基于改进YOLOv5s的交通信号灯检测算法

Improved YOLOv5s-Based Traffic Signal Detection Algorithm

摘要: 针对交通信号灯检测在背景环境复杂、小目标检测困难和实时性要求高的问题，基于YOLOv5s提出一种YL-YOLOv5s方法用于交通信号灯检测。首先，将Neck特征融合网络中残差结构替换为多层次跨通道DenseNet连接方式，以增强特征提取能力与减少梯度消失问题；然后，将DenseNet中的卷积操作替换为倒残差结构MobileNetV2，以加快推理速度并降低内存占用；最后在模块中嵌入通道注意力机制ECA (Efficient Channel Attention)，有效建立了通道之间的相关性，从而提高了交通信号灯检测的准确率。实验结果表明，在法国巴黎信号灯LaRa和博世小型交通信号灯BSTLD上进行测试，与YOLOv5s算法相比，mAP@0.5:0.95、mAP@0.5等关键指标都有提升，在LaRa数据上mAP@0.5:0.95提升8.5%，并且模型计算量减少了23.2%，说明YL-YOLOv5s模型能够有效提升复杂环境下小目标的检测精度，且轻量化效果明显。

Abstract: In order to address the challenges of complex background environments, difficult detection of small targets, and high real-time requirements in traffic signal detection, this paper proposes a YL-YOLOv5s network based on the YOLOv5s architecture. Firstly, the residual structures in the Neck feature fusion network are replaced with a multi-level cross-channel DenseNet connection to en-hance feature extraction capabilities and reduce the problem of gradient vanishing. Next, the con-volution operations in DenseNet are replaced with inverted residual structures from MobileNetV2 to accelerate inference speed and reduce memory usage. Finally, an Efficient Channel Attention (ECA) mechanism is embedded in the modules to effectively establish correlations between chan-nels, thereby improving the accuracy of traffic signal detection. Experimental results show that when tested on the LaRa signal lights in Paris, France, and the Bosch Small Traffic Lights Dataset (BSTLD), the proposed YL-YOLOv5s model achieves improvements in key indicators such as mAP@0.5:0.95 and mAP@0.5 compared to the YOLOv5s algorithm. Specifically, it achieves an 8.5% increase in mAP@0.5:0.95 on the LaRa dataset, while reducing the computational cost by 23.2%. This demonstrates that the YL-YOLOv5s model effectively improves the detection accuracy of small targets in complex environments and exhibits significant lightweight effects.