Violent Behavior Monitoring In Aerial Scenes Based On Human Pose Estimation

Violent crimes have always been the focus of crackdowns by law enforcement agencies.At present,law enforcement agencies are committed to using video surveillance systems to remotely monitor violence.Traditional monitoring methods have many disadvantages,such as fixed location of surveillance cameras,difficulty in achieving full coverage of the surveillance area and easy to fatigue when surveillance personnel watch the surveillance screen for a long time.Considering that the use of Unmanned Aerial Vehicle(UAV)for monitoring can expand the field of view of monitoring,computer vision technology for video surveillance is more intelligent and the advantages of skeleton information has clear and simple features,making it less susceptible to environmental factors,This paper studies violent behavior monitoring in aerial scenes based on human pose estimation.Firstly,a UAV system was set up to monitor violence in public areas.The monitoring UAV system includes a monitoring UAV platform and a ground monitoring center.The monitoring UAV platform is the core of the system,which selects the quadrotor body as the flight platform,uses Pixhawk to realize the flight control of the UAV,and uses the embedded processor Jetson TX2 and camera form the on-board vision system.In addition,the design of monitoring programs and software is completed according the monitoring needs.Secondly,for the embedded processing platform,a lightweight multi-person pose estimation method is designed and implemented to realize the rapid acquisition of multiperson poses in aerial photography scenarios.The Part Affinity Fields(PAFs)is adopted to achieve rapid matching of key-points,by using deep separable convolutions to achieve the decomposition of traditional convolutions,thereby the amount of network calculations is reduced.At the same time,the network subnet is automatically optimized and the parameters are quantified by TensorRT to improve the real-time and practicability of the algorithm.Thirdly,a method for identifying violent behavior in aerial video based on dual-branch classification network and Spatial Temporal Graph Convolutional Network(ST-GCN)is designed.Aiming at the problem of video violence location,using human posture information,a two branch classification network is designed to realize the detection of violence key frame based on human posture,so as to preliminarily judge the time of occurrence of violence and extract adjacent multiple frames of human skeletons based on key frame positions information.In order to realize the classification and recognition of violent behaviors,ST-GCN is used to analyze the human skeleton information of multiple frames,thus completing the classification and recognition of violent behaviors based on human skeletons.Aiming at the problem of joint point missing,to reduce the influence of human body’s position in the image and the distance from the camera lens on recognition rate,Multi-frame skeleton data is normalized,filtered and completed to construct a violent behavior skeleton Temporal-Spatial representation matrix as ST-GCN’s input.Aerial data sets were created to validate the method.Finally,systematic monitoring experiments are performed for typical violent behaviors,such as strangling,stabbing,kicking and shooting in different scenarios.The results show that the UAV monitoring platform flies steadily along the cruise path and the on-board vision system quickly locates the violent behavior in the video.The ground monitoring center implements the classification and identification of violent behaviors based on the skeleton sequence.