Real-time laser stripe robust detection for robotic weld seam tracking based on temporal image sequences.

Area of Science:

• Robotics
• Computer Vision
• Welding Technology

Background:

• Automated robotic welding systems require precise seam tracking for quality and efficiency.
• Existing vision-based systems often struggle with real-time performance and noise interference, such as arc flash.

Purpose of the Study:

• To develop a novel, lightweight segmentation algorithm for robust, real-time vision-based seam tracking in robotic welding.
• To enhance the algorithm's efficiency for deployment on edge devices.

Main Methods:

• A novel lightweight segmentation algorithm leveraging temporal continuity was developed.
• A co-optimization framework was designed for automated structured pruning of the segmentation model.
• The system was evaluated for real-time performance and tracking accuracy.

Main Results:

• The proposed algorithm demonstrated superior immunity to noise, including arc flash, by utilizing temporal continuity.
• The co-optimization framework enabled efficient inference on edge devices through automated model pruning.
• Experimental results showed the system operating at 55.2 FPS with a tracking error of 0.154 mm.

Conclusions:

• The developed vision-based seam tracking system offers robust, real-time performance suitable for industrial robotic welding.
• The lightweight algorithm and efficient pruning framework contribute to the advancement of automated welding technologies.