Abstract: To address the issues of false and missed detections in obstacle detection within underground mines caused by difficulties in filtering elevated structural point clouds and insufficient ground segmentation accuracy, a multi-LiDAR obstacle detection method based on grid method for underground mines was proposed. Firstly, conditional filtering was employed to crop and denoise the point clouds. The point cloud data from multiple LiDARs, after conditional filtering, were fused and subsequently downsampled through voxelization to accomplish point cloud preprocessing. Secondly, the preprocessed point clouds were projected onto a two-dimensional grid, which was divided into near and far regions based on the distance to the vehicle. The tunnel ground and ceiling heights for each grid were then calculated separately, and the features of nearby grids and row-wise grids were updated. Based on the point cloud distribution and inter-row grid continuity, a strategy of updating the global grid characteristics was implemented from near to far and row by row, enabling precise estimation of tunnel ground and ceiling characteristics. Finally, ground segmentation was performed based on the grid features, elevated structural point clouds were filtered out from the non-ground point clouds, and Euclidean clustering was applied to the remaining point clouds to detect obstacles. Field test results demonstrated that the proposed method could effectively filter out elevated structural point clouds in uphill/downhill and sparse point cloud conditions. Detection accuracies for low-profile targets, including signal boxes, traffic cones, low supports, and vehicles reached 92.3%, 90.9%, 96.5%, and 100%, respectively. Across various tunnel conditions, the approach significantly reduces false and missed detections, achieving high obstacle detection accuracy.