1. Approximate Initialization of Camera Sensor Networks Purushottam Kulkarni K.R. School of Information Technology Indian Institute of Technology, Bombay Deepak Ganesan, Prashant Shenoy Department of Computer Science University of Massachusetts, Amherst

2. UNIVERSITY OF MASSACHUSETTS, AMHERST 1 Camera Sensor Networks  Wireless network of tetherless imaging sensors ◊Directional camera sensors  Applications ◊Ad-hoc Surveillance ◊Environmental and habitat monitoring  Tasks ◊Object detection, recognition, tracking Field-of -view

3. UNIVERSITY OF MASSACHUSETTS, AMHERST 2 Camera Initialization  Pre-requisite for applications tasks ◊Localization, requires camera coordinates ◊Duty-cycling, requires set/overlap of neighbors ◊Tracking, requires overlap location with neighbors  Initialization parameters: ◊Extrinsic: location, orientation ◊Intrinsic: focal length, skew, principal point ◊Set of neighbors ◊Degree of overlap

4. UNIVERSITY OF MASSACHUSETTS, AMHERST 3 Factors Effecting Initialization  Computation Capability  Infrastructure Support ◊Range Estimation ◊Landmarks sync range estimation pulse Cricket Mote  Camera Sensor Networks  Landmarks hard to find  Resource-constraints  Estimation of accurate parameters not possible

5. UNIVERSITY OF MASSACHUSETTS, AMHERST 4 Problem Statement  Given a CSN with, ◊ Limited computation capability ◊ No/minimal infrastructure support is it possible to initialize cameras to enable applications?  Proposed solution: Approximate Initialization ◊ Estimate relative relationships between cameras ◊ Use only picture taking capability and local processing of camera

6. UNIVERSITY OF MASSACHUSETTS, AMHERST 5 Outline  Introduction & Problem Statement  Approximate Initialization Parameters  Estimation Techniques  Experimental Evaluation

7. UNIVERSITY OF MASSACHUSETTS, AMHERST 6 Approximate Initialization  Degree of Overlap ◊Fraction of viewing region that overlaps with neighboring cameras ◊k-overlap: fraction of viewing region overlapping by k cameras  Approximates level of sensing redundancy with neighboring cameras

8. UNIVERSITY OF MASSACHUSETTS, AMHERST 7 Approximate Initialization  Region of Overlap ◊spatial volume within viewing region that overlaps with another camera ◊Degree of overlap does not estimate which portion overlaps with neighbors  Approximates location of neighbors and spatial region of overlap Approximate estimates can support application requirements

9. UNIVERSITY OF MASSACHUSETTS, AMHERST 8 Duty-Cycling  Operate in ON-OFF cycles  d:duty-cycling parameter (ON fraction)  O i k : k-overlap of camera  Parameter in proportion to degree of overlap (extent of redundant coverage)

10. UNIVERSITY OF MASSACHUSETTS, AMHERST 9 Triggered Wakeup  Wakeup scenarios ◊Object tracking ◊Reliable detection  Region of overlap can determine potential cameras C1 C2 C3 Object

11. UNIVERSITY OF MASSACHUSETTS, AMHERST 10 Estimating k-overlap  k-overlap: ratio of randomly placed reference objects viewed simultaneously by k cameras  cameras take pictures  determine if object can be viewed simultaneously by other cameras Camera 3 Camera 2 Camera 1 reference points viewed at camera i reference points viewed by k cameras

12. UNIVERSITY OF MASSACHUSETTS, AMHERST 11 Skewed Distributions  Fraction of points does not represent fraction of overlap ◊Points in sparse region actually represent larger region ◊Error in estimation due to non-uniform distribution Camera 3 Camera 2 Camera 1 : 2/3 : 1/9 : 2/9 : 1/2 : 1/4 Estimated Exact

13. UNIVERSITY OF MASSACHUSETTS, AMHERST 12 Handling Skewed Distributions  Assign area of each polygon as weight to corresponding reference point ◊Weight in proportion to density of neighbors Total weight of reference points viewed at camera i Total weight of reference points viewed by k cameras

14. UNIVERSITY OF MASSACHUSETTS, AMHERST 13 Approximate 3D Voronoi Tessellation  Accurate 3D tessellation ◊Compute intensive  Approximation ◊Discretize volume into cubes ◊Calculate closest reference point ◊ Add volume to closest ◊Points in spare regions will have higher weights

15. UNIVERSITY OF MASSACHUSETTS, AMHERST 14 Determining Region of Overlap  where the overlap exists between cameras  region of overlap is the union of cells containing all simultaneously visible points C1 C2

16. UNIVERSITY OF MASSACHUSETTS, AMHERST 15  Estimate d r using object size, image size, focal length  & have same orientation  Use unit vector along and d r to estimate location Estimating Reference Point Location f s s’ Lens P (-x,-y,-f) O R (unknown location) Image plane

17. UNIVERSITY OF MASSACHUSETTS, AMHERST 16 Outline  Introduction & Problem Statement  Approximate Initialization Parameters  Estimation Techniques  Experimental Evaluation

18. UNIVERSITY OF MASSACHUSETTS, AMHERST 17 Experimental Evaluation  Simulation ◊150 x 150 x 150 ◊Two scenarios ◊ 4 cameras ◊ 12 cameras ◊Non-uniform distribution ◊ Fraction of objects restricted area

19. UNIVERSITY OF MASSACHUSETTS, AMHERST 18 Experimental Evaluation  Implementation ◊8 Cyclops camera sensors ◊Crossbow Micaz nodes ◊8ft x 6ft x 17ft Image Grabber Object Detection Bounding Box Cyclops View Table Initialization procedure HostMote trigger view information

20. UNIVERSITY OF MASSACHUSETTS, AMHERST 19 Weighted Approximation  Demonstrates non-weighted scheme shortcoming ◊Performs 4-6 times worse than weighted

21. UNIVERSITY OF MASSACHUSETTS, AMHERST 20 Effect of Skew  Weighted scheme can correct for skew better ◊Non-weighted scheme worse by a factor of 6

22. UNIVERSITY OF MASSACHUSETTS, AMHERST 21 Region of overlap  Error decreases with #reference points ◊~22% with 12 pts/camera ◊10% with 37 pts/camera  Error ~10% in region of overlap estimation

23. UNIVERSITY OF MASSACHUSETTS, AMHERST 22 Applications  Duty-cycling ◊Weighted scheme outperforms non-weighted  Triggered wakeup ◊80% positive wakeups with 10 pts/camera with 2 triggers Duty-Cycling Triggered Wakeup

24. UNIVERSITY OF MASSACHUSETTS, AMHERST 23 Implementation Results  k-overlap estimation error: 2-9%  Region of overlap error: 1-11%  Approximate techniques feasible in real deployments (~10% error)

25. UNIVERSITY OF MASSACHUSETTS, AMHERST 24 Related Work  Camera calibration ◊Accurate Extrinsic and Intrinsic parameters [Tsai 86], [Tsai 87], [Zhang 00]  Multimedia Sensor Networks ◊Panoptes: A vision sensor [Feng 03] ◊Audio sensors [Raykar 03]  Localization ◊Sensor Localization [He 03], [Savvides 01], [Whitehouse 02] ◊Active Badge [Harter 94], RADAR [Bahl 00], Cricket [Priyantha 00], Active Bat [Ward 97], GPS ◊Relative Locationing [Rao 03]

26. UNIVERSITY OF MASSACHUSETTS, AMHERST 25 Conclusions  Proposed approximate techniques to estimate associations between cameras ◊Degree and region of overlap  Demonstrated use of estimates to enable applications ◊Error in estimations tolerable http://sensors.cs.umass.edu

27. UNIVERSITY OF MASSACHUSETTS, AMHERST 26 Technology Trends  Sensors/platforms span a large spectrum  Enable heterogeneous camera networks Stargate Functionality Cyclops CMUcam Webcam Mote Telos XYZ Image Sensors Sensor platforms PTZ Energy Functionality Energy

28. UNIVERSITY OF MASSACHUSETTS, AMHERST 27 Approximate Initialization  Degree of overlap ◊Extent of overlapping coverage ◊k-overlap: fraction of viewing area covered by k cameras  Region of overlap ◊where is the overlapping coverage ◊spatial region of overlap with neighboring cameras  Above estimates can support application requirements

29. UNIVERSITY OF MASSACHUSETTS, AMHERST 28 Triggered Wakeup  Wakeup scenarios ◊Object tracking ◊Reliable detection  Determine best camera ◊Projection line ◊ Object along this line ◊Reference points within distance threshold ◊Extent of overlap determines best camera Image Projection line Object Distance threshold