Crowd Abnormal Behavior Detection Based on Label Distribution Learning

In general, some abnormal crowd behaviors are associated, for example, fight causes tumble or panic and tumble causes stampede. And those abnormal behaviors often happened at the same time. However, most researchers consider those mixed abnormal behaviors as only one behavior and ignore the other behaviors appearing in the video. To analyze those behaviors better, this paper proposes a method using label distribution learning to detect the crowd abnormal behavior such as stampede, fight, panic and tumble. We consider that every behavior sequence associated with some behavior labels, and the behavior label distribution covers a series of behavior labels, representing the describe degree that each behavior labels describe the behavior sequence. Then a label distribution learning algorithm named BFGS can be used to learn the behavior label distribution. Through this way, we not only can obtain which behavior happened, but also all behaviors are taken into account for each behavior sequence. The experimental results show that our approach achieves better performance for crowd abnormal behaviors detection.