Indoor Localization Based on Response Rate of ... - Semantic Scholar
Indoor Localization Based on Response Rate of Bluetooth Inquiries
Mortaza S. Bargh & Robert de Groote Telematica Instituut The Netherlands
19 September 2008
Outline • • • •
Motivations Approach/solution Results Conclusion
Motivations •
Colleague Radar™ application – locate employees in the building for the colleagues
•
Indoor localization – No GPS – Ongoing research
•
Bluetooth being pervasive – Cell phones – (Always) with people – Have Bluetooth – Being discoverable
Indoor localization •
Successful indoor localization systems – Integrate smoothly with existing infrastructures – Preferably require no upgrade of user devices – Need no excessive hardware installation – Use existing technologies – Impose low power consumption on mobile devices – Use low cost infrastructure
•
Bluetooth based approaches – Based on RSSI – Based on LQ (link quality)
Bluetooth Inquiry Response Rate
•
IRR (Inquiry Response Rate) = the percentage of inquiry responses to total inquiries in a given observation window
An experiment •
Each row: – 240 sliding windows (slides every ~ 5 seconds) – Window size = 50 inquiries position
IRR average out of 50 inquiries
IRR variance out of 50 inquiries
1
48.4
0.9
2
48.5
0.6
3
49.4
0.5
4
46.9
0.7
5
43.1
2.3
6
36.8
4.8
7
33.3
2.3
8
NULL
NULL
9
NULL
NULL
Our setting
detected by
A classification problem: location fingerprint • •
Obtain location fingerprint L Compare it with training fingerprints Tk (of room k=1, 2, …) – Kullback-Liebler (KL) measure – Jensen-Shannon (JS) distance measure
! " # #$
%
Typical outputs of the classification process choose a PMF (or room) that minimizes the divergence
Test 1: full coverage rooms with dongles rooms without dongles test room ref device unknown device
7
at 16:00
6
3
at 15:00
at 14:00
2 at 11:00
at 13:00
D1
5
at 12:00
4
1
at 10:00 D3
Test 2: partial coverage rooms with dongles rooms without dongles
G12
test room ref device unknown device
at 16:34
5
at 17:33
4 6
D1
at 15:32 at 11:00
at 12:06
1 Mortaza
at 13:31
2
at 14:38
3
Location estimation results – (1) Full coverage Using Kullback-Liebler (KL) divergence measure Training window: 30’ and 5’
accuracy (%)
• • •
102 100 98 96 94 92 90 88 86 84
30'training 5'training
1
2
3
licalization window (minutes)
4
Location estimation results – (1) 2 problems with basic KL method: – sensitivity to the timing of training data: a drop of accuracy to 83% (WT=30’) or to 77% (WT=5’) – sensitivity to BT dongle coverage: accuracy 15…45%
accuracy (%)
•
102 100 98 96 94 92 90 88 86 84
30'training 5'training
1
2
3
licalization window (minutes)
4
Location estimation results – (2) •
Using Jensen-Shannon (JS) distance measure 105
accuracu (%)
100 95
KL 30'training training KL 5' 30' training JS 30' training KL 5'training JS 5'training
90 85 80 75 1
2
3
localization window (minutes)
4
Location estimation results – (3) JS measure: (1) change of training data
• 105
100
accurac (%)
95
JS 30'training JS 5'JS training 30'training
90
JSfresh 5'training JS 30' training
JS 5'fresh training 85
80
75
11
22
3 3
localization window localization window
4
4
Location estimation results – (4) •
JS measure: – (2) partial dongle coverage – (3) training window
Conclusions • IRR is a valid approach • Robust with respect to device change • Time consuming, but acceptable for some application domains • Good training fingerprints are not necessarily the most recent ones • Accuracy of two best estimates is almost 100% • Increasing observation window size increases accuracy up to a limit • Better performance requires a dedicated Bluetooth network "& ' ' #$
$
$
Measured network characteristics: Response Rate •
Response Rate (RR): – “the percentage of times that a given Access Point was heard in all of the WiFi scans at a specific distance from that AP” [CHE05] – “the frequency of received measurements over time from a given base station” [KJA 07]
Some formulas • •
L
PMFs of observed location and room k: and Kullback-Liebler (relative entropy) measure:
D( L || Tk ) =
M m =1
D( L(d m ) || Tk (d m ))
D( L(d m ) || Tk (d m ) ) = pL ,m log •
pL , m pTk ,m
+ (1 − pL ,m ) log
Jensen-Shannon distance:
1 1 D( L || M ) + D(Tk || M ) 2 2 1 M = ( L + Tk ) 2
JSD( L || Tk ) =
Tk
1 − pL , m 1 − pTk ,m
Location estimation results – (1) Full coverage Using Kullback-Liebler (KL) divergence measure Training window: 30 minutes (30’)
accuracy (%)
• • •
100,5 100 99,5 99 98,5 98 97,5 97 96,5 96 95,5
top-1 top-2
1
2
3
4
localization window (minutes)
Summary • • • • •
•
Localization of stationary users (at this stage) Indoor localization for multi floor buildings with dense deployment of BT sensors Infrastructure-based and network-based Direct location (without any transformation) Network characteristics used: response rate – the frequency of received measurements over time from a specific base station (we did not address privacy issues)
Test result summary •
•
JS measure – WL=3 minutes – WT=10 minutes Performance: – Good coverage • Top-1: 97.82% same device 99.84% • Top-2: 100% same device 100% – Partial coverage • Top-1: 75% same device 99.27 • Top-2: 99.88 same device 100
System overview room fingerprint collection
radio map (room RR fingerprints)
location fingerprint detection
location estimation
movement detection
location estimate(s)
Mortaza S. Bargh & Robert de Groote Telematica Instituut The Netherlands
19 September 2008
Outline • • • •
Motivations Approach/solution Results Conclusion
Motivations •
Colleague Radar™ application – locate employees in the building for the colleagues
•
Indoor localization – No GPS – Ongoing research
•
Bluetooth being pervasive – Cell phones – (Always) with people – Have Bluetooth – Being discoverable
Indoor localization •
Successful indoor localization systems – Integrate smoothly with existing infrastructures – Preferably require no upgrade of user devices – Need no excessive hardware installation – Use existing technologies – Impose low power consumption on mobile devices – Use low cost infrastructure
•
Bluetooth based approaches – Based on RSSI – Based on LQ (link quality)
Bluetooth Inquiry Response Rate
•
IRR (Inquiry Response Rate) = the percentage of inquiry responses to total inquiries in a given observation window
An experiment •
Each row: – 240 sliding windows (slides every ~ 5 seconds) – Window size = 50 inquiries position
IRR average out of 50 inquiries
IRR variance out of 50 inquiries
1
48.4
0.9
2
48.5
0.6
3
49.4
0.5
4
46.9
0.7
5
43.1
2.3
6
36.8
4.8
7
33.3
2.3
8
NULL
NULL
9
NULL
NULL
Our setting
detected by
A classification problem: location fingerprint • •
Obtain location fingerprint L Compare it with training fingerprints Tk (of room k=1, 2, …) – Kullback-Liebler (KL) measure – Jensen-Shannon (JS) distance measure
! " # #$
%
Typical outputs of the classification process choose a PMF (or room) that minimizes the divergence
Test 1: full coverage rooms with dongles rooms without dongles test room ref device unknown device
7
at 16:00
6
3
at 15:00
at 14:00
2 at 11:00
at 13:00
D1
5
at 12:00
4
1
at 10:00 D3
Test 2: partial coverage rooms with dongles rooms without dongles
G12
test room ref device unknown device
at 16:34
5
at 17:33
4 6
D1
at 15:32 at 11:00
at 12:06
1 Mortaza
at 13:31
2
at 14:38
3
Location estimation results – (1) Full coverage Using Kullback-Liebler (KL) divergence measure Training window: 30’ and 5’
accuracy (%)
• • •
102 100 98 96 94 92 90 88 86 84
30'training 5'training
1
2
3
licalization window (minutes)
4
Location estimation results – (1) 2 problems with basic KL method: – sensitivity to the timing of training data: a drop of accuracy to 83% (WT=30’) or to 77% (WT=5’) – sensitivity to BT dongle coverage: accuracy 15…45%
accuracy (%)
•
102 100 98 96 94 92 90 88 86 84
30'training 5'training
1
2
3
licalization window (minutes)
4
Location estimation results – (2) •
Using Jensen-Shannon (JS) distance measure 105
accuracu (%)
100 95
KL 30'training training KL 5' 30' training JS 30' training KL 5'training JS 5'training
90 85 80 75 1
2
3
localization window (minutes)
4
Location estimation results – (3) JS measure: (1) change of training data
• 105
100
accurac (%)
95
JS 30'training JS 5'JS training 30'training
90
JSfresh 5'training JS 30' training
JS 5'fresh training 85
80
75
11
22
3 3
localization window localization window
4
4
Location estimation results – (4) •
JS measure: – (2) partial dongle coverage – (3) training window
Conclusions • IRR is a valid approach • Robust with respect to device change • Time consuming, but acceptable for some application domains • Good training fingerprints are not necessarily the most recent ones • Accuracy of two best estimates is almost 100% • Increasing observation window size increases accuracy up to a limit • Better performance requires a dedicated Bluetooth network "& ' ' #$
$
$
Measured network characteristics: Response Rate •
Response Rate (RR): – “the percentage of times that a given Access Point was heard in all of the WiFi scans at a specific distance from that AP” [CHE05] – “the frequency of received measurements over time from a given base station” [KJA 07]
Some formulas • •
L
PMFs of observed location and room k: and Kullback-Liebler (relative entropy) measure:
D( L || Tk ) =
M m =1
D( L(d m ) || Tk (d m ))
D( L(d m ) || Tk (d m ) ) = pL ,m log •
pL , m pTk ,m
+ (1 − pL ,m ) log
Jensen-Shannon distance:
1 1 D( L || M ) + D(Tk || M ) 2 2 1 M = ( L + Tk ) 2
JSD( L || Tk ) =
Tk
1 − pL , m 1 − pTk ,m
Location estimation results – (1) Full coverage Using Kullback-Liebler (KL) divergence measure Training window: 30 minutes (30’)
accuracy (%)
• • •
100,5 100 99,5 99 98,5 98 97,5 97 96,5 96 95,5
top-1 top-2
1
2
3
4
localization window (minutes)
Summary • • • • •
•
Localization of stationary users (at this stage) Indoor localization for multi floor buildings with dense deployment of BT sensors Infrastructure-based and network-based Direct location (without any transformation) Network characteristics used: response rate – the frequency of received measurements over time from a specific base station (we did not address privacy issues)
Test result summary •
•
JS measure – WL=3 minutes – WT=10 minutes Performance: – Good coverage • Top-1: 97.82% same device 99.84% • Top-2: 100% same device 100% – Partial coverage • Top-1: 75% same device 99.27 • Top-2: 99.88 same device 100
System overview room fingerprint collection
radio map (room RR fingerprints)
location fingerprint detection
location estimation
movement detection
location estimate(s)
