Abstract: Autonomous driving of rail locomotives in coal mines is regarded as a critical link in achieving intelligent auxiliary transportation. However, the operation environment of underground autonomous locomotives is highly complex. Affected by low illumination and severe occlusion, issues such as missed detection of long-distance small targets and feature confusion caused by densely populated personnel are easily triggered. Furthermore, the synergistic optimization of detection accuracy and real-time performance is difficult to achieve under limited embedded computing power. To address the aforementioned issues, a lightweight algorithm named GSE-YOLO, which balances detection accuracy and deployment efficiency, is proposed. First of all, a global context convolution (GCConv) is designed as the key feature layer of the backbone network. By fusing local features and global context information—where local features focus on target detail information, while global context information captures the global semantic and background information of the target—the identification capability of target features in complex environments is enhanced. Secondly, a Space-to-Depth convolution (SPDConv) is introduced to optimize the downsampling process. By performing depth transformation in the spatial dimension, the loss of detailed information is minimized, thereby improving the feature representation capability for long-distance small targets. Finally, a C2f-EACF module integrating dual channel and spatial attention is constructed to strengthen the attention towards key target regions and suppress background interference. Consequently, the feature discrimination capability and bounding box localization accuracy in occluded scenes are prominently improved. Experimental results demonstrate that the mAP@0.5 and mAP@0.5:0.95 of the GSE-YOLO model reach 85.2% and 48.5%, respectively. Under the condition of a parameter scale of only 2.98×10⁶, the metrics are improved by 3.0% and 2.4%, respectively, compared with the baseline model YOLOv8n. Furthermore, when tested on the Jetson Xavier NX platform combined with a TensorRT acceleration strategy, the total inference time per frame under the FP16 quantization mode is contained at merely 31.02 ms. While maintaining the lightweight nature of the model, the detection robustness and real-time performance under complex real-world conditions are effectively enhanced, meeting the engineering requirements for edge deployment in underground auxiliary transportation systems, and providing technical support for the practical application of autonomous driving technology for underground locomotives.