Abstract: Object detection in traffic scenes is a critical technology for intelligent transportation systems and autonomous driving. To address the issues of missed detections, false alarms, and insufficient detection accuracy in multi-scale targets when using the YOLOv5s algorithm in complex traffic scenarios, this paper proposes an improved CRD-YOLOv5s model. First, a convolutional block attention module (CBAM) mechanism is introduced into the backbone network to enhance the feature extraction capability for small targets through channel and spatial attention mechanisms. Second, the spatial pyramid pooling-fast (SPPF) module is replaced with an receptive field block (RFB) module, which leverages dilated convolutions to expand the receptive field and capture global features. Additionally, the SIOU loss function is adopted to optimize the training convergence process, and a decoupled head structure is introduced to separate classification and regression tasks. The experimental results demonstrate that the improved model achieves a 0.7% increase in mAP for single-target detection tasks and a 1.2% improvement in mAP for multi-target detection tasks. These enhancements not only significantly boost the detection accuracy but also effectively mitigate occurrences of missed and false detections.

Key words: object detection; YOLOv5s; attention mechanism; decoupled head

摘要： 交通场景目标检测是智慧交通与自动驾驶系统的关键技术。针对 YOLOv5s 算法在复杂 交通场景下存在的多尺度目标漏检、误检及检测精度不足等问题，文章提出一种改进的 CRDYOLOv5s 模型。首先，在主干网络引入卷积注意力模块（CBAM），通过通道与空间注意力 增强小目标特征提取能力；其次，将空间金字塔池化翻转（SPPF）模块替换为感受野（RFB） 模块，利用空洞卷积扩大感受野以捕获全局特征；同时采用 SIOU 损失函数优化训练收敛过 程，并引入解耦头结构分离分类与回归任务。实验结果表明，改进后的模型在单目标检测任 务中，mAP 提升了 0.7%，在多目标检测任务中，mAP 提升了 1.2%，不仅显著提升了目标检 测的精度，还有效减少了漏检与误检现象。

关键词: 目标检测；YOLOv5s；注意力机制；解耦头