基于YOLOv5s和DeepSORT的改进多目标跟踪算法  

Improved Multi-object Tracking Algorithm Based on YOLOv5s and DeepSORT

作　　者：王红林[1] 黄浩 孙彩云[1] 毛煜鑫 WANG Hong-lin;HUANG Hao;SUN Cai-yun;MAO Yu-xin(School of Artificial Intelligence/School of Future Technology,Nanjing University of Information Science and Technology,Nanjing Jiangsu 210044,China)

机构地区：[1]南京信息工程大学人工智能学院(未来技术学院),江苏南京210044

出　　处：《计算机仿真》2025年第3期263-269,303,共8页Computer Simulation

基　　金：国家自然科学基金委员会青年项目(62101274)。

摘　　要：针对多目标跟踪过程中因漏检和遮挡等因素导致的跟踪准确率降低的问题,提出一种改进的YOLOv5s和DeepSORT的实时多目标跟踪算法。在检测部分将主干网络融合Transformer编码块来优化YOLOv5s,具有比传统卷积神经网络更强的特征提取能力。在跟踪部分,用ECA注意力模块,RepVGG网络,Resnet网络和eSE Block设计了一种新的表观模型结构来缓解因遮挡等因素而出现的跟踪失败的情形,提升了模型鲁棒性。在VisDrone和MOT16数据集上进行实验。结果表明,上述算法检测器在Visdrone数据集上AP50提升了4.07%,算法在MOT16数据集上MOTA提升7.9%。,改进后的DeepSORT算法与目标检测算法相结合能够有效提高跟踪性能。Aiming at the problem of low tracking accuracy caused by factors such as missing detection and occlusion in the process of multi-target tracking,an improved real-time multi-target tracking algorithm of YOLOv5s and DeepSORT is proposed.In the detection part,the backbone network is fused with the Transformer coding block to optimize YOLOv5s,which has a stronger feature extraction ability than the traditional convolutional neural network.In the tracking part,a new appearance model structure is designed with ECA attention module,RepVGG network,Resnet network and eSE Block to alleviate the tracking failure caused by occlusion and other factors,and improve the robustness of the model.Tfhe experiments are performed on the VisDrone and MOT16 datasets.The results show that the algorithm detector improves the AP50 by 4.07%on the Visdrone dataset,and the algorithm improves the MOTA by 7.9%on the MOT16 dataset,which can effectively improve the tracking performance,the improved DeepSORT algorithm combined with the target detection algorithm can effectively improve the tracking performance.

关 键 词：通道注意力 特征融合网络 多目标 跟踪算法 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]