Research on foreign object intrusion detection in railway tracks based on MSL-YOLO

Abstract Railway foreign object intrusion detection poses significant challenges due to complex backgrounds, variable lighting conditions, and the need for real-time, multi-scale object detection. To address these issues, this paper proposes MSL-YOLO, a lightweight and accurate object detector optimized for railway applications. Specifically, a Multi-scale Shared Convolution Module (MSCM) is designed to replace SPPF, enhancing feature extraction while reducing parameters and computational cost. StarBlocks from StarNet are introduced to construct a novel C2f-Star structure, which is further combined with Efficient Multi-scale Attention (EMA) to form C2f-Star-EMA. This integration improves multi-scale feature representation and model efficiency. In addition, a Lightweight Shared Convolutional Detection Head (LSCD) is employed to replace the original head, reducing complexity while maintaining detection accuracy. Experiments on a custom railway intrusion dataset demonstrate that MSL-YOLO achieves a mAP of 94.3% with only 2.35 M parameters and 6.7 GFLOPs, reaching 277 FPS and 3.3 ms latency. Compared with several mainstream lightweight models and SOTA detectors, MSL-YOLO offers the best trade-off between accuracy, speed, and computational cost. Combining high precision with low complexity, the proposed method meets the dual requirements of real-time performance and robustness in practical railway detection scenarios.