Dynamic SLAM method for underground coal mines integrating zero-shot semantic priors and dual geometric constraints

Scarce semantically annotated data for underground coal mines, missing depth information, and complex and diverse motion patterns of dynamic objects such as personnel and equipment increase feature mismatches and reduce pose estimation stability in Visual Simultaneous Localization and Mapping (SLAM) in dynamic environments. To address these problems, a dynamic SLAM method for underground coal mines integrating zero-shot semantic priors and dual geometric constraints was proposed. First, to address scarce semantically annotated underground data and long-tailed equipment categories, a zero-shot semantic prior module was constructed. The module generated initial semantic masks for potential dynamic objects through open-vocabulary detection, sparse optical flow estimation, and interactive segmentation, and optimized mask boundaries through morphological processing, thereby providing semantic priors for dynamic feature detection without training on dedicated annotated underground data. Then, to address missing depth regions caused by high reflectance and dust occlusion underground, an adaptive depth image inpainting strategy based on edge-first filling was used to complete missing depth information and improve the completeness of depth information and the reliability of geometric constraints. Finally, to address diverse motion patterns of dynamic objects underground, epipolar geometry and depth reprojection consistency constraints were integrated to identify dynamic features in terms of two-dimensional epipolar consistency and three-dimensional depth consistency, enabling robust identification and removal of dynamic features under different motion patterns. Experiments were conducted on public TUM RGB-D dynamic sequences and self-collected data from typical underground coal mine scenes. The results showed that the proposed method reduced dynamic feature mismatches and improved pose estimation accuracy and trajectory stability in dynamic environments. In localization experiments on TUM RGB-D dynamic sequences, both the root mean square error and standard deviation of absolute trajectory error for trajectories estimated by the proposed method were lower than those of the comparison methods. In localization experiments in underground coal mine scenes, trajectories estimated by the proposed method were broadly consistent in shape with the reference paths and exhibited good localization stability and scene adaptability.