Abnormal Behavior Recognition Algorithm Based On Deep Learning

With the rapid development of the theory and research of computer vision and deep learning,a series of intelligent algorithms based on deep learning have been gradually applied in the field of public security.Convolutional Neural Network(CNN)is a representative network of deep learning.Compared with traditional neural networks,the recognition effect of convolutional neural networks has been greatly improved,and has in many areas of computer vision achieved success.Therefore,based on deep learning,it is of great significance to efficiently and accurately identify abnormal behavior in video.This thesis focuses on the research and implementation of two aspects: a combined optical flow network and a two-stream structure based on deep residual network.In this part of the extraction of optical flow characteristics,a method based on deep learning is chosen to perform optical flow estimation on moving objects in the video.In order to accurately extract the motion information of the foreground target in the video,the concept of end-to-end optical flow is introduced in the combined optical flow network including multiple optical flow networks,and at the same time,a sub-network for capturing small displacements is added.To describe the changes in the details of the action.The optical flow image of the adjacent video frames is finally calculated by the combined optical flow network,and used as the input of the time flow network.In design of the two-stream structure based on the deep residual network,we divide the video into two parts: space and time,and adopt spatial flow network and time flow network respectively by imitating the ventral and dorsal channels of the brain to process visual signals.To achieve two-stream feature extraction,we process image and motion information.In order to accurately identify abnormal behaviors,we propose a 101-layer deep residual network as the spatial stream network and the time stream network to solve the problem,which prone to gradient explosion caused the degradation of the overall network.We adopt strategies such as data gain,migration learning,and Dropout to solve the over-fitting problem,which enhance the learning and generalization ability.Finally,experiments were conducted on the UCF101 behavioral dataset and the CASIA multi-view behavior database.The experimental results show that the combined optical flow network can effectively extract the characteristics of optical flow and improve the accuracy of the recognition of the time flow network.The dataset gain is obtained by mirroring the original dataset,rotation direction,and random clipping.The deep residual dual-flow network is more effective than the shallow-layer network in anomaly classification tasks and improves the recognition accuracy.