Research On Fuzzy Vehicles Detection And Tracking Algorithm

Intelligent driving obtains the information of the target vehicles from the video through the detection algorithm and the tracking algorithm.The small target,congestion and occlusion lead to the unclear expression of the position of the target vehicles,and the high speed and darkness lead to the unclear expression of the shape of the vehicles.The above two fuzzy conditions lead to the low detection accuracy and tracking accuracy of the vehicles,which is difficult to ensure the safety of the intelligent vehicles.In order to improve the detection and tracking ability of fuzzy vehicles,the shortcomings of Yolov5 s in detecting fuzzy vehicles and the problems of Ocsort tracker are analyzed.Then,based on the shortcomings of Yolov5 s and Ocsort,the detector and tracker are improved respectively to construct a high-precision fuzzy vehicles tracking system.The main research contents are as follows :(1)The HCD-DETECT is fuzzy vehicles detection dataset that Collected by myself,which is made for the five fuzzy conditions of occlusion,congestion,small target,darkness and motion.HCD-DETECT is used to train HCD-yolov5 s to enhance the detection ability of the detector to fuzzy vehicles.The HCD-MOT is tracking dataset which is made for four kinds of videos that are easy to produce fuzzy vehicles,such as foggy days,nights,high speeds and congestion.HCD-MOT is used to evaluate the tracking performance of the tracker.(2)The HCD-yolov5 s detector is designed to enhance the detector’s ability to detect fuzzy vehicles.Firstly,aiming at the ambiguity of vehicles location information,the detection head is introduced from the shallow network to retain the location information of the fuzzy vehicles.Then,the DConv module is designed to separate the regression and classification tasks and enhance the generalization ability of the model.Finally,a CCAT feature fusion module is designed to enhance the expression of model position information and semantic information for different feature layers suitable for position blur and shape blur.The m AP50 of HCD-yolov5 s on Visdrone and Dota datasets is increased by 10.2% and 1.8%,respectively,indicating that HCD-yolov5 s is more suitable for detecting location information and shape information.The m AP of HCDyolov5 s on HCD-DETECT is increased by 8.0%,which verifies that HCD-yolov5 s is more suitable for detecting fuzzy vehicles.(3)The SG-ocsort tracker is designed to enhance the tracker’s ability to track fuzzy vehicles.Firstly,using HCD-DETECT training HCD-yolov5 s which is used as the detector of SG-ocsort to reduce the missing value of trajectory.Then,considering that Ocsort uses linear interpolation method to deal with the trajectory blank value,which cannot fit the nonlinear motion state,this thesis uses Gaussian smoothing interpolation to fit the target trajectory.Finally,Io U only considers intersection and combination ratio.This thesis uses SIo U to replace Io U matching,which adds information such as shape,distance,and angle to the Cost Matrix.The tracking ability of SG-ocsort in HCD-MOT dataset is better than Ocsort and Deepsort.MOTP and MOTA are improved by 2.9% and1.3% respectively on the basis of Ocsort,and IDSW is reduced by 9.0%.MOTP and MOTA are increased by 0.8% and 1.5% respectively on the basis of Deepsort,and IDSW is reduced by 15%.which verifies that SG-ocsort is more suitable for tracking fuzzy vehicles.The experimental results show that the designed HCD-yolov5 s is more suitable for detecting fuzzy vehicles,reducing the missed detection rate of the detector and reducing the missing value of the tracking trajectory.The SG-ocsort improves the problems of detector failure,occlusion,and lack of trajectory information caused by nonlinear motion,and enhances the tracking ability of the tracker to blurred vehicles,thereby improving the safety of intelligent vehicles.