Research on Weakly Supervised Anomaly Event Detection Method Based on Video Temporal Features

In response to data imbalance and high false alarm rates in surveillance video anomaly event detection, this paper proposes a weakly supervised anomaly event detection model based on video temporal features to enhance detection accuracy. The model adopts a video temporal feature extraction module based on causal relationships (VTF-CNN), which effectively captures critical video information through an encoder-decoder structure, enhancing the learning ability of temporal dependency features of abnormal events. Additionally, by introducing a weakly supervised strategy under the multiple instance learning framework, the model effectively addresses the issue of data imbalance and significantly reduces false alarm rates. Experimental validation on the UCF-Crime and ShanghaiTech datasets demonstrates significant advantages of the proposed model over existing methods in improving anomaly detection accuracy and reducing false alarms, proving its effectiveness in enhancing the efficiency and accuracy of surveillance video anomaly event detection.