Identification of unmanned obstacles of railway electric locomotive in coal mine under complicated working conditions

Self-sensing and accurate obstacle positioning is the core technology to realize unmanned driving of railway electric locomotive, the harsh conditions such as high dust, low light and strong occlusion in the underground have a great influence on the obstacle recognition results. Aiming at the problems of missed detection of distant targets, false detection of occluded targets and complicated calculation parameters in this scene, an accurate identification method of track obstacles in the safe operation area of electric locomotives MUV-YOLOv11 is proposed. First of all, the L-Backbone module introduces the detail-enhanced attention block (DEAB) and designs the channel separation and fusion convolution block (CSFCB) to efficiently extract the salient features of obstacles under complex backgrounds. Meanwhile, a depth and width of spatial pyramid pooling-fast block (SPPF-DW) is constructed to enhance effective features and solve the problem of false detection of occluded targets. Secondly, the residual of semantics and detail infiltration module (SDI-R) is designed into the C-Neck module, which breaks through the disadvantage that it is difficult to mine the detailed features caused by the traditional cascade and downsampling fusion methods. Then, the attention of contrast-driven feature aggregation module (A-CDFA) is constructed to improve the detection effect of small target area and enhance the feature expression ability with the help of internal and external circulation structure, thus solving the problem of missing detection of distant targets in complex environment. Finally, a three-branch detector structure is used to extract the feature information of different positions to realize the accurate positioning of obstacles. The obstacles that affect the safe operation of electric locomotive are tested on the self-made underground roadway data set, and the proposed MUV-YOLOv11 achieves mAP50 of 91.4% and mAP50-95 of 67.9%. Compared with the latest low-light detection algorithm WTEFNet, the recognition accuracy of MUV-YOLOv11 improves of 4.3% when the number of parameters reduces of 36.79%, which effectively solves the obstacles that may affect the future of unmanned rail electric locomotive and the difficulty of shielding obstacles.