Research On Multi-Objective Vehicle Tracking Technology Based On DeepSORT Algorithm

The rapid development of road transportation in our country directly leads to a significant increase in the number of vehicles.To improve the throughput efficiency of road transportation,it is necessary to detect and track vehicles in road monitoring systems to monitor the traffic operation status in real time.However,in practical scenarios,the detection and tracking processes are constrained by objective environmental factors and terminal device deployment issues.Therefore,this paper focuses on studying the aforementioned problems and the specific research work is outlined as follows:Firstly,an introduction is provided on deep learning techniques,elaborating on the theoretical aspects of convolutional neural networks,including convolutional layers,pooling layers,fully connected layers,and other common network layers and related network architectures.Secondly,a comparison of various object detection algorithms is conducted,and a one-stage object detection algorithm,represented by YOLO,is selected as the multi-object detector.To improve the issues of large model parameter size and slow detection speed,the Mobile Net V3 network improved with the ECA attention mechanism is adopted as the backbone network,and the original PANet network is replaced with the Bi FPN network to enhance the fusion of feature maps across different layers.Evaluation is performed using a mixed dataset of vehicles.The experimental results demonstrate a 54% reduction in the parameter size of the improved YOLOv5 algorithm,with minimal impact on detection accuracy,validating the advantages of the improved detection algorithm over the original algorithm in edge device deployment and real-time detection requirements.To address the problem of the large volume of the re-identification network model in the DeepSORT algorithm,this study introduces the lightweight network Shuffle Net V2.Evaluation on a subset of the Veri-Wild dataset reveals that the lightweight vehicle appearance feature extraction network achieves a significant reduction in model size to 6% of the original,without a noticeable decrease in detection accuracy.The experiments confirm that the lightweight network effectively reduces the memory usage of the model while maintaining accuracy,making it a viable replacement for the original network structure in the re-identification task.Furthermore,the traditional Intersection over Union(IOU)method,used to calculate the overlapping area between detection boxes and trajectory boxes,yields an IOU value of 0 when there is no overlap between the two,making it unable to quantify the loss between them using the intersection over union ratio.To address this issue,this paper adopts the Distance Intersection over Union(DIOU)distance metric as a replacement for the original IOU distance metric,aiming to improve the accuracy and stability of object tracking and effectively reduce the number of ID swaps.In summary,the improved YOLOv5 achieves a detection speed of 77 frames per second(FPS)in the mixed dataset of vehicle object detection.Compared to YOLOv5 s,it exhibits an approximately 0.12% improvement in m AP@0.5 and a 0.34%improvement in m AP@0.5:0.95.In the multi-object tracking of vehicles,the adoption of the lightweight Shuffle Net V2 and the more accurate DIOU distance metric calculation leads to a reduction in the parameter size of the ReID network from 45 M to 2.5M,with an approximately 0.2% improvement in detection accuracy.Validation on the mixed dataset demonstrates that the improved algorithm maintains almost the same detection accuracy as the original DeepSORT algorithm,while achieving a 7FPS increase in detection speed.Compared to the SORT algorithm,the number of ID swaps is reduced by nearly 65 times.Therefore,the improved DeepSORT algorithm can better fulfill the requirements of multi-vehicle tracking tasks in road environments.