Abstract: Recovering the 3D geometric structure of an observed object from its two perspective views is one of the most important tasks in the computer vision research. It’s well known that the accuracy of 3D reconstruction is mainly decided by the performance of motion estimation method. In the paper, the fundamental model of two-view imaging system is introduced and the principles and disadvantages of some motion estimation methods are analyzed. To improve the estimation accuracy, a modified non-linear iterative optimization method is proposed based on the projective error minimization. The numerical simulations prove that the estimation accuracy is improved either under large translation and small rotation or under small translation and large rotation. Furthermore, the proposed method is verified by reconstruction from two view images of simulated object and real object. The result shows that the metric error is below 2% and the angle error below 3°.

Key words: computer vision, 3D reconstruction, motion estimation, Levenberg-marquard algorithm