Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
1/23
Indoor Localization Without InfrastructureUsing the Acoustic Background Spectrum
Stephen P. Tarzia∗ Peter A. Dinda∗
Robert P. Dick† Gokhan Memik∗
∗Northwestern University, EECS Dept.
†University of Michigan, EECS Dept.
Presented at MobiSys 2011Bethesda, MD, USA
June 30, 2011
http://empathicsystems.org
2/23
Video demonstration of Batphone app
Current acoustic
fingerprint
Location estimate
3/23
Definition: indoor localization without infrastructure
Given:
X A smartphone
X A building composed of many rooms
X At least one prior visit to each room for training
Without:
× Specialized hardware
× Anything installed in the environment
× Cooperation from the building owner
Goal:
I Determine which room the smartphone is currently located in
4/23
Summary
Motivation:
I Indoor localization is important
I Wi-Fi is imperfect and not always available
I Improved accuracy is desired
Distinctive elements of our method:
I Listen to background sounds
I Look at frequency domain
I Rank-order filter for noise
Results:
I 69% accuracy for 33 rooms using sound alone
I Publicly-available app
I Effectively combined Wi-Fi and sound
5/23
Related Work: mobile acoustic sensing
M. Azizyan, I. Constandache, and R.R. Choudhury.SurroundSense: mobile phone localization via ambiencefingerprinting. MobiCom’09.
I Characterized rooms by loudness distribution
I Did not use sound exclusively
H. Lu, W. Pan, N.D. Lane, T. Choudhury, and A.T. Campbell.SoundSense: scalable sound sensing for people-centric applicationson mobile phones. MobiSys’09.
I Focused on transient sounds
I Activity detection, not localization
6/23
Acoustic Background Spectrum (ABS)
A location fingerprint should be:
I Distinctive
I rEsponsive
I Compact
I Efficiently-computable
I Noise-robust
I Time-invariant
X 69% matching accuracy
X 4–30 second sample
X ∼1 kB per fingerprint
X ∼12% mobile CPU usage
∼ sometimes can adapt
X tested on different days
6/23
Acoustic Background Spectrum (ABS)
A location fingerprint should be:
I Distinctive
I rEsponsive
I Compact
I Efficiently-computable
I Noise-robust
I Time-invariant
X 69% matching accuracy
X 4–30 second sample
X ∼1 kB per fingerprint
X ∼12% mobile CPU usage
∼ sometimes can adapt
X tested on different days
7/23
Signal Processing
Discard rows > 7 kHz
Record audio samples
Divide samples into frames
Compute power spectrum of each frame
time
time
fre
q.
Sort each remaining row
spectrogram
audio sample time series
fre
q.
increasing magnitude Extract 5th percentile columnand take logarithm
[ ]= Acoustic Background Spectrum
microphone input
* * * * * * * * * * * * * * * * * *Multiply frames by a window function
Sta
nd
ard
sp
ectr
al a
nal
ysis
AB
S f
ing
erp
rin
t ex
trac
tio
n
8/23
ABS Fingerprints
Various rooms
10-2
100
102
104
106
108 Room 15
10-2
100
102
104
106
108
norm
aliz
ed, lo
g-s
cale
ener
gy
Room 16
10-2
100
102
104
106
108
0 1 2 3 4 5 6 7
frequency (kHz)
Room 17
Different positions and days
10-2
100
102
104
106
108 Room 15
10-2
100
102
104
106
108
norm
aliz
ed, lo
g-s
cale
ener
gy
Room 16
10-2
100
102
104
106
108
0 1 2 3 4 5 6 7
frequency (kHz)
Room 17
8/23
ABS Fingerprints
Various rooms
10-2
100
102
104
106
108 Room 15
10-2
100
102
104
106
108
norm
aliz
ed, lo
g-s
cale
ener
gy
Room 16
10-2
100
102
104
106
108
0 1 2 3 4 5 6 7
frequency (kHz)
Room 17
Different positions and days
10-2
100
102
104
106
108 Room 15
10-2
100
102
104
106
108
norm
aliz
ed, lo
g-s
cale
ener
gy
Room 16
10-2
100
102
104
106
108
0 1 2 3 4 5 6 7
frequency (kHz)
Room 17
9/23
Experimental Platforms
(a) Zoom H4n (b) Apple iPod Touch
10/23
Experimental Rooms
0
5
10
15
20
officelounge
computer lab
classroomlecture hall
Inst
ance
s
Room type
0 2 4 6 8
10 12 14 16
4 8 16 32 64 128 256 512
Inst
ance
s
Maximum room capacity
11/23
Fingerprint-based localization
Supervised learning with two phases:
I Training – gather labeled fingerprints
I Testing/operation – observe new, unlabeled fingerprints
I Experiments use leave-one-out simulation
Our classifier:
I Euclidean distance metric for comparing fingerprints(equivalent to RMS error)
I Nearest-neighbor classification
In summary
To guess the current location find the “closest” fingerprint in adatabase of labeled fingerprints.
12/23
Accuracy Scaling
0 10 20 30 40 50 60 70 80 90
100
2 4 8 17 33
Acc
ura
cy (
%)
Number of rooms in database (log scale)
Proposed Acoustic Background SpectrumSurroundSense [Azizyan et al.]
Random chance
I SurroundSense is used in a way not intended by the authors:using the microphone alone
13/23
ABS Parameters
Presented now:
I Filter rank
I Listening time
I Fingerprintsize/resolution
In paper:
I Frequency band
I Distance metric
I Spectrogram window
14/23
Rank-order Filtering
0
20
40
60
80
100
meanmin
p05p10
p25median
p95max
Acc
ura
cy (
%) Fingerprint type
standard spectrumproposed rank-order filtered spectrum
I 33 rooms in database
I Rank-order filters outperforms simple mean⇒ our transient noise filtering technique is effective
15/23
Listening time
0
10
20
30
40
50
60
70
80
1 2 4 8 15 30
Acc
ura
cy (
%)
Sample time, in seconds
16/23
Frequency resolution
0
10
20
30
40
50
60
70
80
1 10 100 1000
Acc
ura
cy (
%)
Frequency bins / fingerprint vector length
17/23
Batphone app in iTunes store
I Uses a 10 secondsliding window
I Streaming signalprocessing
I Combines Wi-Fi withacoustic fingerprint
18/23
Batphone results
0
20
40
60
80
100
Linear combination
ABSCommercial W
i-Fi
Random
Acc
ura
cy (
%)
Batphone localization accuracy
proposed methods
I 43 rooms in databaseI Similar ABS accuracy for iPod and audio recorderI Linear combination of Wi-Fi and ABS works wellI Didn’t compare to state-of-the-art Wi-Fi localization
19/23
Orthogonality of Wi-Fi and Acoustics
2D histograms of physical and fingerprint distances
0 20 40 60 80
100 120 140 160
0 10 20 30 40 50 60
Wi-
Fi
dis
tan
ce (
m)
Real physical distance (m)
0 100 200 300 400 500 600 700 800
0 10 20 30 40 50 60
AB
S d
ista
nce
(d
B)
I Wi-Fi fingerprints from distant rooms are always different
I ABS fingerprints from nearby rooms can be quite different
21/23
Conclusion
ABS fingerprint can be used for indoor localizationand it requires no infrastructure
See the paper for:
I Full parameter study
I Noise robustness experiment
I More Wi-Fi combination results
I Battery-drain measurements
22/23
Future work
I Improved noise robustnessI Train the various noise statesI Adaptively chose fingerprint frequency band
I Use floorplan and history: Markov movement model
I Isolate factors that contribute to the ABS
I Add other sensors, as in SurroundSense
I In-pocket detection
23/23
Thanks!
For your enjoyment:
I App on the iTunes store:search for Batphone
I Listening demo athttp://stevetarzia.com/listen
I Data and Matlab scripts athttp://stevetarzia.com
I See our other projects athttp://empathicsystems.org
24/23
Room 1: Ford 2221 (office)
Accuracy:100%
Room 2: Ford 2227 (lounge)
Accuracy: 38%
Room 3: Ford 2230 (office)
Accuracy: 25%
Room 4: Ford 3317 (lounge)
Accuracy:100%
Room 5: Tech F235 (classroom)
Accuracy: 0%
Room 6: Tech F252 (computer lab)
Accuracy:100%
Room 7: Tech L158 (classroom)
Accuracy: 88%
Room 8: Tech L160 (classroom)
Accuracy:100%
Room 9: Tech L168 (classroom)
Accuracy: 0%
Room 10: Tech L170 (classroom)
Accuracy: 0%
1 2 3 4 5 6 7
Room 11: Tech L211 (lecture hall)
Accuracy: 50%
Room 12: Tech L221 (classroom)
Accuracy:100%
Room 13: Tech L251 (classroom)
Accuracy: 0%
Room 14: Tech L361 (lecture hall)
Accuracy: 75%
Room 15: Tech LG62 (classroom)
Accuracy:100%
Room 16: Tech LG66 (classroom)
Accuracy:100%
Room 17: Tech LG68 (classroom)
Accuracy:100%
Room 18: Tech LG76 (classroom)
Accuracy:100%
Room 19: Tech LR2 (lecture hall)
Accuracy: 88%
Room 20: Tech LR3 (lecture hall)
Accuracy:100%
Room 21: Tech LR4 (lecture hall)
Accuracy: 88%
1 2 3 4 5 6 7
Frequency (kHz)
Room 22: Tech LR5 (lecture hall)
Accuracy: 63%
Room 23: Tech M120 (classroom)
Accuracy:100%
Room 24: Tech M128 (classroom)
Accuracy: 50%
Room 25: Tech M152 (classroom)
Accuracy: 63%
Room 26: Tech M164 (classroom)
Accuracy:100%
Room 27: Tech M166 (classroom)
Accuracy: 88%
Room 28: Tech M338 (computer lab)
Accuracy: 0%
Room 29: Tech M345 (lecture hall)
Accuracy: 0%
Room 30: Tech M349 (classroom)
Accuracy:100%
Room 31: Tech MG51 (computer lab)
Accuracy:100%
Room 32: Tech RYAN (lecture hall)
Accuracy:100%
1 2 3 4 5 6 7
Room 33: Tech XPRS (lounge)
Accuracy: 75%
25/23
Parameter Study
(a) Frequency band Accuracy
full (0–48 kHz) 59.8%audible (0–20 kHz) 64.8%
low (0–7 kHz)* 69.3%very low (0–1 kHz) 61.0%
(0–600 Hz) 51.5%(0–400 Hz) 44.3%(0–300 Hz) 40.9%(0–200 Hz) 30.7%(0–100 Hz) 15.5%
high (7–20 kHz) 28.4%ultrasonic (20–48 kHz) 25.0%
26/23
Parameter Study (cont.)
(b) Distance metric Accuracy
Euclidean* 69.3%city block 66.7%
(c) Spectrogram window Accuracy
rectangular 65.2%Hamming* 69.3%
Hann 68.2%Blackman 67.4%
27/23
Optimal Parameters
symbol meaning optimal value Batphone
Rs sampling rate 96 kHz 44.1 kHznspec spectral resolution 2048 bins 1024 binsnfp ABS size 299 bins 325 binstspec frame size 0.1 s 0.1 stsamp sampling time 30 s 10 s
frequency band 0–7 kHz 0–7 kHzwindow function Hamming rectangular
28/23
Dealing with Noise by changing Frequency band
Occupancy Noise State
Frequency band Quiet Conversation Chatter
(a) Tech LR5 lecture hall
low (0–7 kHz) 89.2% 2.5% 0.0%(0–300 Hz) 75.7% 63.4% 0.0%
(b) Ford 3.317 lounge
low (0–7 kHz) 98.2% 47.2% —(0–300 Hz) 87.7% 79.2% —
29/23
0
0.2
0.4
0.6
0.8
1
1 10
Cum
ula
tiv
e p
robab
ilit
y
Ranking of correct room (log scale)
Error characteristics of localization methods
Linear combinationTwo-step combination
ABSWi-Fi
Random chance
I Batphone (ABS) beats Wi-Fi at fine granularityI Wi-Fi beats Batphone (ABS) at coarse granularity.I Combination is best overall