Smart object detection and tracking for video surveillance

点击翻译按钮获取中文摘要

The research that my collaborators and I pursue is in the general area of computer vision, the attempt to endow machines with a sense of sight. From a theoretical perspective, this is an important endeavor as it bears on basic issues in complex information processing and may yield results that bear on our understanding of natural vision systems (e.g., humans). From a practical perspective, machines that can see have the potential to interact with humans in a more natural and useful fashion than is currently the case. The research addressed in the current project attacks an outstanding problem in video surveillance. It is aimed at automated detection and tracking of "interesting" objects in surveillance video. Objects of interest will be defined as those useful for alerting security personnel of the need for intervention, e.g., humans running where they normally would walk or vehicles moving in the wrong direction. A key limitation of extant technology for object detection and tracking in video is its inability to encompass the wide range of variability that is present in real-world surveillance. This challenge will be attacked by leveraging recent work in our lab in modeling dynamic video patterns that previously has proven useful for a variety of tasks, including target tracking and human motion analysis. The key to our success in these related areas is our ability to model complicated real-world dynamics in video in a concise and accurate fashion. Our work contributes to Canada in 3 ways. 1) It paves the way for wider applications of video surveillance as more real-world scenarios can be encompassed. This advance will further the success of Canadian industry in the growing area of video surveillance. 2) It more generally