Kernkonzepte
A robust cross-modality point cloud registration framework that combines feature filtering and local-global optimization to achieve state-of-the-art performance.
Zusammenfassung
The paper proposes a cross-modality point cloud registration framework called FF-LOGO that addresses the challenges in aligning point clouds from different sensor modalities. The key components of the framework are:
-
Cross-Modality Feature Correlation Filtering Module:
- Extracts geometric transformation-invariant features from cross-modality point clouds using a Geometric Self-Attention mechanism.
- Performs feature matching and point selection to obtain an initial pose estimation.
-
Local Adaptive Key Region Aggregation Module:
- Identifies dispersed and geometrically representative key points in the point cloud using Farthest Point Sampling.
- Aggregates neighboring points around the key points to form local adaptive key regions.
-
Global Modality Consistency Fusion Optimization Module:
- Matches the local adaptive key regions with the cross-modality feature-coupled point set to compute point-to-plane residuals.
- Performs local-to-global optimization to refine the initial pose estimation and obtain the final optimized transformation.
The proposed method significantly outperforms the current state-of-the-art on the 3DCSR dataset, improving the recall rate from 40.59% to 75.74%. The authors also demonstrate the practical application of FF-LOGO for cross-modality localization on a bipedal wheeled robot.
Statistiken
The 3DCSR dataset contains point clouds from three different modalities: LiDAR, Kinect, and camera sensors.
LiDAR point clouds are relatively sparse, while Kinect point clouds are dense and uniform.
The dataset provides ground truth transformations for aligning either LiDAR or SfM geometry with dense Kinect geometry.
Zitate
"Our method fully leverages the advantages of deep learning in fuzzy correspondence and traditional optimization in pose fine-tuning for cross-modality registration and achieve the state-of-the-art with an improvement from 40.59% to 75.74%."