Research On Intelligent Monitoring Technology Of Violent Behavior Based On Skeleton Information

In recent years,the international security situation has become increasingly severe,with terrorist attacks and violent conflicts occurring at a high rate.Video surveillance technology is widely used in schools,hospitals,shopping malls,streets and other public places,which is a main method to ensure social security and stability in China.In addition,with the rapid development of Internet technology,online video has become an important carrier of information transmission,while the violence,pornography and other bad information that is rife in it has a bad impact on users,especially teenagers.With the extensive deployment of monitoring terminals and the explosive growth of online video,video surveillance technology needs to be more intelligent.Traditional monitoring system can only realize the function of video collecting and storing,while the examination of the data is mostly completed by manpower which is not only timeconsuming and laborious,but also prone to omission and misdetection.Therefore,the intelligent surveillance technology for violent behaviors in public places and online videos has become a research content needed by the people.According to data types,monitoring methods can be divided into three categories: RGB,depth and skeleton.The intelligent surveillance technology based on RGB and depth information is easy to be affected by occlusion,background,illumination and other interference factors,and it is difficult to extract features from large-scale or lowresolution images.Combined with practical problems,this paper focuses on the intelligent monitoring technology of violent behaviors based on skeleton information.The main work is as follows:(1)To solve the problem that skeleton information extracted by Kinect is limited by the collection distance,we use large zoom webcam as the monitoring equipment,build the tf-pose deep network of Carnegie Mellon University to extract human skeleton information from the collected RGB video,and establish a data set of violent behaviors.(2)Considering the requirements of security environment for intelligence and function variety of monitoring system,an intelligent surveillance system for violence is built,which integrates the functions of data collection and extraction,violent behavior location,violent behavior recognition and abuser tracking.The related hardware and software design ideas of modules in system are introduced.(3)Aiming at the low efficiency of traditional monitoring methods,a time series location method of violence is proposed.The method consists of two stages.In the first stage,some prior knowledge is adopted to remove the security fragments that account for the majority of the long video sequence with the help of distance between bodies and motion characteristics of joints.In the second stage,BP neural network is used to locate the suspicious fragments accurately,so as to facilitate the subsequent use of violence recognition network for targeted identification.(4)An end-to-end violence recognition network based on deep learning is built.The network is composed of data enhancement module,training module and fusion module.Firstly,a new way of skeleton information segmentation is used to divide human skeleton into four subspaces to reduce the difficulty of feature extraction.Then,the data enhancement module enriches the violence data set in time and space dimensions,so that the overfitting caused by the increase of network layers can be solved.After that,four subspace models are trained via the depth network that is based on bidirectional long and short-term memory model in the training module.Finally,to further improve recognition accuracy,the results of four subspace models are weighted in the fusion module.(5)In order to capture more information about abuser intelligently,a size adaptive target tracking algorithm based on KCF is proposed.Combining with skeleton information,the position of perpetrator in the image is detected quickly,and the targeted monitoring is carried out by adjusting the electric pan-tilt to solve the problem of target missing caused by significant change of the abuser's attitude and size.