Abstract:

Aiming at the problems of the large error rate, slow detection speed, and manual participation required for assisted identification of traffic signs in autonomous driving in China, an improved YOLOv3 traffic sign detection and recognition method is proposed. Firstly, the network structure of Darknet53 is improved to reduce the forward inference calculation during network iteration and increase the network iteration speed. Secondly, the direct evaluation index GIoU of target detection is introduced to guide the positioning task to improve the detection accuracy. Finally, the [k]-means++ clustering algorithm is used to obtain the anchor size and match the corresponding feature layer. The experimental results show that compared with the original YOLOv3, the proposed method has an average accuracy improvement of 8% on the standard data set Lisa, and the detection speed reaches 76.9 f/s. The average accuracy on the homemade data set CQ-data can reach 94.8%. Compared with traditional recognition and other algorithms, it not only has better real-time performance and accuracy, is more robust to various environmental changes, but also can recognize multiple types of traffic signs.

Key words: traffic sign detection, YOLOv3, GIoU, dimension clustering

摘要：

针对我国自动驾驶的辅助识别交通标志误差率大、检测速度慢、需人工参与等问题，提出一种基于改进YOLOv3的交通标志检测识别方法。通过改进Darknet53网络结构来减少网络迭代过程中前向推理计算，提升网络迭代速度。引入目标检测的直接评价指标GIoU指导定位任务来提高检测精度。使用[k]-means++聚类算法获取anchor尺寸并匹配到对应的特征层。实验结果表明，提出的方法相较于原始YOLOv3在标准数据集Lisa上的平均精度提升了8%，检测速度达到了76.9 f/s；在自制数据集CQ-data上平均精度可达94.8%，与传统识别以及其他算法相比，不仅具有更好的实时性、准确性，对各种环境变化具有更好的鲁棒性，而且可以识别多种交通标志的类型。

关键词: 交通标志检测, YOLOv3, GIoU, 维度聚类