Optimised calibration of machine vision system for close range photogrammetry based on machine learning

Real-time inspection of large mechanical parts manufacturing using camera-based scanning systems are increasingly adopted in industry 4.0. It leads to take preventive actions during the manufacturing process and then to fabricate mechanical parts right-first-time with respect to specified tolerances. Therefore, the use of camera-based scanners requests a preliminary calibration process. It consists on estimating the intrinsic and extrinsic parameters required to relate the 3D world point to its projection on the image plane. Since selection of the calibration grid poses affect the calibration quality, one approach-based machine learning (ML-approach) is proposed including the polynomial approximation of the reprojection errors function of 6 degree of freedom (DoF) combined with particle swarm optimization (PSO). Synthetic and experimental evaluations have been performed while assessing the performance of the proposed ML-approach. The synthetic evaluation reveals a better convergence of the intrinsic and extrinsic parameters in comparison to recent published calibration methods by Wizard (CW-method) and Rojtberg (R-method). The experimental evaluation of the ML-approach shows an average error R_E < 12 µm and a sub-micrometre repeatability, which confirm the benefit of using machine vision-based scanning systems for the inspection of large volume parts in real time.