Visual Surveillance of Human Activity

In this paper we provide an overview of recent research conducted at the University of Maryland's Computer Vision Laboratory on problems related to surveillance of human activities. Our research is motivated by considerations of a ground-based mobile surveillance system that monitors an extended area for human activity. During motion, the surveillance system must detect other moving objects and identify them as humans, animals, vehicles. When one or more persons are detected, their movements need to be analyzed to recognize the activities that they are involved in. Ideally, the surveillance system would be able to accomplish this even while continuing to move; alternatively, the system could stop and stare at that part of the scene containing people. In Section 1 we describe a novel approach to the problem of detecting independently moving objects from a moving ground camera, and illustrate the approach on sequences taken in very cluttered environments. Current research focuses on the problem of classifying those independently moving objects as people based on a combination of their appearance and movement. In Section 2 we describe a system that can track multiple moving people using sequences taken from a stationary camera. This system of algorithms, which has been implemented on a PC and can process 10-30 frames per second (depending on the number of people within the eld of view and the resolution of the imagery) uses a hierarchy of tracking modules to identify and follow people's heads, torsos, feet, ... Finally, in Section 4 we explain how the recovered motion of these people can be classi ed into various activity classes using a principal component model of the time variation of motion of the body parts.