Real-Time 6-DOF Pose Estimation of Known Geometries in Point Cloud Data

Area of Science:

• Robotics and Automation
• Computer Vision
• Geometric Perception

Background:

• Object pose estimation from point clouds is crucial for robot perception and control systems.
• The Iterative Closest Point (ICP) algorithm is a common method but suffers from practical limitations.
• Existing methods often struggle with measurement uncertainty, clutter, and computational efficiency.

Purpose of the Study:

• To introduce a novel, robust, and efficient algorithm for pose-from-point cloud estimation.
• To address the limitations of current methods like ICP in real-world robotic scenarios.
• To enable real-time pose tracking for applications in field robotics and mining.

Main Methods:

• Development of the Pose Lookup Method (PLuM), a probabilistic reward-based objective function.
• Utilization of lookup tables to replace computationally intensive geometric operations like raycasting.
• Benchmarking against state-of-the-art ICP-based methods using triangulated geometry models.

Main Results:

• PLuM demonstrates millimetre accuracy and high-speed pose estimation in benchmark tests.
• The method significantly outperforms existing ICP-based approaches in accuracy and robustness.
• Real-time application successfully tracked haul trucks at 20 Hz using LiDAR data in a mining environment.

Conclusions:

• PLuM provides a dependable and efficient solution for pose estimation from point clouds.
• The algorithm's resilience to uncertainty and clutter makes it suitable for demanding environments.
• PLuM's straightforward implementation and real-time performance offer significant advantages for robotic perception.