Maximum Consensus based Localization and Protection Level Estimation using Synthetic LiDAR Range Images

A core task of autonomous vehicles is the ego-localization. On the one hand, it needs to be accurate and robust towards outliers, and on the other hand, it needs to provide a high integrity. For an accurate localization, usually the likelihood as the most common objective function is maximized with respect to the sensor measurements. More precisely, based on the assumption of normally distributed random variables, usually a least squares minimization is conducted, e. g. in a recursive manner in a Kalman Filter. However, this approach severely lacks robustness since least squares is inherently sensitive to outliers, which, using LiDAR data, arise in a high proportion due to changing environments and dynamic traffic participants. Therefore, more robust loss functions have been introduced with the outlier count, also referred to as maximum consensus optimization, as the most robust one since the estimation result is independent of the outlier magnitude. Whereas localization itself has been investigated thoroughly, localization integrity, which describes the ability of a system to correctly and timely warn the user when specific error limits are exceeded, has received comparatively few attention in the domain of autonomous vehicles. In this paper, on the one hand, we propose a robust localization approach based on the maximum consensus criterion to ensure a high reliability even in challenging environments, and on the other hand, we propose an estimation of a grid-based probability distribution as a first step towards a new integrity framework for LiDAR based localization. Both outcomes are based on the same pipeline using GPU computed range images for the expected measurements. The introduced localization and probability estimation are tested for structured, semi-structured, and unstructured environments using overall around 2,500 epochs from two different LiDARs.