An Experimental Evaluation of Learning-Based Methods for Loop Closure Detection in Simultaneous Localization and Mapping

Abstract

Simultaneous Localization and Mapping (SLAM) is the capability to estimate a robot’s trajectory in an initially unknown environment while reconstructing the geometry of the environment. In order to bound the accumulation of localization error in SLAM, it is crucial to recognize previously seen locations, a process called "loop closure." This allows the robot to make corrections to its localization and map estimates. This project evaluates ORB feature extraction and matching, a state-of-the-art technique to detect loop closures, against recently developed learning-based approaches. In particular, our first contribution is to benchmark established techniques based on hand-crafted descriptor matching against novel learning-based approaches based on neural networks (i.e., SuperPoint and SuperGlue). As a second contribution, we integrate a learning-based loop closure detection method as part of Kimera, a SLAM system, and demonstrate its performance in both simulated and real benchmarking datasets. Finally, we collect data on long trajectories using a Jackal robot to compare the different approaches on real-world situations beyond available datasets. Our evaluation shows that, while learning-based approaches detect many more loop closures across wider baselines, when integrated in a SLAM system, they do not lead to substantial performance improvements compared to standard ORB feature matching.