Surveillance Video Mining

We propose occlusion primitives to dene a set of time-varying predicates on trackers for heterogeneous objects moving in unknown environments. Input pixel information categories for agents and actions. The scene background is learned online at the lowest layer, using feedback from the tracking level to robustly identify multiple agents. Agent shape and color features, status history and trajectories are clustered to discover categories for agents as well as their actions. Unlike existing surveillance systems, the proposed approach does not assume any prior model and aims at learning the scene/agent/event models from the acquired visual informa-tion. Results are demonstrated for trafc videos involving humans, vehicles and animals. I