Video Surveillance is a large application domain with many technical challenges. On the application side, our focus has been on how to incorporate privacy protection in surveillance without sacrificing its primary security functions. We have made novel contributions in moving object segmentation, camera placement, camera fusion, video obfuscation, and video data hiding for privacy data preservation.
Moving object segmentation refers to the techniques in identifying the pixels corresponding to moving objects in video sequences. Missed detections occur due to color confusion with background while false alarms can be caused by “uninteresting” moving background such as moving trees and rains. One of our early work provide a survey study and is one of the top cited work on this topic [1]. To increase the robustness of the segmentation algorithms, we have incorporated rapid illumination change adaptation [19], and developed novel fusion techniques in joint thermal-RGB segmentation [13,21] as well as combining object classifiers at different adaptation rates [3] and within a boosting framework [4].
Providing privacy protection to a large camera network needs to overcome a number of technical challenges. Before summarizing our technical contributions, the following video illustrates the technical challenges and highlights our solution:
Specifically, we have made contributions in five different areas. First, most applications require protection only for trusted individuals and a robust subject identification is thus required. We have developed optimal camera placement techniques to maximize the performance in triangulating each subject in the surveillance area under a binary integer programming framework [7,10,11,15,20]. Second, to obfuscate sensitive imaginary in the surveillance video, we propose an efficient object-based in-painting technique that completely removes the protected individuals from the scene while preserving the motion in the background [5][14][16]. Third, it is imperative that the obfuscation process is authenticated and reversible to maintain the integrity of the surveillance record. We are the first to propose the use of data hiding in preserving privacy data [2] and have developed a rate-distortion optimization framework in embedding privacy data in compressed video [8][9][17]. Fourth, we have constructed a privacy data management system in which the access chain of privacy information can only be authorized by the very individuals in the video [14][18][22][23][24]. Finally, we have also extended the privacy framework to other multimedia domains including life-log capture [6] and video conference [12].
Publications: