Abnormal event detection by a weakly supervised temporal attention network

Abstract

Abnormal event detection aims to automatically identify unusual events that do not comply with expectation. Recently, many methods have been proposed to obtain the temporal locations of abnormal events under various determined thresholds. However, the specific categories of abnormal events are mostly neglect, which are important to help in monitoring agents to make decisions. In this study, a Temporal Attention Network (TANet) is proposed to capture both the specific categories and temporal locations of abnormal events in a weakly supervised manner. The TANet learns the anomaly score and specific category for each video segment with only video-level abnormal event labels. An event recognition module is exploited to predict the event scores for each video segment while a temporal attention module is proposed to learn a temporal attention value. Finally, to learn anomaly scores and specific categories, three constraints are considered: event category constraint, event separation constraint and temporal smoothness constraint. Experiments on the University of Central Florida Crime dataset demonstrate the effectiveness of the proposed method.