近年来，随着矿用卡车自动驾驶技术的兴起，使得矿区道路行车障碍物检测变得至关重要，基于深度学习的目标检测模型应用于矿区道路障碍检测取得了显著的效果，为矿用卡车自动驾驶技术的完善提供了可能。为解决现有模型应用于矿区障碍物检测，往往存在算法庞大与部署成本较高的问题，提出一种面向边缘计算平台的改进YOLOv8矿区道路障碍检测模型，该模型针对资源有限的边缘计算设备进行优化部署，以实现对障碍物的快速、精准检测。该模型在特征提取阶段，引入深度可分离卷积和通道注意力机制，提高模型对障碍物整体特征提取能力，从而提升对不同尺寸障碍物的检测精度；特征融合阶段采用BiFPN网络结构，轻量化颈部网络并自适应地调整融合权重，减少冗余信息，提高特征的表达能力；使用局部卷积PConv对检测头进行重新设计，减少网络参数量以提高检测效率；最后，通过引入Inner-CIoU函数对边界框损失进行优化，加快模型收敛速度并提升边界框定位效果。实验结果显示，该网络在所使用的矿区障碍物数据集上，mAP@0.5仅下降0.05的前提下，模型参数减少了44%，推理时间缩短了34%。相比其他轻量型检测网络，该模型在实验硬件设备上的检测速度更快，且在精度和轻量化之间实现了更好的平衡，为障碍物检测模型的实际部署提供了可行方案。

In recent years, with the rise of autonomous driving technology for mining trucks, detecting obstacles on mining roads has become crucial. Object detection models based on deep learning have been applied to significant effect in detecting obstacles on mining roads, thereby providing possibilities for the improvement of autonomous driving technology for mining trucks. To address the issues of large algorithms and high deployment costs associated with existing models for mining obstacle detection, an improved YOLOv8 model tailored for edge computing platforms is proposed. This model is optimized for deployment on resource-constrained edge computing devices to achieve rapid and accurate obstacle detection. In this model, during the feature extraction stage, depthwise separable convolutions and channel attention mechanisms are introduced to enhance the model’s ability to extract overall features of obstacles, thereby improving the detection accuracy of obstacles of various sizes. In the feature fusion stage, a BiFPN network structure is employed to lightweight the backbone network and adaptively adjust fusion weights, reducing redundant information and enhancing feature representation. The detection head is redesigned using local convolution PConv to reduce network parameter size and improve detection efficiency. Finally, by introducing the Inner-CIoU function for bounding box loss optimization, the model convergence speed is accelerated, and bounding box localization effectiveness is enhanced. Experimental results demonstrate that on the mining obstacle dataset used, while maintaining a decrease of only 0.05 in mAP@0.5, the model parameters are reduced by 44%, and the inference time is reduced by 34%. Compared to other mainstream detection networks, this model exhibits faster detection speed on the hardware devices used in the experiments and better balances the requirements of accuracy and lightweight, providing a feasible solution for the practical deployment of obstacle detection models.