Defocused calibration for large field-of-view binocular cameras

Abstract

In tunnel boring machine construction sites, on-site calibration of large field-of-view binocular cameras is still a challenging task due to factors such as the narrow spaces of the machines' mechanical structures, high pressure, or falling rocks. Therefore, a calibration method using a small plate in proximity to the camera (a defocused position safe for workers) is proposed in this paper to deal with the above challenges. First, a defocus-insensitive feature point encoding pattern is developed. Then feature points are extracted through a phase recovery network based on the Swin Transformer, and finally camera parameters are optimized based on virtual symmetric transfer errors. The relevant simulation and experimental data show that the relative error between the proposed method and the phase-shift method in feature extraction is 0.15 pixels, while the relative error between the proposed method and a high-precision laser tracker in three-dimensional measurement is 1.3%. The results demonstrate the feasibility of the method in a large field-of-view binocular camera for tunnel boring machine construction sites. Increasing the diversity of data sets and lightweight models will be the focus of future research.