Self-Supervised Depth Correction of Lidar Measurements from Map Consistency Loss

TL;DR

The paper tackles lidar depth bias caused by incidence angle and related factors by introducing two map-consistency losses and a differentiable depth-correction model that learns from multi-view scans in a self-supervised manner. It defines polynomial bias models and depth-corrected coordinates, and enforces global map consistency via local covariance-based losses tied to eigenvalues and trace. A joint optimization updates depth-bias parameters and per-scan pose corrections, using a carefully filtered subset of map points observed from multiple viewpoints. Experiments on indoor FEE Corridor data and KITTI-360 demonstrate reduced mapping drift and improved SLAM localization, with practical CPU-time performance and released datasets to support reproducibility.

Abstract

Depth perception is considered an invaluable source of information in the context of 3D mapping and various robotics applications. However, point cloud maps acquired using consumer-level light detection and ranging sensors (lidars) still suffer from bias related to local surface properties such as measuring beam-to-surface incidence angle, distance, texture, reflectance, or illumination conditions. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned depth sensors error while preserving geometric and map consistency details. Despite the effort, depth correction of lidar measurements is still an open challenge mainly due to the lack of clean 3D data that could be used as ground truth. In this paper, we introduce two novel point cloud map consistency losses, which facilitate self-supervised learning on real data of lidar depth correction models. Specifically, the models exploit multiple point cloud measurements of the same scene from different view-points in order to learn to reduce the bias based on the constructed map consistency signal. Complementary to the removal of the bias from the measurements, we demonstrate that the depth correction models help to reduce localization drift. Additionally, we release a data set that contains point cloud data captured in an indoor corridor environment with precise localization and ground truth mapping information.

Figures (9)

• Figure 1: Mobile robot with lidar mapping a corridor-like environment. Points color denote lidar beams incidence angle (darker color - higher angle). Measurements with higher incidence angle have higher bias Laconte-2019-ICRA. The snapshot is taken from the FEE Corridor data set (Sec. \ref{['subsection:dataset']}). The data set is recorded for learning the lidar measurements correction models as well as to evaluate SLAM and reconstruction algorithms.
• Figure 2: Self-supervised pipeline to learn lidar measurements correction model and optimize sensor poses. Point cloud map is constructed from corrected lidar measurements and sensor poses. We compute the map consistency loss on the resultant map and propagate the error back to model parameters as well as to corresponding poses correction terms. The depth correction model removes the bias from raw lidar scans.
• Figure 3: Lidar scan points representation. A ray represented by vector $\mathbf{r}$ falls at an oriented surface (drawn as a blue curve); $\mathbf{v}$ is the vector pointing to the ray origin (sensor location), $\mathbf{x}$ is a vector defining the measured point location. The surface normal $\mathbf{n}$ at the beam heating point forms the lidar ray incidence angle $\gamma$ with the vector $-\mathbf{r}$.
• Figure 4: Filters applied to select from a set of scans the points used in the optimization. Example from the KITTI-360 data set Liao2022PAMI. (a) Outliers with not enough neighboring points are shown in red. (b) Red points have a smaller eigenvalues ratio, $\frac{\lambda_1}{\lambda_2}$ (\ref{['eq:M_plane']}) (c) The rainbow color gradient from red to violet describes the points in the map observed from different sensor poses (view-points). Blue and violet points are measured from higher number of view-points. (d) Resulting filtering mask, purple points are selected in the optimization, Algorithm \ref{['alg:optimization']}.
• Figure 5: Dependence of the point-to-plane distance between measured cloud (by Ouster OS0-128 lidar) and ground-truth plane (given by the calibration board surface) on incidence angle. Results are provided for experiments with the board placed at $5.3 m$ and $8.6 m$ distance from the lidar. Dashed lines correspond to experiments without bias removal, while solid ones represent the error with depth correction applied to the same point clouds. For one experiment its own color is assigned.