Automatic Selection Of Camera Calibration Images By Object Posture

Camera calibration is one of the most classic and important problems in computer vision, which has been studied extensively for decades. Not only newly produced camera must run calibration to correct its radial distortion and intrinsic parameters, it is also the first step towards many important computer vision applications, such as re-constructing3D structures from multiple images.During the past several decades, researchers have studied many different approaches for developing more convenient, practical, and accurate algorithms for camera calibration. To the best of our knowledge, the existing algorithms rely on extracting local certain features such as corners, edges and SIFT features for camera calibration and assembling them to establish correspondences for calibrating the intrinsic parameters, the extrinsic parameters and the radial lens distortion. One important class of these solutions requires a specially designed calibration object, with3-D geometric information known explicitly [1,2,3,4,5]. The calibration objects include3-D [4,11],2-D plane [5], and1-D [6] line targets. These approaches effectively solved the camera calibration problem. They start with an analytical solution, followed by a nonlinear optimization technique based on the maximum likelihood criterion. Finally, they take into account lens distortion, giving both analytical and non-linear solutions. However, all the above calibration schemes require a large number of unselected or random selected or manually selected images.Such as, camera calibration with one-dimensional object consisting of three or more collinear points with known relative positioning moved the stick around while trying to fix one end with the aid of a book [6]. Where a video of150frames was recorded, the calibration algorithm was applied to the random four sample images from the150observation. For the two-dimensional object [5,7], five images of the plane under different orientations were taken from a video, the algorithm was applied to the first2,3,4and all5images. For the three-dimensional object [4], some images were selected from all different viewpoints using an identical camera with same setting. Those kinds of image selections are either time costly, or inconvenient, or might not accurate.Notice that most calibration methods share one thing in common which exclusively relying on whether points or lines can be reliably obtained from local corner or edge features. Whereas feature extraction or labeling often becomes a bottleneck of the process, affecting robustness, accuracy and convenience. For those calibration images including poor features will lead to bad results. Most current calibration techniques compute with a large number of unselected images or random selected images or manually selected images, which is often time consuming or possibly existing poor features or might containing uncertain factors. So we propose an algorithm to standardize, classify and select pictures from a large number of the camera calibration images. In doing so, our method can save time, improve accuracy and be convenient for end users. Computer simulation and real image experiments show our methods are robust and results in high resolution.