Neural-Network-Based Model-Free Calibration Method for Stereo Fisheye Camera.

Area of Science:

• Computer Vision
• Robotics
• Metrology

Background:

• Fisheye cameras offer a wide field of view (>180°) beneficial for medical and precision measurement applications.
• Standard pinhole models are inadequate for severe barrel distortion in fisheye lenses.
• Nonlinear geometric models are computationally intensive for calibrating fisheye cameras.

Purpose of the Study:

• To propose a model-free stereo calibration method for binocular fisheye cameras using neural networks.
• To overcome the computational complexity and limitations of traditional fisheye camera calibration techniques.
• To enable accurate spatial coordinate calculation in the common field of view of binocular fisheye systems.

Main Methods:

• A neural network is employed to implicitly learn the nonlinear mapping between image and spatial coordinates.
• A feature extraction method utilizing a three-step phase-shift technique is incorporated.
• The method avoids the need for image correction and matching, simplifying the calibration process.

Main Results:

• The neural network effectively models the complex distortion inherent in fisheye cameras.
• Spatial coordinates of points within the common field of view are accurately calculated.
• The proposed method preserves the wide field-of-view advantage of fisheye cameras.

Conclusions:

• The model-free, neural-network-based stereo calibration is highly suitable for fisheye cameras, especially those with significant distortion.
• This approach offers a computationally efficient and accurate alternative to conventional methods.
• The technique enhances the applicability of fisheye cameras in precision measurement and medical imaging.