Event Detection Algorithms For Video Surveillance

In the contemporary society, surveillance video systems are playing an increasingly significant role. As a core component of intelligent sur-veillance video systems, and an interdisciplinary research area in com-puter vision, pattern recognition, artificial intelligence, machine learning and etc., event detection algorithms have a great significance both theo-retically and practically. This thesis focuses on the research of these algo-rithms.In this thesis, several classic algorithms for background (BG) sub-traction are first of all implemented and analyzed. Then the general structure of BG subtraction is generalized. To cope with the common problems for BG subtraction, we proposed a novel algorithm for BG sub-traction based on Integrated Foreground Mask (IFM). We also enhanced it with a novel illumination change detection scheme based on the statis-tics of temporal differential image of the sequence, which strengthened the robustness of our algorithm. Experimental results showed that our al-gorithm works well on standard testing database. A running detection system for the crowd is also implemented and improved in this thesis. Detection is made possible via a detect-ing-tracing-judging structure for blobs that contain human head and shoulders.Finally we designed and implemented a fall detection algorithm for indoor scenes. In this algorithm, the quantification of human movement is first of all computed from local Movement History Image (MHI); the quantification of human orientation and shape is then computed from the slant angle and ratio of the approximate ellipse of the human. By analyz-ing both data, fall event detection is implemented.