Abstract: In the application of visual SLAM(Simultaneous Localization and Mapping) in coal mines, lighting changes and low-texture scenes seriously affect the extraction and matching of feature points, resulting in the failure of pose estimation and affecting the positioning accuracy. Therefore, a binocular vision localization algorithm SL-SLAM for underground mobile robots in coal mines based on the improved ORB (Oriented Fast and Rotated Brief)-SLAM3 algorithm is proposed. For the lighting change scenario, the original ORB feature point extraction algorithm is replaced by the SuperPoint feature point extraction network with lighting stability at the front end, and a feature point grid limitation method is proposed to effectively eliminate the invalid feature point area and increase the stability of pose estimation. For the low-texture scene, a stable LSD(Line Segment Detector) line feature extraction algorithm is introduced at the front end, and a point-line joint algorithm is proposed, which groups the line features according to the feature point grid, and matches the line features according to the matching results of the feature points, so as to reduce the matching complexity of the line features and save the pose estimation time. The reprojection error model of point features and line features is constructed, the angle constraints are added to the line feature residual model, the Jacobian matrix of the pose increment of point features and line features is derived, the unified cost function of the reprojection error of point features and line features is established, the local mapping thread uses the ORB-SLAM3 classic local optimization method to adjust the pose of points, line features and keyframes, and performs loop correction, subgraph fusion and global BA(Bundle Adjustment) in the back-end thread. The experimental results on the EuRoC dataset show that the APE(Absolute Pose Error) index of SL-SLAM is better than other comparison algorithms, and the trajectory prediction results closest to the true value are obtained, and the root mean square error is reduced by 17.3% compared with ORB-SLAM3. The experimental results of simulating the underground scene of coal mine show that SL-SLAM can adapt to the scene of light change and low texture, and can meet the positioning accuracy and stability of the mobile robot in the underground coal mine.