Camera parameters estimation and evaluation in active vision system

摘要

Active vision system requires accurate camera intrinsic and extrinsic parameters for its sensor scheduling. The recovery of these camera parameters is an inverse problem in this case. We argue that the ability to cope with noises in input data is vital for the camera calibration method since solving the inverse problem is very sensitive to the noises. Random Sample Consensus (RANSAC) technique which makes no assumption about the distribution of the noises is used to obtain stable and accurate results. Perfect assessment rules for the camera calibration in the active vision system are suggested. These rules not only test the 2D-3D mapping relation but also measure the precision of the individual camera parameters. Experiments with real data and a comparison among several camera calibration approaches are included.