Abstract: For the SLAM system based on feature point matching, the problem of pose estimation failure is due to the inability to extract enough feature points in weak texture regions lacking corner points. Applying direct visual odometry method LDSO for indoor robot visual localization and combined with depth estimation or depth camera acquired key frame depth image, key frame camera pose, original key frame image data, point cloud stitching to generate 3D point cloud dense map, experimental results shows that the robot can be in a complex environment Accurate and fast localization, and the algorithm significantly reduces the cumulative error of rotation, translation and proportional drift by pose optimization without global Bundle Adjustment, so that the overall performance is comparable to SLAM systems based on feature point matching, the algorithm takes less time, has better real-time performance and is more robust in areas lacking corner points.