Automotive Target Tracking Method Based On Machine Vision

With the economic and social development,the total quantity of vehicles had been increasing in China during the last few years.Nevertheless,the growth of the highway total mileage is slow at the same time.This causes that the traffic supply and demand are unbalanced,and safety along with other traffic contradictions have been deteriorated.In recent years,the intelligent technology has been emerging one after another and developing,the intelligent vehicle has gradually become an effective means to alleviate and even solve traffic contradictions,and has regarded as one of the key support objects by national policies such as "Made in China 2025" and "The 14 th Five-Year Plan for the Development of modern comprehensive transportation system".The intelligent vehicle is a synthetic system that integrates multi functions including the environment perception,the planning decision,and the multi-level assisted driving.Among these functions,environment perception is the basic one for the planning decision and automatous driving,which is directly related to driving safety and efficiency.As one of most vital parts of the intelligent vehicle environment perception,the vehicle target tracking along with its accuracy and real-time performance is critical for the driving safety and efficiency of the intelligent vehicle.Therefore,"vehicle target tracking method" was studied in this paper with the utilization of the machine vision technology,and the main research contents are(1)Hardware selection and system construction of vehicle target tracking.Firstly,considering the complexity of traffic scene,the specification and performance requirements of front-end sensor and back-end processor for vehicle target tracking were analyzed systematically.Secondly,based on the analysis results,the hardware part of the system composed of sensors and processors was built.Thirdly,aiming at the problem of insufficient samples of some types of vehicles in the existing open data set,the method of collecting vehicle images and self-labeling was used to expand the types and sample size,and the self-built data set was constructed.Finally,the deep learning environment was deployed on the processor to complete the system construction.(2)Vehicle target detection method based on improved YOLOv5 s algorithm.Detection is a precondition for tracking.Target tracking performance is directly affected by target detection method.In the field of vehicle target detection,the methods based on YOLOv5s were gaining popularity.However,the existing vehicle target detection methods based on YOLOv5s still have room for improvement in terms of detection accuracy,false detection rate and robustness in complex scenes.aiming at this problem,two improvement strategies were proposed in this paper:one was to replace C3 module in the trunk structure with Co T3 module;the other was to replace the loss function with EIo U or SIo U,and the improvement effect was analyzed by ablation experiment.The results showed that the improved YOLOv5s algorithm based on Co T3 module and SIo U loss function had better comprehensive performance.Meanwhile,the experimental results compared with the YOLOv4-tiny algorithms and YOLOv7-tiny algorithms prove the superiority of the improved YOLOv5s algorithm in detection performance.Compared with the unimproved method,the accuracy of the proposed method was increased by 5.7%and the average accuracy was increased by 7.4%.(3)Vehicle target tracking method based on Strong SORT algorithm.Deep SORT algorithm was widely used in target tracking.However,the existing vehicle target track methods based on Deep SORT algorithm generally have the defects of unstable continuous tracking and low retracking implementation rate after the targets were occluded.Aiming at this problem,the Strong SORT algorithm for vehicle target tracking was proposed in this paper.Firstly,the enhanced correlation coefficient algorithm was used to compare and register two consecutive frames.Secondly,NSA Kalman filter and Bo T feature extractor were used to re-adapt the motion features and appearance features of target vehicles respectively.Finally,global linear matching was used to replace the cascade matching mechanism in Deep SORT algorithm.By comparing and evaluating the experimental results,the performance of Strong SORT algorithm for target tracking in this paper was proved,which was as follows: Compared with the target tracking method based on Deep SORT,the Strong SORT method proposed in this paper achieved4.9%,7.4% and 3.6% performance improvement on MOTA,IDF1 and HOTA,respectively.(4)Verification and evaluation of vehicle target tracking methods considering multiple vehicle types.In order to comprehensively test and evaluate the overall practical application performance of the vehicle target tracking method based on the improved YOLOv5 s and Strong SORT,on the one hand,the video data set collected by dashcam was used in this paper.The average missed detection rate,average false detection rate,average retracking rate,total number of ID switches and average speed were used as performance indexes to test and evaluate the performance of the proposed method.The results showed that the proposed method achieves an average processing time of 70.05 ms,an average missed detection rate of 14.1%,an average false detection rate of 4.0%and an average retracking rate of 29.2%in various types of vehicle detection and tracking tasks in the front area that has an impact on the normal running of vehicles.On the other hand,natural driving experiments in six typical traffic scenes including normal traffic flow,crowded traffic flow,lane change behavior,turning behavior,illumination change and retracking after occlusion were organized to verify and evaluate the performance of the proposed method in actual deployment to the vehicle platform.The results showed that the proposed method can realize the target detection and continuous and stable tracking of various vehicle types in typical traffic scenes,and had high robustness.The above results showed that the vehicle target tracking method based on machine vision proposed in this paper could realize the target detection and tracking of various vehicle types in typical urban road traffic scenes,including small passenger cars,trucks,buses,etc.The research results were robust and real-time,which could provide theoretical and methodological references for the development and application of intelligent vehicles,advanced driver assistance systems,autonomous driving and other technologies.