Traffic Video Frame Prediction And Application Based On Deep Multi-branch Mask Network

In recent years,because the video frame prediction task can predict the action behavior and scene state,it has a wide range of application prospects in the field of intelligent driving and security.In this paper,we will study the future frame image generation problem in the traffic scene and the anomaly detection problem in the monitoring scene.According to the different number of future frame prediction,this paper divides the video frame prediction problem into single frame prediction and multi frame prediction,and defines these two kinds of problems.In the part of single frame prediction,aiming at the shortcomings of current video frame prediction algorithms,this paper proposes a deep multi branch mask network(DMMNet)which combines the algorithm of optical flow warping prediction with the algorithm of direct prediction,trained and tested in Caltech traffic data set.Through the experimental comparison,the function of each mask layer and fusion operation in the model is proved,and the DMMNet model has excellent single frame prediction effect.For the multi frame prediction part,this paper improves the original DMMNet model,encoding the input multi frame image successively through the shared motion encoder.the feature map is input into multiple Conv LSTM for spatiotemporal modeling.A comparative test is carried out on Caltech data set.It is proved that the improved DMMNet model can delay the performance degradation of the original model in the multi frame prediction problem,and the multi frame prediction ability of the improved DMMNet model is verified.In this paper,based on the analysis of anomaly detection related models and the idea of anomaly detection using video frame prediction,DMMNet model is applied to the problem of anomaly detection,and DMMNet-Ab Det algorithm is proposed.Aiming at the problem of the algorithm in anomaly detection,this paper designs an anomaly detection algorithm DMMNet-Det-Ab Det,which first performs image prediction and then uses target detection to detect specific areas.The experiment results in UCSD data set show that DMMNet-Ab Det algorithm can solve the problem of anomaly detection,and DMMNet-Det-Ab Det algorithm improves the accuracy of anomaly detection.