Abnormal behavior detection based on dense tracklets

This paper presents an approach for detecting abnormal events in videos by analyzing tracklets extracted in dense optical field. To begin with, this method codes a video as a set of spatio-temporal volumes in order to obtain the contextual information. Then, the dense optical fields is established to extract the dense tracklets, and the descriptor histogram of oriented tracklets is utilized to quantize tracklets in terms of angle and magnitude. Finally, the abnormal behaviors is detected based on probability mechanism using the topology of an ensemble of tracklets. Unlike standard approaches that use optical flow, dense trajectories have shown to be efficient for representing videos. The effectiveness ofthe proposed method is evaluated on public datasets: UCSD, subway and mall. The experimental results demonstrates that this method outperformed former approaches based on the dense tracklets and optical flow.