Surveillance video mining
Prithwijit Guha, Arindam Biswas, Amitabha Mukerjee, Prashanth Sateesh, K. Subramanian Venkatesh · 2006
This paper approaches the problem of surveillance event characterization at two levels. First, the information characterizable at the single camera image-plane level, where we use a set of occlusion primitives to define a set of time-varying predicates on heterogeneous objects moving in unknown environments. After learning the scene background online, tracking data generates overlapping agent trajectories. Agent appearance and trajectories are clustered to discover both agent and object categories. Second, the information characterizable at the 3D scene ground-plane level, where scene priors (such as object names) and camera calibration information is used to index and retrieve the event characterizations. The proposed approach does not assume any shape or other priors, and names are assigned a posteriori. Results are presented from a traffic video which can answer queries such as which motorcycles turned left from IIT onto GT Road, or list the red cars which overtook rickshaws while going West at a speed exceeding 50 kmph.