[Paper Review] Enhance Accuracy: Sensitivity and Uncertainty Theory in LiDAR Odometry and Mapping

Zeyu Wan, Yu Zhang+5 more|arXiv (Cornell University)|Nov 15, 2021

Robotics and Sensor-Based Localization4 citations

TL;DR

This paper proposes a novel sensitivity and uncertainty theory to enhance LiDAR odometry and mapping accuracy by selecting high-sensitivity, low-uncertainty point residuals. By formulating sensitivity and uncertainty in six dimensions and applying a threshold-based selection, the method achieves global statistical optimality, reducing residual terms and improving pose estimation accuracy in both indoor and outdoor LiDAR odometry and LiDAR-inertial odometry systems without sacrificing real-time performance.

ABSTRACT

Currently, the improvement of LiDAR poses estimation accuracy is an urgent need for mobile robots. Research indicates that diverse LiDAR points have different influences on the accuracy of pose estimation. This study aimed to select a good point set to enhance accuracy. Accordingly, the sensitivity and uncertainty of LiDAR point residuals were formulated as a fundamental basis for derivation and analysis. High-sensitivity and low -uncertainty point residual terms are preferred to achieve higher pose estimation accuracy. The proposed selection method has been theoretically proven to be capable of achieving a global statistical optimum. It was tested on artificial data and compared with the KITTI benchmark. It was also implemented in LiDAR odometry (LO) and LiDAR inertial odometry (LIO), both indoors and outdoors. The experiments revealed that utilizing selected LiDAR point residuals simultaneously enhances optimization accuracy, decreases residual terms, and guarantees real-time performance.

Motivation & Objective

• To address the critical need for improved pose estimation accuracy in LiDAR-based SLAM systems, especially in long-term operation.
• To identify that not all LiDAR points contribute equally to pose estimation accuracy, with varying sensitivity and uncertainty.
• To develop a theoretically grounded, generalizable method for selecting optimal point residuals to enhance front-end accuracy in LiDAR odometry and mapping.
• To ensure the selected method maintains real-time performance while reducing computational load and residual error.

Proposed method

• The paper formulates sensitivity as a six-dimensional vector (three for rotation, three for translation), quantifying how much a registration residual changes under small pose disturbances.
• Uncertainty is modeled as a three-dimensional Gaussian distribution, reflecting the reliability of a LiDAR point's measurement and its geometric model (e.g., plane or line).
• Sensitivity and uncertainty are decoupled and combined into a score vector for ranking all LiDAR point residuals.
• A threshold-based selection rule is applied to retain only high-sensitivity, low-uncertainty residuals for optimization.
• Theoretical proof demonstrates that this selection method achieves global statistical optimality in minimizing pose estimation error.
• The method is implemented as a plug-in module compatible with any optimization-based LiDAR SLAM system, including LO and LIO.

Experimental results

Research questions

• RQ1 Can sensitivity and uncertainty theory be used to systematically identify and select the most informative LiDAR point residuals for improved pose estimation?
• RQ2 Does selecting high-sensitivity, low-uncertainty residuals lead to a statistically optimal reduction in pose estimation error?
• RQ3 Can this selection method be generalized across different LiDAR SLAM algorithms and hardware configurations without compromising real-time performance?
• RQ4 How does the proposed method compare to using all valid planar feature points in terms of accuracy and computational efficiency?

Key findings

• The proposed method achieved significantly lower translation errors on the KITTI sequence 03 benchmark, despite using approximately half the number of planar feature points compared to the baseline.
• The selected residuals reduced the number of residual terms in optimization while simultaneously improving pose estimation accuracy in both indoor and outdoor environments.
• Theoretical analysis confirmed that the selection strategy achieves a global statistical optimum in minimizing pose estimation error.
• The method demonstrated real-time performance when integrated into both LiDAR odometry (LO) and LiDAR-inertial odometry (LIO) systems.
• High-sensitivity points were found to be predominantly located on middle-building walls, while near-ground and tree points exhibited low sensitivity and high uncertainty.
• The method is independent of specific LiDAR hardware and SLAM algorithm design, making it a general-purpose accuracy enhancement module.