基于改进YOLOv5和SORT算法的车辆跟踪系统

Vehicle Tracking System Based on Improved YOLOv5 and SORT Algorithms

研究了复杂道路环境下识别模型中的参数量大,以及在遇到遮挡时利用SORT算法追踪目标时易出现的ID切换问题。在识别模型上,设计了一款基于YOLOv5的轻量化模型,使用MobileNetV3轻量级网络优化原网络结构,以减少模型参数。参数的减少可能会导致模型精度的下降,在特征融合网络的末端加入CA注意力机制,以提高模型的特征提取能力,并在Head层将原本CIOU Loss函数替换为EIOU Loss函数。在追踪算法上,SORT算法引入马氏距离公式与最小余弦公式,以实现SORT模型在遇遮挡时保持前后ID信息一致。结果显示,优化后的YOLOv5平均精度(mAP50)为89.3%,与优化前相比,模型平均精度在下降2.7%的同时,模型参数量减少了49.4%,大幅降低了模型复杂度。优化后的SORT算法在遇遮挡时能顺利完成目标的重检测,实现了一致的跟踪ID。

Studied a large number of parameters in recognition models in complex road environments,as well as the ID switching problem that is prone to occur when using SORT algorithm to track targets when encountering occlusion.On the recognition model,a lightweight model based on YOLOv5 was designed,using MobileNetV3 lightweight network to optimize the original network structure and reduce model parameters.The reduction of parameters may lead to a decrease in model accuracy.To improve the feature extraction ability of the model,a CA attention mechanism is added at the end of the feature fusion network,and the original CIOU Loss function is replaced with an EIOU Loss function in the Head layer.In terms of tracking algorithms,the SORT algorithm introduces the Mahalanobis distance formula and the minimum cosine formula to achieve consistency in the front and back ID information of the SORT model when encountering occlusion.The results showed that the average accuracy(mAP50) of YOLOv5 after optimization was 89.3%.Compared with before optimization,the average accuracy of the model decreased by 2.7%,while the number of model parameters decreased by 49.4%,significantly reducing model complexity.The optimized SORT algorithm can successfully complete target re detection and achieve consistent tracking IDs when encountering occlusion.