Research On Abnormal Behavior Detection Based On Human Key Points

With the rapid development of computer technology and video capture hardware equipment,the amount of video data has grown explosively.How to filter and remove invalid information from massive data,automatically analyze video content,and intelligently detect abnormal behavior in video is the research topic of this paper.Due to the non rigidity and randomness of human motion,the modeling of human abnormal behavior becomes particularly difficult.Research on the modeling of general characteristics of pedestrian abnormal behavior in intelligent surveillance video to achieve the detection of pedestrian abnormal behavior in video.The main research contents include:(1)The human posture estimation method and target tracking technology are introduced to obtain the key point information of the pedestrian’s whole body in the video.Due to the spatial complexity of surveillance video and the difference of human behavior,it will have a greater impact on behavior anomaly detection.Human skeleton key points have the advantages of compact,rich semantic information and strong feature structure.Therefore,human key point data is used to express human behavior characteristics.The current Alpha Pose human posture estimation algorithm with good performance is used to obtain the key point data of the human body.At the same time,target tracking is performed for different pedestrians,so that the key point information of the human body can accurately correspond to different pedestrians in the video,providing basic data for modeling and analysis of abnormal behavior.(2)Research feature engineering based on human key point data,and build feature attribute set.Firstly,the human body key point coordinates under the world coordinate system are converted into the coordinates under the local human body coordinate system,and then the corresponding feature attribute values are calculated according to the coordinates to obtain a series of feature attribute sets based on key point information.Taking the human body center of mass as the center,a circular key point representation method with different distances from the origin is designed.Key point data at different locations are treated differently.Feature points are extracted according to the rings with different radii,and key point feature subsets at different locations are constructed.Attention mechanism is used to apply different weights to key point data at different locations,so as to further complete the modeling of abnormal behavior.(3)Using the research method of comparative analysis,the abnormal behavior detection is carried out by taking into account the inter-frame and intra-frame data of video respectively.When detecting abnormal behavior between frames,the attention mechanism selection strategy is integrated to focus on different feature attribute subsets for different anomaly detection tasks,so as to achieve abnormal behavior modeling based on inter frame data.When detecting abnormal behavior in the frame,enhance or suppress part of key point subset information,conduct horizontal contrast in the frame,and use unsupervised classification algorithm to detect abnormal behavior in the frame.The experimental results show that the accuracy of anomaly detection is significantly improved after the fusion of attention enhancement,and the abnormal pedestrians in the video are effectively detected.