Rslo: self supervised lidar odometer (real time + high precision, icra2022)

Paper title ：Robust Self-Supervised LiDAR Odometry Via Representative Structure Discovery and 3D Inherent Error M

Code link ：https://github.com/SamuelYale/RSLO

Lidar is a necessary sensor in automatic driving , However, due to the sparsity and signal-to-noise ratio of lidar data, the robustness of motion estimation will be challenged , Here's the picture , Even two adjacent LIDAR point clouds , There will also be a mismatch due to dynamic objects and misalignment , This mismatch is very fatal to lidar odometer , So how to find and reduce the weight of these unreliable regions is a problem worth studying .

On the other hand , Although based on 3D CNN The model has been proved to be able to solve the above problems to a certain extent , But now some are based on 3D CNN The self-monitoring lidar odometer is not explicitly designed to force the network to focus adaptively on reliable areas , And the self-monitoring method usually uses the geometric consistency between scans to achieve self motion estimation . As shown in the figure above , Geometric consistency may not be fully maintained in practice . And most of the existing methods model and optimize the inherent error at the class level by estimating the uncertainty scalar . But as shown in the figure below , The magnitude of the inherent error is not only different in different categories , And in different directions , Therefore, a more detailed inherent error model is needed to distinguish the uncertainty differences in different directions in the training process .

Contributions

1. A two-stage self motion estimation framework is proposed , Focus the network on two frames LiDAR On the representative structure of odometer .

2. adopt 3D CNN Estimate the pointwise covariance , And the point uncertainty is modeled , On this basis, the inherent error of self - monitoring loss can be reduced .

3. An uncertain perceptual mapping module is introduced , The module will discover representative structures and CNN Point by point covariance estimation of is used for map construction and scan2map Pose optimization ..

4. stay KITTI Translation of data set / Rotation error , The proposed self-monitoring odometer is higher than the previous technical level 16%/12%, stay Apollo Southbay The performance is also very good . Based on the mapping module and more training data , The proposed system can even compete with supervised methods in real time and efficiently .

chart 1 Model frame