A High Accuracy Camera Self-calibration Method Based On The Essential Matrix

Photogrammetry system aims to extract precise3D metric information from2D images. And its measurement precision is determined by the calibration accuracy of the camera parameters. Generally speaking, camera calibration methods may be divided into two categories:traditional algorithms which make use of high-precision calibration pattern and self-calibration methods that aim to calibrate the camera directly from unknown images. Traditional calibration algorithms are in most cases very simple, easy to implement and efficient due to the presence of high-accuracy calibration patterns. As the calibration accuracy is dependent on the precision of the calibration pattern, the calibration pattern must be high-precision, which increases the cost of the algorithm. Besides, when working environment changes a lot, the transportation of the calibration may cause trouble, too. In conclusion, though traditional calibration methods can provide high-precision calibration results, it’s expensive and inconvenient to apply. Self-calibration methods aim to calibrate the camera through totally unknown scenes, that is, there isn’t any control points in the scene.Self-calibration methods usually make use of the external constraints extracted through projection geometry. The principles are in general more complicated that those of the traditional methods.In this paper, we propose a high accuracy self-calibration method which may calibrate the camera in the absence of the calibration patterns. Our method bypasses the ambiguous concepts such as the absolute dual quadric or the absolute conic, and is thus simple. We utilize the properties of the essential matrix, which is, two of the three singular values of the essential matrix must be equal and the third one must be zero, to self-calibrate the intrinsic parameters of the camera. And the extrinsic parameters are then obtained through factorizing the essential matrix. The calibration accuracy is guaranteed by the final bundle adjustment which optimizes all the camera parameters and the3D coordinate of the feature points simultaneously. Synthetic data as well as real data has been used to test the method, and the simulation result shows that the mean absolute relative calibration errors of the focal lengths and the principal point are about4.5e-4,8e-4respectively, under zero-mean Gaussian noise with0.1 pixels standard deviation. Compared with existing calibration methods, the proposed method is cost-effective and simple with comparable accuracy.