Design And Implementation Of Wandering Behavior Detection Method Based On Pedestrian Trajectory

With the development of computer vision,related technologies such as image processing and video analysis have begun to be applied to video surveillance systems,accelerating the development of video surveillance systems in the direction of intelligence.Due to the rapid increase in the number of surveillance videos,traditional processing methods are no longer suitable for processing massive amounts of data,while intelligent surveillance systems are highly efficient and accurate,and are rapidly replacing traditional surveillance systems,which are of great significance to the development of social security towards intelligence.Abnormal behavior detection can use computer vision methods to analyze and process continuous video frames in the video stream to detect and track pedestrians in motion in the video,and further determine whether pedestrians have abnormal behaviors.Wandering behavior is a kind of abnormal behavior.The content of the research here is to analyze and process the video stream,obtain pedestrian trajectories through detection,and judge whether pedestrians have wandering behaviors based on the trajectory.In the detection process of wandering behavior based on pedestrian trajectory,it mainly includes three parts:pedestrian detection,pedestrian tracking and wandering behavior detection.The main tasks and research contents are as follows:(1)Pedestrian detection.Wandering behavior detection requires real-time analysis of pedestrians in the video as much as possible to determine whether the detected pedestrians have wandering behavior.YOLOv5(You Only Look Once)has the characteristics of high detection accuracy and fast detection speed.Here,YOLOv5 detector is used for pedestrian detection and simple optimization is made.In order to further improve the accuracy and speed of pedestrian detection,YOLOv5 is trained on the organized pedestrian data set.Comparing the trained model with the original model,through the experimental test,its speed and accuracy have been improved,and the expected effect has been achieved.(2)Pedestrian tracking.A multi-pedestrian tracking algorithm based on YOLOv5 is proposed here.By tracking the pedestrians detected in the video,the detected pedestrian trajectory is obtained.In the process of pedestrian tracking,occlusion will cause the interruption of the pedestrian trajectory.This is mainly aimed at the problem of pedestrian trajectory interruption caused by short-term occlusion in the video.By introducing pedestrian appearance feature information,the defect of single use of location information is changed,and Res Net18 is used as the pedestrian appearance feature extraction network.At the same time,measurement and matching are used twice to improve the accuracy of tracking.In the experiment,through the real scene video pedestrian tracking test,it is verified that the introduction of appearance features and twice measurement and matching can effectively improve the problem of pedestrian trajectory interruption caused by short-term occlusion,and reduce the number of pedestrian identity label switching in the tracking process.(3)Wandering behavior detection.A detection method of wandering behavior based on pedestrian trajectory is proposed here,and the direction of slope change of the trajectory point is used as the basis for whether the direction of pedestrian movement changes.When the pedestrian movement direction has not changed,by counting the number of trajectory points,it is judged whether it is a trajectory segment that constitutes a wandering behavior trajectory.When the trajectory segment that constitutes the wandering behavior meets certain conditions,it is determined that the pedestrian has a wandering behavior,otherwise it is a non-wandering behavior.Combined with the pedestrian trajectory-based wandering behavior detection method,a pedestrian trajectory-based wandering behavior detection system is designed.Through the graphical interface,users can conveniently "add,delete,and modify" the video stream.The system analyzes and processes the video stream,obtains the detected pedestrian motion trajectory through pedestrian detection and pedestrian tracking,and then analyzes and processes the pedestrian motion trajectory to determine whether the pedestrian has wandered.If in the current video frame,a pedestrian is detected to have wandering behavior,the pedestrian is marked,the current frame is saved and sent to the front-end interface for early warning display.