铁路外部环境的常见隐患以及未知风险（包括泥石流、落石、动物入侵等）严重威胁着铁路安全运行，需要巡检人员耗时费力地频繁检查，但巡检范围仍十分有限。目前，低空经济已成为国家新质生产力代表，无人机凭借其高空作业能力、远距离覆盖优势，以及不受地形限制和铁路维修天窗影响的特点，在巡检领域具有得天独厚的技术优势。针对铁路外部环境未知风险样本稀疏且具有随机不确定性的挑战，利用无人机进行沿线遥感图像采集，并基于Faster R-CNN提出了一种未知风险检测框架。首先，设计了一种目标性与多分类解耦训练策略，并融合在未知风险检测框架中，显著提升了通用目标检测性能，避免将未知风险目标错分为背景。其次，改进了VOS的虚拟特征合成方法，提出了基于相似度的特征空间采样，在构建实例级目标特征空间基础上进行多元高斯分布参数估计与重采样，获得泛化性的未知风险目标特征表示。再次，利用基于能量的不确定性度量，对实例级特征进行不确定性度量，并据此计算损失以诱导网络优化常见类别和未知风险类别的决策边界。最后，在采集的铁路外部环境数据集、开源无人机数据集以及泛化性测试数据上进行了定量与定性实验分析，本文方法在常见隐患识别上取得了95.7%的mAP₅₀，同时在测试集和泛化性测试数据上分别取得了98%和80.8%的Recall₅₀，实验结果表明本文方法在保证常见隐患类别高识别率的基础上对于未知风险目标也有较高的检测能力。

Common hazards as well as unknown risks （including mudslides， rockfalls， animal intrusion， etc.） in the external environment of railroad seriously threaten the safe operation of railroads， requiring frequent time-consuming and laborious inspections by inspectors， but the scope of inspections is still very limited. At present， the low-altitude economy has become China’s new quality productivity representative， and the UAV has innate inspection advantages of high altitude， long distance， and small impact from the terrain and railroad maintenance windows. To overcome the challenge of sparse samples and random uncertainty of unknown risks in the external environment of the railroad， this paper utilizes UAVs for remote sensing image acquisition along the railroad， and proposes an unknown risk detection framework based on Faster R-CNN. Firstly， a novel targeted and multi-classification decoupling training strategy is designed and integrated in the unknown risk detection framework， which significantly improves the performance of general object detection and avoids misclassifying unknown risk objects as background. Secondly， the virtual feature synthesis method of VOS（Visual Object Segmentation） is improved， and similarity-based feature space sampling is designed to obtain a generalized unknown risk object feature representation by performing multivariate Gaussian distribution parameter estimation and resampling based on the construction of instance-level object feature space. Subsequently， an energy-based uncertainty measurement is utilized to measure the uncertainty of instance-level features， and losses are calculated accordingly to induce the network to optimize the decision boundaries for common and unknown risk categories. Finally， quantitative and qualitative experimental analyses are conducted on the collected railroad external environment dataset， open-source drone dataset， and generalization test data. The proposed method achieves 95.7% mAP₅₀ in common hazard identification， while achieving 98% and 80.8% Recall₅₀ in the test set and generalization test data， respectively. The experimental results show that the proposed method has high detection ability for unknown risk objects， while ensuring high recognition rate of common hazard categories.