Accurate Robot Arm Attitude Estimation Based on Multi-View Images and Super-Resolution Keypoint Detection Networks.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Robot arm monitoring is crucial for intelligent industrial automation.
• Accurate estimation of robot arm attitude is essential for precise control and operation.
• Existing methods may face challenges with data acquisition and generalization.

Purpose of the Study:

• To propose a novel two-stage method for robot arm attitude estimation using multi-view images.
• To enhance keypoint detection accuracy and reduce computational load.
• To achieve precise joint angle estimation with reduced reliance on real-world data.

Main Methods:

• A super-resolution keypoint detection network (SRKDNet) with a subpixel convolution module for high-resolution heatmap generation.
• A training strategy combining virtual and real data for improved generalization.
• A coarse-to-fine dual-SRKDNet detection mechanism for refined keypoint localization.
• An equation system integrating camera and robot kinematic models with confidence-based keypoint screening for attitude estimation.

Main Results:

• Achieved 96.07% keypoint detection accuracy (PCK@0.15) on a real robot arm.
• Demonstrated an average joint angle estimation error of 0.53 degrees on a UR10 robot arm using three views.
• The proposed method shows superior performance compared to existing approaches.

Conclusions:

• The two-stage method effectively estimates robot arm attitude from multi-view images.
• SRKDNet and the confidence-based screening scheme significantly improve accuracy and efficiency.
• The approach offers a practical solution for robot arm monitoring in intelligent industrial scenarios.