Abstract: In underground coal mine unmanned vehicle operations, traditional object detection models often suffer from issues such as missed detection, false detection, and insufficient real-time performance due to factors like low light, high dust, and varying object shapes. To address these challenges, a highly efficient and accurate object detection model for coal mine underground unmanned vehicles, named YOLOv8-CBD, is proposed in this paper. First, the Dual approach is applied to fuse CSPNet and HetConv, and a new CSPConv module is introduced to replace the C2f module in the original YOLOv8 network, reducing computational complexity while maintaining detection accuracy. Second, a weighted bidirectional feature pyramid network (BiFPN) is introduced to improve the neck network of YOLOv8. Different weights are assigned based on the importance of various features, effectively fusing target features across multiple scales and ensuring efficient utilization of target feature information. Additionally, the traditional detection head is replaced with a Dynamic Head, which significantly enhances the model's performance in complex backgrounds and strengthens its ability to detect small targets. Finally, CIoU is replaced with Focaler-IoU, reconstructing the linear interval mapping of the IoU loss and enabling dynamic adjustment of the model's loss value across regression samples of varying difficulty. This enhances the model's localization and detection capability in complex underground environments. The proposed YOLOv8-CBD model was tested on the coal mine underground unmanned vehicle object detection dataset. The experimental results show that YOLOv8-CBD achieves a mAP@50 of 91.5%, mAP@50:95 of 65.9%, and a Recall of 80.9%, F1-score of 0.861, FPS of 106, which represents improvements of 3.3%, 4.7%, and 2.9%, 3.1% and 16%, respectively, compared to the baseline model YOLOv8. Its parameter count is only 2.63M, and its computational complexity is 6.4 GFLOPs, which are 12% and 22% lower than the baseline model YOLOv8n, respectively. Compared with Faster-RCNN, SSD, YOLOv5n, YOLOv6n, YOLOv7-tiny, and YOLOv8s detection models, as well as RT-DETR and Deformable-DETR, YOLOv8-CBD outperforms all in terms of accuracy, recall, parameter size, and computational complexity. YOLOv8-CBD not only improves detection accuracy but also reduces model size and computational load, making it more suitable for deployment on various unmanned vehicles operating in underground coal mine environments. This model effectively addresses the challenges in object detection caused by insufficient lighting, smoke, and dust in such environments.