Robustness of neural network calibration model for accurate spatial positioning

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Camera calibration is crucial for 3D spatial positioning in machine vision.
• Traditional models like pinhole and polynomial have limitations in robustness to calibration errors.
• Understanding error propagation is key to improving calibration accuracy.

Purpose of the Study:

• To systematically evaluate the robustness of neural network (NN)-based camera calibration.
• To propose a metric for quantifying calibration model robustness against input errors.
• To assess NN model performance under various optical configurations and non-normal scenarios.

Main Methods:

• Developed a dimensionless error attenuation coefficient to measure model robustness.
• Analyzed NN model robustness via synthetic simulations considering noise, distortion, and viewing angle.
• Deployed NN models for asymmetric camera layouts and multiple camera joint calibration.

Main Results:

• NN models demonstrate superior robustness over pinhole and polynomial models.
• Error attenuation coefficient effectively quantifies model robustness.
• NN models significantly improve 3D positioning accuracy in challenging, non-ideal conditions.

Conclusions:

• Neural network-based camera calibration offers enhanced robustness and accuracy for 3D spatial positioning.
• The proposed error attenuation coefficient is a valuable metric for evaluating calibration models.
• NN models are highly effective in complex machine vision applications with non-ideal camera setups.