Abstract: With the rapid development of artificial intelligence, driverless driving has become one of the main research directions in the current automotive industry, and vision-based vehicle detection has become an irreplaceable part of driverless car technology. This paper proposes a new target detection method based on the YOLOv3 detection algorithm. First, an augmented data set is made on the basis of the COCO data set, and the characteristics of vehicles such as cars, trucks, motorcycles and wheels were marked, and then the obtained COCO data set and augmented data set are used to train the YOLOv3 network and test the experimental results. Then use the network structure to compare the detection results between different detection algorithms. The method proposed in this paper effectively improves the accuracy of vehicle feature extraction, but also improves the detection speed and increases the robustness.

Key words: Vehicle detection; YOLOv3; Unmanned driving; COCO

摘要： 随着人工智能的迅速发展，无人驾驶成为当前汽车行业的主要研究方向之一，基于视 觉的车辆检测成为了无人驾驶汽车技术中不可替代的一部分。文章基于单阶段检测算法 YOLOv3 提出了一种新的目标检测方法。首先在 COCO 数据集基础上制作增广数据集，并且 标注轿车，卡车，摩托车和车轮等车辆特征，之后用得到的 COCO 数据集与增广数据集对 YOLOv3 的网络进行训练，得到新的目标检测模型后将检测实验结果进行对比。然后利用该 网络结构对不同检测算法之间的检测结果进行对比。通过对比可知，文章提出的方法有效地 提高了车辆特征提取的准确度，同时也提高了检测速度，增加了鲁棒性，有效地解决了无人 驾驶环境感知模块检测精度低的问题，为无人驾驶决策模块提供精准的感知结果。

关键词: 车辆检测；YOLOv3；无人驾驶；COCO