Research And Application Of An Improved Algorithm For Pedestrian Multi Target Tracking Based On Deep Learning

In recent years,computer vision technology has been widely used in public place monitoring,automatic driving,garbage classification and other fields.Among them,target detection and tracking are an important application field of computer vision technology.With the increasing development of deep learning theory,target detection based on deep learning target detection and tracking better than traditional target detection and tracking.However,the existing detection and tracking algorithms have low detection efficiency and slow speed in public places with high crowd density and high traffic.Besides,the detection targets are often intertwined,similar appearance and background color of the detection targets are similar,leading to issues such as missed detection,false detection,and disordered ID allocation of target tracking.In response to the above problems,this paper proposes an improved algorithm for pedestrian multi-target tracking based on YOLOv5 and Deep SORT,which can effectively improve the effect of target detection and tracking.The research work of this paper mainly focuses on the following three aspects:(1)Optimized the YOLOv5 target detection network to improve the accuracy of pedestrian detection and reduce the number of missed detections.First of all,the YOLOv5 detector combined with the global attention mechanism replaces the Faster R-CNN detector.Secondly,WIo U loss function is introduced and DIo U-NMS nonmaximum suppression replaces non-maximum suppression to further improve the training model and solve the problem that the model cannot be optimized when Io ULoss is 0.(2)A pedestrian detection method based on head frame is proposed,which filters the data of the detection frame in the detection section to select data that is more relevant to the actual situation and improve the accuracy of the detection.Data filtering is dominated by the head frame and supplemented by the body frame.After the detection frames are matched by custom rules,redundant and messy detection frames are removed to obtain more accurate detection frames,which are put into Deep SORT for tracking.(3)Optimized the Deep SORT target tracking model.This model can greatly reduce the index of identity transformation and improve the robustness of the entire target detection and tracking model.Firstly,reconstruct the feature extraction network model of Deep SORT,which the model extracts deeper semantic information for pedestrians by removing one convolutional layer,adding four residual layers and adopting an adaptive average pooling layer;Secondly,after the cascade matching,the Io U matching is replaced with a more realistic DIo U matching,so as to improve the robustness of the tracking algorithm model.The improved algorithm proposed in this paper was tested on the MOT16 data set.The evaluation indicators of the MOT Challenge have been improved,and the disorder of pedestrian ID allocation has been greatly reduced.It can effectively detect in complex places and has a certain practical value.