TSO-BoW: Accurate Long-Term Loop Closure Detection With Constant Query Time via Online Bag of Words and Trajectory Segmentation

Abstract

This letter presents TSO-BoW, a lightweight trajectory segmentation-based Bag-of-Words algorithm for loop closure detection, utilizing intermittent online training for collected segments. In the online training phase, segments of collected data form sub-trajectories that are used for online training based on their features, ultimately creating corresponding sub-databases for querying. In the querying phase, we use a multiple-level querying approach. Initially, candidate sub-databases are selected based on geometric distance using prior pose information. Subsequently, a lower bound criterion is applied to filter out some sub-databases, followed by PnP-RANSAC for geometric verification and precise relative pose estimation. Our algorithm mitigates the pose drift issue in prior pose selection-based loop detection algorithms by using a segmented Bag-of-Words and lower bound elimination. It maintains constant query time and memory cost without compromising query performance in long-term (Simultaneous localization and mapping) SLAM. Evaluations on large-scale public datasets demonstrate our algorithm's excellent computational and memory efficiency, query time efficiency, and superior query performance in long-term SLAM system.