摘要： 智能驾驶中道路交通标志的检测与识别极其重要,针对交通标志被遮挡时存在目标小、检测精度低的问题,提出一种基于改进YOLOv5s的被遮挡交通标志检测算法。通过自制中国遮挡交通标志数据集(COTSD),解决交通标志在被遮挡情况下数据集缺乏的问题,并根据细粒度物体检测需要模型提取丰富判别特征,提出一种基于MobileNetv2改进的主干网络。针对交通标志较小、被遮挡且分辨率较低,普通线性插值方式无法捕捉更高阶、更细节特征的问题,设计一种改进的动态权重上采样模块(DEUM),整合经过通道注意力加权之后的通道信息进行像素重排,生成高分辨率图像;针对综合交并比(CIoU)等损失函数对小目标位置变化敏感的问题,使用归一化高斯Wassertein距离(NWD)来优化边界框回归损失。实验结果表明,该算法在自制被遮挡COTSD上的准确率为93.60%,召回率为72.50%,F1值为81.71%,mAP@0.5为79%;在包含少量被遮挡交通标志的公开CCTSDB上的准确率为92.2%,召回率为78.8%,F1值为85%,mAP@0.5为88.5%。该算法有效提高了被遮挡后交通标志的检测精度。

关键词: 交通标志检测, 遮挡条件, MobileNetv2模型, 动态权重上采样模块, 归一化高斯Wassertein距离

Abstract: The detection and recognition of road traffic signs are extremely important in intelligent driving. When traffic signs are obstructed, problems such as small targets and low detection accuracy occur. This paper proposes an improved model for obstructed traffic sign recognition based on YOLOv5s. To address the issue of a lack of dataset for traffic signs under occlusion, a self-made Chinese Occlusion Traffic Sign Dataset (COTSD) is developed. A backbone network based on MobileNetv2 improvement is proposed to extract rich discriminative features for fine-grained object detection. For small and obstructed traffic signs with low resolution, ordinary linear interpolation methods cannot capture higher-order and more detailed features; therefore, an improved Dynamic weight Upsampling Module (DEUM) is designed to integrate channel information weighted by channel attention for pixel rearrangement and generate high-resolution images. To address the sensitivity of loss functions such as CIoU (Complete IoU) to small target position changes, Normalized Gaussian Wasserstein Distance (NWD) is used to optimize the bounding box regression loss. For the self-made occluded COTSD dataset, the accuracy is 93.60%, recall is 72.50%, F1 value is 81.71%, and mAP@0.5 is 79%. For the publicly available CCTSDB dataset containing a small number of occluded traffic signs, the accuracy is 92.2%, recall is 78.8%, F1 value is 85%, and mAP@0.5 is 88.5%. The experimental results for the two datasets demonstrate that the improved algorithm can effectively improve the detection accuracy of traffic signs after occlusion.

Key words: traffic sign detection, occlusion conditions, MobileNetv2 model, Dynamic weight Upsampling Module (DEUM), Normalized Gaussian Wasserstein Distance (NWD)

中图分类号: 

• TP391