BPS‐YOLO: A Localisation‐Robust Detector for Foreign Objects in Railway Overhead Catenary Systems

ABSTRACT

Under conditions of complex backgrounds, low illumination and long‐distance imaging, foreign object detection in railway catenary suffers from localisation inaccuracies due to insufficient high‐quality positioning information. To address this, we introduce BPS‐YOLO based on YOLOv11. Our model simultaneously strengthens geometric alignment capability in foreign object detection and improves localisation accuracy while maintaining a compact parameter count. Specifically, we adapt the C3K2_PLA module, which integrates a progressive local attention mechanism to strengthen local structural cues and boundary detail preservation, thereby enhancing the discriminative representation of small‐scale foreign objects. Secondly, we introduce the attention‐guided bidirectional feature pyramid network (AG‐BiFPN), which effectively suppresses background interference and enhances the spatial representation of the target regions by introducing the high‐resolution P2 feature layer and applying the global‐local spatial attention (GLSA) mechanism before feature fusion, thereby enhancing cross‐scale feature interaction and precise localisation. In addition, Shape‐IoU is used as an auxiliary training‐stage localisation constraint; this loss itself does not increase inference parameters or GFLOPs. Extensive experiments demonstrate that BPS‐YOLO achieves the highest mAP@50–95 among the compared detectors. Qualitative feature visualisation reveals concentrated target activations under complex backgrounds and reduced spurious responses, indicating its potential for real‐time GPU‐based inspection scenarios.