Spatiotemporal heterogeneous information fusion model for loitering anomaly detection

In the field of video surveillance security in public places, loitering anomaly detection plays a crucial role. Currently, the complexity of public scenes and the difficulty in extracting apparent features due to limitations in the resolution of surveillance videos make tracking, which serves as the foundation for loitering anomaly detection, challenging. In order to solve the problem of low robustness of the tracker in low-resolution complex scenes, a reassessment module based on gait features and a matching algorithm based on spatial information are proposed. To locate loitering abnormal frames more sensitively and accurately, and to better distinguish normal and abnormal samples, an adaptive loitering detection algorithm based on motion states is proposed. The spatiotemporal heterogeneous information fusion model for loitering anomaly detection is tested on the IITB-Corridor dataset and compared with the most effective deep learning method, Semi-supervised Video Anomaly Detection and Anticipation, showing an increase in accuracy by 1.2% and recall by 1.8%.