Human Pose Estimation Based Abnormal Action Recognition In Schools

With the different types of social dilemmas,the school security has been intensified.With the consistently appearance of fighting,accidentally falling of teachers and students,and suiciding in school,not only the physical robustness but also the mental firmness of students and faculty will be affected,which can not only bring the concerns from parents and society,but also interfere with the regular teaching process,scientific researches and the routine life order of students and faculty.Those situations highlight that there is a large ocean of inadequacies and weaknesses in the school management work.With the ever-accelerating expand and application of the information skills and the computer vision based artificial intelligence,instead of the traditional and natural human vision,the cameras and computers can locate,track and measure task targets.With the emergence of intelligent video surveillance,behavior recognition results can be automatically analyzed,so abnormal behavior detection has become a reality.Applying intelligent video surveillance to the campus environment can greatly reduce the dependence on eyes,automatically analyze campus scenes and identify abnormal behavior types on campus,so as to minimize the adverse impact of unexpected accidents and loss of life and property.Combined with action recognition and the demand of school surveillance video,this paper studies four specific abnormal behaviors of falling,fighting,running and entering the dangerous area in schools.The main achievements of this paper contain four aspects:(1)This paper sorts out the related work of the human pose estimate,the abnormal behavior detection in schools,and the studies of relevant algorithms.(2)We proposed an improved feature extraction method of human bone key points to detect and locate the bone key points of personnel in schools,then extract the features for action recognition.Specifically,we introduced the channel attention mechanism to enhance the representation of features.(3)We regard the skeleton key points as features of action recognition.Then model the key points in spatial and temporal dimensions respectively.Combining with the spatial temporal graph convolutional networks,which can capture the spatial configuration and time change of the key points,the human behavior is identified.(4)We adopt Kalman filter based multi-object tracking algorithm to identify and judge whether the pedestrians enter the danger areas in the school or not.Our method can detect both single and multi-person situations.