1. TelosCAM: Identifying Burglar Through Networked Sensor-Camera Mates with Privacy Protection Presented by Qixin Wang Shaojie Tang, Xiang-Yang Li, Haitao Zhang Jiankang Han, Guojun Dai, Cheng Wang, Xingfa Shen

2. Introduction Video surveillance are already widely used in Airports, Border, Railways, Underground, and Roadway… Wireless Sensors also attract significant attention these days for event monitoring and communication…

3. What is Sensor-Camera Network? Integrates wireless module nodes (such as TelosB nodes) with legacy surveillance cameras Wireless module node is able to detect event and trigger tracking process Surveillance camera is used to capture video of interest

4. Why is Sensor-Camera Network? TelosCAM system is able to track and identify the burglar who stole the property TelosCAM has the following advantages: Camera Storage Efficiency High Reliability Privacy Protection Light Modification

5. Frame Work

6. Design Requirement Computation Efficient High Accuracy Reliable Energy Efficient Privacy Aware Storage Efficient High Hitting Ratio

7. Privacy Aware Triggering Scheme I Naïve triggering scheme may cause serious privacy issue The tracking process is triggered if the property is out of transmission range of the owner.

8. Privacy Aware Triggering Scheme II However, the above design suffers from potential security issues if not designed carefully. We thus design a secured message exchanging scheme to prevent potential attacking.

9. TRAJECTORY BASED VIDEO EXTRACTION I Video extraction aims to archive only those videos which contain the burglar with high probability. When a burglar passes through a surveillance point, the surveillance wireless module (s) will receive some alarm messages sent by the secondary wireless module. A naive scheme is to let each camera start extract the video once they detect the appearance of the object, however, it suffers from poor storage efficiency.

10. TRAJECTORY BASED VIDEO EXTRACTION II We propose a trajectory based extraction scheme to ensure high storage efficiency without scarifying reliability. The basic idea is to 1.Reconstruct burglar’ trajectory based on sensing result; 2.Estimate the time when the object entered and left the camera sensing range; 3. Filter out those videos which are less likely to contain the burglar based on above information.

11. Burglar Identification Through Video Processing I To identify the burglar from set of retrieved videos. Among all the objects ever appeared in the extracted videos from all relevant cameras, the burglar tends to have the most occurrences.

12. Burglar Identification Through Video Processing II Step I. Suspicious Objects Selection From Single Camera Object Classification

13. Burglar Identification Through Video Processing II Step I. Suspicious Objects Selection From Single Camera We select top-k objects with longest appearance durations as k most suspicious objects for each video.

14. Burglar Identification Through Video Processing II Step II. Collaborative Burglar Identification Through Networked Cameras Inter-Camera Calibration: to construct pair wise camera color mapping that maps the color histogram from one camera to the other. We formulate the mapping problem as a maximum weighted matching problem A histogram, h, is a vector {h[0], …, h[N]} in which each bin h[i] contains the percentage of pixels corresponding to the color range color_i in this object. We compute a weighted bipartite graph between two histograms as the positive weighted edge represent the bin-wise histogram Distances. Finding a maximum weighted matching of this bipartite graph.

15. Burglar Identification Through Video Processing II Step II. Collaborative Burglar Identification Through Networked Cameras Burglar Identification: identify the object that has the most occurrences across all videos from those suspicious objects. We formulate the mapping problem as a clique problem First evaluate the likelihood of possible object matches between videos from two cameras based color, height, speed similarity. We then build a similarity graph G = (O,E), where O is the set of objects from all surveillance cameras. Consider two objects from different cameras, we add a edge between them if their similarity is greater than a pre-defined threshold. Finding a maximum clique in this graph.

16. Evaluation Results TelosB node [24] as wireless module node. Canon PowerShot A3300 IS (16 megapixels ) as camera. The cameras sample the visual information of the surveillance regions at a frame rate of 15 Hz, and the resolution of the captured video sequence is 360×240 pixels. The video processing algorithm was carried out on the platform of VC++.NET 2005 combined with OpenCV (the open source computer vision library supported by Intel Corporation).

17. Evaluation Results

20. Thank you!