Crowd Anomaly Detection Based On Unsupervised Three-dimensional Pyramid Network

With the increasing public security issues and the popularity of video surveillance equipment,the detection of abnormal events based on video surveillance can detect crowd anomalies in time and avoid unnecessary losses,so it has important research significance in the field of public security.In the reality scenario,there is a large amount of monitoring of video data without labeling.The unsupervised learning algorithm is more suitable than the supervised learning algorithm.However,the current abnormal event detection based on unsupervised learning has low detection rate and missed detection of abnormal events.This paper studies the detection of crowd anomaly events based on video surveillance,and proposes a crowd abnormal detection method based on unsupervised 3D pyramid generation network.The method fuses the spatio-temporal features of image sequences and local information of different scales,and optimizes the loss function of the network to obtain a model for generating normal images.Anomaly detection is performed in the detection phase by the difference between the image generated by the model and the image to be detected.The main work and innovations of this paper are as follows:A three-dimensional pyramid image generation network is proposed.In this paper,by studying the image generation algorithm of Auto Encoder and generating the anti-network and Pyramid Scene Parsing Network(PSPNET),the image generation algorithm based on unsupervised learning in the current abnormal detection is not considered.Motion information and local information loss at different scales of the image,the pyramid pooling model and 3D Convolutional Neural Network(3DCNN)in PSPNET are proposed to generate images similar to normal images,and 3DCNN is used to extract image sequences.The temporal and spatial characteristics of the pyramid pooling model extract local information of different scales.An anomaly detection network based on 3D pyramid image generation is constructed.The network is mainly composed of a three-dimensional pyramid image generation module and a VGG16 feature extraction module,and is used to obtain a model for generating a normal image in the network training phase;in order to further reduce the gap between the generated image and the real image,the loss function is improved,and the generated image is generated.The Euclidean distance between the real image and the depth feature distance extracted by the VGG16 added to the gradient distance together constitute a loss function.In the USCD ped,ped2 and Avenue Datasets,this method is compared with the Auto Encoder,GAN and PSPNET methods.The experimental results show that the unsupervised method proposed in this paper can improve the detection and recognition rate of abnormal and reduce the missed detection rate of abnormal.