Research On Crowd Abnormal Behavior Detection In Video Surveillance

Public safety issues have always been the focus of attention in society.In recent years,the development of monitoring equipment technology has made monitoring equipment all over the place.As a research hotspot in the field of computer vision,intelligent video detection technology has also developed rapidly with the popularity of surveillance equipment.The detection of abnormal behavior in surveillance video is a very important research topic in intelligent monitoring technology.The abnormal behavior detection is mainly to detect the abnormal behavior that may occur in the surveillance video through computer vision correlation detection algorithm.The difficulty lies in the complex polygons of the video scene and the accuracy,real-time and stability required for the practical application of the abnormal behavior detection technology.Due to the wide application of convolutional neural networks in computer vision,this paper mainly studies the anomaly detection algorithm based on convolutional neural networks.Two different anomalous behavior detection algorithms based on convolutional neural networks are proposed for different scenarios.Among them,the detection of common violent behaviors in monitoring scenarios and some specific abnormal behavior detection in specific monitoring scenarios are the main research focuses of this paper.The main research contents of this paper are:(1)The method of extracting dynamic information in traditional video scenes is studied.It mainly includes the interframe difference method,the background subtraction method and the optical flow method.Different methods have their own advantages and disadvantages and applicable scenarios.This paper also analyzes the advantages and disadvantages of different methods,and selects appropriate dynamic feature extraction algorithm for feature extraction in the subsequent anomaly detection tasks.(2)The development history of convolutional neural network model and the basic structure of the network are studied.In addition,several commonly used convolutional neural network structures are briefly introduced.The algorithm framework of violent behavior detection in surveillance scenarios is proposed.By adopting two-steam network,the algorithm extracts the dynamic features and static features of the video content separately,and uses support vector machine to classify and fuse.Compared with other algorithms,it can effectively detect the violent behaviors in the scene,and achieve realtime performance while ensuring accuracy.(3)This paper also proposes an optical flow-FCN network structure for the detection and location of specific anomalous behaviors in the surveillance scene,including cyclists,skaters,small carts,and people walking across the sidewalk or surrounding grassland.By combining the optical flow feature and the FCN semantic segmentation feature by1?1convolution,it is possible to effectively detect and locate the abnormal behavior in the monitoring scene.