Learning-Based Distortion Correction and Feature Detection for High Precision and Robust Camera Calibration

Abstract

Camera calibration is a crucial technique which significantly influences the performance of many robotic systems. Robustness and high precision have always been the pursuit of diverse calibration methods. State-of-the-art calibration techniques, however, still suffer from inexact corner detection, radial lens distortion and unstable parameter estimation. Therefore, in this paper, we improve the precision and robustness of calibration by widening these bottlenecks. In particular, effective distortion correction is performed by a learning-based method. Then, accurate sub-pixel feature location is achieved by the combination of robust learning detection, exact refinement and complete post-processing. To obtain stable parameter estimation, an image-level RANSAC-based calibration procedure is proposed. Ultimately, we assemble these methods into a novel and practical calibration framework. Compared with state-of-art methods, experiment results on both real and synthetic datasets under noise, bad lighting and distortion manifest the better robustness and higher precision of the proposed framework.