Efficient deep learning based rail fastener screw detection method for fastener screw maintenance robot under complex lighting conditions

For the rail fastener replacement operation at night in the wilderness, the lighting conditions on the rail fastener screws are complex, due to the multiple illuminants like headlamps and flashlights at the site, making some parts of the objects appear dark or low light status in the camera. These complex lighting conditions (CLCs) interfere with the fastener recognition ability of the fastener screw detection algorithm since it can hardly maintain fixed and optimized lighting conditions of the fastener screw. We propose the LFGB-YOLO, a novel YOLO-based model that contains two principal parts: the Light-Fast part and the GB-Neck part. The Light-Fast part can reduce the network Params, FLOPs, and memory access frequency in feature extraction while keeping a high precision. The GB-Neck part can lighten the feature fusion network while maintaining the ability of accurate feature information extraction operation. Experimental results demonstrate that the LFGB-YOLO performs excellently in metrics like Recall, mAP@0.5, F1 score, and FPS, better than the performance of competitive models like YOLOv5n, YOLOv7-miny, and YOLOv8. The performance metrics of the proposed model, Recall, mAP@0.5, F1-score, and FPS are increased by 8.9%, 4%, 4.8%, and 8.1% compared with the baseline model. It shows that our work not only performs satisfactorily in detecting fastener screws under CLCs but also inspires new studies that focus on the fastener screw detection affected by environmental factors.