LRMO: A Lightweight and Redundant Multi-Modal Odometry Framework for Robust Intelligent Vehicle Localization

Abstract. Reliable and robust self-localization is the essential component of intelligent vehicles (IV). Many scholarly works have been focused on developing accurate multi-modal integrated pose estimation schemes. Such single estimation engine design lacks consideration of potential individual sensor failures. In this paper, we present a resilient framework that exploits the redundancy of different sensors using a stack of odometry algorithms. The multiple pose estimation algorithms run in parallel with a general adaptivity and lightweight design. Specifically, we integrate the vehicle wheel encoder and the vehicle dynamics data to the filter-based LiDAR-inertial odometry. In contrast to most of the odometry algorithms which may fail entirely against temporary failures, the redundant system enables self-recovery of individual odometry through reinitialization. The most promising odometry is selected at each timestamp through weighting metric evaluation. In this way, our method can exploit the robustness and advantages of individual estimating engines. We evaluate our method on both research purpose IVs and mass-produced IVs. The experimental results suggest that our approach is resilient to various failure cases and achieves better performance than individual methods.