Object recognition and localization from 3D point clouds by maximum-likelihood estimation.

Area of Science:

• Computer Vision
• Robotics
• Geometric Modeling

Background:

• Object recognition and pose estimation are crucial in robotics and computer vision.
• Existing methods often rely on interest points or separate algorithms for initial estimation and refinement.
• Surface segmentation from depth images offers rich feature information often overlooked by other techniques.

Purpose of the Study:

• To develop a unified, efficient algorithm for automated object recognition and pose estimation from 3D point clouds.
• To leverage surface segmentation features for improved performance compared to interest point-based methods.
• To provide a single approach for both initial pose estimation and subsequent pose refinement.

Main Methods:

• A maximum-likelihood analysis algorithm is proposed.
• Surfaces segmented from depth images are used as primary features.
• The algorithm is demonstrated with a 2 degrees of freedom (d.f.) example and a full 6 d.f. analysis.

Main Results:

• The algorithm exhibits negligible memory requirements compared to the 6D Hough transform.
• It is computationally efficient, outperforming iterative closest point algorithms.
• Achieved an RMS alignment error as low as 0.3 mm in a cluttered scene analysis.

Conclusions:

• The presented maximum-likelihood algorithm offers a unified and efficient solution for object recognition and pose estimation.
• Utilizing surface features provides a robust alternative to interest point-based methods.
• The method's applicability spans from initial recognition to precise pose refinement with high accuracy.