system performance throughexperiments

(4) In three scenarios where the multipath effect rangesfrom weak to strong (meeting room corridor andoffice) the performance of CSI-HC is tested eexperimental results show that CSI-HC has highrobustness and in three different scenes the accu-racy of the XingYiQuan motion can be more than85

e main contents and organizational structure of thispaper are as follows We first introduce the related work inSection 2 and then we describe the proposed researchmethodology in Section 3 e specific implementationprocess of the method is provided in Section 4 Section 5explores parameter settings and demonstrates performanceFinally we summarize all the work of this paper in Section 6

2 Related Works

In this section we introduce the advantages and disad-vantages of existing motion recognition works from twoperspectives that is device-based and device-free

21 Device-Based Human Motion Recognition As we allknow most human perception systems need additionalhardware support to complete the recognition of relatedmotions ese hardware devices are divided into fourcategories special sensors infrared devices optical camerasand smartphones e special sensor realizes the perceptionof different actions by collecting the relevant physical in-formation of human body movements Skinput [10] uses awearable bioacoustic sensor array to analyze the mechanicalvibrations that travel in the body to identify themovement ofthe arms and fingers FEMO [11] is a human motion de-tection system based on RFID which uses the backscatteringsignal of the passive RFID tag installed on the training

2 Mobile Information Systems

equipment to detect the motion state of the user eFisio-Track [12] is a telemedicine assistance system that detectspatientsrsquo rehabilitation training movements by using theaccelerometer equipment e GrandCare [13] systemdetects patient behavior by calling a motion sensor installedon the door Although the special sensor can realize thehigh-precision perception of the fine-grained movement ofthe human body users need to wear sensors or deployspecial equipment and it is difficult to carry and invade theprivacy of users e infrared device images the humanbody through infrared rays to realize the perception of anindependent light source e representative product isMicrosoftrsquos Kinect [14] Infrared rays have a limited de-tection range due to problems such as their frequencybands and transmission distances and require expensiveadditional equipment and it is difficult to achieve large-scale deployment e optical camera captures the imagesequence of the human motion through the camera an-alyzes the human motion characteristics and the motiontrajectory in the image sequence and senses the state of thehuman body e method based on optical cameras can beused in many scenes such as gesture recognition [15] gaitrecognition [16] and target tracking [17] However thismethod still has shortcomings and it cannot work in low-light conditions and where privacy is involved Smart-phones use the built-in sensors (accelerometer gyroscopemagnetometer and pressure gauge) to detect human ac-tivities ey have the advantages of being easy to use andnot disturbing the userrsquos normal activities Smartphonesalso have some research results in detecting human mo-tions Gu et al [18] combined the accelerometer andpressure gauge in a smartphone to monitor 7 differentstates of motion PerFallD [19] uses the accelerometer inthe smartphone to detect the fall of the human bodyUnfortunately although smartphones have become pop-ular and have many advantages they are not applicable insome scenarios In particular when it comes to detectingfalls among the elderly it is often impractical to have themcarry around smartphones

22 Device-Free Human Motion Recognition

221 RSS-Based Human Motion Recognition In the pastfew decades due to the rapid development of sensingtechnology device-based human perception has beenwidely used in daily life However device-based perceptionrequires special equipment because of its high hardwareand maintenance costs It is difficult to deploy effectively ona large scale To solve this problem researchers have begunto focus on device-free human perception technology thatis the detection of human behavior without the need forthe human body to wear any physical device [20] ewidespread deployment of wireless networks makes itpossible to realize device-free human perception based onWiFi signals Seifeldin et al [21] realized simple humanmotion detection by analyzing RSS changes in WiFi signalscaused by the human motion and Sigg et al [22] used asoftware-defined radio to transmit RF signals and to

determine the human motion based on changes in thereceived RSS With the maturity of this technology thehuman body motion that the WiFi signal can detect is moreand more detailed Wi-Vi [23] and WiSee [24] systems usethe WiFi signal to realize gesture recognition AlthoughRSS-based technology has made great progress the essenceof RSS is to reflect the strength of signal reception espe-cially susceptible to multipath effects and narrowbandinterference resulting in its own flaws with low accuracyIn order to overcome the inherent defects of RSS a finergranularity WiFi channel feature CSI based on the physicallayer is discovered Compared with RSS CSI can estimatethe channel characteristic information on each subcarrierby orthogonal frequency-division multiplexing (OFDM)technology e frequency attenuation variation of theWiFi channel can be better described to reduce the in-fluence of multipath components and narrowband inter-ference In addition CSI contains rich frequency-domaininformation such as amplitude and phase which can betterreflect the influence of the human body on the WiFi signalto achieve higher sensing accuracy

222 CSI-Based Human Motion Recognition e researchof human motion perception based on CSI originated froma tool CSI Tool [25] which can obtain CSI informationfrom the commercial WiFi network card CSI Tool greatlyfacilitates the extraction of CSI sensing data so that the useof more fine-grained CSI signals for sensing has become anew trend Since CSI has excellent characteristics of beingrelatively stable in an indoor environment and sensitive tohuman motions CSI signals are widely used in variousmotion recognition scenarios and FIMD [26] attempts touse CSI to detect user location changes without activelycarrying any physical devices DeMan [27] is a noninvasivedetection scheme which can judge the motion and staticstate of the human body e motion state is judged byamplitude combined with phase information and thestatic state of the human body is judged by the periodicmodel caused by human respiration in the wireless signalR-PMD [28] is a passive motion detection scheme that usesPCA to extract the covariance of CSI data as a motionfeature and the human motion state is judged by themapping between covariance and human motion CRAM[29] constructed a CSI speed model which correlates themovement speed of different parts of the human body withthe dynamic changes of CSI and detects specific humanactivities such as walking running and sitting down RT-Fall [30] is a device-free fall detection system based on CSIdata which can effectively identify the normal walkingstate and abnormal fall state of the human body Wi-Finger[31] recognizes finger gestures by establishing user ges-tures and CSI signal changes caused by different gestures(for example numbers 1 to 9 in ASL) E-eye [32] is adevice-independent activity detection system which dis-tinguishes a large number of daily activities by matchingthe measured values of CSI with known features and re-alizes the recognition of different actions such as sleepingcooking and watching TV

Mobile Information Systems 3

3 Preliminaries and Program

In this section we introduce the WiFi-based recognitionmodel and the reasons why CSI can perform human motionperception and give the specific solution and the overview ofthis solution

31 Research eory

311 WiFi-Based Recognition Model e principle of hu-man motion perception of WiFi signals is depicted inFigure 1 When a person is in a signal link the propagationof wireless signals will be reflected scattered and diffractedby the influence of the human body e signal received atthe receiving end is a composite signal that is propagated bythe direct path and the human body reflection path as well asthe reflection path of the floor ceiling e influence of thehuman body on the propagation of the WiFi signal will becharacterized by the wireless signal arriving at the receivingend Assume that the line of sight (LOS) length from thetransmitter to the receiver is d then the distance between thereflection point of the ceiling and the floor and the LOS is RCombining the Friis free-space propagation equations andsignals with the reflection scattering generated by the humanbody the impact of the human body on wireless signalpropagation can be defined as [33]

Prx(d) PtxGtxGrxλ

2

(4π)2(d + 4R + ε)2 (1)

where Ptx is the transmitting power of the transmitting endPrx(d) is the receiving power of the receiving end Gtx is thetransmitting gain Grx is the receiving gain λ is the wave-length of theWiFi signal and ε is the approximate change ofthe path length caused by the scattering of the signal by thehuman body Because the signal scattering paths caused bydifferent actions are different according to the above for-mula different human actions will cause the difference inreceiver receiving power and by establishing the mappingrelationship between these differences and different humanactions it lays the basic idea of WiFi human motionperception

312 Channel State Information eCSI signal can achieveuniversal low-cost fine-grained human perception andthere are three reasons for this (1) e maturity of WiFitechnology and the widespread use of WiFi devices (2) CSIdata can be easily extracted from commercial WiFi devicesusing tools released by Halperin (3) e WiFi signaltransmits a modulation scheme using OFDM under theIEEE 80211N protocol and OFDM can encode the CSI datato a plurality of subcarriers of different frequencieserefore the radio channel information at the subcarrierlevel can be obtained from the original CSI measured in theWiFi data link In OFDM transmission systems it is as-sumed that a general model of channel state information canbe represented as

Yrarr

HXrarr

+ noise (2)

where Yrarr

and Xrarr

represent the received signal vector and thetransmitted signal vector respectively noise is additivewhite Gaussian noise andH is the channel impulse response(CIR) complex matrix in the CSI frequency domainreflecting the channel gain information at the subcarrierlevel Assuming that we obtain the measured CSI values H ofthe 2times 3 frames received by the Atheros AR9380 NIC (thatis 2 transmit antennas and 3 receiving antennas) under thecondition that the channel bandwidth is 20MHz and thetime is T we can obtain the CSI values of 336 subcarriers

H H f1( 1113857 H fN( 1113857 H f336( 11138571113858 1113859T 1leNle 336

(3)

erefore the CIR can characterize the frequency-do-main information of each different subcarrier and the i-thsubcarrier can be defined as

H(i) |H(i)|ej sinangH(i)

(4)

where |H(i)| represents the amplitude value of the i-thsubcarrier and angH(i) represents the phase value of the i-thsubcarrier

32 Research Program

321 Specific Solution CSI-HC proposed in this paper is ascheme to identify complex human motions SpecificallyCSI-HC uses CSI amplitude characteristics to identify hu-man motions on commercial WiFi devices filters envi-ronmental noise through low-pass filtering and waveletfunction and uses RBM and SoftMax machine learningclassification algorithms to identify human motions Takingthe ancient Chinese martial art Xingyi boxing as thebackground of action recognition XingYiQuan is a tradi-tional Chinese fitness martial art which is composed of sixmotions such as QiShi BengQuan HuQuan MaXingQuanZuanQuan and ShouShi Each action consists of differentcombinations of hands arms head torso and legs Wecollect the data of six XingYiQuan motions as shown inFigure 2 and show that these motions correspond to thecharacteristics of CSI signals generated in the frequency

LOS

Ceiling reflection

Ground reflectionReflection by human

d

r1

r2

Figure 1 WiFi human motion recognition model


QiShi the body is upright and the two feet are close together

BengQuan the body squats and presses under both palms

HuQuan the left foot is forward and the fists are punched forward by the tigerrsquos claws

MaXingQuan the rightfoot is half step forward and the two fists are up

ZuanQuan legs are kept tight body slightlysquats and the left fist is up

ShouShi both hands are everted and the arms are lifted up

50

45

40

35

30

25

50

45

40

35

30

25

20

464442

3840

3634323028

504540353025

50

45

40

35

30

25

201510

40

35

30

25

20

20

15

10

5

0 10 20 30 40 50 600 10 20 30 40

Channel 1Channel 2Channel 3






50 60

0 10 20 30 40 50 600 10 20 30 40 50 60

0 10 20 30 40 50 600

Subcarrier index Subcarrier index



Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

10 20 30 40 50 60

Figure 2 XingYiQuan motion and the corresponding amplitude feature


domain from which we can see that there are differences inCSI amplitude characteristics between different motions Byusing these differences to establish the mapping with theaction and carry on the pattern recognition the concreteXingYiQuan motions can be recognized

322 Overview We describe in detail the overall archi-tecture of the CSI-HC system which consists of three phasesthe data collection phase the offline motion fingerprintestablishment phase and the online motion recognitionphase as shown in Figure 3 In the data collection phase theWiFi device with the Atheros AR9380 NIC is used to collectCSI data of human motion perception in a variety ofscenarios

In order to remove the influence of the multipath effectand environmental noise we filter the outliers of the dataobtained in the data collection phase e Butterworth low-pass filter is used to remove the high-frequency outliers andthen combined with the wavelet function of the Sym8 wavebase to filter the low-frequency outliers en we get theCSI data which can better reflect the XingYiQuan motionsuse these data as the sample set of RBM training adjust thelearning parameters classify the data and construct thestandard fingerprint information of each XingYiQuanmotion In the online motion recognition phase the sameexception handling and machine learning classificationmethods as in the offline phase are adopted the dataconsistency is maintained and the RBM classification re-sults are corrected using the SoftMax classifier en itmatches the data with the fingerprint database of theconstructed motion and recognizes the specific XingYi-Quan motion

4 Methodology

In this section we mainly describe in detail the two core dataprocessing processes in the CSI-HC method (1) Effective

outlier filtering is performed on the collected CSI data usinga Butterworth low-pass filter and wavelet function (2) Byconstructing the RBM training model the CSI data ofcomplex motions are classified and the RBM classificationresult is modified by SoftMax to achieve higher precisionmotion recognition

41 CSI Outlier Filtering When using CSI data as a featureof motion perception filtering outliers due to environ-mental noise interference and multipath components iscritical to the accuracy of motion perception In this paperthe amplitude of the acquired CSI data is taken as thefeature value Figure 4 shows the original CSI amplitudeimage of XingYiQuan in which it can be seen that there aremany abnormal values ese outliers may reduce theaccuracy of the recognition of the motion method For thisreason the Butterworth low-pass filter is used to filter thehigh-frequency interference in the outliers and Sym8 isused as the wave-based wavelet function to filter the low-frequency interference in the outliers e collected orig-inal CSI amplitude data are filtered by the outliers topreserve the signal feature integrity to the greatest extentso as to establish the feature fingerprint information of eachXingYiQuan motion

411 Filtering Effect Evaluation As shown in Figure 5 weselected four different filters to filter the abnormal data inFigure 4(a) to evaluate the performance of different filtersHere Figure 5(a) shows abnormal data which we use as theoriginal input data of various filters Figure 5(b) shows thedata processed by the mean filter Figure 5(c) shows thedata processed by the Butterworth low-pass filterFigure 5(d) shows the data processed by the median filterand Figure 5(e) shows the data processed by the thresholdfilter

By comparing the processing effects of four differentfilters we can find that the processing effect of themean filter

TX

RX

LOS

Raw CSI data

Wavelet function

6 Xingyi fist actions

Butterworth low-pass filter

RBM training and classification

Online collection of fistdata

Perform the same data processing as in the

offline phase

SoftMax classification

Match with the fingerprint

Establish fingerprint information for each fist

Data collection

Reflection

Output

Offline phase Online phase

Input

Figure 3 CSI-HC system flow chart


is poor and it cannot filter the anomalies effectively is isbecause the original data become smaller and the datawaveform changes after averaging the CSI data Howeveralthough the median filter and threshold filter can filter outthe obvious noise they ignore the detailed noise in the high-frequency part It is obvious that the Butterworth low-pass

filter has the best filtering effect on the data It can preservethe integrity of the data to the greatest extent that is removethe detailed noise in the abnormal data without changing thedata size and data waveform By comprehensive comparisonand consideration we use the Butterworth low-pass filter inthis paper to complete the data exception processing

Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)

0 10 20 30 40Subcarrier index

50 60

(b)

Figure 4 Original CSI data

4442403836343230

0 10 20 30 40 50 60Subcarrier index

Original data

Am

plitu

de (d

B)

(a)

After mean filtering

45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)

After Butterworthlow-pass filtering

4442403836343230


Am

plitu

de (d

B)

(c)

After median filtering

44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)

After threshold filtering

44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)

Figure 5 Evaluation of four different filtering effects (a) Original data (b) After mean filtering (c) After Butterworth low-pass filter (d)After median filtering (e) After threshold filtering


412 High-Frequency Outlier Processing e high-fre-quency outliers in the CSI data of the XingYiQuan motionare filtered by the Butterworth low-pass filtere processingsteps are as follows

Step 1 the amplitude-frequency transfer function H(ε)of the Butterworth low-pass filter is designed as|H(jΩ)|2 (1(1 + (ΩΩC)2N)) where N is the filterorder ΩC is the cutoff frequency (in rads) and j is thedenominator coefficient of each orderStep 2 the relevant parameters of the Butterworth low-pass filter η(fn fs fp Rp Rs) are set where fn is thesampling frequency fs is the stopband cutoff fre-quency fp is the passband cutoff frequency Rp is theminimum attenuation of the fluctuation in the pass-band and Rs is the minimum attenuation in thestopband Using the influence frequency of the humanbody on the signal as the passband cutoff frequency thepart of the environmental noise higher than the averagefrequency of the human body is filtered out generating

relevant parameters suitable for processing amplitudedataStep 3 the amplitudematrix of XingYiQuan is regardedas the original signal and the Butterworth low-passfilter designed in Step 2 is used to filter out the high-frequency outliers in the amplitude information

After the Butterworth low-pass filter filters out the ab-normal XingYiQuan motion data before and after com-parison as shown in Figures 6(a)ndash6(d) it can be seen that theCSI data become smoother and high-frequency anomaliesare filtered out

413 Low-Frequency Outlier Processing e waveletfunction is used to filter out low-frequency interference datain the CSI data Taking the XingYiQuan motion featureinformation as the original input signal of the wavelettransform and Sym8 wavelet as the wave base of wavelettransform the CSI signal is decomposed by five layers and

44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)

Figure 6 Data before and after comparison by the Butterworth low-pass filter (a) channel 1 outlier 1 (b) after the Butterworth low-passfiltering of outlier 1 (c) channel 2 outlier 2 (d) after the Butterworth low-pass filtering of outlier 2


the corresponding approximate part and detailed part areobtained which approximately represent the low-frequencyinformation e detail represents the high-frequency in-formation and the sure threshold mode and scale noise areselected for the detail coefficient

Figures 7(b) and 7(d) are CSI amplitude diagramsgenerated by Sym8 wavelet function It can be seen that afterfiltering out low-frequency anomalies CSI data becomemore stable and retain the integrity to a large extent so thatthe features of each motion can be extracted more easily andthey can be used as the sample set of XingYiQuan datatraining so as to better realize the classification of motiondata in the next step

42 XingYiQuan Motion Classification

421 Offline XingYiQuan Motion Fingerprint Establishmente XingYiQuan data filtered by the outliers are used as thetraining sample e RBM training requires the learningparameter θprime It is assumed that a training set λ

λ(1) λ(t)1113966 1113967 is given where t is the number of training

samples e maximum likelihood method is used to solvethe likelihood function P(v) of RBM training

maximize L θprime( 1113857 maximizeWijaibj1113864 1113865

1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)

where v is the visible layer h is the hidden layer Wij is theconnection weight between the visible layer and the hiddenlayer l is the number of visible and hidden layer neuralunits when solving the maximum likelihood function and t

is the number of training samples In this paper the k-stepcontrast divergence (k-CD) algorithm is used to train theRBM network that is it only needs M steps of Gibbssampling to obtain an approximation suitable for thesample model to be trained According to the establishmentof the likelihood function when the visible layer data areconstant the probability of the activation state of all hidden

Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)

Figure 7 Sym8 wavelet function filters out low-frequency outliers (a c) After the Butterworth low-pass filter (b d) After the Sym8 waveletfilter


layer neurons can be deduced as shown in formula (6)Similarly when the hidden layer data are constant theactivation state of the visible layer neurons can also bededuced from formula (7)

P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)

where ai is the bias of the visible layer bj is the bias of thehidden layer and σ(x) is a nonlinear sigmoid activationfunction of the neuron

After collecting a large amount of action data of theXingYiQuan motion the data are preprocessed e data ofeach motion after processing are used as the fingerprintinformation S (S1 S2 S3 SW)T and S is used as thesample set to be trained and the RBM is trained etraining process is as follows

Step 1 we first initialize the RBM and assign initialvalues to the visible neurons v0 vStep 2 Gibbs sampling on the training sample set S andthe M-step Gibbs sampling are carried out when thetime is l 1 2 MStep 3 the Gibbs sampling method is used to samplethe visible layer neurons and process them By settingup a likelihood function to sample the visible layer fromthe hidden layer we can use P(h(l) | v(lminus 1)) to samplethe visible layer and calculate the corresponding vlSimilarly to Gibbs sampling of hidden layer neurondata we can use P(hlminus 1 | vlminus 1) to sample the hiddenlayer and calculate hlminus 1Step 4 the corresponding expected values are calcu-lated for the hidden layer and the visible layer tocomplete the training

After the M-step Gibbs sampling of the training sampleset the system can obtain an approximate sampling valuewhich is suitable for the training sample model At the sametime the system can also determine the activation state ofthe neurons in the visible layer In this paper the fingerprintinformation after training is recorded as Sprime (S1prime S2prime S3prime

SWprime )Tprime to establish the characteristic fingerprint database of

each XingYiQuan motion

422 Online SoftMax-Modified RBM Classification In theonline phase the data are collected in real time in the ex-perimental environment and the same abnormal filteringand RBM classification are carried out in the offline phaseand the RBM is trained to get θprime (W a b) as the test inputof the SoftMax functione logistic regression cost functionis used as the cost function of the SoftMax classifier and the

SoftMax function is set as Uη(θprime) e probability P(yi

j | θprimei) of each class of boxing samples is given to obtain thecategory label If the input classification parameter is θprime (θ1prime θ2prime θ3prime θm

prime) the model parameter is η1 η2 ηk andthe sum of all probabilities is 1 by normalization

Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)

Next the maximum likelihood estimation is used toobtain the cost function of SoftMax In order to simplify thecost function the indicator function ind(middot) is introduced

ind yi j( 1113857 1 true

0 false1113896 1113897 (10)

en the cost function can be expressed as

J(η) minus 1113944k

i1ind y ji( 1113857ln

eηT

iθprime

1113936ki1 e

ηiθprime⎛⎝ ⎞⎠ (11)

By constructing the SoftMax classification model thedata sample sets of different actions are input in turn andeach time the data representing different actions are se-lected a one-dimensional matrix containing six categories isreturned and select the category with the highest probabilityeach time to correspond to the six motions to be classifiede six motions of the classification further correct theclassification accuracy of the RBM through the SoftMaxclassification function and ensure that the CSI-HC methodhas higher motion recognition accuracy

43 Online XingYiQuan Motion Recognition In the onlinerecognition phase the transmitter is used to collect the dataof each XingYiQuan motion to be identified the collecteddata are sent to the receiver and the amplitude of the CSIdata is selected as the feature And the frequency in the signalis represented by the rate of multipath change caused bybody motion while the amplitude represents the energy ofthe signal erefore by analyzing the amplitude informa-tion in the signal we can obtain the amplitude character-istics of each XingYiQuan motion and then discriminate thedifferent motions e online phase recognition process is asfollows

Step 1 the real-time data of the XingYiQuan motionare collected between the receiver and the transmitterequipped with the Atheros network cardStep 2 the amplitude in the CSI data is selected as afeature


Step 3 the outliers are filtered out by data pre-processing which is easy to match with the establishedstandard fingerprint databaseStep 4 the RBM classification model is constructed byusing the trained learning parameter θprime (W a b)and the detection data are classifiedStep 5 the SoftMax classifier is used to modify the RBMclassification results to improve the classificationaccuracyStep 6 the classified amplitude data are matched withthe offline fingerprint database in real timeStep 7 according to the above steps the specificXingYiQuan motion performed by the tester is finallyrecognized

5 Experiments and Evaluation

In this section in order to evaluate the performance of CSI-HC a large number of experiments have been carried outFirst of all we have introduced the experimental configu-ration in detail en combined with three typical indoorscenarios the key parameters that affect the recognitionaccuracy and the diversity of users are analyzed and finallythe overall performance of CSI-HC is shown

51 Experimental Design

511 Hardware Configuration In order to verify the fea-sibility of the CSI-HC method in the actual scenario thispaper adopts the Atheros AR9380 network card solutionbased on the IEEE 80211N protocol e required equip-ment is two desktop computers equipped with the AtherosAR9380 network card CPU model is Intel Core i3-4150operating system is Ubuntu 1604 LTS4110 Linux kernelversion and two 15m long 5 dB high-gain external an-tennas one of which acts as the transmitting end and theother as the receiving end which respectively connect theantenna contacts of the Atheros NIC of the transmitter andreceiver with a 15m external antenna e Atheros AR9380NIC and external antenna are shown in Figure 8

512 Experimental Scenarios We choose the testbed inthree classical indoor scenarios such as the meeting room(65mtimes 10m) corridor (2mtimes 46m) and office(65mtimes 13m) in order to correspond to the change of themultipath effect from low to high e experimental sce-narios and experimental scenariosrsquo plane structure areshown in Figures 9 and 10 First of all keeping the height anddistance of the receiver and the transmitter and thetransmitter sending a certain rate in the three scenarios thetesters are arranged to stand at the midpoint of the deployedtransmitter and receiver to make a fixed motion of Xin-gYiQuan Each motion collects 10000 packets of CSI dataand saves them to the receiver PC After processing by theCSI-HCmethod the RBM is used to train and learn the dataof each XingYiQuan motion and the standard feature

fingerprint information of each XingYiQuan motion isestablished

In the online recognition phase the testers stand in themiddle of the transmitter and receiver in the deployed ex-perimental scenario to do the XingYiQuan motion collectand process CSI data in real time and classify them using theconstructed RBM model rough the established standardfeature fingerprint database for real-time matching thespecific motions of the testers are identified In Table 1 wegive the relevant steps for a XingYiQuan motion recogni-tion as well as the relevant hardware used in each step andthe specific time it takes to process these operations

52 Analysis of Influencing Factors of the Experiment

521 TX-RX Distance Analysis In the process of estab-lishing the offline motion fingerprint database two im-portant device parameters (TX contract rate and RX and TXdistance) played a key role in the effect of online XingYi-Quan motion recognition In order to analyze the effect ofthe distance between TX and RX on the motion recognitionmethod proposed in this paper we take the method ofcontrolling variables during the offline fingerprint databaseconstruction phase For the same user the sampling time iskept constant (100 s) the contract rate is 10 ps and the TX-RX distances of 06m 08m 1m 12m 14m 16m 18mand 2m are set in three scenarios such as the office corridorand meeting room to analyze the impact of different TX andRX distance settings on the effect of XingYiQuan motionrecognition in the online phase during offline training inorder to find the most suitable distance setting for offlinefingerprint training Figure 11 shows how the recognitionaccuracy of six XingYiQuan motions varies with TX-RXdistance in three scenarios

As can be seen from Figure 11 with the increase of thedistance between the transmitter and the receiver in threedifferent scenarios the accuracy of recognition of the sixXingYiQuan motions generally shows an upward trend eoverall data except the MaXingQuan motion indicate whenthe distance between the transmitter and the receiver is14m the accuracy of motion recognition reaches a peakand when the distance is between 14m and 2m the ac-curacy of motion recognition begins to decline slowly Afterthe distance exceeds 14m the accuracy of motion recog-nition begins to decrease slowly Because of the complexityof the MaXingQuan motion and the ZuanQuan motion themultipath interference and the interference within the en-vironment are large so there are fluctuations that are in-consistent with other XingYiQuan motions And theexperimental results also verify the previously set multipatheffect-increasing scene so the distance between the trans-mitter and the receiver is 14m which is most suitable forthe XingYiQuan motion recognition in this question

522 TX Contracting Rate Analysis e transmitterrsquospacket rate determines the number of samples collectedduring the same time period as well as the sensitivity of theXingYiQuan motion to be acquired in the data link Since


RX

TX

10M

65M

(a)

C501 C505

C512 C506Meeting room

C507C508C509C510Office

C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)

Figure 10 Floor plan of the experimental scenarios

TX RX

(a)

TX RX

(b)

TX RX

(c)

Figure 9 Experimental scenarios (a) meeting room (b) corridor (c) office

(a) (b)

Figure 8 (a) WiFi commercial Atheros AR9380 NIC (b) 5 dB high-gain external antenna


the signal propagates in the air with a certain attenuation thenumber of sample data collected at different packet rates isdifferent at the same time Assuming that q is the presetcontract rate T is the sampling time and N is the actualnumber of samples then the estimated number of samplingpoints N1 is N1 Tlowast q and the packet loss rate loss can bedefined as loss (N minus Tlowast q)N According to the definitionof the packet loss rate we test the packet loss number of thetransmitter under the 10 ps 20 ps 50 ps and 100 ps

packet loss rates under the condition that the same user has asampling time of 10 s and calculate the corresponding packetloss rate Table 2 reflects the relationship between thecontract rate and the packet loss rate of the transmitter

What is obvious in Table 2 is that when the contract rateof the transmitter is 10 ps the actual number of packetscollected accounts for the highest proportion of the esti-mated number of sampled packets that is the lowest packetloss rate Because the minimum packet loss rate can get the

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688

Meeting roomCorridorOffice

(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)

06 08 1 12TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

Figure 11 Effect of TX-RX distance on the accuracy of motion (a) QiShi motion (b) BengQuan motion (c) HuQuan motion(d) MaXingQuan motion (e) ZuanQuan motion (f ) ShouShi motion

Table 2 Relationship between the transmitter contract rate and the packet loss rate

Contract rate (ps) Estimated number of samplesp (10 s) Actual number of samplesp (10 s) Packet loss rate ()10 100 91 920 200 173 13550 500 418 164100 1000 813 187

Table 1 Hardware processing time design

Related steps Data collection Offline fingerprint building Online motion recognitionHardware type TP-Link WDR4310 PC TP-Link WDR4310 + PCProcessing time 028 hours 011 hours 005 hours


most true sampling number when the total collectionnumber is fixed and as many sampling points as possible canmore truly reflect the motion state of the human body andthen achieve a high-precisionmotion recognition we choosethe contract rate of 10 ps in the construction phase of theoffline fingerprint database

In order to verify the effectiveness of the contract rate setduring the offline phase we tested different offline contractrate settings Under the condition that the distance betweenTX and RX is set to 14m and the total number of samples is1000 packets the training users in the offline phase conductoffline training on the contract rate settings of 10 ps 20 ps50 ps and 100 ps in three scenarios including the officecorridor and meeting room and conduct the correspondingXingYiQuan motion recognition test

By analyzing and comparing the contract rate andmotion recognition accuracy data in Figure 12 we find thatthe accuracy of motion recognition is the highest when thepacket rate is 10 ps in the experimental scenario designed bythis method and the action rate is 20 ps When the packetsending rate is 50 ps and 100 ps the accuracy of recog-nition action is greatly reduced because of excessive packetloss erefore the most suitable contract rate for the CSI-HC method is 10 ps

523 User Diversity Analysis In order to explore the in-fluence of user diversity on the accuracy of motion recog-nition two different testers were arranged in the experimentUnder the condition that the distance between TX and RXwas set to 14m the total number of samples was 1000packets and the contract rate was 10 ps six XingYiQuanmotions were performed many times to collect and processdata en the average recognition rates of the six Xin-gYiQuan motions of the two testers were calculated eresult of the accuracy distribution of two usersrsquo motionrecognition is shown in Figure 13

As can be seen from Figure 13 the accuracy of themotion recognition of both testers was maintained above80 and the highest accuracy of the first testerrsquos Xingyiaction was 923 is is because the first tester had a longperiod of XingYiQuan motion practice and performed thestandard motion In addition the second tester as he was anovice had a low degree of proficiency in the XingYiQuanmotion e lower the accuracy of the motion due to thenonstandard motion (808) the higher the standard of themotion performed by the tester in the box diagram and theshorter the length of the box such as the BengQuan andShouShi motions of tester 1 e lower the standard of thetesterrsquos motion the longer the length of the box such as theBengQuan and ShouShi motions of tester 2

53 Overall Performance Evaluation

531 Robustness Testing In order to test the robustness ofthe CSI-HC method we cut through the environment andthe user and arrange two different testers First let a tester dothe same XingYiQuan motion in the meeting room cor-ridor office etc e CSI data of the action are collected the

data processing is trained and the difference of CSI-HCrecognition accuracy in different scenarios is analyzed andthen another tester is arranged to do the same motion in thethree scenarios e motion data are collected and processedand compared with the CSI characteristics of the previousperson Figure 14 shows the amplitude of the QiShi motionof the two testers in the meeting room corridor office etcTable 3 shows the specific accuracy of motion recognition

e amplitude features of CSI motions of differenttesters in three scenarios (taking the QiShi motion as anexample) are compared as shown in Figure 14 On the onehand according to the similar amplitude trend of (a) (c)and (e) in Figure 14 it is shown that CSI-HC is less affectedby environmental factors and can have higher performanceeven in the case of serious multipath interference On theother hand by comparing (a) with (b) (c) with (d) and (e)with (f) in Figure 14 we can see that the trend of amplitudecharacteristics collected by different people in the sameenvironment is approximately the same Since the principleof motion recognition of CSI-HC is based on the motion andthe corresponding amplitude characteristics the amplitudechange trend of different testers is consistent which showsthat CSI-HC is not sensitive to user differences and has highrobustness

Table 3 shows the accuracy of the CSI-HC detection ofsix motions of XingYiQuan in three different scenarios eaverage detection accuracy in the meeting room is 8933However the average detection accuracy in the office is only8123 In addition the average detection rate in the cor-ridor and meeting room is higher than that in the officebecause there are more multipath components and humaninterference in the office

532 Overall Performance In order to evaluate the overallperformance of the CSI-HC proposed in this paper wedefine the following three metrics compare them with twoclassic human motion detection methods R-PMD andFIMD and compare their detection results in the meetingroom corridor and office e experimental results areshown in Figure 15 and Table 4

(i) Precision precision TP(TP + FP) (also definedas sensitivity) refers to the probability of correctlyrecognizing the human motion

(ii) Recall recall TP(TP + FN) (also defined asparticularity) refers to the probability of correctlyidentifying nondetected human motions

(iii) F1 score F1score 2lowastprecisionlowast recall(precision+recall) e F1 indicator is a comprehensiveevaluation standard combining precision and recallBy calculating the F1 score index of differentmethods the stability of the method can be effec-tively evaluated

Here TP true positives FP false positives andFN false negatives

Figure 15 shows the comparison of CSI-HC R-PMDand FIMD methods in terms of precision recall and F1score From Figure 15 we can see that the precision recall


QiShi BengQuan HuQuan MaXingQuan ZuanQuan ShouShi80

82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)

Figure 13 User diversity and recognition accuracy analysis (a) Tester 1 (b) Tester 2

10 20 50 100Contract rate (ps)

0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan

MaXingQuanZuanQuanShouShi

(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)

Figure 12 Effect of the contract rate on motion accuracy in three testbeds (a) meeting room (b) corridor (c) office


and F1 score index of CSI-HC are obviously better thanthose of the other twomethodsis is because the denoisingmethod of CSI-HC is a combination of low-pass filtering and

wavelet transform which greatly filters multipath interfer-ence and compares the complete retention of the motionfeatures However R-PMD uses low-pass filtering and

0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )

Figure 14 Amplitude characteristic maps of two testers performing the QiShi motion in three scenarios (a) Meeting room (tester 1)(b) Meeting room (tester 2) (c) Corridor (tester 1) (d) Corridor (tester 2) (e) Office (tester 1) (f ) Office (tester 2)

Table 3 Robustness evaluation of CSI-HC in three scenarios

Different scenariosDetection accuracy of different XingYiQuan actions ()

QiShi BengQuan HuQuan MaXingQuan ZuanQuan ShouShiMeeting room 882 896 910 878 871 923Corridor 843 859 866 836 855 880Office 809 801 810 810 812 832


principal component analysis (PCA) filtering out moremotion features retaining only a few representative featuresand affecting the recognition accuracy However FIMDadopts the method of false alarm is method can onlyeliminate the erroneous data with large deviation from theCSI feature and cannot filter out some interference data withsmall deviation

Table 4 shows the motion detection accuracy of CSI-HCR-PMD and FIMD in the three scenarios of this paper eexperimental results show that the detection accuracy ofCSI-HC in the meeting room reaches 893 e approachhas high environmental adaptability and robustness

533e Impact of Sample Size We find that the number oftraining samples affects the detection accuracy of the motionrecognition method To this end we test different samplesizes in the real environment we have built and the ex-perimental results are shown in Figure 16

Figure 16 reflects the relationship between the recog-nition accuracy of the three motion detection methods andthe number of training samples It can be seen that thedetection accuracy of the motion recognition methods in-creases with the number of training samples Compared withthe other two methods the CSI-HC method has a highmotion recognition rate e motion detection accuracy ofthe R-PMD method is slightly lower than that of CSI-HCand FIMD has the worst effect is is because CSI-HC usesthe RBM training method based on contrast divergence

After setting the training weight and learning rate it canquickly fit with the sample by adjusting the weight pa-rameters reduce the system overhead by repeated reuseimprove the classification training efficiency of motions andincrease the accuracy of motion recognition HoweverR-PMD uses the PCA approach It means that with theincrease of data principal component analysis should becarried out on all data to select characteristic data whichincreases the system overhead In contrast FIMD uses thedensity-based spatial clustering of applications with noise(DBSCAN) method to determine the clustering center Byconstantly calculating the Euclidean distance betweensample points and updating clustering points the timecomplexity of the algorithm is greatly increased and with theincrease of the number of training samples the

Precision Recall F1-scoreEvaluation metrics

0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD

Figure 15 Comprehensive evaluation of three methods

Table 4 Motion detection method performance in three indoorenvironments

Method Meeting room () Corridor () Office ()CSI-HC 893 857 812R-PMD 882 840 805FIMD 761 718 706

0 200 400 600 800 1000Number of samples (p)

40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD

Figure 16 Impact of the number of samples

2 3 4 5 6 7 8Feature number

65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD

Figure 17 Impact of the number of features


computational burden of the system continues to increaseand the training and recognition effects on the motion dataare not good

534 e Impact of the Number of Features As shown inFigure 17 the motion detection accuracy increases as thenumber of features used increases Specifically when thenumber of features increases the detection accuracy of theCSI-HC method proposed in this paper will also increaseWhen the number of feature values reaches 6 the detectionaccuracy reaches the highest and remains stable In contrastthe accuracy change trend of the R-PMD method is roughlythe same as that of CSI-HC but it is significantly lower thanthat of CSI-HC However the turning point of FIMD de-tection accuracy is 5 and the detection rate decreases sig-nificantly when the number of features is 2ndash4

e test results in Figure 17 show that the detection rateof our proposed CSI-HC method is higher and more stableis is because the CSI-HC method uses the RBM networkto train the model suitable for CSI data and the charac-teristics are concentrated in the training data set whichpreserves the data integrity to a large extent so it has a highdetection rate and stability However the R-PMD methoduses the PCAmethod to extract the most representative dataas features e data integrity is not as high as that of CSI-HC which causes the detection accuracy to decrease eFIMD method uses the correlation matrix of CSI data as thefeature and uses the DBSCAN method for clustering edata are scattered in the feature which makes it unstableand the detection accuracy is low

535 Comparison with the Optical Sensor Method Wecompare the CSI-HC method proposed in this paper withthe optical sensor-based method e CSI-HC method usesCSI data as a recognition feature while the optical sensor-based method generally uses an optical signal composed ofinherent characteristics of light as a feature for motionrecognitione average detection accuracy of our proposedCSI-HC for the motion is 854 while the accuracy ofmotion detection based on optical sensor-based methods isas high as 989 [34] Although in terms of recognitionaccuracy the CSI-HC method is lower than the opticalsensor-based method However in terms of applicationscenarios the optical sensor-based method does not workproperly in low-light and dark environments or in scenariosinvolving privacy e CSI-HC method overcomes thisdefect and can be used in all indoor environments In thefuture research we will further improve the accuracy of the

CSI-HC method to make up for its lack of recognitionaccuracy

536 Comparison with Previous Motion RecognitionMethods In order to reflect the unique advantages andcomprehensive performance of the CSI-HC method pro-posed in this paper we compare the CSI-HC method withthe previous three human motion recognition methods(computer vision method infrared method and specialsensor method) on multiple parameters [35] e results areshown in Table 5

As can be seen from Table 5 the CSI-HCmethod has theadvantages of non-line-of-sight low deployment cost notlimited by environmental conditions and low algorithmcomplexity while the other three methods have high de-ployment cost and computational complexity Among themcomputer vision methods cannot work effectively in darkand privacy-related scenes However although the infraredmethod and the dedicated sensor method have achievedhigh motion detection accuracy the deployment cost is toohigh and they are not suitable for widespread deployment inindoor scenarios

6 Conclusion

In this paper we propose a new complex human motionrecognition method namely CSI-HC which is verified bythe XingYiQuan motion which is a complex motion ecore part is the denoising of CSI signals and the classificationof complex motions We collect CSI amplitude data in theoffline phase use the Butterworth low-pass filter combinedwith the Sym8 wavelet function method to filter the outliersand use the RBM to train and classify the XingYiQuanmotion to establish standard fingerprint information for theXingYiQuan motion Next in the online phase the Xin-gYiQuan motion is collected in the experimental scenariosto collect real-time data the abnormal value filtering andRBM training classification are performed the SoftMaxclassification model is constructed to correct the RBMclassification result and the offline phase XingYiQuanmotion fingerprint database is used for data matching toachieve recognition of different XingYiQuan motionsrough a large number of experiments this paper exploresthe influence of parameter setting and user diversity on theaccuracy of motion recognition e experimental resultsshow that the CSI-HC achieves an average motion recog-nition rate of 854 in three practical scenarios It has goodperformance in robustness motion recognition rate prac-ticability and so on

Table 5 Comparison of parameters between CSI-HC and various human motion recognition methods

ParametersMethods

CSI-HC Computer vision Infrared Dedicated sensorPerceived distance Non-line-of-sight Line of sight Non-line-of-sight Short distanceDeployment costs Low Medium High HighEnvironmental limitations No Yes No YesComputational complexity Low High High MediumMotion detection accuracy Relatively high High High High


Data Availability

e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

is work was funded by the National Natural ScienceFoundation of China (61616070 and 61762079)

References

[1] Z Wang B Guo Z Yu and X Zhou ldquoWi-fi CSI basedbehavior recognition from signals actions to activitiesrdquo IEEECommunications Magazine vol 56 no 5 pp 109ndash115 2017

[2] Y Lu S Lv X Wang X Zhou et al ldquoA survey onWiFi basedhuman BehaviorAnalysis technologyrdquo Chinese Journal OfComputers vol 1ndash222018 in Chinese

[3] X Dang Y Huang Z Hao and X Si ldquoPCA-Kalman device-free indoor human behavior detection with commodity Wi-Firdquo EURASIP Journal on Wireless Communications andNetworking vol 2018 no 1 2018

[4] J Liu Y Wang Y Chen et al ldquoTracking vital signs duringsleep leveraging off-the-shelf WiFirdquo in Proceedings of theACM International Symposium on Mobile Ad Hoc Networkingamp Computing ACM pp 267ndash276 Hangzhou China June2015

[5] X Liu J Cao S Tang J Wen and P Guo ldquoContactlessrespiration monitoring via off-the-shelf WiFi devicesrdquo IEEETransactions on Mobile Computing vol 15 no 10pp 2466ndash2479 2016

[6] L Sun K H Kim S Sen et al ldquoWiDraw enabling hands-freedrawing in the air on commodity WiFi devicesrdquo in Pro-ceedings of the 21st Annual International Conference onMobileComputing and Networking pp 77ndash89 Paris France Sep-tember 2015

[7] Y Zeng P Patha and P Mohapatra ldquoWiWho WiFi-basedperson identification in smart spacesrdquo in Proceedings of the2016 15th ACMIEEE International Conference on Informa-tion Processing in Sensor Networks (IPSN) pp 1ndash12 ViennaAustria April 2016

[8] L Chen X Chen L Ni Y Peng and D Fang ldquoHumanbehavior recognition using Wi-Fi CSI challenges and op-portunitiesrdquo IEEE Communications Magazine vol 55 no 10pp 112ndash117 2017

[9] D Zhang H Wang and D Wu ldquoToward centimeter-scalehuman activity sensing withWi-Fi signalsrdquoComputer vol 50no 1 pp 48ndash57 2017

[10] C Harrison D Tan and D Morris ldquoSkinput appropriatingthe body as an input surfacerdquo in Proceedings of the SigchiConference on Human Factors in Computing Systems ACMpp 453ndash462 Atlanta GA USA April 2010

[11] H Ding L Shangguan Z Yang et al ldquoFemo a platform forfree-weight exercise monitoring with rfidsrdquo in Proceedings ofthe 13th ACM Conference on Embedded Networked Sensor

Systems ACM pp 141ndash154 Seoul Republic of KoreaNovember 2015

[12] D Ruiz-Fernandez O Marın-Alonso A Soriano-Paya andJ D Garcıa-Perez ldquoeFisioTrack a telerehabilitation envi-ronment based on motion recognition using accelerometryrdquoe Scientific World Journal vol 2014 Article ID 49539111 pages 2014

[13] S Herath M Harandi and F Porikli ldquoGoing deeper intoaction recognition a surveyrdquo Image and Vision Computingvol 60 pp 4ndash21 2017

[14] Z Zhang ldquoMicrosoft kinect sensor and its effectrdquo IEEEMultimedia vol 19 no 2 pp 4ndash10 2012

[15] H-M Zhu and C-M Pun ldquoAn adaptive superpixel basedhand gesture tracking and recognition systemrdquo e ScientificWorld Journal vol 201412 pages 2014

[16] D Xu S Yan D Tao S Lin and H-J Zhang ldquoMarginal fisheranalysis and its variants for human gait recognition andcontent- based image retrievalrdquo IEEE Transactions on ImageProcessing vol 16 no 11 pp 2811ndash2821 2007

[17] M Fiaz and B Ijaz ldquoVision based human activity trackingusing artificial neural networksrdquo in Proceedings of the In-ternational Conference on Intelligent and Advanced Systems(ICIAS) Kuala Lumpur Malaysia June 2010

[18] F Gu A Kealy K Khoshelham and J Shang ldquoUser-inde-pendent motion state recognition using smartphone sensorsrdquoSensors vol 15 no 12 pp 30636ndash30652 2015

[19] J Dai X Bai Z Yang Z Shen and D Xuan ldquoPerFallD apervasive fall detection system using mobile phonesrdquo inProceedings of the 2010 8th IEEE International Conference onPervasive Computing and Communications Workshops(PERCOM Workshops) pp 292ndash297 IEEE MannheimGermany March 2010

[20] M Youssef and M Mah ldquoChallenges device-free passivelocalization for wirelessrdquo in Proceedings of the ACM Mobi-Com pp 222ndash229 Montreal Canada September 2007

[21] M Seifeldin A Saeed A E Kosba A El-Keyi andM Youssef ldquoNuzzer a large-scale device-free passive local-ization system for wireless environmentsrdquo IEEE Transactionson Mobile Computing vol 12 no 7 pp 1321ndash1334 2013

[22] S Sigg S Shi F Busching Y Ji and L C Wolf ldquoLeveragingrf-channel fluctuation for activity recognition active andpassive systems continuous and rssi-based signal featuresrdquo inProceedings of the 11th International Conference on Advancesin Mobile Computing amp Multimedia p 43 Vienna AustriaDecember 2013

[23] F Adib and D Katabi ldquoSee through walls with WiFirdquo ACMSIGCOMM Computer Communication Review vol 43 no 4pp 75ndash86 2013

[24] Q Pu S Gupta S Gollakota and S Patel ldquoWhole-homegesture recognition using wireless signalsrdquo in Proceedings ofthe 19th Annual International Conference on Mobile Com-puting and Networking pp 27ndash38 New York NY USADecember 2013

[25] D HalperinW Hu A Sheth and DWetherall ldquoTool releasegathering 80211n traces with channel state informationrdquoACM SIGCOMM Computer Communication Review vol 41no 1 p 53 2011

[26] J Xiao K Wu Y Yi L Wang and L M Ni ldquoFIMD fine-grained device-free motion detectionrdquo in Proceedings of the2012 IEEE 18th International Conference on Parallel andDistributed Systems vol 90 no 1 pp 229ndash235 SingaporeDecember 2012


[27] C Wu S Member Z Zhou X Liu Y Liu and J Cao ldquoNon-invasive detection of moving and stationary human withWiFirdquo IEEE Journal on Selected Areas in Communicationsvol 33 no 11 pp 2329ndash2342 2015

[28] H Zhu F Xiao L Sun X Xie P Yang and RWang ldquoRobustand passive motion detection with cots wifi devicesrdquo TsinghuaScience and Technology vol 22 no 4 pp 345ndash359 2017

[29] W Wang A X Liu M Shahzad et al ldquoUnderstanding andmodeling ofWiFi signal based human activity recognitionrdquo inProceedings of the 21st Annual International Conference onMobile Computing andNetworking pp 65ndash76 NewYork NYUSA September 2015

[30] H Wang D Zhang Y Wang J Ma Y Wang and S Li ldquoRT-fall a real-time and contactless fall detection system withcommodity WiFi devicesrdquo IEEE Transactions on MobileComputing vol 16 no 2 p 1 2016

[31] H Li W Yang J Wang X Yang and L Huang ldquoWiFingertalk to your smart devices with finger-grained gesturerdquo inProceedings of the ACM International Joint Conference onPervasive amp Ubiquitous Computing ACM pp 250ndash261Heidelberg Germany September 2016

[32] YWang J Liu Y Chen M Gruteser J Yang and H Liu ldquoE-eyes device-free location-oriented activity identification us-ing fine-grained WiFi signaturesrdquo in Proceedings of theACMMobiCom pp 617ndash628 Maui HI USA September 2014

[33] C Han KWu YWang and L M Ni ldquoWiFall devicefree falldetection by wireless networksrdquo in Proceedings of the IEEEConference on Computer Communications IEEE INFOCOMpp 271ndash279 Toronto ON Canada May 2014

[34] M Pierre D Yohan B Remi P Vasseur and X Savatier ldquoAstudy of vicon system positioning performancerdquo Sensorsvol 17 no 7 p 1591 2017

[35] A F M Saifuddin Saif A S Prabuwono andZ R Mahayuddin ldquoMoving object detection using dynamicmotion modelling from UAV aerial imagesrdquo e ScientificWorld Journal vol 2014 Article ID 890619 12 pages 2014


requirement and obscurity Not only can it run on the cheapcommercial WiFi device but also it has lower hardware costand maintenance cost than the motion sensor and opticalcamera e traditional indoor human behavior detectionmainly depends on the received signal strength (RSS) methodbut the experimental results show that the RSS method haspoor motion recognition and low stability Compared withRSS CSI is a finer measure of the physical layer which de-scribes the amplitude attenuation and phase shift of a wirelesssignal based on which it can effectively identify a variety ofbehaviors from vital signs to basic behavior as well ascomplex activities [4ndash7] In reference [4] Liu developed asystem to track vital signs such as the heart rate and respi-ration rate during sleep by analyzing CSI signals in real timeeWi-Sleep system proposed in reference [5] can extract therespiratory information of users in various sleep positions byidentifying the rhythm patterns associated with breathing Inreference [6] WiDraw is a hand motion tracking systemwhich uses the angle of arrival (AOA) value of theWiFi signalon the mobile device to track the hand motion trajectory eWiWho framework proposed in reference [7] can use CSI forhuman gait recognition

According to the different recognition algorithms andapplication scenes a WiFi-based human motion recognitionsystem can be divided into two categories one is the model-based recognition system and the other is the fingerprintrecognition system e main difference between them iswhether a priori learning is required [8] e model-basedrecognition system can recognize human motion withouttraining For example the Fresnel model of WiFi humanrecognition is established in reference [9] However themodel-based recognition system needs to access a largenumber of access points (APs) to accurately identify humanbehavior which leads to a significant increase in hardwarecost andmaintenance cost Based on the fingerprint databaserecognition scheme through offline phase training andonline phase recognition for pattern matching humanmotion recognition can be realized

However the existing WiFi-based human motion rec-ognition scheme recognizes the action is relatively simplethe actual scene availability is not strong or it is a dailybehavior such as walking opening the door sleeping or asingle human body standing picking up and sitting downSome complex motions cannot be accurately identified andthe application scenarios are relatively simple and theexisting application scenarios are diversified (special scenedetection such as prison hospitals wildlife behavior detec-tion indoor elderly activity detection and construction sitesafety detection) erefore it is imperative to find acomplex motion recognition scheme Considering the ap-plication scene in real life we use WiFi to identify theChinese traditional martial art XingYiQuan motion in theindoor environment We guide users to carry out correctfitness motions so as to maintain human health

e main contributions of our work are as follows

(1) In this paper a complex human motion recognitionscheme CSI-HC based on WiFi is proposed andverified with the background of the Chinese

traditional martial art XingYiQuan Its advantage isthat the detection motion is relatively complex anddoes not need human wearing equipment It canwork effectively on cheap commercial devices and isof great value in guiding and monitoring humanhealth campaigns

(2) CSI-HC uses the amplitude of the signal received bythe receiver array antenna to establish a mappingrelationship with the different actions of the humanbody It uses the Butterworth low-pass filter andSym8 wavelet function to construct a powerfuldenoising method to filter outliers in motion datae offline fingerprint construction is completedthrough RBM training In the online phase SoftMaxregression is used to modify the RBM classificationto accurately sense the complex motions of differenthuman bodies

(3) We analyze the key factors that affect the perceptualeffect such as the distance between transceiver de-vices and transmitter contracting rate find the ap-propriate parameter settings and explore the impactof user diversity on system performance throughexperiments

(4) In three scenarios where the multipath effect rangesfrom weak to strong (meeting room corridor andoffice) the performance of CSI-HC is tested eexperimental results show that CSI-HC has highrobustness and in three different scenes the accu-racy of the XingYiQuan motion can be more than85

e main contents and organizational structure of thispaper are as follows We first introduce the related work inSection 2 and then we describe the proposed researchmethodology in Section 3 e specific implementationprocess of the method is provided in Section 4 Section 5explores parameter settings and demonstrates performanceFinally we summarize all the work of this paper in Section 6

2 Related Works

In this section we introduce the advantages and disad-vantages of existing motion recognition works from twoperspectives that is device-based and device-free

21 Device-Based Human Motion Recognition As we allknow most human perception systems need additionalhardware support to complete the recognition of relatedmotions ese hardware devices are divided into fourcategories special sensors infrared devices optical camerasand smartphones e special sensor realizes the perceptionof different actions by collecting the relevant physical in-formation of human body movements Skinput [10] uses awearable bioacoustic sensor array to analyze the mechanicalvibrations that travel in the body to identify themovement ofthe arms and fingers FEMO [11] is a human motion de-tection system based on RFID which uses the backscatteringsignal of the passive RFID tag installed on the training










31 Research eory


Prx(d) PtxGtxGrxλ

2

(4π)2(d + 4R + ε)2 (1)



Yrarr

HXrarr

+ noise (2)

where Yrarr

and Xrarr


H H f1( 1113857 H fN( 1113857 H f336( 11138571113858 1113859T 1leNle 336

(3)



(4)


32 Research Program


LOS

Ceiling reflection


d

r1

r2









50

45

40

35

30

25

50

45

40

35

30

25

20

464442

3840

3634323028

504540353025

50

45

40

35

30

25

201510

40

35

30

25

20

20

15

10

5

0 10 20 30 40 50 600 10 20 30 40







50 60

0 10 20 30 40 50 600 10 20 30 40 50 60

0 10 20 30 40 50 600




Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

10 20 30 40 50 60






4 Methodology






TX

RX

LOS

Raw CSI data

Wavelet function






offline phase




Data collection

Reflection

Output


Input





Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References















































31 Research eory


Prx(d) PtxGtxGrxλ

2

(4π)2(d + 4R + ε)2 (1)



Yrarr

HXrarr

+ noise (2)

where Yrarr

and Xrarr


H H f1( 1113857 H fN( 1113857 H f336( 11138571113858 1113859T 1leNle 336

(3)



(4)


32 Research Program


LOS

Ceiling reflection


d

r1

r2









50

45

40

35

30

25

50

45

40

35

30

25

20

464442

3840

3634323028

504540353025

50

45

40

35

30

25

201510

40

35

30

25

20

20

15

10

5

0 10 20 30 40 50 600 10 20 30 40







50 60

0 10 20 30 40 50 600 10 20 30 40 50 60

0 10 20 30 40 50 600




Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

10 20 30 40 50 60






4 Methodology






TX

RX

LOS

Raw CSI data

Wavelet function






offline phase




Data collection

Reflection

Output


Input





Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References









































31 Research eory


Prx(d) PtxGtxGrxλ

2

(4π)2(d + 4R + ε)2 (1)



Yrarr

HXrarr

+ noise (2)

where Yrarr

and Xrarr


H H f1( 1113857 H fN( 1113857 H f336( 11138571113858 1113859T 1leNle 336

(3)



(4)


32 Research Program


LOS

Ceiling reflection


d

r1

r2









50

45

40

35

30

25

50

45

40

35

30

25

20

464442

3840

3634323028

504540353025

50

45

40

35

30

25

201510

40

35

30

25

20

20

15

10

5

0 10 20 30 40 50 600 10 20 30 40







50 60

0 10 20 30 40 50 600 10 20 30 40 50 60

0 10 20 30 40 50 600




Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

10 20 30 40 50 60






4 Methodology






TX

RX

LOS

Raw CSI data

Wavelet function






offline phase




Data collection

Reflection

Output


Input





Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References













































50

45

40

35

30

25

50

45

40

35

30

25

20

464442

3840

3634323028

504540353025

50

45

40

35

30

25

201510

40

35

30

25

20

20

15

10

5

0 10 20 30 40 50 600 10 20 30 40







50 60

0 10 20 30 40 50 600 10 20 30 40 50 60

0 10 20 30 40 50 600




Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

Am

plitu

de (d

B)A

mpl

itude

(dB)

Am

plitu

de (d

B)

10 20 30 40 50 60






4 Methodology






TX

RX

LOS

Raw CSI data

Wavelet function






offline phase




Data collection

Reflection

Output


Input





Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References










































4 Methodology






TX

RX

LOS

Raw CSI data

Wavelet function






offline phase




Data collection

Reflection

Output


Input





Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References









































Outliers

Outliers

Channel 1

50

48

46

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Outliers

Outliers

Outliers

Channel 2

50

52

48

46

44

42

40

38

36

34

Am

plitu

de (d

B)


50 60

(b)


4442403836343230


Original data

Am

plitu

de (d

B)

(a)


45

40

35

30

25

200 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(b)


4442403836343230


Am

plitu

de (d

B)

(c)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(d)


44

42

40

38

36

34

320 10 20 30 40 50 60

Subcarrier index

Am

plitu

de (d

B)

(e)








44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References












































44

42

40

38

36

34

32

300 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Outlier 1

(a)

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

Filter outlier

(b)

Outlier 2

44

46

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(c)

Filter outlier

44

46

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)










1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References














































1t

1113944

t

i1ln 1113944

h

P vl h

l1113872 1113873⎛⎝ ⎞⎠

1t

1113944

t

l1ln P v

(l)1113872 11138731113872 1113873

(5)



Butterworth filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(a)

Sym8 wavelet filter

44

42

40

38

36

34

320 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(b)

Butterworth filter

32

44

42

40

38

36

34


Am

plitu

de (d

B)

50 60

(c)

Sym8 wavelet filter

44

42

40

38

36

340 10 20 30 40

Subcarrier index

Am

plitu

de (d

B)

50 60

(d)




P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References








































P vi 1 h θprime11138681113868111386811138681113872 1113873 σ 1113944

j

Wijhj + ai⎛⎝ ⎞⎠ (6)

P hj 1 v θprime11138681113868111386811138681113872 1113873 σ 1113944

i

Wijvi + bj⎛⎝ ⎞⎠ (7)

σ(x) sigmoid(x) 1

1 + eminus x (8)











Uη θprime(i)1113874 1113875

P y(i)( 1113857 1 θprime(i)1113874 1113875η1

1113868111386811138681113868111386811138681113868

P y(i)( 1113857 2 θprime(i)1113874 1113875η2

1113868111386811138681113868111386811138681113868

⋮

P y(i)( 1113857 k θprime(i)1113874 1113875ηk

1113868111386811138681113868111386811138681113868

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

1

1113936kj1 e

ηTjθprime

(i)

eηT1 θprime

(i)

eηT2 θprime

(i)

⋮

eηTkθprime

(i)

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(9)



0 false1113896 1113897 (10)




eηT

iθprime

1113936ki1 e



















RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References




















































RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References







































RX

TX

10M

65M

(a)

C501 C505



C511

ToiletC503

LaboratoryC504C5

02

TXRX

46M

2M

(b)

RXTX

13M

65M

(c)


TX RX

(a)

TX RX

(b)

TX RX

(c)


(a) (b)






7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References










































7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

7476788082848688


(a)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

747678808284868890


(b)

90

7506 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

80

85


(c)

7206 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82

84


(d)

72

70

6806 08 1 12

TX-RX distance (m)

Mot

ion

accu

racy

()

14 16 18 2

74

76

78

80

82


(e)


Mot

ion

accu

racy

()

14 16 18 278

80

82

84

86

88

90

92


(f )

























82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References

























































82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References








































82

84

86

88

90

92

94

96

Mot

ion

accu

racy

()

(a)


74

76

78

80

82

84

86

88

90

92

Mot

ion

accu

racy

()

(b)



0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(a)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(b)


0

20

40

60

80

100

Accu

racy

()

QiShiBengQuanHuQuan


(c)





0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References









































0 10 20 30 40 50 60


Subcarrier index

30

35

40

45

50

Am

plitu

de (d

B)

(a)


20

25

30

35

40

45

50

Am

plitu

de (d

B)


(b)


25

30

35

40

45

50

Am

plitu

de (d

B)


(c)


3436384042444648

Am

plitu

de (d

B)


(d)



30

35

40

45

50

Am

plitu

de (d

B)

(e)



36

38

40

42

44

46

48A

mpl

itude

(dB)

(f )












0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References













































0

20

40

60

80

100

Rate

()

CSI-HCR-PMDFIMD





40

50

60

70

80

90

Det

ectio

n ac

cura

cy (

)

CSI-HCR-PMDFIMD



65

70

75

80

85

90D

etec

tion

accu

racy

()

CSI-HCR-PMDFIMD










6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References














































6 Conclusion



ParametersMethods



Data Availability




Acknowledgments


References







































Data Availability




Acknowledgments


References

















































Documents

CSI-HC: A WiFi-Based Indoor Complex Human Motion