MagneticPillars++: Efficient LiDAR Odometry Via Deep Frame-To-Keyframe Point Cloud Registration

Downstream applications for point cloud registration, like LiDAR Odometry, often conduct Iterative Closest Points (ICP) in the initial frame-to-frame matching and/or subsequent map refinement. However, due to its distance-based processing nature, ICP relies on an accurate pose initialization while i...

Full description

Saved in:
Bibliographic Details
Main Authors: Kai Fischer, Martin Simon, Stefan Milz, Patrick Mäder
Format: Article
Language:English
Published: Taylor & Francis Group 2025-12-01
Series:Applied Artificial Intelligence
Online Access:https://www.tandfonline.com/doi/10.1080/08839514.2025.2472105
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Downstream applications for point cloud registration, like LiDAR Odometry, often conduct Iterative Closest Points (ICP) in the initial frame-to-frame matching and/or subsequent map refinement. However, due to its distance-based processing nature, ICP relies on an accurate pose initialization while implicating increased computational complexity with a growing number of points. To meet specific runtime requirements, methods often apply the extensive mapping step at low frequencies, e.g. every 10 frames, which in turn leads to increased noise on the calculated trajectory. To tackle the discrepancy between runtime and accuracy, we present MagneticPillars++, an extension of our previous point cloud registration approach optimized for LiDAR Odometry, introducing novel intermediate cell correspondence filtering and accelerated match normalization. Furthermore, we propose a frame-to-keyframe matching technique replacing the simple frame-to-frame matching within a LiDAR Odometry pipeline. This can tremendously reduce noise without the need for expensive ICP corrections. We conduct extensive experiments for various tasks like point cloud registration, LiDAR Odometry, and loop closure estimation, demonstrating the versatility of our approach, where we are able to outperform state-of-the-art approaches in terms of accuracy and runtime, resulting in residual translation and rotation errors of up to 4.7 cm and 0.231 with an average runtime of.
ISSN:0883-9514
1087-6545