Point cloud registration from local feature correspondences-Evaluation on challenging datasets.

Area of Science:

• Robotics
• Computer Vision
• Geomatics

Background:

• Accurate point cloud registration is essential for mobile robot localization and 3D object modeling.
• Coarse alignment is a prerequisite for precise registration algorithms like Iterative Closest Point (ICP).

Purpose of the Study:

• To propose and evaluate a novel feature-based approach for accurate point cloud registration.
• To compare the proposed method against state-of-the-art registration techniques.

Main Methods:

• A feature-based method utilizing points as underlying features for keypoint detection and local reference frame establishment.
• A novel sign disambiguation technique for creating repeatable local reference frames.
• Evaluation on challenging real-world laser scan datasets with moderate overlap.

Main Results:

• The proposed method achieves superior registration accuracy compared to Generalized ICP, 3D Normal-Distribution Transform, Fast Point-Feature Histograms, and 4-Points Congruent Sets.
• Using points as features yields better performance in keypoint detection and local reference frame creation than surface normals.
• The novel sign disambiguation method significantly improves the repeatability of local reference frames.

Conclusions:

• The proposed feature-based registration method offers enhanced accuracy and repeatability for point cloud alignment.
• Points as features and robust sign disambiguation are critical for high-performance 3D registration.
• This work advances the capabilities of mobile robot localization and 3D object modeling pipelines.