Using a Light DBSCAN Algorithm for Visual Surveillance of Crowded Traffic Scenes
Amar El Maadi, Mohand Saïd Djouadi · IETE Journal of Research · 2015
Behaviour analysis in visual surveillance has become a very active issue for the computer vision research community, particularly when crowded scenes are concerned. In this perspective, motion analysis and tracking is challenging due to significant visual ambiguities which incite to look into more alternative solutions. In this work, we introduce a new framework for recognizing various motion patterns, extracting abnormal behaviours, and tracking them over crowded traffic scenes. The proposed approach exploits a novel density-based clustering method and highlights three traffic density levels. It performs in two modes: an “offline” mode for motion patterns learning and modelling, and an “online” mode for distinguishing irregular motions and tracking them separately. We developed a light density-based clustering technique to “online” cluster motion vectors, produced by optical flow, and compare them with motion pattern models previously defined. Non-identified clusters are treated as suspicious and tracked over the scene. To evaluate and validate our work, experiment results are discussed and compared with the most recent approaches.