Scalable Change Detection from 3D Point Cloud Maps: Invariant Map Coordinate for Joint Viewpoint-Change Localization

Abstract

This study addresses the problem of visual change detection using a 3D point cloud (PC) map acquired by a car-like robot. With recent advances in long-term autonomous navigation, change detection under global viewpoint uncertainty has become a topic of considerable interest. In our study, we extend the traditional two-level pipeline of change detection: (1) scene registration and (2) scene comparison, to enable scalable and efficient change detection. In the traditional pipeline, the registration stage is required to align a given scene pair (i.e., query and reference PC maps) that are taken at different times into the same coordinate system, before comparing the two PCs. However, the registration stage is a time-consuming step, which makes it harder to realize a scalable change detection. Our key concept is to transform every query or reference PC beforehand into an invariant coordinate system, which should be predefined and invariant to environment changes (e.g., dynamic objects, clutters, the mapper vehicle's trajectories), so as to enable a direct comparison of spatial layout between the two different maps. The proposed framework employs an efficient bag-of-local-features (BoLF) scene model and realizes a scalable joint viewpoint-change detection. Change detection experiments using a publicly available cross-season NCLT dataset validate the efficacy of the approach.