GeoDyn-Stereo: a geometry-guided dynamic fusion network for disparity estimation

Area of Science:

• Computer Vision and Robotics
• Optical Metrology
• Industrial Automation

Background:

• Stereo vision is crucial for 3D perception in high-precision manufacturing.
• Dynamic processes like laser machining challenge stereo vision due to accuracy-speed trade-offs.
• Existing methods struggle with real-time monitoring in optically complex industrial environments.

Purpose of the Study:

• To develop a stereo vision network that overcomes the accuracy-speed limitations in dynamic manufacturing.
• To enable real-time, high-resolution 3D monitoring for processes like laser machining.
• To provide a robust vision-based solution for in-process optical inspection.

Main Methods:

• Introduced the geometry-guided dynamic fusion network (GeoDyn-Stereo).
• Integrated explicit geometric constraints into cost-volume representation.
• Employed a dynamic fusion mechanism for robustness and a GRU-based module for disparity map refinement.

Main Results:

• Achieved a 10.96% reduction in end-point error (EPE) and a 6.45% decrease in disparity outlier rate (D1-error).
• Improved processing speed by 22%, increasing frame rate from 0.37 FPS to 0.45 FPS.
• Demonstrated superior performance on standard benchmarks and a custom laser machining dataset.

Conclusions:

• GeoDyn-Stereo successfully balances reconstruction accuracy and computational speed for dynamic 3D perception.
• The method meets the dual requirements of real-time feedback and sub-millimeter spatial resolution for in-process inspection.
• Presents a practical vision-based solution applicable to laser machining and similar industrial applications.