Research On Multi-object Tracking Algorithm Based On Joint Detection And Tracking Paradigm

Multi-object tracking in dense pedestrian scenes has always been a difficult problem in pedestrian monitoring.Such dense pedestrian motion scenes not only make models face challenges in detection,but also a large number of object intersections and occlusions seriously affect the effectiveness of data association in traditional tracking algorithms.Therefore,for dense pedestrian scenes,this paper adopts an algorithm structure based on a joint detection and tracking paradigm,and designs a multi-object tracking network that integrates detection and tracking.It consists of an improved object detection model based on Center Net and an improved multi-object tracking model based on Center Track.It can effectively reduce the identity switches in the tracking process in dense pedestrian scenes,and effectively improve the tracking accuracy of the multi-object tracking network in such scenes.Firstly,aiming at the inaccurate object detection and localization of Center Net network in dense pedestrian scenes,the Gaussian effective radius generation method and loss function are improved.A new generation method of Gaussian effective radius of heatmap is proposed,which simplifies the three different situations in the generation process into one unified situation.Through the demonstration of the L1 norm and L2 norm loss function forms,the object size regression loss function and the center point offset loss function are improved to improve the detection accuracy in dense pedestrian scenes.Secondly,aiming at the problem of identity switches caused by only using motion information tracking in the Center Track,a re-identification feature extraction method based on the cost matrix is proposed,and the displacement information of the object is extracted from the cost matrix to replace the direct regression of displacement in Center Track.At the same time,a feature fusion method based on deformable convolution is used to convert the predicted displacement and feature frame difference into the offset of each point of the deformable convolution kernel,and a channel and spatial attention mechanism is added to the convolution layer.In order to enhance the feature extraction ability for the key position of the feature frame difference.In addition,in the data association part,an embedded feature stacking method is proposed to process the trajectory regeneration.Finally,the dense pedestrian tracking data set MOT20 is selected,and a large number of ablation comparison experiments are carried out on the detection model and tracking model in the designed multi-object tracking network respectively.Compared with the original Center Track network,the MOTA score of the improved multi-target tracking network increased by 5.72%,the IDF1 score increased by 56.87%,and the number of identity switches decreased by 92.91%.