Abstract

Network architecture

Loop closure detection is an essential component of Simultaneous Localization and Mapping (SLAM) systems, which reduces the drift accumulated over time. Over the years, several deep learning approaches have been proposed to address this task, however their performance has been subpar compared to handcrafted techniques, especially while dealing with reverse loops. In this paper, we introduce the novel LCDNet that effectively detects loop closures in LiDAR point clouds by simultaneously identifying previously visited places and estimating the 6-DoF relative transformation between the current scan and the map. LCDNet is composed of a shared encoder, a place recognition head that extracts global descriptors, and a relative pose head that estimates the transformation between two point clouds. We introduce a novel relative pose head based on the unbalanced optimal transport theory that we implement in a differentiable manner to allow for end-to-end training. Extensive evaluations of LCDNet on multiple real-world autonomous driving datasets show that our approach outperforms state-of-the-art techniques by a large margin even while dealing with reverse loops. Moreover, we integrate our proposed loop closure detection approach into a LiDAR SLAM library to provide a complete mapping system and demonstrate the generalization ability using different sensor setup in an unseen city.


How Does It Work?

We present the novel LCDNet for loop closure detection which performs both loop detection and point cloud registration. Our method combines the ability of DNNs to extract distinctive features from point clouds, with algorithms from the transport theory for feature matching. LCDNet is composed of a shared backbone that extracts point features, followed by the place recognition head that extracts global descriptors and the relative pose head that estimates the transformation between two point clouds. One of the core components of our LCDNet is the Unbalanced Optimal Transport (UOT) algorithm that we implement in a differentiable manner. UOT allows us to effectively match the features extracted from the two point clouds, reject outliers, and handle occluded points, while still being able to train the network in an end-to-end manner. As opposed to existing loop closure detection methods that estimate the relative yaw rotation between two point clouds, our proposed LCDNet estimates the full 6-DoF relative transformation between them which significantly helps the subsequent ICP refinement to converge faster.


Network architecture
Figure: Overview of our proposed LCDNet which is composed of a shared feature extractor (green), a place recognition head (blue) that generates global descriptors, and a relative pose head (orange) that estimate the transformation between two point clouds. We use three loss functions to train LCDNet (triplet loss, aux loss, and pose loss) which are depicted in purple.

Videos

Code

A software implementation of this project based on PyTorch can be found in our GitHub repository for academic usage and is released under the GPLv3 license. For any commercial purpose, please contact the authors.

Publications

Daniele Cattaneo, Matteo Vaghi, Abhinav Valada,
LCDNet: Deep Loop Closure Detection and Point Cloud Registration for LiDAR SLAM
IEEE Transactions on Robotics, 2022.

(Pdf) (Bibtex)


Acknowledgement

This work was partly funded by the Federal Ministry of Education and Research (BMBF) of Germany under SORTIE, and by the Eva Mayr-Stihl Stiftung.

People