AGHL: Anchor-Guided Point Cloud Registration Network With Hybrid Local Feature Perception.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Point cloud registration is crucial for 3D applications.
• Existing methods lack hybrid local feature extraction and are susceptible to irrelevant data interference.
• Transformers, while effective for global context, can be hindered by noisy regions in point cloud data.

Purpose of the Study:

• To propose AGHL, a novel detector-free approach for accurate point cloud registration.
• To address limitations in feature extraction and global context integration in existing methods.
• To improve the robustness and accuracy of point cloud recognition and alignment.

Main Methods:

• Introduced a hybrid local feature perception module with parallel branches for low-level and high-level feature extraction.
• Developed an anchor-guided cross-attention mechanism to focus on geometrically consistent regions.
• Utilized Euclidean and high-dimensional feature spaces for encoding point-neighborhood correlations.

Main Results:

• AGHL achieved superior point cloud registration accuracy on synthetic, indoor, and outdoor datasets.
• The method effectively encodes correlations between points and their neighbors.
• Demonstrated strong generalization ability in real-world robot localization experiments.

Conclusions:

• AGHL significantly enhances point cloud registration by improving feature representation and attention mechanisms.
• The proposed method offers a robust solution for various 3D data alignment tasks.
• AGHL shows promise for real-world applications requiring precise 3D spatial understanding.