Subset-based stereo calibration method optimizing triangulation accuracy

Calibration of vision systems is essential for performing measurement in real world coordinates. For stereo vision, one performs stereo calibration, the results of which are used for 3D reconstruction of points imaged in the two cameras. A common and flexible technique for such calibration is based on collection and processing pairs of images of a planar chessboard calibration pattern. The inherent weakness of this approach lies in its reliance on the random nature of data collection, which might lead to better or worse calibration results, depending on the collected set of image pairs. In this paper, a subset-based approach to camera and stereo calibration, along with its implementation based on OpenCV, is presented. It utilizes a series of calibration runs based on randomly chosen subsets from the global set of image pairs, with subsequent evaluation of metrics based on triangulating the features in each image pair. The proposed method is evaluated on a collected set of chessboard image pairs obtained with two identical industrial cameras. To highlight the capabilities of the method to select the best-performing calibration parameters, a principal component analysis and clustering of the transformed data was performed, based on the set of metric measurements per each calibration run.