Research on lightweight rail foreign object intrusion detection based on shallow feature fusion

HOU Tao; LI Junchang; NIU Hongxia

doi:10.11860/j.issn.1673-0291.20250056

Abstract：To address the issues of low detection accuracy,slow detection speed,and frequent missed or false detections in railway track foreign object intrusion detection,this study proposes a lightweight railway track foreign object intrusion detection algorithm based on shallow feature fusion(YOLO-LSF).First,building on the YOLOv8n feature extraction network,the C2f module is improved based on GhostConv to construct the C2f_Ghost module,thereby reducing both the parameter count and computational cost of the model.Second,the MLCA attention mechanism is introduced at the end of the backbone network to en-hance the feature representation of the target area and optimize the feature extraction efficiency of the model.Third,deformable convolution DCNv2 is employed to replace some ordinary convolutions in the C2f module of YOLOv8n,constructing the C2f_DCNv2 module and further strengthening the model's feature extraction capacity.Finally,shallow feature information from the backbone network is integrated into the neck network,effectivelt mitigating detail loss caused by multiple convolution operations and en-hancing the model′s ability to detect distant foreign objects(small targets).Experimental results show that on a self-constructed railway track foreign object intrusion detection dataset,compared with the original YOLOv8n algorithm,the YOLO-LSF algorithm achieves an improvement of 5.2%in average precision,3.37%in FPS,a reduction of 20.1%in the number of parameters,and a decrease of 22.2%in computa-tional complexity.These results verify that the proposed algorithm significantly enhances detection accu-racy and speed in complex environments while reducing the likelihood of missed and false detections.