AthleteBehaviorRecognitionTechnologyBasedonSiamese-RPN

Research ArticleAthlete Behavior Recognition Technology Based on Siamese-RPNTracker Model

Changhui Gao

Public Physical Education Department Xinyang University Xinyang Henan 464000 China

Correspondence should be addressed to Changhui Gao wdh0300xynueducn

Received 2 August 2021 Accepted 24 September 2021 Published 19 October 2021

Academic Editor Syed Hassan Ahmed

Copyright copy 2021 Changhui Gao 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

With the rapid development of deep learning algorithms it is gradually applied in UAV (Unmanned Aerial Vehicle) drivingvisual recognition target tracking behavior recognition and other fields In the field of sports many scientists put forward theresearch of target tracking and recognition technology based on deep learning algorithms for athletesrsquo trajectory and behaviorcapture Based on the target tracking algorithm a regional proposal network RPN algorithm combined with the twin regionalproposal network Siamese algorithm is proposed to study the tracking and recognition technology of athletesrsquo behavior en theadaptive updating network is used to track the behavior target of athletes and the simulation model of behavior recognition isestablishedis algorithm is different from the traditional twin network algorithm It can accurately take the athletersquos behavior asthe target candidate box in model training and reduce the interference of environment and other factors on model recognitione results show that the Siamese-RPN algorithm can reduce the interference from the background and environment whentracking the athletesrsquo target behavior trajectory is algorithm can improve the training behavior recognition model ignore thebackground interference elements of the behavior image and improve the accuracy and overall performance of the modelCompared with the traditional twin network method for sports behavior recognition the Siamese-RPN algorithm studied in thispaper can perform offline operations and distinguish the interference factors of athletesrsquo background environment It can quicklycapture the characteristic points of athletesrsquo behavior as the data input of the tracking model so it has excellent popularization andapplication value

1 Introduction

In recent years the target behavior tracking model andbehavior recognition technology are the important di-rection of the development of the artificial intelligencedeep learning network [1] With the distinction betweensupervised and unsupervised deep learning algorithmsdeep learning algorithms are widely used in securityprotection driverless smart home public transportationsports and other fields [2 3] In the field of sports thedemand for athletesrsquo behavior trajectory analysis andbehavior pattern recognition is gradually increasing [4]In the traditional single-target tracking model manyresearchers use the athletesrsquo position of the initial videoframes per time for behavior trajectory analysis [5] andaccording to the athletersquos characteristic points to identify

the action behavior However the traditional single targettracking algorithm cannot accurately capture the movingposition and the tracking target model often shows de-formation speed of control many interference factorsand other reasons [6] is situation makes the trackingmodel cannot accurately capture the trajectory and theposition of feature points in the process of athletesrsquo be-havior tracking and recognition [7 8] It cannot completethe purpose of accurate behavior recognition Follow-upresearchers proposed a mean-based tracking algorithmwhich does not need data validation and only needstracking modeling of the athletersquos target to achieve iter-ative search to determine the target position [9] isalgorithm has fast iteration speed and high computationalefficiency but its persistence is poor and its accuracycannot be guaranteed [10] e results of behavior

HindawiComputational Intelligence and NeuroscienceVolume 2021 Article ID 6255390 10 pageshttpsdoiorg10115520216255390

recognition cannot be accurately captured when the targetmotion is deformed One after another the method ofusing correlation over filtering to calculate target trackingand recognition has gradually become the mainstreammode [11] In this way the information data is combinedwith the tracking model for the first time and the cor-relation filter is used to detect the responsiveness in theframe number of the behavior of the moving person andthe feature points of the data are captured It furtherimproves the recognition efficiency of the tracking modeland the speed can exceed 100 frames [12 13]

In recent years with the rapid development of deeplearning algorithms in the field of computers vision andrecognition [14] many researchers apply the convolu-tional neural network algorithm to classification functionwhich can extract a large number of feature points to trainthe tracking model so as to improve the performance ofbehavior recognition model [15] In the deep learningalgorithm the twin network has the characteristics ofsharing data which is suitable for repetitive similarprocessing tasks Siamese FC a twin network targettracking algorithm can combine the number of videoframes with the feature points extracted by the convo-lution network to obtain the predicted position of targetbehavior [16] Subsequently based on the above trackingmodel behavior recognition technology development thispaper proposes the Siamese-RPN algorithm which canaccurately identify the candidate box of athletesrsquo behaviortarget and reduce the influence of environmental factorson Athletesrsquo behavior recognition model [17ndash19]

Contributions in this paper are summarized as follows(1) a regional proposal network RPN algorithm combinedwith the twin regional proposal network Siamese algorithmis proposed to study the tracking and recognition technologyof athletesrsquo behavior (2) the proposed model can accuratelytake the athletersquos behavior as the target candidate box inmodel training and reduce the interference of environmentand other factors on model recognition and (3) comparedwith the traditional twin network method for sports be-havior recognition the proposed model in this paper canperform an offline operation and distinguish the interferencefactors of athletesrsquo background environment It can quicklycapture the characteristic points of athletesrsquo behavior as thedata input of the tracking model so it has excellent pop-ularization and application value

e following content is divided into four sections efirst section introduces the related work about the devel-opment of tracking model technology and behavior rec-ognition technology e second section discusses theSiamese-RPN algorithm to study the target tracking modeland athlete behavior recognition technology Firstly theproposed tracking algorithm by probability regression andthe proposed athlete behavior tracking recognition modelare optimized Finally an adaptive updating algorithm isproposed to build a tracker model and the simulation modelof athlete behavior recognition is established e thirdsection is the result analysis on the Siamese-RPN algorithmfor athlete target tracking e fourth section is the con-clusion to output some important results

2 Related Work

Athlete behavior recognition is a part of the application ofthe tracker model in the field of artificial intelligence rec-ognition [20] Firstly the target tracking algorithm in thetracker model is used to capture the behavior trajectory ofthe moving person and finally the simulation model isconstructed by using behavior recognition technology [21]In athlete target tracking it is necessary to predict the changeof target position in each frame number but in the trackingprocess there will be occlusion size change deformationcomplex background illumination and other influencingfactors [22] To realize the recognition of behavior we mustfirst solve the problem of target tracking and capture Inrecent years target tracking is based on the Siamese twinnetwork framework e core content is to capture therepetitive similar actions in the search area according to thelearning training template and matching rules [23] How-ever the common network algorithm will consume trackingtime and the tracking target is not accurate Based on thecommon network many subsequent researchers proposedthe introduction of the region proposal algorithm to obtain amore accurate target range [24] e tracker model con-structed by the Siamese-RPN algorithm can accuratelycapture the target behavior box and reduce the influencefactors of complex background rough the classificationand regression branch of target tracking finally build thebehavior recognition model of athletes

Xu et al invested a lot of research energy and funds intracking technology and behavior recognition technology[25] With the large demand for target tracking technologythe deep learning algorithm is also improving ey focusedon the research and development of target tracking andmonitoring projects mainly applied in the battlefield andcivil environment After the real-time monitoring model isestablished human behavior tracking and capturing is re-alized And it can identify whether people carry dangerousgoods and other functions

Deng et al mainly applied target tracking and behaviorrecognition technology in the public transport monitoringsystem aiming at tracking and interactive recognition ofpeoplersquos traffic behavior and vehicle driving [26] eycombined the deep learning algorithm with the feature pointanalysis technology to form the ability to accurately trackillegal vehicle behavior in a complex environment as well asthe realization of the illegal personnel behavior identifica-tion After analysis and identification the violation infor-mation is uploaded to the management and control system

Jia et al also applied tracking technology to trafficmanagementey mainly implanted the tracking algorithmand recognition algorithm into the radar system [27]rough the target tracking and behavior recognitioncontrol of traffic vehicles it can achieve the measurement ofvehicle distance angle speed and other functions Laterthey also used this technology in the VRmodel In the virtualenvironment eye movement is tracked and real-time in-teraction is realized

Visual tracking and behavior recognition are applied inautomation medical image artificial intelligence and other

2 Computational Intelligence and Neuroscience

fields [28] is paper mainly studies the classification andbehavior understanding of the target in the tracking modele function of human body recognition and motiontracking is realized e development of this technology isconducive to the optimization of the public security envi-ronment of the whole society According to the applicationand development of target tracking and behavior recogni-tion technology in the above countries this paper studies thetracking technology and finally realizes the establishment ofthe athletersquos behavior recognition model In the trackingmodel the target personnel is locked and tracked and thenthe Siamese-RPN algorithm is optimized to improve thetracking speed and performance Finally the Siamese-RPNalgorithm adaptive update tracking technology combinedwith the behavior recognition of sports personnel is used tobuild the simulation model

3 Methods on Athlete Behavior RecognitionTechnology based on Siamese-RPN NetworkTracking Model

31Method onAthletesrsquo Behavior Locking andTarget TrackingTechnology based on Siamese-RPNNetwork Twin network iscomposed of two or more network structures in the deeplearning network algorithm which can be calculated accordingto a variety of input variables And it can share the weight in thenetwork structure e core content of the twin network is tofind a group of variables and find the variables of similarity byspecifying the spatial range between target distances [29] Itmakes the difference between the degree values of variables inthe same class smaller and the difference between differentclasses larger e traditional target tracking behavior recog-nition technology uses the Siamese FC algorithm a twinnetwork structure tracking model based on all connectedlayers e core of the algorithm is to compare the similaritybetween the original sample image and the target imagethrough the training function If the similarity of targetjudgment is consistent the score is higher otherwise it islower However this algorithm cannot guarantee high per-formance in accuracy and it will show the problem of trackingtarget deformation and will be interfered with by variousbackgrounds and environments Based on the above situationthis paper proposes the Siamese-RPN algorithm to integratethe candidate networks into an architecture and uses the RPNattention mechanism and frame regression classification toscreen out background and environmental interference factorse structure of the twin network area proposal algorithm isshown in Figure 1

As can be seen from Figure 1 the Siamese-RPN networkcan realize offline training operation between end-to-endmodels is structure is divided into many branches theoriginal target and the tracking target are classified into twopaths and the twin network is used to extract network featurepoints to achieve the function of target recognition emultirange evaluation mechanism and online strategy of tra-ditional tracking technology are changed and the trackingtechnology is improved to a detection process for a certaintarget e performance and detection efficiency of the whole

tracking model are improved In the offline operation thetarget position is predicted and located by the loss function andthe repeated range of the set threshold value and the target areais screened e loss function formula is as follows

ltotal α11139443

i1lcls + β1113944

3

i1lreg

lcls minus 1113936

ni labelsi middot lg scoresi( 1113857

n

lreg minus 1113936

ni 1113936

4j R tj minus t

lowastj1113872 1113873

n

(1)

e weight coefficient of loss is defined in the formulaand the background environmental factors are classified intotwo categories e loss coefficient of the bounding box isupdated and the number of effective mechanism variables inthe candidate region is selected for calculation ethreshold of intersection and union differentiation is set asthe range of matching parameters Different thresholds willaffect the prediction range and border regression of thetracking target

IOU area(C)cap area(G)

area(C)cup area(G) (2)

e process of athlete behavior recognition is a processof target location and tracking Firstly each video material isdivided into any number of frames And the motion state ofeach frame position is predictede simplest operation is totrack the range of the target according to the shape box eflow of the target tracking algorithm is shown in Figure 2

When tracking and identifying athletesrsquo behavior if thetarget has other positioning the tracking task can be definedas a regression calculation process We need to build a modelto track the position and analyse the state of the target etraditional tracking model uses confidence regression tooptimize the operation that is to predict the similaritybetween the feedback state feature and the target featureis method is mainly used in the deep learning algorithmand mainstream tracking method and our regression al-gorithm can predict the trajectory of an independent target[30] In the regression algorithm the background inter-ference in the traditional tracking can be ignored and theappropriate prediction method can be improved In theprocess of athletersquos behavior locking if the target changesthe position of the center shape box will be different fromthat of the target center as shown in Figure 3

It can be seen from Figure 3 that the general regressionalgorithm can only focus on a single target ignoring the gapbetween the shape box positions which is not conducive tothe target tracking task erefore this paper proposes tosolve the location difference problem in Siamese-RPN byusing the shape box of the probability regression algorithme node with the largest probability density is selected asthe prediction output that is the target center positionWhen training the Siamese-RPN network the predictionprobability distribution of input data and output data iscalculated firstly

Computational Intelligence and Neuroscience 3

p(y | x θ) 1

zθ(x)e

sθ(yx)

zθ(x) 1113946 esθ(yx)dy

(3)

Because the sθ(y x) parameter is essentially convertedfrom the confidence number that is to say the relationshipbetween input data and output data is a standard value thepurpose of the above formula is to convert the value intoprobability density Finally it is adjusted by function nor-malization In view of the difference between the conditionaldistribution described by the negative likelihood numberand the conditional probability distribution predicted by thenegative likelihood number we use the divergence trans-formation to define it [31]

DKL(p q) 1113944N

i1p xi( 1113857log

p xi( 1113857

q xi( 11138571113888 1113889 (4)

In the formula q(xi) is the target approximate distri-bution and p(xi) is the real location distribution [32] Afterthe divergence variable is defined the training networkparameters can be calculated

KL p middot | yi( 1113857 p middot | xi θ( 1113857( 1113857 1113946 p y | yi( 1113857logp y | yi( 1113857

p y | xiθ1113872 1113873dy

KL p middot | yi( 1113857 p middot | xi θ( 1113857( 1113857 log 1113946 esθ yxi( )dy1113874 1113875

minus 1113946 sθ y xi( 1113857p y | yi( 1113857dy

(5)

By minimizing the value of output divergence the mostaccurate value of probability prediction distribution can beobtained e comparison between the actual accuracy andthe prediction accuracy of the probability regression algo-rithm in the Siamese-RPN network is shown in Figure 4

127times127times3

255times255times3

Detection frame

CNN

CNN

22times22times256

Template frame

6times6times256

Conv

Conv

Conv

Conv

classification

regression

1 group

Positive Negtive

K groupsdh dx dy dw

Figure 1 Structure chart of twin network area proposal algorithm

model training TrackingstartsData

preparation

Networktraining

Outputmodel

end

Y

Is it the lastframe

Locate thetarget

Firstframe

Y

N

N

target tracking

featureextraction

initialization

Trackingtemplate

Generatenew template

Determinethe next

frame area

Figure 2 Flowchart of the target tracking algorithm


As can be seen from Figure 4 with the increase of thenumber of training iterations the accuracy also graduallyincreases e final result accuracy of the Siamese-RPNalgorithm is basically close to the prediction result Inconclusion the tracker model based on the Siamese-RPNnetwork can improve the accuracy of athletesrsquo behaviortracking and recognition

32 Method on Athlete Behavior Tracking and RecognitionTechnology based on Adaptive Update of Siamese-RPNNetwork Compared with the traditional twin network al-gorithm the Siamese-RPN network can take into accountthe influence factors of the background environment in-formation in the process of behavior tracking and recog-nition and has a stronger discrimination ability in the face ofinterference information erefore this paper proposes anadaptive target classification updating behavior trackingalgorithm based on the Siamese-RPN network To accuratelyobtain the target motion position the cross-correlationfeature monitoring module is introduced e offlinetraining and learning method is used to supervise the be-havior recognition feature points of athletes Finally theattention mechanism is introduced to analyse the back-ground environment information in the video image toextract the accurate target feature point information eupdate mechanism is mainly divided into two modules thejudgmental target classification module and the adaptivemodule rough feature point information extraction Si-amese-RPN algorithm model classification and supervisionfeature point model adaptive model and so on the athletebehavior recognition model is constructed

Based on the Siamese-RPN algorithm the best matchingbox position in the region to be searched is calculated bytraining and learning embedded spatial feature point ex-traction model e calculation formula is as follows

fM(x z) φ(x)lowastφ(Z) + blowast 1 (6)

In the formula the branch variable is used to learn thefeature point representation of the training target moduleand the other branch variable is used in detection enetwork parameter weights of the two branches are shared

Based on this formula the defined variables in the candidatenetwork are used to calculate the predicted target positionand boundary box independently e formula is as follows

fclsM(x z) h

cls φcls(x)lowastφcls(z)1113858 1113859

fregM (x z) h

reg φreg(x)lowastφreg(z)1113960 1113961(7)

e independent variables are used to learn the featurepoint representation of the original template frame and thedetection frame and store the background information ofeach definition box f

regM (x z) stores the width height ratio

information between the offset coordinates of the center boxpoint of the prediction definition and the real box It can getmore accurate background score information of the movingtarget On this basis the boundary box regression calcula-tion is carried out for the center point with the highest score

Bpredict xpre

ypre

wpre

hpre

( 1113857

xpre

xandan

+ dxreg

middot wandan

ypre

yandan

+ dyreg

middot handan

wpre

wandan

+ edwreg

hpre

handan

+ hdhreg

(8)

After the final boundary frame coordinates of targetposition prediction are calculated according to the aboveformula we need to compare the local similarity between thetemplate frame and detection frame based on the Siamese-RPN network target tracking algorithm e top position ofthe cross-correlation feature map is the actual coordinate ofthe target to be trackede RPNmodule can get an efficientand accurate image when it is modified

e image change is related to the corresponding weightvalue in the RPN module In the training we also need totrain and learn the weight value In the athlete behaviorrecognition model it is necessary to set the target classifi-cation module and extract the target feature points fortraining e attention mechanism is used to analyse thetwo-dimensional spatial feature map and the weight coef-ficient of coordinate position is obtained Finally the featureinformation of the corresponding target is extracted to

Actual center position of target box

Target actual center position

Figure 3 Comparison between target box and target center


separate the tracking target from other interference factorsin the region e comparative analysis of the number offeature points extracted by the Siamese-RPN algorithm andtraditional tracking algorithm on the accuracy coefficient ofthe target position is shown in Figure 5

It can be seen from Figure 6 that the tracking modelusing Siamese-RPN is more accurate than the traditionaltracking algorithm in target behavior recognition andtracking with the increase of the number of feature pointsextracted In the above adaptive update strategy the originalinformation of the target is accumulated according to theinitial frame number and the predicted current position iscompared with the extracted template information eresults show that the adaptive updating model can captureand transmit the deep feature points in the current regione real motion behavior of the first time is fed back to thetarget bounding box which is based on the twin regionproposed algorithm to predict the final target behaviormodel e results show that the adaptive updating strategybased on Siamese-RPN can recognize and track the behaviorof athletes under the interference of many factors

In the process of athlete behavior recognition it isnecessary to analyse the contour of the human body It willchange regularly according to the time factor e de-scription of athletesrsquo behavior has obvious representative-ness and the analysis of behavior contour can be carried outaccording to the shape movement rules and other aspectsAfter extracting the behavior contour from the video imagedata the boundary extraction method is used to capture thecoordinate points After getting the center coordinates of thedetailed position the maximum pixel is taken as the startingpoint and the whole scope box is obtained by rotating in thedirection To eliminate the influence factors in the process ofrecognition it is necessary to standardize the image enumber of data points affects the distance distribution of theoverall feature points e comparison of the results beforeand after the standardized processing is shown in Figure 6

It can be seen from Figure 6 that the distance distributioncurve of the feature points after the standardization treat-ment is obviously more simplified than that before thetreatment It can eliminate the interference feature factors inthe whole recognition process

4 Results Analysis on Athlete BehaviorRecognition Technology Based on Siamese-RPN Network Tracking Model

41 Results of Athletesrsquo Behavior Locking and Target TrackingBased on Siamese-RPN Network e whole algorithm usesthe data source set of athletesrsquo sports as training data videoinformation and image information of totally more than5000 groups of data e sample data is trained and testedrespectively and the edge of each frame in the video data iscut and filled e pixel definition of image information isguaranteed Firstly the Siamese-RPN algorithm is used totrain the model offline and then target tracking is carriedout In the training environment the data is preprocessedand the network structure is simplified It is necessary tomodel from the center coordinates of the first frame and thetarget box when capturing the athletersquos behavior targetFinally the output data is trained by a regression model toget the behavior goal of athletes after removing the inter-ference factors In the whole experiment we detect andanalyse the center position error and overlap rate and cal-culate the accuracy difference between the Siamese-RPNalgorithm and the traditional tracking algorithm echange of accuracy under the influence of the center positionthreshold is shown in Figure 7 e change of accuracyunder the influence of overlap rate is shown in Figure 8

It can be seen from Figure 8 that the number of targetframe center position errors calculated by the Siamese-RPNalgorithm has little effect on the accuracy With the increaseof the error threshold the accuracy of the traditional al-gorithm decreases obviously As can be seen from Figure 9

Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000

Actual accuracy of output resultsPrediction accuracy of output results

Figure 4 Comparison chart of actual accuracy and prediction accuracy of calculation results in the Siamese-RPN network


the influence of overlap rate on both algorithms is obviousCompared with the expected curve the accuracy based onthe Siamese-RPN algorithm can keep in the same range ofthe prediction curve under the influence of the overlap rateIn conclusion the athlete behavior locking and targettracking technology based on the Siamese-RPN algorithmhas good performance

42 Results of Athletesrsquo Behavior Tracking and RecognitionBased on Adaptive Update of Siamese-RPN Network Toverify the effectiveness of the tracking method based on theadaptive update of the Siamese-RPN network the artificially

labelled image data set is supplemented with tracking pointsmainly from the camera movement athletesrsquo behavior tracklight changes shelter and other aspects of supplement Fordifferent motion recognition using different parameters tocalculate can get the maximum performance comparisonresults erefore this paper uses the control variablemethod to evaluate the parameter data and find the mostsuitable parameter range for the experiment e optimizeddata set can reduce the waste of time of the identificationmodel e evaluation data is introduced into the cross-correlation feature map to update the monitoring moduleclassification module and adaptive module is paperanalyses the data quantity and recognition success rate of the

250

225

200

175

150

125

Accu

racy

fact

or

100 200 300 400 500 600 700Number of feature points

Precision coefficient curve of Siamese RPN algorithmThe precision coefficient curve of traditional trackingalgorithm

Figure 5 Comparison of precision coefficients between the Siamese-RPN algorithm and traditional tracking algorithm

60

50

40

30

20

10

Feat

ure p

oint

dist

ance

50 100 150 200 250 300Number of data points

Distance distribution curve of featurepoints after standardizationDistance distribution curve of featurepoints before standardization

Figure 6 Comparison of results before and after standardization


adaptive strategy model using the Siamese-RPN algorithmand finally compares it with the ordinary tracking algorithmwithout the Siamese-RPN algorithm e change of contrastcurve is shown in Figure 9

Siamese-RPN can still maintain the success rate ofrecognition when the amount of input data increasesCompared with the traditional tracking algorithm theperformance has been greatly improved In the athletebehavior recognition model the feature points extracted

from the contour can ensure that the processed image isclear and accurate which ensures the recognition effi-ciency and accuracy of the model In the analysis ofathletesrsquo behavior whether there is contour analysis candetermine the result of behavior recognition ereforethis paper proposes to use the Siamese-RPN algorithm totrack athletesrsquo behavior and obtain athletesrsquo feature pointsaccording to the contour to supplement the recognitionmodel

60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50Center position threshold

Siamese RPN algorithm curveTraditional tracking algorithm curve

Figure 7 Accuracy change of two algorithms under the influence of center position threshold

60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()

Expectation curveSiamese RPN algorithm curveTraditional tracking algorithm curve

Figure 8 Comparison of the accuracy of the two algorithms under the influence of overlap rate


5 Conclusions

With the extensive application of deep learning algorithmsin automobile driving traffic management artificial intel-ligence and pattern recognition many tracking algorithmshave been produced for pattern recognition In manytracking algorithms the traditional Siamese FC networkalgorithm cannot meet the needs of target tracking andrecognition is algorithm will be affected by many factorssuch as background and environment and cannot guaranteethe accuracy of target recognition and detection Based onthe target tracking algorithm this paper proposes an al-gorithm based on the RPN tracker model to track andrecognize athletesrsquo behavioris algorithm can optimize theoriginal tracking model and eliminate the influence of in-terference factors By changing the training model to offlinemode the original target area can be distinguished and thebounding box is defined to track the target behavior Firstlythe probability regression algorithm is used to classify thebackground factors to optimize the location accuracy oftarget tracking in the athlete behavior recognition e re-sults show that the Siamese-RPN algorithm has more ac-curate ability to capture the target than the traditionaltracking algorithm and can clearly identify the behavior ofathletes Finally this algorithm is combined with theadaptive update strategy to change the capture mode ofbehavior recognition feature points e results show thatthe efficiency of feature point capture based on the Siamese-RPN algorithm increases with the increase in input data Inthe process of athlete behavior recognition the overallrecognition success rate is improved

Data Availability

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

Conflicts of Interest

e author declares that there are no conflicts of interest

Acknowledgments

is article was supported by Xinyang University

References

[1] Y Yuan ldquoA scale-adaptive object-tracking algorithm withocclusion detectionrdquo EURASIP Journal on Image and VideoProcessing vol 1 no 2020 pp 1ndash15 2020

[2] Z H Li and Y Yu ldquoSurvey of multi-target tracking algorithmbased on detectionrdquo Internet of 7ings Technology vol 11no 4 pp 20ndash24 2021

[3] L S Jin Q Hua B C Guo X Y Xie F G Yan and B T WuldquoMulti-target tracking of vehicles based on optimizedDeepSortrdquo Journal of Zhejiang University (Engineering Sci-ence) vol 55 no 6 pp 1056ndash1064 2021

[4] F Liu and S Sun ldquoImproved twin convolution network targettracking algorithm based on adaptive template updatingrdquoComputer Applications and Software vol 38 no 4 pp 145ndash151 2021

[5] M Ota H Tateuchi T Hashiguchi et al ldquoVerification ofreliability and validity of motion analysis systems duringbilateral squat using human pose tracking algorithmrdquo Gait ampPosture vol 80 pp 62ndash67 2020

[6] B Sun and L Zhang ldquoResearch on light vision tracking ofunmanned ship target based on improved deep learningrdquoShip Science and Technology vol 43 no 6 pp 64ndash66 2021

[7] D Cheng Z W Lv and Y Zhu ldquoTwin network targettracking algorithmrdquo Fujian Computer vol 37 no 2pp 85-86 2021

[8] L Sun S Chen S Chen T Liu C Liu and Y Liu ldquoPig targettracking algorithm based on multi-channel color featurefusionrdquo International Journal of Agricultural and BiologicalEngineering vol 13 no 3 pp 180ndash185 2020

[9] Q Yu B Wang and Y Su ldquoObject detection-tracking al-gorithm for unmanned surface vehicles based on a radar-photoelectric systemrdquo IEEE Access vol 9 pp 57529ndash575412021

[10] Y J Xu ldquoVideo multi-target pedestrian detection andtracking based on deep learningrdquo Modern InformationTechnology vol 4 no 12 pp 6ndash9 2020

[11] X Fang and L Chen ldquoNoise-aware manoeuvring targettracking algorithm in wireless sensor networks by a novel

100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)

100 200 300 400 500 600Number of data input

Siamese RPN algorithm success rate curveGeneral tracking algorithm success rate curve

Figure 9 Comparison of recognition success rate between the Siamese-RPN algorithm and ordinary tracking algorithm


adaptive cubature Kalman filterrdquo IET Radar Sonar amp Nav-igation vol 14 no 11 pp 1795ndash1802 2020

[12] B Yang M Tang S Chen G Wang Y Tan and B Li ldquoAvehicle tracking algorithm combining detector and trackerrdquoEURASIP Journal on Image and Video Processing vol 2020no 1 pp 1ndash20 2020

[13] Y Han C Fan Z Geng et al ldquoEnergy efficient buildingenvelope using novel RBF neural network integrated affinitypropagationrdquo Energy vol 209 Article ID 118414 2020

[14] W Huo J Ou and T Li ldquoMulti-target tracking algorithmbased on deep learningrdquo Journal of Physics Conference Seriesvol 1948 no 1 Article ID 012011 2021

[15] N An and W Qi Yan ldquoMultitarget tracking using Siameseneural networksrdquo ACM Transactions on Multimedia Com-puting Communications and Applications vol 17 no 2pp 1ndash16 2021

[16] A Agrawal and M K Rohil ldquoSpecific motion pattern de-tection state-of-the-art and challengesrdquo in Proceedings of the2021 6th International Conference for Convergence in Tech-nology (I2CT) pp 1ndash6 IEEE Mumbai India 2021

[17] M Yang and Y S Jiaxu ldquoTarget tracking algorithm based onjoint attention twin networkrdquo Journal of Instrumentationvol 42 no 1 pp 127ndash136 2021

[18] Y Sun J Xu and H Wu ldquoDeep learning based semi-su-pervised control for vertical security of maglev vehicle withguaranteed bounded airgaprdquo IEEE Transactions on IntelligentTransportation Systems vol 22 no 7 pp 4431ndash4442 2021

[19] M Wong C Mayoh L M S Lau et al ldquoWhole genometranscriptome and methylome profiling enhances actionabletarget discovery in high-risk pediatric cancerrdquo Nature Med-icine vol 26 no 11 pp 1742ndash1753 2020

[20] J Huang C Chen F Ye W Hu and Z Zheng ldquoNonuniformhyper-network embedding with dual mechanismrdquo ACMTransactions on Information Systems vol 38 no 3 pp 1ndash182020

[21] A F Klaib N O Alsrehin W Y Melhem H O Bashtawiand A A Magableh ldquoEye tracking algorithms techniquestools and applications with an emphasis on machine learningand internet of things technologiesrdquo Expert Systems withApplications vol 166 Article ID 114037 2021

[22] Y M Tang Y F Liu and H Huang ldquoOverview of visualsingle target tracking algorithmrdquo Measurement and ControlTechnology vol 39 no 8 pp 21ndash34 2020

[23] J Jiao X Sun Y Zhang et al ldquoModulation recognition ofradio signals based on edge computing and convolutionalneural networkrdquo Journal of Communications and InformationNetworks vol 6 no 3 pp 280ndash300 2021

[24] H Zhang and L Chen ldquoCalibration and recognition methodof human motion posture feature points based on nearestneighbor specific pointrdquo Laser Journal vol 42 no 4pp 183ndash186 2021

[25] X D Xu X F Zhao and J W Liao ldquoResearch and appli-cation of adaptive edge feature recognition for moving videoimagesrdquo Aeronautical Computing Technology vol 51 no 2pp 41ndash44 2021

[26] P Deng R X Zhao and F F Zhu ldquoA pedestrian trackingalgorithm based on human motion recognitionrdquo ChineseJournal of Inertial Technology vol 29 no 1 pp 16ndash22 2021

[27] PWang J Cheng F Hao LWang andW Feng ldquoEmbeddedadaptive cross-modulation neural network for few-shotlearningrdquo Neural Computing amp Applications vol 32 no 10pp 5505ndash5515 2020

[28] H Lee and S Youm ldquoDevelopment of a wearable camera andAI algorithm for medication behavior recognitionrdquo Sensorsvol 21 no 11 p 3594 2021

[29] A Ben Mabrouk and E Zagrouba ldquoAbnormal behaviorrecognition for intelligent video surveillance systems a re-viewrdquo Expert Systems with Applications vol 91 pp 480ndash4912018

[30] G Batchuluun J H Kim H G Hong J K Kang andK R Park ldquoFuzzy system based human behavior recognitionby combining behavior prediction and recognitionrdquo ExpertSystems with Applications vol 81 pp 108ndash133 2017

[31] K K Santhosh D P Dogra and P P Roy ldquoAnomaly de-tection in road traffic using visual surveillance a surveyrdquoACM Computing Surveys vol 53 no 6 pp 1ndash26 2020

[32] M Sajjad S Zahir A Ullah Z Akhtar and K MuhammadldquoHuman behavior understanding in big multimedia datausing CNN based facial expression recognitionrdquo MobileNetworks and Applications vol 25 no 4 pp 1611ndash1621 2020


recognition cannot be accurately captured when the targetmotion is deformed One after another the method ofusing correlation over filtering to calculate target trackingand recognition has gradually become the mainstreammode [11] In this way the information data is combinedwith the tracking model for the first time and the cor-relation filter is used to detect the responsiveness in theframe number of the behavior of the moving person andthe feature points of the data are captured It furtherimproves the recognition efficiency of the tracking modeland the speed can exceed 100 frames [12 13]

In recent years with the rapid development of deeplearning algorithms in the field of computers vision andrecognition [14] many researchers apply the convolu-tional neural network algorithm to classification functionwhich can extract a large number of feature points to trainthe tracking model so as to improve the performance ofbehavior recognition model [15] In the deep learningalgorithm the twin network has the characteristics ofsharing data which is suitable for repetitive similarprocessing tasks Siamese FC a twin network targettracking algorithm can combine the number of videoframes with the feature points extracted by the convo-lution network to obtain the predicted position of targetbehavior [16] Subsequently based on the above trackingmodel behavior recognition technology development thispaper proposes the Siamese-RPN algorithm which canaccurately identify the candidate box of athletesrsquo behaviortarget and reduce the influence of environmental factorson Athletesrsquo behavior recognition model [17ndash19]

Contributions in this paper are summarized as follows(1) a regional proposal network RPN algorithm combinedwith the twin regional proposal network Siamese algorithmis proposed to study the tracking and recognition technologyof athletesrsquo behavior (2) the proposed model can accuratelytake the athletersquos behavior as the target candidate box inmodel training and reduce the interference of environmentand other factors on model recognition and (3) comparedwith the traditional twin network method for sports be-havior recognition the proposed model in this paper canperform an offline operation and distinguish the interferencefactors of athletesrsquo background environment It can quicklycapture the characteristic points of athletesrsquo behavior as thedata input of the tracking model so it has excellent pop-ularization and application value

e following content is divided into four sections efirst section introduces the related work about the devel-opment of tracking model technology and behavior rec-ognition technology e second section discusses theSiamese-RPN algorithm to study the target tracking modeland athlete behavior recognition technology Firstly theproposed tracking algorithm by probability regression andthe proposed athlete behavior tracking recognition modelare optimized Finally an adaptive updating algorithm isproposed to build a tracker model and the simulation modelof athlete behavior recognition is established e thirdsection is the result analysis on the Siamese-RPN algorithmfor athlete target tracking e fourth section is the con-clusion to output some important results

2 Related Work

Athlete behavior recognition is a part of the application ofthe tracker model in the field of artificial intelligence rec-ognition [20] Firstly the target tracking algorithm in thetracker model is used to capture the behavior trajectory ofthe moving person and finally the simulation model isconstructed by using behavior recognition technology [21]In athlete target tracking it is necessary to predict the changeof target position in each frame number but in the trackingprocess there will be occlusion size change deformationcomplex background illumination and other influencingfactors [22] To realize the recognition of behavior we mustfirst solve the problem of target tracking and capture Inrecent years target tracking is based on the Siamese twinnetwork framework e core content is to capture therepetitive similar actions in the search area according to thelearning training template and matching rules [23] How-ever the common network algorithm will consume trackingtime and the tracking target is not accurate Based on thecommon network many subsequent researchers proposedthe introduction of the region proposal algorithm to obtain amore accurate target range [24] e tracker model con-structed by the Siamese-RPN algorithm can accuratelycapture the target behavior box and reduce the influencefactors of complex background rough the classificationand regression branch of target tracking finally build thebehavior recognition model of athletes

Xu et al invested a lot of research energy and funds intracking technology and behavior recognition technology[25] With the large demand for target tracking technologythe deep learning algorithm is also improving ey focusedon the research and development of target tracking andmonitoring projects mainly applied in the battlefield andcivil environment After the real-time monitoring model isestablished human behavior tracking and capturing is re-alized And it can identify whether people carry dangerousgoods and other functions

Deng et al mainly applied target tracking and behaviorrecognition technology in the public transport monitoringsystem aiming at tracking and interactive recognition ofpeoplersquos traffic behavior and vehicle driving [26] eycombined the deep learning algorithm with the feature pointanalysis technology to form the ability to accurately trackillegal vehicle behavior in a complex environment as well asthe realization of the illegal personnel behavior identifica-tion After analysis and identification the violation infor-mation is uploaded to the management and control system

Jia et al also applied tracking technology to trafficmanagementey mainly implanted the tracking algorithmand recognition algorithm into the radar system [27]rough the target tracking and behavior recognitioncontrol of traffic vehicles it can achieve the measurement ofvehicle distance angle speed and other functions Laterthey also used this technology in the VRmodel In the virtualenvironment eye movement is tracked and real-time in-teraction is realized

Visual tracking and behavior recognition are applied inautomation medical image artificial intelligence and other







ltotal α11139443

i1lcls + β1113944

3

i1lreg

lcls minus 1113936


n

lreg minus 1113936

ni 1113936

4j R tj minus t

lowastj1113872 1113873

n

(1)








p(y | x θ) 1

zθ(x)e

sθ(yx)


(3)


DKL(p q) 1113944N

i1p xi( 1113857log

p xi( 1113857

q xi( 11138571113888 1113889 (4)



p y | xiθ1113872 1113873dy



(5)


127times127times3

255times255times3

Detection frame

CNN

CNN

22times22times256

Template frame

6times6times256

Conv

Conv

Conv

Conv

classification

regression

1 group

Positive Negtive

K groupsdh dx dy dw



preparation

Networktraining

Outputmodel

end

Y

Is it the lastframe

Locate thetarget

Firstframe

Y

N

N

target tracking

featureextraction

initialization

Trackingtemplate


Determinethe next

frame area









fclsM(x z) h


fregM (x z) h





Bpredict xpre

ypre

wpre

hpre

( 1113857

xpre

xandan

+ dxreg

middot wandan

ypre

yandan

+ dyreg

middot handan

wpre

wandan

+ edwreg

hpre

handan

+ hdhreg

(8)














Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000







250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)

































ltotal α11139443

i1lcls + β1113944

3

i1lreg

lcls minus 1113936


n

lreg minus 1113936

ni 1113936

4j R tj minus t

lowastj1113872 1113873

n

(1)








p(y | x θ) 1

zθ(x)e

sθ(yx)


(3)


DKL(p q) 1113944N

i1p xi( 1113857log

p xi( 1113857

q xi( 11138571113888 1113889 (4)



p y | xiθ1113872 1113873dy



(5)


127times127times3

255times255times3

Detection frame

CNN

CNN

22times22times256

Template frame

6times6times256

Conv

Conv

Conv

Conv

classification

regression

1 group

Positive Negtive

K groupsdh dx dy dw



preparation

Networktraining

Outputmodel

end

Y

Is it the lastframe

Locate thetarget

Firstframe

Y

N

N

target tracking

featureextraction

initialization

Trackingtemplate


Determinethe next

frame area









fclsM(x z) h


fregM (x z) h





Bpredict xpre

ypre

wpre

hpre

( 1113857

xpre

xandan

+ dxreg

middot wandan

ypre

yandan

+ dyreg

middot handan

wpre

wandan

+ edwreg

hpre

handan

+ hdhreg

(8)














Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000







250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)




























p(y | x θ) 1

zθ(x)e

sθ(yx)


(3)


DKL(p q) 1113944N

i1p xi( 1113857log

p xi( 1113857

q xi( 11138571113888 1113889 (4)



p y | xiθ1113872 1113873dy



(5)


127times127times3

255times255times3

Detection frame

CNN

CNN

22times22times256

Template frame

6times6times256

Conv

Conv

Conv

Conv

classification

regression

1 group

Positive Negtive

K groupsdh dx dy dw



preparation

Networktraining

Outputmodel

end

Y

Is it the lastframe

Locate thetarget

Firstframe

Y

N

N

target tracking

featureextraction

initialization

Trackingtemplate


Determinethe next

frame area









fclsM(x z) h


fregM (x z) h





Bpredict xpre

ypre

wpre

hpre

( 1113857

xpre

xandan

+ dxreg

middot wandan

ypre

yandan

+ dyreg

middot handan

wpre

wandan

+ edwreg

hpre

handan

+ hdhreg

(8)














Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000







250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)


































fclsM(x z) h


fregM (x z) h





Bpredict xpre

ypre

wpre

hpre

( 1113857

xpre

xandan

+ dxreg

middot wandan

ypre

yandan

+ dyreg

middot handan

wpre

wandan

+ edwreg

hpre

handan

+ hdhreg

(8)














Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000







250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)



































Acc

urac

y ra

te (

)

Iterations1000 2000 3000 4000 5000 6000

10

20

30

40

50

60

70

7000







250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)































250

225

200

175

150

125

Accu

racy

fact

or




60

50

40

30

20

10

Feat

ure p

oint

dist

ance








60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)































60

50

40

30

20

10

Accu

racy

()




60

50

40

30

20

10

Accu

racy

()

10 20 30 40 50 60Overlap rate ()




5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)




























5 Conclusions


Data Availability




Acknowledgments


References












100

80

60

40

20Reco

gniti

on su

cces

s rat

e (

)



















































Documents

AthleteBehaviorRecognitionTechnologyBasedonSiamese-RPN