Abstract: To address the issues of low detection accuracy of target and poor real-time performance caused by severe environmental interference in rail track foreign object detection under conditions such as rain, fog and low light, an improved YOLOv12 railway track intrusion foreign object detection algorithm (SF-YOLO) was proposed. Firstly, to address low detection accuracy under environmental interference, an enhanced multi-dimensional collaborative attention mechanism (MCAM) was employed, which enhanced feature expression through three-dimensional collaborative modelling. Secondly, to improve detection accuracy while maintaining real-time performance, a new lightweight feature fusion module, SFNet-Rail, was established, which preserved the details of foreign object boundaries while strengthening multi-scale feature consistency. Thirdly, a FastPartial-Detect detection head network based on partial convolutions (PConv) was designed to reduce memory access redundancy and further enhance the detection performance of the model. Finally, a unified IoU loss function (UIoU) was introduced to dynamically optimize the weight distribution of prediction boxes. Experimental results demonstrate that on the self-built rail track foreign object dataset, the proposed algorithm achieves a mean accuracy (PmA@0.5) of 85.8%, with a 2.7% improvement over YOLOv12. The detection speed of the proposed algorithm is increased by 12%, and the computational load is decreased by 21.25%. The accuracy and real-time performance of multi-scale railway foreign object detection in environments such as rain, fog, and low light conditions can be effectively enhanced.

Key words: railway engineering; railway track foreign object detection; SF-YOLO; lightweight detection; feature fusion

摘要： 针对雨雾、弱光等条件下铁轨异物检测中因环境干扰严重导致的目标检测精度较低、实时性较差等问题，提出一种改进YOLOv12的铁路轨道入侵异物检测算法(SF-YOLO)。首先，针对环境干扰下检测精度低的问题，采用改进的多维协作注意力机制(MCAM)，通过三维空间协同建模增强特征表达效果；其次，为提升检测精度的同时保持检测的实时性，新建轻量化特征融合模块SFNet-Rail，在保留异物边界细节的同时强化多尺度特征一致性；然后，设计基于部分卷积(PConv)的FastPartial-Detect检测头网络，减少内存访问冗余，进一步提升模型的检测性能；最后，引入统一IoU损失函数(UIoU)，动态优化预测框权重分配。实验结果表明：在自建铁轨异物数据集上，该算法平均精度(PmA@0.5)达85.8%，较YOLOv12 提升了2.7%，检测速度提升了12%，计算量减少了21.25%，可有效提升雨雾、弱光照等环境下多尺度铁轨异物检测精度与实时性。

关键词: 铁道工程；铁轨异物检测；SF-YOLO；轻量化检测；特征融合