Abstract: In view of the complex environment, low light, dust and other interference factors in the auxiliary transportation roadway in the coal mine, and the reflected light interference in the detection of workers, the mining unmanned vehicles using vehicle-mounted cameras have low detection accuracy and poor real-time detection of workers in the roadway. In addition, existing target detectors based on deep learning have the problems of large number of parameters and high computational complexity of the model. A detection model of mine unmanned vehicle assisted transportation roadway workers based on YOLOv8, namely GCDB-YOLOv8, was proposed. Firstly, lightweight modules Ghost Convolution (GhostConv and GhostC2f) are introduced into the feature extraction network to achieve lightweight network design and reduce the number of model parameters and computational complexity. Secondly, the C2F-EMA module is designed and used to replace the C2f model of the neck, so as to enhance the attention of key areas in low light and complex background, so that the model can make efficient use of the feature information of the staff. At the same time, the DicPSA module is designed and used to replace the spatial pyramid pooling module (SPPF) in the backbone network to enhance the ability of the model to capture, extract and utilize key feature information. Finally, the weighted bidirectional feature pyramid (BiFPN) mechanism is designed and improved, and the original FPN+PAN structure is replaced by BIFPN to reduce the problem of feature information loss, achieve the full fusion and utilization of deep feature map target semantic information and shallow feature map target location information, and improve the detection accuracy. On the Underground Transportation Roadway Workers Detection Dataset, the experimental results show that compared with YOLOv8n, the detection accuracy of GCDB-YOLOv8 model reaches 80.64%, which is improved by 6.06%. The detection speed reaches 112 f/s, which is faster than the baseline model and meets the requirement of real-time detection. The number of model parameters is 2.70 M, and the computational complexity is 7.50 GFLOPs, which is 0.31 M and 0.70 GFLOPs less than the baseline model, respectively. Compared with Faster R-CNN, SSD, YOLOv3-tiny, YOLOv5s, YOLOv7-tiny, YOLOv8s, YOLOv9s, IAT-YOLO, RT-DETR, MLFE-YOLOX, CDD-YOLO, YOLO_GD detection models, GCDB-YOLOv8 is superior to other comparison models in terms of detection accuracy, detection speed, number of parameters, and computational complexity. On the Miner Action Detection Dataset, the mAP@0.5 and the mAP@0.5～0.95 of GCDB-YOLOv8 reach 87.69% and 64.77%, respectively, which are 3.43% and 2.26% higher than that of the baseline model YOLOv8n. GCDB-YOLOv8 model improves the detection accuracy of underground transportation roadway workers while taking into account the lightweight and real-time performance of the model, which is easy to deploy on mining unmanned vehicles. The model can meet the detection requirements of mining unmanned vehicles for workers in the roadway and reduce safety hazards. In addition, the accurate real-time detection of the workers in the underground transportation roadway by GCDB-YOLOv8 can provide security for the subsequent autonomous obstacle avoidance, path planning, decision control and other tasks of the mining unmanned vehicle, and promote the application of unmanned driving technology in the field of intelligent coal mines.