LiTell: Robust Indoor Localization Using Unmodified Light Fixtures

Transcription

1 LiTell: Robust Indoor Localization Using Unmodified Light Fixtures Chi Zhang, Xinyu Zhang Department of Electrical and Computer Engineering University of Wisconsin-Madison MobiCom'16

2 Indoor Localization: Enabling Technology 1% 1% discount discount near near you you

3 RF-based Localization Multiple factors compromise robustness Multipath propagation Environmental dynamics Device heterogeneity

4 Visible Light Localization LoS propagation: robust, multipath-free Densely deployed landmarks: high accuracy Existing works show 1cm to 1m accuracy1 [1] Y.-S. Kuo, et al., Luxapose, MobiCom'14; L. Li, et al, Epsilon, NSDI'14

5 Deployment Challenges LED beacons needs extra circuits Adds to manufacturing cost Image of lights: S. Schmid, et al. Linux Light Bulbs, VLCS'15

6 Deployment Challenges LED beacons needs extra circuits Adds to manufacturing cost Huge retrofitting effort Image of lights: S. Schmid, et al. Linux Light Bulbs, VLCS'15

7 Deployment Challenges Source: US DoE

8 LiTell Landmark: unmodified fluorescent light fixtures Sensor: smartphone cameras Zero retrofitting cost!

9 Localization Using Incumbent Lights Fluorescent light driver operates at high frequency This frequency is reflected in the light emission It is unique due to manufacturing variations, hence we call it characteristic frequency (CF)

10 Localization Using Incumbent Lights Fluorescent light driver operates at high frequency This frequency is reflected in the light emission It is unique due to manufacturing variations, hence we call it characteristic frequency (CF) 2 Lab FL GE CFL Aisle FL GE CFL RSS (1dB/DIV) Normalized Intensity Lab Tube Aisle Tube 4 6 Time ( s) Frequency (khz) 11 12

11 Characteristic Frequency (CF) lights in an office building CDF Frequency (khz) 7 lights in our lab Time (week) All Lights 1 1 Same Model 1 1k f (Hz) 1k 1k

12 Characteristic Frequency (CF) e 84.4 m i t r ve o 84.2 e l tab S Time (week) 5 lights in an office building All Lights CDF Frequency (khz) 7 lights in our lab 14 Same Model 1 Hz.8 2 > f Δ.6 se i w r ai p.4 % 9 9 : e.2ivers D k 1k 1k f (Hz)

13 Capturing CF on Phones COTS phones do not have high speed light sensors. Instead, we can use cameras. Key Challenges: Low sampling rate Low SNR

14 Sampling CF Intensity Rows Intensity Rolling shutter: a primer Time Columns Key benefits: higher sampling rate, reliable timing Time

15 Sampling CF Typical smartphone has sampling rate of 56 ~ 15 khz RSS CFs are around 8 ~ 1 khz, need Nyquist sampling rate > 2 khz Frequency

16 Sampling CF Typical smartphone has sampling rate of 56 ~ 15 khz CFs are around 8 ~ 1 khz, need Nyquist sampling rate > 2 khz Solution: leverage aliasing effect (If the analog bandwidth suffices!) RSS Observation: CF is extremely sparse Frequency

17 Sampling CF Adaptive exposure to avoid notches De-aliasing mechanism to disambiguate CF Frequency CF Range 1/37k 1/75k -1 RSS (db) Camera's analog bandwidth: up to 2 khz RSS Frequency (khz)

18 Boosting SNR Noise from dark areas / ambient sunlight reduces SNR. SNR (db) ISO Number

19 Boosting SNR Noise from dark areas / ambient light reduces SNR. SNIS: fast spatial noise removal tailored for lights SNR (db) 8 No SNIS SNIS ISO Number

20 Boosting SNR Interleaving: from camera Intensity Intensity Artifacts: Time Spatial patterns: from light cover Time

21 Localization Identifying individual lights: collision rate grows as more lights are added in database Collision Probability (%) Error control mechanism: using pairs of consecutive lights single 1 4 pair 8 12 Number of Lights 16 2

22 Localization with Respect to a Specific Light Observation: images are scaled from physical structure Light size in pixels h L C Relative position in pixels Light physical size Scaling factor Physical relative position Light position Global position

23 Implementation LiTell Android App LiTell Server Send ceiling light samples Return current location

24 Experimental Evaluation Robustness Accuracy User study

25 Robustness Distance: up to 2m with typical smartphone camera

26 Robustness Distance: up to 2m with typical smartphone camera Ambient light: works by the window with direct sun light Phone

27 Robustness Distance: up to 2m with typical smartphone camera Ambient light: works by the window with direct sun light User behavior: unaffected by normal pose and walking Phone

28 Light Identification Accuracy Landmark Identification: 9.3% accuracy in typical places Office Building Grocery Store Underground Parking Lot

29 Localization Precision Localization w.r.t. each light: 1cm in ideal case, ~25cm when walking Stable 1 Walking CDF Error in Euclidean distance (cm) 35

30 User Study 1 office building, 2 floors, 6 targets, 1 volunteers. Find room by number. Control Group LiTell Group

31 User Study 1 office building, 2 floors, 6 targets, 1 volunteers. Find room by number. Control Group ge a r e v a n si s e l m m) s l 1 e 5 v 3 Tra th a p t s e t (shor ime t % 5 er v o e v a Can s le c a r o o le t b a r a p Com LiTell Group

32 Conclusion LiTell uses fluorescent lights as zero-cost location landmarks LiTell turns smartphone cameras into high speed light sensors Deterministic visible light channel assures LiTell's accuracy and robustness LiTell brings accurate visible light localization to today's buildings

33 Thanks!