Abstract: The road target detection is one of the key technologies in the field of autonomous driving. With the development of artificial intelligence, the application of road target detection is gradually widespread. This paper proposes a new target detection method based on the YOLOv5 network. The improvements include the integration of modules in ShuffleNet V2 and the use of GhostConv to transform the traditional Conv module. First, real-time video streams are collected in different road environments, and pictures and video streams are annotated. The modules in ShuffleNet V2 are integrated into the backbone network and the GhostConv module is used to improve the Conv module, which reduces the weight of the model and has less impact on the target detection accuracy. The marked images are input to the improved YOLOv5 network for training, and the obtained model is combined with the Deep SORT algorithm for target detection and tracking. The experimental results show that the weight of the obtained results decreases a lot, while the target detection accuracy increases. The improved network is lighter and easier to deploy on edge embedded devices.

Key words: Road target detection; YOLOv5; Model compression; Target tracking

摘要： 道路目标检测环节是自动驾驶领域的关键技术之一，随着人工智能的发展应用逐渐广 泛。文章基于 YOLOv5 网络提出一种新的目标检测方法，改进包括融合了 ShuffleNet V2 中的 模块，使用 GhostConv 改造了传统的 Conv 模块等。先在不同道路环境中实时采集视频流，并 进行图片和视频流的标注。在主干网络中融入 ShuffleNet V2 中的模块并使用 GhostConv 模块 改进 Conv 模块，在降低模型权重的同时对目标检测精度影响较小。将标注完成后的图片输入 改进后的 YOLOv5 网络进行训练，并将得到后的模型与 Deep SORT 算法结合，进行目标检测 追踪。实验结果表明，所得结果权重大小下降许多，而目标检测精确度有所上升。改进后的 网络更加轻便，易于部署在边缘嵌入式设备上。

关键词: 道路目标检测；YOLOv5；模型压缩；目标追踪