Research On Abnormal Behavior Detection Algorithm Based On Human Key Points

With the popularization of surveillance cameras in daily life,surveillance video data is growing explosively.It can no longer meet the needs of social public security to only rely on manual screening of abnormal behavior of video surveillance.Therefore,the automatic abnormal behavior detection based on deep learning has drawn people’s attention.Its task is to identify the events that deviate significantly from the normal events from videos,which can be subdivided into two categories of detection algorithms based on videos and human key points.The abnormal behavior detection algorithms based on videos have made a lot of achievements,but this kind of algorithm uses the whole video frame area to learn features.The information in the frame contains dynamic environment and complex background,which increases the processing overhead of the detection system.Besides,the features learned are often not focused on the behavior.Abnormal behavior detection based on human key points abstracts the key parts of the human body as human key points,which can describe human behavior more compact,more specific and more structured,and does not contain redundant information such as environmental background.Therefore,abnormal behavior detection algorithms based on human key points have gradually attracted the attention of researchers.Aiming at the problem that abnormal behavior is unknown,which makes it difficult for the system to define and detect abnormal behavior,a DGHKP algorithm is proposed.The algorithm constructs dynamic human key point spatial temporal graphs to characterize human behavior,and uses graph convolutional network to extract behavioral features.After clustering all samples,the abnormal behavior samples will be far away from the cluster center,and the outlier samples will be screened according to the different local density to complete the definition and detection of abnormal behavior.The experimental results show that the DGHKP algorithm has the advantage of lightweight deployment and has a good detection effect on abnormal behavior.As the occurrence frequency of abnormal behavior is much lower than that of normal behavior,the sample data amount of normal and abnormal behavior is extremely unbalanced,which leads to the poor effectiveness of the detection system.To solve this problem,this thesis proposes a RHKPD algorithm.On the basis of preserving the main features of the original data,the algorithm adds noise information to different key points according to the dynamic random probability,and reconstructs the new behavior samples to achieve the balance of the data amount of normal and abnormal behavior samples.Then,the graph attention network is used for feature learning and abnormal behavior detection.The experimental results show that the RHKPD algorithm can improve the effectiveness of the detection system in the case of highly unbalanced data volume of normal and abnormal behavior samples.It can reduce the false detection rate and missed detection rate of abnormal behavior detection system.