Abstract: In underground mining environments, unstructured terrains, poor lighting conditions, and repetitive features lead to insufficient accuracy in single-sensor simultaneous localization and mapping (SLAM). Although tight coupling fusion of multi-sensor data can improve accuracy to some extent, challenges remain, including high computational load and poor adaptability to sudden illumination changes. To address these issues, an improved LiDAR-inertial-visual tightly-coupled SLAM algorithm was proposed based on the Fast Tightly-Coupled Sparse-Direct LiDAR-Inertial-Visual Odometry (FAST-LIVO) algorithm. For multi-sensor tight coupling fusion, the Lucas-Kanade (LK) optical flow method was adopted to replace the original sparse direct method, tracking stable feature points and constructing a visual reprojection error. Meanwhile, the random sample consensus (RANSAC) algorithm was applied to eliminate outliers and retain high-quality visual constraints. By combining Inertial Measurement Unit (IMU) prior estimation with LiDAR point-to-plane residuals, the iterative error-state Kalman filter was employed to achieve tight coupling fusion of multi-sensor data, outputting high-precision pose estimates. For map construction, an incremental k-d tree (ikd-Tree) was employed to dynamically manage point clouds for building the LiDAR local map. Visual feature points were extracted through grid filtering and Shi-Tomasi score calculation, while an array was utilized to manage and dynamically remove features outside the field of view, constructing the visual local map. The LiDAR point clouds were projected onto corresponding images to extract RGB color information, generating colored point cloud frames. These frames were then stitched based on optimized poses to construct a colored point cloud map. Experimental results on the Gazebo simulation platform demonstrated that, compared to the FAST-LIVO algorithm, the proposed method reduced both absolute trajectory error (ATE) and relative pose error (RPE) by over 20%, with clearer features such as tunnel sidewalls, internal pile contours, and ground surfaces. Tests on the public M2DGR dataset showed that the proposed algorithm achieved higher localization accuracy than LEGO-LOAM, FAST-LIO, and FAST-LIVO, exhibited no significant drift at turns, and maintained superior trajectory stability. Additionally, the average data processing time of the proposed method was reduced. Test results in a long corridor simulation environment demonstrate that the proposed algorithm achieves clearer reconstruction of spatial structures, more accurate details such as lines and contours, better noise suppression, and more precise reflection of the actual environmental layout.