Learning-based calibration of a high-precision monocular vision system for macro-micro manipulation

Calibration is critical for the optical vision measurement system (OVMS) that requires high measuring accuracy within a large field of view. In this paper, we propose a practical two-step calibration method that improves measuring accuracy for the multi-scale vision measurement system based on a learning-based approach. In this method, an OVMS combined with a degenerated perspective-n-points algorithm is established for macro-micro manipulation systems. After the OVMS is coarsely calibrated by the analytic technique, the fine calibration is performed using the Artificial Nerual Network to compensate for both geometric and non-geometric measuring errors. Simulations, experiments and comparisons are carried out with a laser tracker, and the results show that the proposed two-step calibration method can remarkably minimize measuring errors and shows stronger robustness and stability, which enables the OVMS to be capable of macro-micro manipulation tasks.