Application of Subdefinite Models in the Global Localization of a Mobile Robot

Abstract. This paper considers the application of subdefinite (SD) models, a variation of constraint programming, to the localization problem of a mobile robot. A complex technology with semantic maps and point cloud maps is proposed. The technology is intended to accelerate and increase the accuracy of global localization in large, symmetric, and periodic environments. The conventional localization approach is based on data from rangefinders generating point clouds; the idea proposed instead is, first, to match the objects observed by the robot to those on the semantic map (recognize the scene), and then apply SD computations to perform localization via visual landmarks. SD computations are used to determine interval constraints on the robot’s positions, represented by several sets for each hypothesis obtained during the scene recognition. Within the interval constraints, the robot is localized using rangefinder data based on a particle filter initialized within these constraints. According to the experiments conducted on the open KITTI-360 dataset, localization based on SD computations can reduce the search space to 0.2% of the original map size. The complex technology shows a significant advantage compared to approaches involving point clouds or visual landmarks only, especially in scenarios with multiple hypotheses about the matches of observed objects and those on the semantic map.

Keywords: global localization, subdefinite models, constraint programming, scene recognition, semantic maps, mobile robot.