Crowd anomaly detection and localization using histogram of magnitude and momentum

Video anomaly detection is an important and crucial area from security point of view. Anomaly detection means identifying unusual activities. It is a tedious task to recognize abnormal activities due to its infrequent occurrence in the crowd. Surveillance cameras are installed in crowded places, but manual analysis of video data gathered from these cameras is a cumbersome process and becomes almost impossible if cameras are in large number. In this work, an automated approach is proposed to detect and locate anomalies. A concept of momentum from Physics is used to connect foreground occupancy and motion of object. The proposed work is divided into three major steps: (a) background removal, (b) feature extraction and behavior recognition, and (c) anomaly detection and localization. The background removal step separates the background from each frame. To detect anomalies, appearance and motion characteristics of foreground objects are incorporated by histogram of magnitude and momentum features. Behavior of objects is learned through unsupervised clustering technique. In order to locate anomalies, positional features are used. The proposed approach is verified on benchmark datasets like UCSD and UMN devised for anomaly detection and crowd analysis. Experimental results are validated to contemporary methods.