Abstract: Video anomaly detection is one of the important research areas in computer vision, widely applied in fields such as transportation and public safety. However, the current field of video anomaly detection faces issues such as susceptibility to noise interference in individual prediction models and generalization anomalies in individual reconstruction models. To address these problems, a video anomaly detection method combining reconstruction and prediction models is proposed. A reconstruction network with an attention mechanism and a memory enhancement module is trained on normal optical flow data. The reconstructed optical flow and original video frames are then simultaneously input into a future frame prediction network, where the reconstructed optical flow serves as a conditional aid to assist the frame prediction network in better generating future frames. To extract more effective features, a residual convolutional attention module (SRCAM) is proposed to facilitate the reconstruction and prediction networks in effectively learning feature representations of latent spaces at both global and local levels, thereby enhancing the model's ability to detect anomalous events in videos and improving its robustness. Extensive experimental evaluations on two commonly used video anomaly detection datasets, UCSD Ped2 and CUHK Avenue, demonstrate the effectiveness of the proposed method.

Key words: video anomaly detection, attention mechanism, stream reconstruction, frame prediction, auto-encoder