Public security is an important part of national security,and it is also the basic guarantee for the happy life of the people across the country.It has been a key concern of the whole society for many years.In particular,the security issues of key national departments must not be negligent.A little carelessness will cause a great sensation and cause huge losses.In order to ensure "absolute safety",this article focuses on important places that have high requirements for safety issues and cannot accurately enumerate abnormal behaviors.Based on the idea of unsupervised deep learning,abnormal behaviors are not defined,and training can only reconstruct normal behaviors.According to the large difference between abnormal behavior and normal behavior,the model cannot reconstruct abnormal behavior for abnormal detection.Using the two frameworks of convolutional encoder and generative confrontation network,the technology of abnormal behavior detection in video is researched.The main results include the following aspects:1、Propose an abnormal behavior detection algorithm based on the improved U-Net network.First,the input of the network is preprocessed,and then the improved U-Net network is used in the feature extraction network part.On the one hand,a convolution with step size is used.Instead of the pooling layer,the number of jump connections is increased.On the other hand,BN is added after each layer of convolutional layer,which reduces the loss of information in the feature extraction process,so that the network can extract features more fully and converge faster.Fast.In the anomaly detection stage,the trained model can reconstruct the normal behavior,showing that the reconstruction loss is small.On the contrary,if the abnormal behavior cannot be reconstructed by the model,a larger reconstruction loss value will appear.This can be used to judge abnormal behavior.2、Propose an abnormal behavior detection algorithm based on an improved generative adversarial network.In order to take into account the temporal and spatial characteristics of the behavior in the video,the improved U-Net network is first extended from two-dimensional convolution to three-dimensional convolution,and then used as the generator of generative network.The input of the generator is the video frame representing the spatial information,and the input of the discriminator part is the optical flow representing the time information.Through the adversarial training of the generation network and the discriminant network,the spatiotemporal feature conversion model of normal behavior is obtained.Only use the generator part in the test stage,locates anomalies by the pixel-by-pixel difference between the generated image and the target image,and detects abnormal behaviors through reconstruction errors. |