Localization in urban environments. A hybrid interval-probabilistic method

Abstract

Ensuring safety has become a paramount concern with the increasing autonomy of vehicles and the advent of autonomous driving. One of the most fundamental tasks of increased autonomy is localization, which is essential for safe operation. To quantify safety requirements, the concept of integrity has been introduced in aviation, based on the ability of the system to provide timely and correct alerts when the safe operation of the systems can no longer be guaranteed. Therefore, it is necessary to assess the localization's uncertainty to determine the system's operability.

In the literature, probability and set-membership theory are two predominant approaches that provide mathematical tools to assess uncertainty. Probabilistic approaches often provide accurate point-valued results but tend to underestimate the uncertainty. Set-membership approaches reliably estimate the uncertainty but can be overly pessimistic, producing inappropriately large uncertainties and no point-valued results. While underestimating the uncertainty can lead to misleading information and dangerous system failure without warnings, overly pessimistic uncertainty estimates render the system inoperative for practical purposes as warnings are fired more often.

This doctoral thesis aims to study the symbiotic relationship between set-membership-based and probabilistic localization approaches and combine them into a unified hybrid localization approach. This approach enables safe operation while not being overly pessimistic regarding the uncertainty estimation. In the scope of this work, a novel Hybrid Probabilistic- and Set-Membership-based Coarse and Refined (HyPaSCoRe) Localization method is introduced. This method localizes a robot in a building map in real-time and considers two types of hybridizations. On the one hand, set-membership approaches are used to robustify and control probabilistic approaches. On the other hand, probabilistic approaches are used to reduce the pessimism of set-membership approaches by augmenting them with further probabilistic constraints.

The method consists of three modules - visual odometry, coarse localization, and refined localization. The HyPaSCoRe Localization uses a stereo camera system, a LiDAR sensor, and GNSS data, focusing on localization in urban canyons where GNSS data can be inaccurate. The visual odometry module computes the relative motion of the vehicle. In contrast, the coarse localization module uses set-membership approaches to narrow down the feasible set of poses and provides the set of most likely poses inside the feasible set using a probabilistic approach. The refined localization module further refines the coarse localization result by reducing the pessimism of the uncertainty estimate by incorporating probabilistic constraints into the set-membership approach.

The experimental evaluation of the HyPaSCoRe shows that it maintains the integrity of the uncertainty estimation while providing accurate, most likely point-valued solutions in real-time. Introducing this new hybrid localization approach contributes to developing safe and reliable algorithms in the context of autonomous driving.