Graph-Laplacian-Processing-Based Multimodal Localization Backend for Robots and Autonomous Systems

Simultaneous localization and mapping (SLAM) for positioning of robots and autonomous systems (RASs) and mapping of their surrounding environments is a task of major significance in various applications. However, the main disadvantage of traditional SLAM is that the deployed backend modules suffer from accumulative error caused by sharp viewpoint changes, diverse weather conditions, etc. As such, to improve the localization accuracy of the moving agents, we propose a cost-effective and loosely coupled relocalization backend, deployed on top of original SLAM algorithms, which exploits the topologies of poses and landmarks generated either by camera, LiDAR, or mechanical sensors, to couple and fuse them. This novel fusion scheme enhances the decision-making ability and adaptability of autonomous systems, akin to human cognition, by elaborating graph Laplacian processing concept with Kalman filters. Initially designed for cooperative localization of active road users, this approach optimally combines multisensor information through graph signal processing and Bayesian estimation for self-positioning. Conducted experiments were focused on evaluating how our approach can improve the positioning of autonomous ground vehicles, as prominent examples of RASs equipped with sensing capabilities, in challenging outdoor environments. More specifically, experiments were carried out using the CARLA simulator to generate different types of driving trajectories and environmental conditions, as well as real automotive data captured by an operating vehicle in Langen, Germany. Evaluation study demonstrates that localization accuracy is greatly improved both in terms of overall trajectory error as well as loop closing accuracy for each sensor fusion configuration.