LiDAR Odometry by Deep Learning-Based Feature Points with Two-Step Pose Estimation

An accurate ego-motion estimation solution is vital for autonomous vehicles. LiDAR widely adopted in self-driving systems to obtain depth information directly and eliminate the influence of changing illumination environment. In odometry, lack descriptions feature points as well failure assumption uniform motion may cause mismatches or dilution precision navigation. this study, a method perform odometry utilizing bird’s eye view data combined with deep learning-based point proposed. Orthographic projection applied generate image 3D cloud. Thereafter, an R2D2 neural network employed extract keypoints compute their descriptors. Based on those descriptors, two-step matching pose designed keep these tracked over long distance lower mismatch ratio compared conventional strategy. experiment, evaluation proposed algorithm KITTI training dataset demonstrates that can provide more trajectories handcrafted feature-based SLAM (Simultaneous Localization Mapping) algorithm. detail, comparison descriptors demonstrated. The difference between RANSAC (Random Sample Consensus) also demonstrated experimentally. addition, collected by Velodyne VLP-16 evaluated solution. low-drift positioning RMSE (Root Mean Square Error) 4.70 m from approximately 5 km mileage shown result indicates has generalization performance low-resolution LiDAR.