Research And Implementation Of Pedestrian Multi-object Tracking System Based On Re-recognition

For the past few years,people’s working style and living habits have great changes with the progress and development of AI.Computer vision is an important part of AI.Many of experts in the world are focusing on this area.The area of object tracking has always been a popular area in CV,and multi-object tracking has been wide range of applications in area such as drones,automatic driving,traffic management,automatic navigation and military weapons.In many public places,cameras are often used as infrastructure.With the increasing number of cameras,the growth of video data requires the assistance of various data processing software,the pedestrian tracking has become a hot and useful area,so the research of multi-object tracking in this paper is significant.Through the research and analysis of recent object tracking algorithms,pedestrian multi-object tracking is difficult to run in real time.Tracking accuracy is low,tracking effect is not ideal;faced with multiple occlusion environment,it is difficult for the algorithm to match pedestrians correctly.To address the above problems,this paper proposes a pedestrian multi-object tracking system based on re-recognition,and the main research content of the system is as follows.YOLOv5 is a detection algorithm with high accuracy and light weight,it provides reliable detection results for this system without increasing the system performance burden.CIo U loss is introduced to retrain the network so that the detector can focus on the pedestrian object and provide reliable detection results for the system without increasing the burden of system performance by utilizing the characteristics of high precision and lightweight.Then,a lightweight fusion multi-scale weight recognition network LMSR is designed.The parameters of the convolution module are simplified by depth-separable convolution.The fusion multi-scale residual structure is used as the algorithm to improve the recognition accuracy and provide a better pedestrian feature extraction model for the system.Kalman filter is used to estimate the state of pedestrian objects.Combining the pedestrian features and Mahalanobis distance of re-recognition network,object detection data and pedestrian tracks are correlated.The Hungarian algorithm is used to optimally distribute the remaining pedestrian objects and finally complete the multiobject pedestrian tracking task.Finally,according to the demand analysis of this paper,this paper design a system about pedestrian object tracking based on re-recognition network.The main functions of the system include login and registration,object detection and tracking,user information management and system data management.The effectiveness of each module of the system is verified by experiments.The system test and experimental results show that the system is efficient and stable,and requirements analysis description is satisfied.