ANOMALY DETECTION WITH FRAME-GROUP ATTENTION IN SURVEILLANCE VIDEOS

Jinsheng Xiao, Haowen Guo, Yuanxu Wu, Yunhua Chen, Honggang Xie

Keywords: Anomaly detection, attention mechanism, frame-group, spatial-temporal feature

Abstract: The paper proposes an end-to-end abnormal behavior detection network to detect strenuous movements in slow moving crowds, such as running, bicycling, throwing from a height. The algorithm forms continuous video frames into a frame group and uses the frame-group feature extractor to obtain the spatio-temporal information. The implicit vector based attention mechanism will work on the extracted frame-group features to highlight the important features. We use fully connected layers to transform the space and reduce the computation. Finally, the group-pooling maps the processed frame-group features to the abnormal scores. The network input is flexible to cope with the form of video streams, and the network output is the abnormal score. The designed compound loss function will help the model improve the classification performance. This paper arranges several commonly used anomaly detection datasets and tests the algorithms on the integrated dataset. The experimental results show that the proposed algorithm has significant advantages in many objective metrics compared with other anomaly detection algorithms.

One-sentence Summary: The paper proposes an end-to-end abnormal behavior detection network to detect strenuous movements in slow moving crowds.