Real-time Weakly supervised Anomaly Detection with Attention Mechanism

The use of surveillance cameras for ensuring the safety of people in public gatherings, e.g., roads, intersections, banks, shopping malls, etc., is being used increasingly in recent years. To detect violence and other abnormal behaviors from these surveillance videos, the monitoring, and detection need to be intelligent and precise. Manual monitoring of these videos is tiring and prone to errors. So automatic detection of video anomaly is a very important research topic nowadays. Although many methods have been proposed, research is still ongoing for a reliable method for violence detection. Most of the earlier state-of-the-art works suggest that supervised detection is better in anomaly detection than unsupervised one. However, given the difficulty of manually annotation and training large datasets for supervised learning, weakly supervised methods provide an excellent alternate solution. Generally, in weakly supervised learning, anomaly detection is constructed as a Multiple Instance Learning (MIL) problem. The base of that technique is to build the training model from labeled bags (comprising multiple instances) rather than labeled individual instances (in the computer vision branch: videos or frames). A bag is annotated as negative if all the instances are negative, while it is marked positive if at least one instance is positive. But the problem is that the recognition of positive instances is largely biased by the dominance of negative ones. As in the real world, abnormal events are very subtle, and the difference between normal and abnormal instances is minimal. Addressing this issue, this paper proposes a weakly supervised learning algorithm with an attention mechanism for anomalous events detection. Also, a novel ranking loss function is proposed to lessen the number of false negatives in the anomaly detection task and to extend the distance between classification scores of anomalous and normal videos in the learned model. Experimental results of this proposal within UCF-Crime and Shanghai-Tech datasets show that the proposed method outperforms state-of-the-art ones. Moreover, using an attention mechanism enhances feature extraction in the generated model inference. Finally, it is demonstrated that the proposal can be run in real-time, a contribution that can greatly impact real applications in the context of interest.