Abstract: Obstacle perception based on target detection is an important technical support for mine autonomous driving. Affected by class imbalance and long-distance detection problems, the accuracy of 3D object detection in mines is limited. To address this problem, a 3D object detection network based on an improved Voxel-Graph Attention Network (VGAT) was proposed. The graph attention network was used to distinguish node importance and alleviate the dilution of effective features. The nearest-neighbor voxel search method was adopted to cope with the structural sparsity of mine point clouds. A Voxel Feature Encoding Plus (VFE+) module was designed to solve the problem that different point distributions within voxels might be encoded into identical features. A Voxel Feature Encoding Compensation (VFE_C) module was designed to compensate for the loss of spatial information caused by voxel partitioning. The loss function was optimized to alleviate the imbalance between positive and negative samples and the periodic ambiguity in angle prediction. On this basis, a 3D object detection network for autonomous driving in mines integrating Ultra-Wideband (UWB) positioning and long-range Point Cloud Upsampling (PU), namely UWB-PU-VGAT, was proposed. The UWB positioning system was used to obtain prior location information of mine trucks and miners, and LiDAR point clouds were accurately cropped to focus on target regions of interest, thereby alleviating the class imbalance problem. For sparse long-range point clouds, the Gradient-descent-based Point Cloud Upsampling (Grad-PU) network was adopted to generate high-density point clouds and alleviate the long-range detection problem. Experimental results showed that the average precisions of UWB-PU-VGAT for mine trucks and miners reached 90.23% and 83.67%, respectively, and the mean average precision reached 86.95%, outperforming mainstream 3D point cloud object detection networks such as SECOND and PointRCNN. The frame rate reached 32.3 fps, meeting the real-time obstacle avoidance requirements of autonomous driving.