OverlapNet: a siamese network for computing LiDAR scan similarity with applications to loop closing and localization

Abstract Localization and mapping are key capabilities of autonomous systems. In this paper, we propose a modified Siamese network to estimate the similarity between pairs LiDAR scans recorded by cars. This can be used address both, loop closing for SLAM global localization. Our approach utilizes deep neural exploiting different cues generated from data. It estimates using concept image overlap generalized range images furthermore provides relative yaw angle estimate. Based on such predictions, our method is able detect closures in system or globally localize given map. For closure detection, use prediction as measurement find candidates integrate candidate selection into an existing improve performance. localization, novel observation model predictions provided OverlapNet it Monte-Carlo localization framework. We evaluate multiple datasets collected scanners various environments. The experimental results show that effectively surpassing detection performance state-of-the-art methods generalizes well Furthermore, reliably localizes vehicle typical urban environments data seasons.