Research On Abnormal Action Recognition Based On Human Joint Points

As the demand for social security continues to increase,electronic monitoring devices have been widely deployed in various settings and have played a significant role,such as in classrooms,where they can record all student behavior in real-time.Traditional monitoring devices cannot analyze the large amount of information contained in videos and can only detect abnormal behavior through manual supervision,which is no longer sufficient to meet modern society’s security needs.Therefore,analyzing abnormal human behavior in surveillance videos using computer vision technology has gradually attracted attention in various fields.Abnormal human behavior refers to behavior that deviates significantly from other observed results in specific scenarios.The current challenge in detecting abnormal human behavior is that the diversity of human poses makes it challenging to accurately describe human behavior.Detection typically requires a precise estimation of an individual’s posture information.However,in practical scenarios,factors such as occlusion and complex postures can cause a decrease in the accuracy of abnormal behavior detection.At the same time,in practical applications,data collection and labeling are often limited by factors such as time,manpower,and cost,resulting in insufficient data quantity and poor data quality,which affects the performance of abnormal behavior detection.This thesis mainly focuses on detecting abnormal behavior in scenes such as kindergarten classrooms.In-depth research is conducted on the above-mentioned problems,and abnormal behavior detection is based on human joint points.This thesis’ s main contribution is as follows:(1)A behavior feature extraction method based on human joint coordinates was designed.The target tracking and posture estimation network were used to obtain human posture information,and various features were extracted from the 2D joint coordinate information using the correlation between human joint points to describe human behavior and eliminate interference from video image backgrounds.A mixture of Gaussian distribution models was used to fit the feature data,complete feature conversion,and reduce data dimensions.(2)An unsupervised learning and supervised learning combination pattern was constructed to improve the performance of abnormal behavior detection.The unsupervised clustering algorithm was used to process the transformed Gaussian model parameters,and data with significant differences in parameters compared to normal behavior were labeled as abnormal.Additionally,the dataset for abnormal behavior recognition is difficult to obtain using existing technological methods.As a result,the unsupervised learning pattern proposed in this thesis can extract usable abnormal behavior data from a large dataset of video data.The behavior classification model was trained using supervised learning to identify labeled abnormal behavior,including desk crawling,roughhousing,running,and standing.This method combining unsupervised clustering and supervised learning can accurately detect abnormal behavior,and avoid processing a large amount of useless data,and speed up the detection process.This thesis used kindergarten classrooms and meeting rooms as experimental backgrounds,and demonstrated through relevant experimental results that the abnormal behavior detection algorithm proposed in this thesis can detect multiple types of abnormal behavior in complex environments and has high recognition accuracy.