Efficient Center Voting for Object Detection and 6D Pose Estimation in 3D Point Cloud.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Estimating 6D object poses from point clouds is crucial for robotics and computer vision applications.
• Existing methods based on point pair features (PPF) face challenges with complex scenes and similar objects.

Purpose of the Study:

• To develop a novel and efficient framework for accurate 6D object pose estimation using point clouds.
• To improve the robustness and discriminative power of PPF-based methods in challenging environments.

Main Methods:

• Utilized a point pair feature (PPF) matching approach for initial object pose estimation.
• Introduced a new center voting strategy based on relative geometric relationships for improved pose hypothesis generation.
• Implemented a pose verification operator to filter false positives and refine predictions.

Main Results:

• The proposed method effectively estimates 6D object poses in complex scenes represented by point clouds.
• Demonstrated robustness against similar-looking distractors, sensor noise, and simple geometric shapes.
• Achieved superior performance compared to state-of-the-art methods on benchmark datasets.

Conclusions:

• The enhanced PPF-based framework offers a discriminative and robust solution for 6D object pose estimation.
• The novel center voting strategy significantly improves the accuracy and reliability of pose prediction.
• The approach is suitable for multi-instance and multi-object detection tasks.