Multiple Points Localization With Defocused Images

Abstract

3D points reconstruction has attracted increasing attentions both in computer vision and robotic intelligence areas. However, the real depth measurement still relies on depth measurement instruments. Although many measurement methods for depth exist, they usually need additional instruments, which always increase the cost of the measurement system. To better localize the position of 3D points without use of other instruments, a direct method is proposed which acquires depth from defocus of current images in this paper. The method utilizes the property of camera lens system and mechanism of SFM to remove the ambiguity of structure scale and the relative error between these 3D points. In addition, a multiple image setting for improving the robustness of depth estimation is proposed that can further eliminate depth error from some kinds of nature noises. Experiments on the real scene are implemented, which shows that the proposed method outperforms the ordinary 3D points localization method.