With the advances in the video technology, video cameras have become an integral part of the daily life. They are installed in parking lots, traffic intersections, airports, banks, etc. for constant surveillance. Usually a human operator watches them to catch events of interest in the scene, but this is a tedious and time consuming process requiring constant attention, and leads to inadequate surveillance capability. Therefore, there is an urgent need for automated systems for the analysis of surveillance video streams. In the last few years, there has been progress towards satisfying this need. There is a growing interest in the computer vision community towards video understanding, in particular towards visual event recognition. Many recent multi-institutional research projects and individual researchers have explored, and continue to explore the issues in event-based video analysis, action recognition, and related areas. This dissertation surveys different taxonomies of motion understanding problems, identifies the major components in an automated visual event recognition system, and presents the challenges and the significant studies in moving object detection, shadow elimination, and object tracking. Novel schemes for shadow detection and object tracking are proposed and implemented. The proposed shadow detection scheme does not rely on models of scene or objects, which makes it robust for a variety of outdoor surveillance applications, and also successfully eliminates problems due to illumination changes that are common in outdoor sequences. The proposed schemes for object tracking address the problem of correspondence in the presence of multiple moving objects and occlusions in the scene, and involve multi-hypothesis decision making and color appearance models. The experimental results are presented with accompanying discussions. Also, the roles of the above mentioned components in visual event detection systems are presented in a number of selected applications.
展开▼
