Pedestrian Abnormal Behavior Recognition Based On Video Analysis

With the increasing attention to public safety and the rapid development of video surveillance equipment and computer vision technology,the analysis of pedestrian behavior in video surveillance technology has also become one of the research hotspots.Video-based pedestrian behavior analysis includes pedestrian trajectory analysis,pedestrian action gesture recognition,pedestrian flow prediction and other directions.One of the important branches is the analysis of abnormal behaviors based on motion trajectories.Since abnormal behaviors are difficult to specifically define,it is difficult to apply the binary classification method based on supervised learning in this field.The motion trajectories contain rich semantic information,which has a strong correlation in space and time.The correlation of pedestrians can be analyzed through the trajectory.Starting from the pedestrian’s motion trajectory,this paper studies how to quickly and accurately identify pedestrians and track them for a long time.Then,analyze pedestrian motion patterns from these massive trajectory data,and cluster normal patterns to identify abnormal behaviors.The research contents of this paper are as follows:Aiming at the problem of target loss caused by occlusion in target detection and tracking technology,this paper proposes a tracking by detection method.In order to improve the target detection speed and reduce the ID switch,the backbone of YOLOv5 is improved,and the partially redundant convolution operations are replaced with lightweight operations to reduce the amount of parameters and operations.By matching the high-scoring frame and the lowscoring frame respectively and extracting the features of targets,the occluded objects are excavated from the low-scoring frame to maintain the coherence of the trajectory.The experimental comparison verifies the effectiveness of the multi-objective matching algorithm in this paper.Aiming at the problem that the existing trajectory clustering algorithms lack reasonable and effective similarity measurement,this paper proposes a new trajectory segmentation method and trajectory similarity measurement method,which can calculate the features of acceleration,velocity,growth and direction of each trajectory.First,trajectories are clustered through DBSCAN clustering to obtain different trajectory clusters,and then LSTM is used to focus on historical and current trajectory sequence data.Each type of trajectory that appears in the scene is input into the LSTM network for modeling to obtain normal behavioral trajectories model.The trajectory to be detected is divided into sub-trajectories for temporal and spatial anomaly detection,and the overall trajectory uses the spatiotemporal semantic information to measure the similarity and compare with the threshold to determine whether it is an abnormal trajectory,thereby identifying the abnormal behavior of pedestrians.Experiments on the road dataset collected in this paper show that the method in this paper can effectively identify the abnormal behavior of pedestrians.