Anomaly Detection of Predicted Frames Based on U-Net Feature Vector Reconstruction

Anomaly detection in surveillance video scenes is one of the current research hotspots. Due to the small sample collection of anomalous events, the lack of negative sample labeling data training in anomaly detection research adds a lot of difficulties. Therefore, we adopt the method of unsupervised training and improve the method of anomaly detection based on the reconstruction of the potential features of the predicted frame and ground truth based on u-net. We reduce the reconstruction error between the potential features of u-net in the predicted frame and the potential features of the real frame. Then through other constraints, the reconstruction error of the entire predicted frame is minimized according to the generative adversarial training. Due to the use of normal behavior sample training, when the abnormal behavior is detected, the reconstruction error value exceeds the set threshold to judge whether abnormal behavior occurs in the surveillance video. Experiments prove that our improved method is effective and accurate.