Abstract: Due to the narrow underground environment of coal mine, dark and changeable light, the mine image has the characteristics of low illumination, low contrast map and uneven color, which affects the matching result of visual SLAM feature points extraction and makes the positioning performance drop sharply. In order to improve the positioning accuracy of monocular visual positioning algorithm of coal mine mobile robot in low illumination, weak texture and unstructured environment features, the ORB-SLAM3 positioning algorithm is improved. On the basis of the front-end feature point extraction (ORB) algorithm, histogram equalization, non-maximum suppression, adaptive threshold method and feature point homogenization based on quadtree strategy are introduced. In feature point matching, LK optical flow method based on image pyramid is introduced to reduce the number of optimization iterations. After the feature point matching is completed, the RANSAC algorithm is added to remove the mismatched feature points and improve the matching accuracy of the feature points. Through the method of triangulation at the back end, the pixel depth information is obtained, and the 2D-2D pose solving problem is transformed into 3D-2D (pnp) pose solving problem. According to the principle of tight coupling of visual inertial navigation, the residual function of the whole positioning system is constructed by fusing visual residual error and IMU residual error, and the sliding window BA algorithm based on nonlinear optimization is used to iteratively optimize the residual function to obtain accurate pose estimation of the mobile robot. The improved algorithm is compared with ORB-SLAM3 algorithm and VSIN-Mono algorithm in four data sets. The results show that: (1) Compared with the ORB-SLAM3 algorithm and the VMS-MONO algorithm, the motion trajectory of the proposed positioning system is the closest to the true value trajectory; (2) All indexes of APE of the positioning system are better than ORB-SLAM3 algorithm and VMS-MONO algorithm; The root-mean-square error of the positioning system is 0.049m (the mean value of four experiments), which is 31.1% lower than that of ORB-SLAM3 (the mean value of four experiments).