Deep learning-based methods have attracted more attention to the pose estimation research that plays a crucial role in location and navigation. How to directly predict the pose from the point cloud in a data-driven way remains an open question. In this paper, we present a deep learning-based laser odometry system that consists of a network pose estimation and a local map pose optimization. The network consumes the original 3D point clouds directly and predicts the relative pose from consecutive laser scans. A scan-to-map optimization is utilized to enhance the robustness and accuracy of the poses predicted by the network. We evaluated our system on the KITTI odometry dataset and verified the effectiveness of the proposed system.
展开▼
