Design And Implementation Of Pedestrian Multi-target Tracking System Based On DeepSort

With the improvement of Deep Learning theory and the continuous improvement of the accuracy of object detection algorithm,the multiple object tracking algorithm based on object detection has ushered in a new stage of development.The main task of multiple object tracking is to locate multiple targets in continuous images and give the ID and trajectory of each object.Although multiple object tra cking is widely used in video surveillance,human-computer interaction and national defense tasks,two tasks need to be solved for multiple object tracking :(1)accurately maintain the object's ID and trajectory;(2)handle with frequent occlusion,similar appearance occlusion,interclass interference,etc.Solving the above two problems will improve the accuracy of multiple object tracking and promote the commercial application of multi-target tracking.Multiple object tracking system mostly adopts the ide a of tracking-by-detection.First,the object detection algorithm is used to detect the interested target in each frame,and the object central coordinate and dimension information are obtained.Secondly,the feature matching algorithm is used to match the detection results with the objects of the previous frame.This thesis mainly studies the design and implementation of pedestrian multiple object tracking system based on DeepSort,comprehensively summarizes the development process,research status and key theories of multiple object tracking algorithm,and improves tracking algorithm,so as to achieve a more accurate pedestrian tracking system and provide more accurate coordinates of targets and more stable ID.The main tasks of this thesis are:(1)YOLO V4 algorithm with excellent detection performance is used to detect the target.In order to improve the accuracy of detection,the 3 scales of YOLO V4 were changed into 4 scales.(2)The cascade matching algorithm in DeepSort multiple object tracking system and the Hungarian assignment algorithm were used as the basic tracking architecture.In order to improve the distinction between motion information and appearance information,the depth feature and manual feature were introduced,and the multi-features were divided and conquer.(3)In order to avoid over-fitting of detection algorithm,label smoothing operation is introduced to improve detection accuracy.(4)In order to improve the ease of use of the tracking system,an open source framework and Python language are used to make a visual interface for easy operation.The pedestrian multiple object tracking system proposed in this thesis is based on the well-performance YOLO algorithm.The high-precision detection makes the matching algorithm less dependent on the detection result,less interference from factors such as occlusion and illumination changes,and the tracker is more stable.The algorithm is used to delete and select video frames with small changes,which reduces the detection cost and improves the tracking speed of the algorithm.