Deep Bayesian-Assisted Keypoint Detection for Pose Estimation in Assembly Automation.

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Accurate pose estimation is essential for automating assembly tasks but remains challenging and part-specific.
• Existing methods often require extensive training data and part-specific tailoring, limiting scalability in production environments.

Purpose of the Study:

• To present a novel, streamlined pose estimation method for enhancing assembly automation.
• To improve the accuracy and adaptability of pose estimation for industrial assembly tasks.

Main Methods:

• Employs deep learning on limited annotated images to identify keypoints on assembly parts.
• Incorporates a Bayesian updating stage, leveraging part design knowledge to refine network outputs.
• Utilizes high-quality keypoint positions as measurements and network outputs as priors, with geometry data to construct likelihood functions.

Main Results:

• Significantly improved pose estimation accuracy through Bayesian refinement of keypoint locations.
• Demonstrated promising results on a 14-point snap-fit dash trim assembly for a Ford Mustang dashboard.
• Achieved accurate pose estimation for updating nominal assembly trajectories using Maximum A Posteriori (MAP) estimates.

Conclusions:

• The proposed method offers a scalable and adaptable solution for pose estimation in assembly automation.
• It reduces the need for extensive machine learning expertise and large datasets, making it practical for production floors.
• The approach effectively refines pose estimation, enabling more robust and accurate automated assembly processes.