Robust Low-Overlap Point Cloud Registration via Displacement-Corrected Geometric Consistency for Enhanced 3D Sensing.

Area of Science:

• Computer Vision and Robotics
• 3D Perception and Sensing

Background:

• Accurate 3D point cloud alignment is crucial for robotics, autonomous navigation, and environmental reconstruction.
• Low-overlap scenarios, caused by sensor limitations or occlusions, significantly challenge registration robustness and reliability.

Purpose of the Study:

• To introduce GeoCORNet, a novel deep learning network for geometric consistency optimization and rectification in 3D point cloud registration.
• To enhance registration accuracy and robustness, particularly in challenging low-overlap conditions.

Main Methods:

• Developed GeoCORNet, integrating a geometric consistency enhancement module and a bidirectional cross-attention mechanism.
• Implemented a predictive displacement rectification strategy to correct erroneous correspondences dynamically.
• Employed joint optimization of overlap loss and a novel displacement loss function.

Main Results:

• GeoCORNet demonstrated superior performance over existing methods in registration recall and reduced rotation error under low-overlap conditions.
• The Attentive Cross-Consistency module effectively suppresses noise and reinforces geometric coherence in overlapping regions.
• The predictive displacement strategy maximizes the utility of sparse sensor data.

Conclusions:

• GeoCORNet establishes a new paradigm for robust 3D sensing in real-world applications with partial sensor data.
• The proposed method significantly improves robustness and accuracy in complex, low-overlap 3D registration scenarios.