Abstract: In the face of complex road environment, a road environment target tracking optimization algorithm model based on YOLOv5 and DeepSort is proposed. The recognition of road environment targets is based on YOLOv5. The basic structure of MobileNet V3 network is used to replace the recognition network in DeepSort algorithm and the backbone network in YOLOv5, which can compress the model and speed up the detection. This paper adopts the method of combining CIoU and GIoU to modify the original loss function, which makes up for the defect that GIoU degenerates into IoU in some cases. The experimental results show that the model weight decreases while the target tracking accuracy of the original network is improved, which can better achieve the continuous tracking effect.

Key words: Road target tracking; DeepSort; YOLOv5; Convolutional neural network; Optimization algorithm

摘要： 面对复杂的道路环境，提出一种基于 YOLOv5 及 DeepSort 的道路环境目标追踪优化算 法模型。基于 YOLOv5 进行道路环境目标的识别，使用 MobileNet V3 网络中的基本结构对 DeepSort 算法中的重识别网络和 YOLOv5 中的主干网络进行替换，以起到压缩模型，加快检 测速度的效果。在损失函数上，文章采取 CIoU 和 GIoU 相结合的方法对原有损失函数进行改 造，弥补了 GIoU 在一些情况下退化成 IoU 的缺陷。实验结果表明，优化后的网络在模型权 重大小下降的同时目标追踪准确度和精度均提升，可以更好达到持续追踪效果。

关键词: 道路目标追踪；DeepSort；YOLOv5；卷积神经网络；优化算法