Precise and Reliable Localization of Mobile Robots in Crowds Using NDT-AMCL

Abstract

The number of applications for autonomous mobile robots is continuously growing. Recently, there has been a trend towards integrating mobile robots closer to humans and deploying them in environments with high human presence. Particularly in indoor settings, this poses a new challenge for localization algorithms, as most indoor localizers use their environmental sensors to localize themselves in the robot's environment. Humans moving dynamically within the environment disrupt this process. The widely used SLAM (Simultaneous Localization and Mapping) or MCL (Monte Carlo Localization) based algorithms reach their limits when there are many people around the robot or when the robot is in a dense crowd. Therefore, further development of algorithms that can reliably and accurately localize even in crowded environments is necessary. In this paper, we introduce our localizer, NDT-AMCL, which enhances the commonly used Adaptive Monte Carlo Localization (AMCL) by incorporating Normal Distribution Transformations (NDT). We investigate the localization error of NDT-AMCL in private residences and compare it with AMCL through realistic simulations. Compared to AMCL, NDT-AMCL can improve localization error in human-free environments by 56 % and is significantly less disturbed by humans in its vicinity. Thus, the localization error of NDT-AMCL in an apartment with 30 dynamically moving humans is still smaller than the localization error of AMCL in an empty apartment. As systematic evaluation of localization errors in this context has proven challenging, we propose a new error-weighted metric for the objective assessment of localization errors. We make NDT-AMCL available to the research community as an extension of the AMCL package in ROS2.