• Proposed a novel Multi-Modal Frustum PointNet (MMFP) method that fuses laser point cloud and images for accurate coal-rock recognition in mining faces.
• Improved Mask R-CNN with MobileNetV3 backbone, Dilated CBAM, and inception structure to enhance detection accuracy while reducing model parameters.
• Utilized frustum point cloud extraction and self-attention PointNet for efficient and precise segmentation and bounding box prediction.
• Experimental validation on a shearer coal wall cutting platform demonstrated superior performance of MMFP compared to existing advanced models.