RCB-YOLOv5s：基于改进YOLOv5s算法的高速护栏安装立柱检测

RCB-YOLOv5s: Detection of Highway Guardrail Installation Poles Based on Improved YOLOv5s Algorithm

摘要: 随着高速公路里程的快速增长，高速公路立柱的自动化检测已成为智能交通基础设施安全巡检与维护的一项重要任务。然而，高速公路环境极为复杂，存在强烈的背景干扰以及目标尺度变化显著等问题，严重限制了传统检测算法的准确性。为应对这些关键的工程挑战，本文提出了一种基于YOLOv5的改进目标检测算法，这对于实现可靠、实时的高速公路基础设施监测具有重要意义。首先，在模型架构中集成了RepVGG模块，利用结构重参数化技术，在不降低推理速度的前提下，增强了对边缘和纹理特征的提取能力。其次，嵌入了上下文Transformer (CoTAttention)机制以抑制背景噪声，并采用双向特征金字塔网络(BiFPN)重构网络颈部(Neck)，从而加强了针对远处小尺度目标的多尺度特征融合。在自建的高速公路立柱数据集上进行的大量实验验证了所提方法的优越性。与原始YOLOv5模型相比，改进算法的平均精度均值(mAP@0.5)提升了4.7%，精确率(Precision)和召回率(Recall)分别达到了88.4%和85.1%。综上所述，本研究成功实现了高检测精度与实时效率的有效平衡，为高速公路基础设施的智能巡检与安全预警系统提供了坚实的技术支撑与实际的工程应用价值。

Abstract: As the mileage of highways expands rapidly, the automated detection of highway poles has become a crucial task for the safety inspection and maintenance of intelligent transportation infrastructure. However, complex highway environments characterized by strong background interference and significant target scale variations severely limit the accuracy of conventional detection algorithms. To address these critical engineering challenges, this paper proposes an improved target detection algorithm based on YOLOv5, which is of great significance for achieving reliable and real-time highway infrastructure monitoring. First, the RepVGG block is integrated into the model architecture to enhance the extraction of edge and texture features without compromising inference speed through structural re-parameterization. Subsequently, the Contextual Transformer (CoTAttention) mechanism is embedded to suppress background noise, and a Bidirectional Feature Pyramid Network (BiFPN) is employed to reconstruct the network’s neck, thereby strengthening multi-scale feature fusion for distant small-scale targets. Extensive experiments conducted on a self-constructed highway pole dataset demonstrate the superiority of the proposed method. Compared with the original YOLOv5 model, the improved algorithm achieves a 4.7% increase in mean average precision (mAP@0.5), with precision and recall metrics reaching 88.4% and 85.1%, respectively. Consequently, this research successfully balances high detection accuracy with real-time efficiency, providing robust technical support and practical engineering value for intelligent inspection and safety early warning systems in highway infrastructure.