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

OverlapTransformer: An Efficient and Rotation-Invariant Transformer Network for LiDAR-Based Place Recognition

“…Most recently, there are also works exploiting semantic information for place recognition. For example, Chen et al [4], [3] propose OverlapNet to exploit multiple cues generated from LiDAR scan, including depth, normal, intensity, and semantics for LiDAR-based loop closure detection and localization. SGPR by Kong et al [16] exploits the semantics and topological information of the raw point cloud and extracts the semantic graph representation with graph neural networks to find loop closures.…”

“…The range image is a natural representation of a single 3D scan from a rotating LiDAR sensor such as Velodyne or Ouster sensors. It is a compact representation and is especially suitable for online tasks such as online SLAM [2], [6], loop closing [3], [4], or Fig. 1: Query scan (blue) and reference scan (orange) with adjacent locations but opposite viewpoints in our novel Haomo dataset.…”

“…Instead of using handcrafted descriptors [13], [22], [23], we propose a new transformer neural network to extract yaw-rotationinvariant global descriptors from LiDAR scans. We apply a similar overlap concept as described in OverlapNet [3], [4] to supervise the network learning and to estimate the similarity between pairs of scans. Different to OverlapNet, which exploits multiple cues such as normal, intensity, and semantic information as the network input, in this work we only use the depth information of the range image to achieve faster online performance and make the approach easier to generalize.…”

“…We build the architecture of our Overlap-Transformer to ensure that each descriptor is yaw-rotationinvariant, which makes our method robust to viewpoint changes. We train the proposed OverlapTransformer only on a part of the KITTI dataset and evaluate it on both, KITTI and Ford Campus datasets with the loop closure metric in line with OverlapNet [3]. Besides, we recorded and will release a new dataset, which contains three different challenges including place recognition for long time spans, reverses driving, and different appreances scenes to evaluate different methods.…”

Special issue on the 2020 “Robotics: Science and Systems” conference

“…The Lat-ticeNet approach is also used for instance and dynamic object segmentation. "OverlapNet: a siamese network for computing LiDAR scan similarity with applications to loop closing and localization" by Chen et al (2021) proposes a modified version of Siamese Deep Neural Network to estimate similarity between pairs of LiDAR scans collected by autonomous cars. The paper then shows that this similarity can be used effectively to help with loop closing in the context of SLAM and global localization.…”