Abstract: In coal mine inspection scenarios, traditional SLAM algorithms typically face the dual challenges of poor real-time performance and significant trajectory drift during local mapping due to excessive accumulation of redundant frames. These limitations not only compromise the autonomous navigation accuracy of inspection robots but also undermine the reliability of their intelligent decision-making. To address these issues, this paper investigates an environmental perception method for coal mine inspection robots based on a visual SLAM keyframe selection algorithm. The system is equipped with a Kinect v2 camera to capture RGB-D multimodal data from the coal mine environment. Within the ORB-SLAM2 framework, it innovatively integrates optical flow tracking with a geometric mapping model to achieve precise camera pose estimation. A geometric constraint method is introduced, which combines the geometric mapping model with dynamic thresholds to perform keyframe selection for SLAM. Building on map initialization results and referencing the selected keyframes, local mapping is carried out by incorporating pose information from tracking and feature point matching from keyframes. Newly acquired keyframes are continuously integrated to estimate map point positions and update the map. If an update fails, a new local map is constructed during the tracking phase, and local mapping proceeds accordingly. The latest tracking and keyframe selection results are then combined to resume the mapping process. A loop closure detection method is employed, which selects candidate frames and computes SE（3） transformations to achieve loop fusion and optimize the environmental perception map. The current frame is compared with existing keyframes; when a revisited area is identified, a loop closure is established based on forward and backward pose constraints, thereby eliminating accumulated errors and ensuring global map consistency and reliability. Experiments conducted in two inspection scenarios demonstrate that the proposed visual SLAM keyframe selection-based environmental perception method reduces the number of keyframes and trajectory tracking errors, improves perception real-time performance, and yields high-quality local and global maps that align closely with the actual trajectories.