RGB-D Semantic-Guided 3D Geometric Estimation for Intelligent Coal Gangue Separation

Coal and gangue separation is crucial for clean and efficient coal utilization. However, practical sorting environments are highly unstructured, where irregular fragments exhibit random distribution, varying orientations, and occlusion. Conventional 2D vision methods cannot provide the 3D geometric information required for intelligent sorting, while model-based 6D pose estimation is unsuitable for coal gangue fragments due to their irregular morphology and lack of predefined models. This study proposes an RGB-D cascaded fusion framework for partial point cloud-based 3D geometric estimation. The framework follows a detection–segmentation–reconstruction pipeline, where object detection provides semantic priors, zero-shot segmentation generates pixel-level masks, and the masks guide depth extraction and local point cloud reconstruction. Based on the reconstructed point clouds, task-oriented geometric parameters, including centroid position, oriented bounding box dimensions, visible volume, and principal orientation, are estimated for intelligent separation. Experiments on a dedicated RGB-D coal and gangue dataset demonstrate that the proposed method achieves a centroid localization error of 14.22 mm, dimension error of 9.34%, and volume error of 11.3%. Compared with detection-box-based depth reconstruction, the proposed semantic-guided strategy significantly improves geometric estimation accuracy. The proposed framework provides reliable 3D perception for intelligent coal gangue separation and robotic sorting.