Quantifying and Improving Stereo Camera Calibration Robustness: An Outlier-Aware Algorithm for Digital Twin Data...

Area of Science:

• Computer Vision
• Metrology
• Robotics

Background:

• Stereo camera calibration is crucial for accurate 3D reconstruction and digital twin data acquisition.
• Existing pairwise calibration methods lack robust data-quality checks, necessitating improved data selection strategies.

Purpose of the Study:

• To develop a novel outlier-aware stereo calibration algorithm (OutAw) for enhanced accuracy and reliability in 3D modeling.
• To introduce a systematic, criterion-based data selection framework that overcomes limitations of conventional arbitrary selection methods.

Main Methods:

• Developed the OutAw algorithm, integrating multi-stage outlier detection, geometric selection, subset generation, ranking, and stability analysis.
• Compared OutAw against baseline (BSC) and filtered (InterFil) algorithms using 49 stereo pairs captured with a checkerboard target.
• Validated geometric consistency using triangulation-based metrics like square-length standard deviation and square absolute error.

Main Results:

• OutAw achieved superior results using only nine image pairs, with significantly lower epipolar error (0.5119 px vs. 1.3687 px) and stereo RMS error (0.7666 px vs. 1.9385 px) compared to BSC.
• Demonstrated statistically significant improvements of 60.5% and 62.3% in epipolar and stereo RMS errors, respectively.
• Contamination analysis confirmed that increasing outlier rates progressively degrade calibration quality, highlighting the importance of geometric quality-driven selection.

Conclusions:

• The OutAw algorithm provides a robust and systematic approach to stereo camera calibration, significantly enhancing 3D reconstruction accuracy.
• Geometric quality-driven image selection is critical for reliable stereo calibration, particularly for digital twin applications.
• OutAw redefines stereo calibration from arbitrary to criterion-based data selection, offering a more reliable framework.