UWB Range-Assisted Monocular Visual Localization System for Underground Parking Lots

In this paper, we present a novel monocular visual localization system that leverages range information to accurately estimate the global pose of a vehicle in realtime. We address the three key challenges of monocular visual odometry - namely, scale ambiguity, scale drift, and error accumulation - by integrating UWB range measurement. To overcome the UWB non-line-of-sight (NLOS) issue, we design an innovative NLOS identification and culling mechanism that relies on the latest visual relative measurement. Our system is implemented using a factor graph, which incorporates co- visible and adjacent factors to effectively utilize the visual information. We validate the exceptional performance of the proposed localization system by extensively testing it with a large dataset collected in underground parking lots. Additionally, to facilitate further research, we plan to publish this self- collected dataset at https://github.com/fwf-lernen/TJ-UWB-MonoImage-Dataset.gitsoon.