In Computer Vision Point Of The Opposite Sex Regression Techniques

Recently, in using geometry constraint to solve the structure reconstruction problems from the perspective projection views of the interested targets, one of outstanding research achievements is to study the constraint relationship among images by utilizing the geometry properties of different scenes. It plays a vital role in three-dimensional computer vision research field. However, for most of the available algorithms, the incorrect estimations of measurement error cause inaccurate analysis of geometry constraint relationship. Also, these algorithms are very sensitive to the effect of noise and result in biggish error in parameter estimation.The main objective of this project is to study and analyze a great deal of ill-posed problems of computer vision through using various traditional algorithms and heteroscedastic regression technique. The research is focused on three typical applications of computer vision: ellipse fitting, the estimation of the fundamental matrix and camera self-calibration. The applied heteroscedastic regression technique takes the heteroscedastic propertiy of noise into account and creates the Errors-ln-Variables (EIV) model from statistics. Based on the model, from the observation of data vectors, the optimal algorithm is found to obtain the optimal estimations of EIV model parameters and the truth-value of the observed data points.In the application of ellipse fitting, the in-depth theoretical analysis to the usual algorithms as well as heteroscedastic regression algorithm suggested by Leedan et al, is performed, and a new algorithm which computes heteroscedastic EIV model parameters is proposed. The new algorithm uses original and exact minimal Mahalanobis distance, also called as Maximum Likelihood method under Gauss distribution. In numerical computation, more robust Generalized SVD is used, and the optimal solution is obtained by regression computation. The process of data correction is isochronous to the regression process of model parameters.Compared to the existing algorithms, the new algorithm is more accurate and can converge steadily and rapidly, even when original data is far from exact value. In the application of vision fundamental matrix estimation, the experiments prove that the heteroscedastic regression algorithm gives obvious improvement toovercome disadvantages of common algorithms, such as too big sensitive response to noise, weak anti-noise robustness and so on, so that its numerical steady property and robust property are also improved.Combining the new algorithm of camera self-calibration suggested by Zhonggen Yang et al., we apply three algorithms discussed in the application of vision fundamental matrix estimation to camera self-calibration. Because of the computation error and the inherent anti-noise robustness of that algorithm, there is no much difference in the simulation results of three algorithms, and the heteroscedastic regression algorithm is appreciably better than the others. Thus, further research is still necessary.Through analysis vast ill-posed problems of computer vision, this thesis seeks the accurate estimation of parameters by considering EIV measurement error. Also, experiments and corresponding analysis are performed on several robust algorithms. The accomplishment of this project can solve many types of computer vision problems and lead to better performance and practicability.