Research On Algorithms For Recognizing Fence Climbing Behaviors From Surveillance Videos

With the continuous development and popularization of video surveillance technology,it has become possible to obtain large-scale surveillance video data.Research on intelligent security for surveillance video is a hot issue in video analysis.Its task is to recognize violent or illegal behaviors involving public safety,personnel safety,and property safety in videos.It continuously monitors and promptly warns.It is widely used in public places such as buses,subways,shopping centers and communities.This thesis studies the behavior recognition method for fence climbing behavior in surveillance videos,and obtains the following research results.1.Study the latest action recognition network methods in video analysis,including: spatial temporal graph convolutional network,two-stream adaptive graph convolutional network and action structure graph convolutional network.This Thesis compares and analyzes these three kinds of action recognition networks.We mainly consider the natural connection of body joints in spatial temporal graph convolutional network,and does not consider the influence of non-adjacent joints on action recognition.Action structure graph convolutional network considers the relationship between non-adjacent joints within a limited distance and its effect on action recognition.Therefore,the recognition performance of action structure graph convolutional network is better.Two-stream adaptive graph convolutional network introduces an adaptive matrix to learn the relationship between joints within arbitrary distances,and adds bone edge information when constructing the network.So,the recognition performance of the two-stream adaptive graph convolutional network is the best.Finally,experiments on public datasets verify the above conclusions.2.Propose a finite state automaton based on human skeleton information to recognize fence climbing behavior.This thesis introduces fence information as a priori knowledge into the action recognition network,so that the network can learn the key features of the climbing behavior from the change in the relative position between the human skeleton and the fence.First,this thesis studies the relative position changes between the body joints of the action subject and the fence during the climbing,and proposes a general pattern suitable for fence climbing,which mainly involves the stages of fence climbing behavior and transition conditions between the stages.Second,this thesis converts this pattern to the corresponding finite state automaton to recognize fence climbing behavior.Finally,this thesis conducts experiments on real-world climbing dataset.The results show that the automata-based method achieves good accuracy,high precision and relatively low recall under different shooting distances.3.Propose a climbing-graph convolutional network based on human skeleton information to recognize fence climbing behavior.First,this thesis constructs climbinggraph convolutional network based on graph convolutional block and two-stream structure,and provides coordinate information of fence in the video to the network as a priori knowledge.Second,this thesis trains climbing-graph convolutional network to learn the fence climbing action based on the human skeleton information,and then uses this network to recognize fence climbing behavior in test videos.Finally,this thesis conducts experiments on real-world climbing dataset.The results show that climbinggraph convolutional network achieves good results at different shooting distances and scenes,and it achieves better accuracy,precision and recall than the automaton-based method and existing state-of-the-art action recognition networks.