Abstract
Abstract
Traditional SLAM systems often suffer from feature extraction failures and localization drift when operating in degraded underground coal mine environments characterized by dust interference, low illumination, and homogeneous textures. To address these issues, we propose a robust SLAM method based on environmental awareness and adaptive enhancement. First, to overcome the scarcity of visual features caused by low illumination, a LiDAR-assisted pseudo-visual cue generation mechanism is designed. By projecting filtered, high-confidence LiDAR structural points onto the image plane, auxiliary cross-modal geometric constraints are constructed to guarantee the continuous tracking of the visual front-end in dark environments. Second, to mitigate point cloud noise and geometric distortion caused by high-concentration dust, a dual-signal-driven adaptive feature extraction mechanism is proposed. This mechanism fuses the spatial distribution and reflection intensity of the point cloud to construct a physical degradation index, and integrates the cumulative angular displacement within the scanning period to quantify motion distortion. By dynamically adjusting the feature extraction thresholds, the system effectively filters out suspended particle noise while strictly preserving the key geometric structures of the tunnels. Finally, visual, LiDAR, and inertial constraints are tightly coupled and solved within a factor graph optimization framework. Experiments on the public M2DGR dataset and in real underground coal mine tunnels demonstrate that the proposed method exhibits improved robustness compared to mainstream algorithms. Achieving an absolute pose root mean square error (RMSE) of 0.066 m in the underground mine, the results demonstrate its effectiveness in extremely degraded environments.