Multi-level Threat Analysis in Anomalous Crowd Videos

Crowd anomaly detection is a challenging problem in the field of computer vision. An abnormal event in a crowd scene can be labeled as threat in a video. Several existing solutions in this area have marked video frames either normal or abnormal event. Such categorization of frames can be referred as two-class threat labeling problem. However, this notion of two-class threat labeling is not well defined in literature. An event can have multiple aspects as it can be treated as anomalous or non-anomalous based on the situation of occurrence. Based on this argument, we propose a new paradigm of extending this two class threat labeling problem to multi-class labeling. As a solution to this multi-class labeling problem, we cluster frames with low, medium and high threat. We also propose a new feature known as pseudo-entropy for better clustering of threats. Our framework consists of two main components, namely, Earth mover distance (EMD) based anomaly detection system and multi-level threat analysis. As an outcome frame-wise and segment-wise threat representation are also presented to facilitate real time video search for relevant events. Exhaustive internal comparison and statistical analysis over benchmark UCSD and UMN dataset clearly indicates the merit of the proposed framework.