Robust Self-Supervised LiDAR Odometry Via Representative Structure Discovery and 3D Inherent Error Modeling

The correct ego-motion estimation basically relies on the understanding of correspondences between adjacent LiDAR scans. However, given complex scenarios and low-resolution LiDAR, finding reliable structures for identifying can be challenging. In this letter, we delve into structure reliability accurate self-supervised aim to alleviate influence unreliable in training, inference mapping phases. We improve odometry substantially from three aspects: 1) A two-stage network is developed, where obtain by estimating a set sub-region transformations averaging them with motion voting mechanism, encourage focusing representative structures. 2) inherent alignment errors, which cannot eliminated via optimization, are down-weighted losses based 3D point covariance estimations. 3) discovered learned covariances incorporated module robustness map construction. Our two-frame outperforms previous state arts 16%/12% terms translational/rotational errors KITTI dataset performs consistently well Apollo-Southbay datasets. even rival fully supervised counterparts our more unlabeled training data.