Study On Crowd Statistics And Situation Analysis Method Considering Pedestrian Position

Crowd counting has important applications in many fields such as epidemic prevention and control,security protection,commercial operation,and fire emergency.In recent years,with the development of computer vision technology,deep learning algorithms have greater advantages in target detection and tracking.The algorithm of using deep learning-based target detection and tracking technology to count pedestrians through surveillance video has been developed relatively well,but due to the limitations of a single camera,it still cannot meet the current needs of users.And multi-view video surveillance can solve the problems of occlusion and limited view range caused by single view,so the cooperative monitoring of dynamic targets based on multiple cameras is a research hotspot in the field of computer vision in recent years.In this paper,starting from the principle of target detection and tracking based on deep learning,we mainly study the pedestrian tracking and counting algorithm based on feature similarity,the target localization algorithm based on 3D space,the search algorithm of the same target under multiple cameras,and the visualization method of crowd density distribution in the study area,and the main research contents are as follows:(1)A pedestrian tracking counting algorithm based on feature similarity is studied for the problem of high pedestrian counting due to the target frame drift phenomenon in the tracking process caused by large pedestrian distribution density and fast pedestrian movement speed.Based on deep learning target detection and tracking network,the algorithm uses individual pedestrian feature difference information to construct a target feature pool,then judges the matching degree between the detection target and the trajectory feature pool by the feature difference function,and finally obtains the target matching result based on the optimal matching degree.By selecting a typical pedestrian-intensive dataset to train the target detection network and conducting target tracking experiments on video images such as MOT-16,the results show that the improved target detection and tracking algorithm in this paper reduces the ID conversion frequency and can meet the real-time requirements of tracking and curb the occurrence of target frame drift.(2)A three-dimensional space-based target localization algorithm is proposed for the occlusion problem caused by single-view statistics and the missed detection problem caused by limited detection field of view in the current number of statistical calculation method.The algorithm expands the study area by monitoring multiple cameras simultaneously and solves various problems caused by the single-viewpoint limitation.The basic idea is to use the principle of visual localization,that is,to construct a transformation model of camera and world coordinate system,to import the two-dimensional image coordinates into the three-dimensional space,and to instantiate each point,and to complete the regional pedestrian count by counting the points in the current space through the target search algorithm among multiple cameras.At the same time,a 3D coordinate system is constructed,and the point clouds generated by all cameras are represented in the coordinate system,and the study area is processed in blocks using differential correction to realize the mapping of pedestrian 3D spatial location coordinates.(3)The process of crowd statistics and posture analysis method considering pedestrian location is designed by combining two key algorithms.Firstly,the pedestrian target detection and tracking in the video is completed,and the feature matching module of current target frame and historical trajectory is proposed for the target frame drifting problem in the tracking process to strengthen the correlation between pedestrian features.Secondly,a mutual mapping model between 2D image coordinates and 3D spatial coordinates is constructed to convert pixel coordinates into 3D spatial coordinates,and a cross-camera target matching strategy is proposed to associate the same pedestrian target between different cameras by three constraints.Finally,the target point cloud generates a crowd area density distribution map to achieve an integrated control of pedestrian distribution in the study area.WILDTRACK image sequences captured by three cameras at the same time are selected for experiments,and the experimental results show that the algorithm constructs a real-time cognitive model of pedestrian location through a deep learning-based detection and tracking algorithm,restores the location of pedestrians in the 3D scene,and improves the accurate localization and tracking capability of dynamic targets.