改进YOLOv5的中国交通标志检测算法

Chinese Traffic Sign Detection Based on YOLOv5 Algorithm

摘要: 针对实际复杂道路环境下，交通标志像素小而密集导致的检测定位精度较低，较高漏检的问题，本文提出了一种改进YOLOv5的中国交通标志检测算法。通过遗传算法和K-means算法优化先验锚点框来提高小目标的检测准确度；引入Bi-FPN结构对语义信息和位置信息进行特征融合，增强网络表征能力；增加注意力机制GAM，提升网络对于复杂环境下的抗干扰能力和特征提取能力；改变边框回归损失函数SIoU loss，实现模型的快速收敛和准确定位。实验结果表明，与标准的YOLOv5算法相比，本文改进的算法的平均精度均值mAP、精确度P、召回率、F1提升值分别为10.03个百分点、4.7个百分点、2.6个百分点和3.48个百分点，验证了该算法对中国交通标志检测的有效性。FPS达到了208.33FPS，满足实时性检测要求。

Abstract: In order to solve low accuracy and missing detection of traffic signs in sign recognition task, caused by complex road environment, small and dense signs filled of whole picture, we propose an improved Yolov5 algorithm. Firstly, we use genetic algorithm and K-means algorithm to get perfect anchors fit for the dataset which is good for the small objects. Secondly, we introduce the Bi-FPN structure to the neck of mode to combine high level semantical features and low-level features, which improves the representational capacity of the network. Thirdly, embedding the Gam attention into the backbone and neck could enhance feature extraction ability and anti-interference ability to messy background. Furthermore, a modified SIoU loss function is ap-plied to optimize the training process and improve accuracy. Experimental results show that the mean average precision (mAP), precision, recall and F1 score are improved by 10.03%, 4.7%, 2.6 % and 3.48 % compared with the benchmark YOLOv5 algorithm respectively, which shows that the proposed algorithm is effective for Chinese traffic sign detection. In addition, 208.33 FPS could meet the real-time requirement.