Visual-LiDAR-Inertial fusion SLAM algorithm for underground detecting robot in coal mines

Abstract

Abstract In seeking to overcome the challenges of unknown environments in underground coal mines, such as low illumination, weak textures, and complex structures, the localization and mapping accuracy of mobile robots often suffers from image degradation and poor feature-matching robustness. To address these limitations, this paper presents a visual-LiDAR-IMU fusion SLAM framework specifically tailored for subterranean conditions. At the front end, the system employs the HSV color space and Multi-Scale Retinex (MSR) algorithm, enhanced with contrast-limited adaptive histogram equalization (CLAHE), to suppress noise and significantly improve feature extraction in low-texture regions. Structural line features are extracted using the EDLines algorithm, and Line Band Descriptors (LBD) are utilized to enhance the robustness of feature matching. At the back end, a joint optimization model combining visual point and line reprojection residuals with IMU pre-integration residuals is constructed. A sliding-window optimization is then employed to achieve high-precision pose estimation. Concurrently, the Scan Context descriptor is integrated to enable loop closure detection and enhance global consistency. Compared with conventional approaches, the proposed method effectively mitigates LiDAR degradation, reinforces visual constraints, and significantly improves SLAM stability and accuracy in weak-texture and highly dynamic environments. It provides a robust, real-time, and high-precision localization and mapping solution for autonomous operations in underground coal mine environments.