Abstract: Aimed at the problem of data imbalance in video anomaly detection, an anomaly video detection method based on SE-U-Net predictive network is proposed. We extracted the saliency map of the video frame and made it into a mask to preprocess the data. We used the preprocessed data to train the prediction model. In order to make the prediction model pay more attention to the optimization of the foreground area, this paper combined the attention mechanism, and a new set of loss functions were designed to constrain the training of the model. In addition, in the testing phase, this paper designed a new anomaly evaluation score calculation method, and anomaly detection was performed only by calculating the prediction error of the saliency region in the video, which alleviated the problem of data imbalance. Public datasets for comparative experiments and ablation experiments were used to verify the effectiveness of the proposed method for abnormal event detection.