【摘要】 由于隧道施工现场事故频发，通过机器人代替人工进行安全巡检可有效保障工作人员的生命安全。然而，目前该领域的机器人大多需要人工操控，缺乏自主避障能力。针对隧道施工环境下的障碍物检测方法，提出利用改进YOLOv5进行障碍物检测的轻量化模型。首先构建隧道场景下的障碍物数据集；其次修改骨干网络为轻量级Shufflenet v2网络，并将其中的激活函数修改为SiLU函数，以提高检测速度降低计算量；然后引入坐标注意力机制，增强网络学习特征的表达能力；最后将颈部卷积块修改为GSConv,在减小模型计算量的同时提升算法的检测精度。研究结果表明：基于本研究所构建的数据集上进行对比试验，该方法的检测速度较YOLOv5-n原始算法提升了57%,减少了模型对硬件的需求。更多还原

【Abstract】 Due to the frequent accidents at the tunnel construction site, using robots instead of manual safety inspections can effectively protect the safety of staff. However, most of the robots in this field currently require manual control and lack autonomous obstacle avoidance capabilities. Aiming at the obstacle detection method in tunnel construction environment, a lightweight model using improved YOLOv5 for obstacle detection was proposed. Firstly, the obstacle dataset in the tunnel scenario was constructed. Secondly, the backbone network was modified to a lightweight Shufflenet v2 network, and the activation function was modified to the SiLU function to improve the detection speed and reduce the amount of computation. Next, the coordinate attention mechanism was incorporated to improve the network′s ability to learn and represent features. Finally, the neck convolution block was modified to GSConv, which reduced the calculation amount of the model and improved the detection accuracy of the algorithm. Comparative experiments on the dataset constructed in this research showed that the detection speed of the proposed method was increased by approximately 57% compared to the original YOLOv5-n algorithm, which reduced the hardware requirements of the model.更多还原