Visual Odometry with Lidar Depth Enhancement

Aiming at the accuracy defect of single sensor in positioning, this paper puts forward a positioning algorithm based on visual depth enhancement of lidar. In this algorithm, the point cloud is obtained by lidar, and the ground is segmented by depth learning. For the sparse area of point cloud, different depth fitting methods are used for the ground and other areas to obtain the best depth value. At the same time, the Kalman filter method is used to fuse lidar and visual information. Verified on public data set KITTI, the error rate of this method is reduced by 74.26% and 36.79% on average compared with monocular mode and stereo mode of classical ORB-SLAM algorithm.