Post on 28-Feb-2018
FIMD: Fine-grained
Device-free Motion Detection
Jiang Xiao, Kaishun Wu, Youwen Yi, Lu Wang, Lionel M. Ni
Department of Computer Science and Engineering
Hong Kong University of Science and Technology
Dec. 18th, 2012
ICPADS 2012
Outline
• Introduction & Motivation
• Related Work
• System Design
• Performance Evaluation
• Conclusion
2
Device-free Motion Detection is in need!
I won’t be caught
3
Security for intrusion
Introduction
Current Detection Methods Camera Infrared
Environment
constraints
(e.g., dark, smoke)
Line-of-sight
Ultrasonic
High cost for
large environment
Sensors
High cost for
dense deployment
Zhang et. al
[PerCom’07, ’09, ’11]
4
Can’t work well in large scale, complicate,
typical indoor environment!
Motivation
WLAN Advantages
Low Cost
Easy Implementation
5
Outline
• Introduction & Motivation
• Related Work
• System Design
• Performance Evaluation
• Conclusion
6
Related Work
Static
Motion Happens
Idea: RSS becomes anomalous when the environment changes
Received Signal Strength Indicator (RSSI)
Youssef et. al. [MobiCom’07][PerCom’09][PerCom’12]
7
RF Signal
Transmitter Receiver
Limitations of RSS-based Techniques 1. Narrowband interference
high false alarm rate
Hard to distinguish narrowband interference from motion dynamic
8
2. RSS is of high variability
miss detection Node ID
Slow dynamic is easily hidden by the inherent RSS variance
A reliable method is in need.
resist from the narrowband interference
in the 2.4GHz band
temporal stable in static while
sensitive to a motion instantly
9
Challenge 1
Could we find such a reliable method?
10
2.4GHz
CSI Property 1 –
Frequency diversity
Key Insight
single value
RSSI Receiver CSIs
S/P FFT
Baseband
multiple values
RF band
11 CSI-based Indoor Localization: FILA[INFOCOM’12]
Data out OFDM
Transmitter
Channel
Data in OFDM
Receiver × +
Key Insight
In OFDM system, the received signal over multiple subcarriers is
amplitude phase
Channel gain CSI
12
RSS: variant CSI: relatively stable
Key Insight
CSI Property 2 –
Temporal Stability 13
CSI
am
plit
ud
e
RSS
I (d
Bm
)
Time Duration (s) Time Duration (s)
We want to harness fine-grained CSI
for device-free indoor motion detection.
14
Goal
To improve detection rate
To reduce false alarm rate
To improve detection accuracy
with low cost
15
Our Contributions
Exploit the possibility of CSI
Extract suitable CSI features
Propose burst detection alg.
16
Challenges
1.
Could we find a reliable method
for device-free motion detection?
2.
How to resist from
narrowband interference?
3.
How to distinguish motion event
from noise?
Outline
• Introduction & Motivation
• Related Work
• System Design
• Performance Evaluation
• Conclusion
17
Static Map Update AP
RF S
ignal
FIMD Design
CSI Feature Extraction
Burst Detection
Static Map Construction
False alarm
Filter
Static Map Update
DP CSI
Collection
Release Alarm
Server
18
Outline
• Introduction & Motivation
• Related Work
• System Design
– CSI Feature Extraction
– Burst Detection
– False Alarm Filter
• Performance Evaluation
• Conclusion
19
Challenge 2
How to extract the features of CSI that
resist from narrowband interference?
20
1. CSI Feature Extraction
Length n+1 Start
Process CSIs
W
Target: to extract CSI features can reflect static/dynamic patterns
21
H1,1, H2,1, …, Hk,1, Hk+1,1, … Hk+n,1, …
H1,2, H2,2, …, Hk2, Hk+1,2, … Hk+n,2, …
H1,30, H2,30, …, Hk,30 Hk+1,30, … Hk+n,30, …
…
…
…
…
…
subca
rrie
rs
[Hk, Hk+1, … Hk+n]
raw CSIs
1. CSI Feature Extraction
Process CSIs
Feature value vector V V= (eigen (C)/n+1)
Target: to extract CSI features can reflect static/dynamic patterns
22
Compute correlation factor
btw. each column of
[Hk, Hk+1, … Hk+n]
1st max & 2nd max eigenvalues are chosen!
1. CSI Feature Extraction
Target: to extract CSI features can reflect static/dynamic patterns
23
[Hk, Hk+1, … Hk+n]
V= (eigen (C)/n+1)
1. CSI Feature Extraction
CSI feature can reveal normal/motion behavior
Target: to extract CSI features can reflect static/dynamic patterns
24
1st max eigenvalue
2nd m
ax e
igen
val
ue
RSS: more variant
CSI: relatively stable
Robust to narrowband interference
1. CSI Feature Extraction
In presence of narrowband interference, e.g., Bluetooth
25
Outline
• Introduction & Motivation
• Related Work
• System Design
– CSI Feature Extraction
– Burst Detection
– False Alarm Filter
• Performance Evaluation
• Conclusion
26
“burst”
27
0 20 40 60 80 100 120 1400
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
static dynamic static
2n
d
eige
nva
lue
1
st
eige
nva
lue
28
0.4 0.5 0.6 0.7 0.8 0.9 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Feature X
Featu
re Y
Ground Truth
1st max eigenvalue
2nd m
ax e
igen
val
ue
Ground Truth
Challenge 3
How to distinguish motion event
from noise?
29
2. Burst Detection
Method: apply DBSCAN alg. [Ester, et al. (KDD’96)]
Fact: patterns of motion are diff. from static ones
No prior knowledge of the # of clusters
Discovery arbitrary shape clusters
Classify CSI
Patterns
Target: to monitor the “burst” motion occurrence
30
2. Burst Detection
Method: DBSCAN clustering
ε – max. radius of the neighborhood
Nε (Vi): {Vj belongs to Nε | Dist(Vi,Vj) ≤ ε}
minPts – min. # of points in an ε-neighbor to form a cluster
for each point in a cluster, ε-neighbouor has to contain
≥ minPts.
Density = the # of points within a specified radius ε
Target: to monitor the “burst” motion occurrence
31
2. Burst Detection
Euclidean distance for DBSCAN clustering
Target: to monitor the “burst” motion occurrence
32
0.4 0.5 0.6 0.7 0.8 0.9 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Feature X
Featu
re Y
DBSCAN Result
2nd m
ax e
igen
val
ue
1st max eigenvalue
Outline
• Introduction & Motivation
• Related Work
• System Design
– CSI Feature Extraction
– Burst Detection
– False Alarm Filter
• Performance Evaluation
• Conclusion
33
3. False Alarm Filter
Target: to reduce the false alarm rate
Burst
false alarm!
Burst Burst
34
true motion!
Fact: a single motion instance always lasts a short period
Outline
• Introduction & Motivation
• Related Work
• System Design
• Performance Evaluation
• Conclusion
35
Hardware
Commercial NICs, APs
Software
Linux 2.6.38 kernel, Matlab, Python
Testbeds
2 typical indoor
environments in HKUST
Size
–7m × 11m (Lab)
–32.5m×1.5m (Corridor)
Experimental Setup
802.11n Router Intel WiFi Link 5300
Corridor Lab
36
Evaluation Metrics
True Positive
(TP) Rate
False Positive (FP) Rate
False Negative
(FN) Rate
Condition: Motion Condition: Static Te
st o
utco
me:
Motion
Test
out
com
e:
Sta
tic Ignore
37
Effectiveness of FIMD
Detection performance w.r.t false alarm is high!
1%
DR > 70%
9%
DR>90%
Detection performance w.r.t false alarm (ROC Curve)
38
Effectiveness of FIMD
The longer the window length, the less sensitivity
Influence of the length of sliding window
39
Effectiveness of FIMD
The longer the window length, the lower FP rate
Influence of the length of sliding window
40
CSI vs. RSS
Without Narrowband Interference
41
Lab Corridor0
0.2
0.4
0.6
0.8
1D
ete
cti
on
Rate
CSI-based
RSSI-based
Refer to [PerCom’12] Youssef et. al. RASID system
• kernel density-based approach
CSI vs. RSS
No Motion, with Narrowband Interference
Bluetooth
42
SW=5 SW=8 SW=11 SW=14 SW=17 SW=200
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Fals
e P
osi
tiv
e
CSI-based
RSS-based
Outline
• Introduction & Motivation
• Related Work
• System Design
• Performance Evaluation
• Conclusion
43
We presented a novel device-free indoor motion detection
system with commodity hardware in large indoor scenarios.
We leverage both the frequency diversity and temporal
stability of CSI to enhance detection performance.
Experimental results show that CSI-based approach is
superior to RSS-based approach in RF domain.
Conclusion
44
Thanks.
Questions?
jxiao@cse.ust.hk
PhD Candidate @ Hong Kong University of Sci.& Tech.