Dynamic SLAM Method for Underground Coal Mines Integrating Zero-Shot Semantic Priors and Dual Geometric Constraints

To address the problems of scarce semantic annotations, incomplete depth information, and complex motion patterns of dynamic objects such as personnel and equipment in underground coal mine scenes, which increase feature mismatches and reduce pose estimation stability in visual SLAM, this paper proposes a dynamic SLAM method integrating zero-shot semantic priors and dual geometric constraints. First, a zero-shot semantic segmentation module based on open-vocabulary detection and optical-flow prompts is constructed to generate potential dynamic object masks, whose boundaries are further refined by morphological processing, providing semantic priors for dynamic feature detection without mine-specific annotated data. Second, an edge-prior adaptive depth inpainting strategy is introduced to repair depth holes caused by high reflectance and dust occlusion, thereby improving depth completeness and the reliability of geometric constraints. On this basis, epipolar geometry and depth reprojection consistency are jointly employed to identify dynamic features from both two-dimensional matching consistency and three-dimensional depth consistency, enabling robust detection and removal of dynamic features under different motion patterns. Pose estimation is then performed using only static features. Experiments are conducted on public TUM RGB-D dynamic sequences and self-collected typical underground coal mine scenes. The results show that the proposed method effectively removes dynamic features, reduces feature mismatches, and improves pose estimation accuracy and trajectory stability in dynamic environments. On the TUM RGB-D dynamic sequences, compared with DynaSLAM, the proposed method reduces ATE RMSE and S.D. by approximately 23.92% and 30.03% on average, respectively, while maintaining comparable or lower errors than several dynamic SLAM methods. In real underground mine experiments, the method achieves smaller trajectory fluctuations and local drift, demonstrating good scene adaptability and localization stability.