Bag-of-Event-Models based embeddings for detecting anomalies in surveillance videos

Automated monitoring of unconstrained videos is becoming mandatory due to its widespread applications over public and private domains. Especially, research over detecting anomalous human behaviors in surveillance videos has created much attention. Understanding patterns in surveillance videos are always challenging due to the rapid movement of the crowd, occlusions, and cluttered backgrounds. The intra-class variations existing among normal and abnormal events lead to poor performance of anomaly detection system. These issues can be addressed by learning discriminative embeddings for video segments of surveillance videos. We propose an efficient Bag-of-Event-Models (BoEM) based embedding to represent video segments of normal and abnormal behaviors. Proposed BoEM can also be formed using training data of normal events only and the embeddings can be given as input to one-class classifier such as OC-SVM in an outlier detection fashion. The proposed embeddings handle intra-class variations and provide improved discrimination with much reduced dimension. Results over benchmark datasets namely Live Videos (LV), UCF-Crime and Crowd Violence demonstrate that the proposed BoEM based event embeddings in conjunction with SVM Classifier give significantly better performance than the other state-of-the-art methods. In addition, studies prove that the proposed embeddings are appropriate even for imbalanced sequential data such as video segments.