Pose360: Metric-Scale Visual Odometry by Grounding Learned Features with LiDAR Depth
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Metric-scale state estimation is a cornerstone for autonomous systems, yet single-modality solutions often fail in challenging real-world environments. Visual odometry suffers from inherent scale ambiguity, while LiDAR odometry can be fragile in geometrically sparse scenes. To overcome these limitations, we propose Pose360, a novel Visual-LiDAR Odometry (V-LIO) system that robustly fuses a 360° panoramic camera and a 360° LiDAR. Our approach leverages learned features, SuperPoint and LightGlue, to establish strong 2D visual correspondences, which are then lifted into a sparse set of metric 3D-3D point correspondences using depth information from a synchronized LiDAR point cloud. The 6-DoF relative pose is then computed efficiently via a closed-form SVD-based solution. This focused fusion strategy directly resolves visual scale ambiguity without requiring complex non-linear optimization. We perform a rigorous quantitative evaluation on a challenging 450 m real-world dataset, demonstrating that our system achieves high global consistency with a translation Absolute Trajectory Error (ATE) of just 0.177 m RMSE against a high-fidelity LiDAR SLAM ground truth. We further validate its real-world applicability by successfully integrating Lidar360Pose as the core odometry engine in a full robotic navigation stack, proving it is an accurate and reliable solution for metric state estimation.
Fields of Study & Subjects
Robotics and Sensor-Based Localization
3D Surveying and Cultural Heritage
Advanced Vision and Imaging
Publication details
- Journal / source
- International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
- Pages
- 1438–1442
- Publication date
- 24 February 2026
- Language
- EN