Aimed at the problems of insufficient information and poor reusability of the sparse map constructed by ORB_SLAM2, a visual navigation method based on improved ORB_SLAM2 was proposed. It included two stages of building a multi-layer map and navigation. In the stage of building a multi-layer map, a local dense point cloud was calculated by the key frame of ORB_SLAM2, outliers were removed by radius filter and fast itreative closest point (FAST ICP) algorithm was used to register the processed point cloud. After that, 3D occupancy information was calculated by local dense point cloud; 3D occupancy information was extracted by means of the height of mobile robot in 3D space and 2D occupancy information was calculated by 2D mapping; 3D, 2D occupancy information and 3D, 2D features of the key frames were fused to generate a globally consistent multi-layer map. In navigation stage, according to the prior information of positioning layer, 2D features of the key frame were clustered to generate a visual dictionary, the visual dictionary was indexed according to the characteristics of current image to obtain the reference key frame; the initial pose was calculated by perspective-n-point (PnP) algorithm, and then reprojection error was used to construct inter-frame constraints, final result of localization was obtained through Gauss-Newton optimization; in planning layer, A* algorithm was used to plan path so that mobile robot visual navigation was realized. Verified by TUM dataset, the proposed method was about 50% faster than RGB-D SLAM, and the pose estimating was almost improved by 10%, the localization result based on prior map were consistent with the original map. In addition, the experiments on the real robot platform showed that the proposed method can construct a high-precision multi-layer map, and the error between the measured value of lAC and the real value was 6.7%, and the error between the measured value of lBD and the real value was 5.6%, and the fast and accurate navigation was achieved on the basis of multi-layer map.