HALNet: Partial Point Cloud Registration Based on Hybrid Attention and Deep Local Features.

Area of Science:

• Computer Vision
• Robotics
• 3D Geometry Processing

Background:

• Point cloud registration is crucial for 3D reconstruction and target recognition.
• Existing deep learning methods struggle with accurate registration of partial point clouds.
• A need exists for robust algorithms capable of handling incomplete 3D data.

Purpose of the Study:

• To propose a novel deep learning network, HALNet, specifically designed for partial point cloud registration.
• To enhance the accuracy and robustness of point cloud registration in scenarios with partial overlap.
• To improve performance in applications like 3D reconstruction and target recognition.

Main Methods:

• Feature extraction using adaptive graph convolution (AGConv), 2D convolution, and convolution block attention (CBAM).
• Overlapping estimation to filter non-overlapping points between point clouds.
• Hybrid attention mechanism (self-attention and cross-attention) for fusing geometric information.
• Rigid transformation estimation via a fully connected layer.

Main Results:

• HALNet demonstrated superior registration accuracy compared to five state-of-the-art methods.
• Achieved a 10.67% reduction in RMSE(R) and a 12.05% reduction in MAE(R) compared to SCANet.
• Ablation studies confirmed the significant contribution of the hybrid attention mechanism and fully connected layer to performance.

Conclusions:

• HALNet effectively addresses the challenge of partial point cloud registration.
• The proposed network architecture significantly improves registration accuracy and robustness.
• The findings suggest HALNet as a promising solution for real-world 3D computer vision and robotics applications.