RGB-D-Based Pose Estimation of Workpieces with Semantic Segmentation and Point Cloud Registration

Area of Science:

• Robotics
• Computer Vision
• Industrial Automation

Background:

• Industrial robots are crucial for automated production lines.
• Grasping scattered workpieces is a challenging task in robot manipulation research.
• Accurate object recognition and pose estimation are essential for robotic grasping.

Purpose of the Study:

• To propose an efficient and practical solution for object recognition and pose estimation of scattered workpieces.
• To develop a robot vision system using RGB-D information for industrial assembly lines.
• To introduce a novel, cost-effective pose error evaluation method.

Main Methods:

• Fusing semantic segmentation and point cloud registration techniques.
• Utilizing RGB-D (color and depth) information for workpiece analysis.
• Training a modified Fully Convolutional Network (FCN) on a custom industrial dataset for semantic segmentation.
• Estimating real-time workpiece pose using point cloud data derived from depth information.

Main Results:

• The proposed solution achieves accurate object recognition and pose estimation for scattered workpieces.
• The system demonstrates high precision in an industrial scenario, with rotation error under two degrees and translation error below 10 mm.
• The novel pose error evaluation method provides accurate results without expensive equipment.

Conclusions:

• The developed robot vision system offers an efficient and practical approach to workpiece grasping in industrial automation.
• The fusion of semantic segmentation and point cloud registration effectively addresses the challenges of object recognition and pose estimation.
• The proposed evaluation method validates the system's accuracy and practicality for real-world industrial applications.