Crowd behavior detection in videos using statistical physics

Anomaly in the crowds can create a threat to public security. Manual detection of crowd behavior in video surveillance by human operators is a difficult and time-consuming task. There is an urgent need for automated detection of abnormal event in videos. In this process, the calculation of descriptors which expressed as measurements with accurate interpretation is a challenging problem. In this paper, we defined a novel descriptor of crowd behavior which is analogized between human crowds and statistical physics principles. We build a new crowd motion systems and the theory of entropy and internal energy are introduced for evaluating the state of the crowd. Specifically, entropy is calculated with the concept of position density map and pixels in the image are treated as particles. The internal energy of a crowd is obtained with optical flow field. Finally, the entropy and internal energy are collectively as energy flow, and LDA method is used to train the normal crowd behavior. A benchmark with state-of-the-art abnormal detection methods confirms that the effectiveness of the proposed method. Comparison of available techniques for detecting abnormal crowd behavior in surveillance videos has also been done in this paper.