Configuration of the experimental environment.

Ensuring high accuracy and efficiency in foreign object intrusion detection along railway lines is critical for guaranteeing railway operational safety under limited resource conditions. However, current visual detection methods generally exhibit limitations in effectively handling diverse object shapes, scales, and varying environmental conditions, while typically incurring substantial computational overhead. To overcome these limitations, this study proposes a multi-level feature aggregation and context enhancement network (MACE-Net). The network architecture integrates the GOLD-YOLO module, an advanced object detection approach, alongside the updated deformable convolutional networks (DCNv3). The incorporation of DCNv3 allows the model to dynamically adapt its sampling positions according to actual object shapes, significantly enhancing feature extraction accuracy, especially for irregularly shaped intrusions. Additionally, the convolutional block attention module (CBAM) is employed to refine spatial and channel-wise feature representation, enabling the model to emphasize crucial object characteristics without substantially increasing computational complexity. Meanwhile, to improve localization robustness, the generalized intersection over union (GIoU) loss function is implemented, offering more reliable detection across various object sizes and shapes. Furthermore, to address the shortage of domain-specific datasets, we created a railway intrusion dataset comprising 7,200 images. Experimental results demonstrate that MACE-Net achieves superior detection performance, improving mAP@0.5 from 78.9% (baseline YOLOv8) to 83.8%—a notable increase of 4.9%. Meanwhile, the F1-score also rises by 5.2%. Importantly, despite significant accuracy gains, MACE-Net maintains computational efficiency similar to that of the baseline, affirming its suitability for real-time railway foreign object detection tasks under constrained energy and computational environments.

在资源受限的场景下，保障铁路沿线异物入侵检测的高精度与高效率，对于确保铁路运行安全至关重要。然而，当前视觉检测方法普遍存在局限：难以有效适配多样化的物体形态、尺度与多变的环境条件，且通常伴随高昂的计算开销。为克服上述局限，本研究提出一种多级特征聚合与上下文增强网络（Multi-level Feature Aggregation and Context Enhancement Network，MACE-Net）。该网络架构集成了先进目标检测方案GOLD-YOLO模块，以及升级后的可变形卷积网络（Deformable Convolutional Networks，DCNv3）。引入DCNv3可使模型根据实际物体形态动态调整采样位置，显著提升特征提取精度，尤其针对形态不规则的入侵目标。此外，本研究采用卷积块注意力模块（Convolutional Block Attention Module，CBAM）对空间与通道维度的特征表示进行精细化优化，使模型能够在不显著提升计算复杂度的前提下，突出目标物体的关键特征。同时，为提升定位鲁棒性，本研究引入广义交并比（Generalized Intersection over Union，GIoU）损失函数，可在不同物体尺度与形态下实现更可靠的检测效果。进一步而言，为解决领域专属数据集匮乏的问题，本研究构建了包含7200张图像的铁路入侵数据集。实验结果表明，MACE-Net实现了更优异的检测性能：其平均精度均值@0.5从基准模型YOLOv8的78.9%提升至83.8%，增幅达4.9个百分点。同时，F1分数也提升了5.2个百分点。尤为关键的是，尽管MACE-Net的检测精度实现了显著提升，但其计算效率仍与基准模型相当，这证实了其在能源与计算资源受限的环境下，适配实时铁路异物入侵检测任务的可行性。