Video Prediction And Anomaly Detection Methods Based On Deep Learning

In the past decade,the video surveillance market has shown an explosive growth trend.In addition to the sharp drop in the price of high-definition cameras,the diversification of video surveillance application scenarios has also prompted the further upgrade of video surveillance systems.With the rapid development of deep learning,video prediction and anomaly detection technology has become a new research hotspot in computer vision.Use the image information of historical video frames to predict future frames and detect abnormal events in the predicted frames,so as to make early warning and decision control in advance,and have huge application prospects in many fields such as automatic driving,social security management,medical care,robot decision-making,etc.The main innovations of this article include:(1)A video prediction and anomaly detection method based on convolutional neural network and normalized visual attention and adapt layer-instance normalization is proposed.In view of the characteristics of large amount of video data and high redundancy,we use visual attention mechanism and adaptive layer normalization method to process the video frame image.The visual attention mechanism assigns a higher weight to the region of interest in the video,so that it can obtain more information during the feature extraction process;the adaptive layer normalization method processes the data to normalize the data and generates it during the training process The adversarial network is easier to converge and speed up model training.(2)A video prediction and anomaly detection method based on perceptual loss and dual discriminator is proposed.In order to improve the details of image generation,we add perceptual loss function,identity loss function,optical flow loss function and other mixed loss functions to optimize the generator and discriminator to make the generated image more suitable for human sensory system;introducing a dual discriminator structure and using different reward mechanisms to train two discriminators with independent parameters can prevent the model from collapsing against the network.After ablation research and comparative experiments,both methods have achieved expected results in video prediction and abnormal event detection tasks.