3D robust reconstruction using a hand-held digital camera

Abstract

Conventional vision-based reconstruction methods comprise three main steps: feature extraction and matching, estimation of camera matrices, and triangulation method for 3D point computation. Standard factorization algorithms provide an effective means for computing the 3D object shape and camera models simultaneously from the multiple views without going through the process of pairwise reconstructions. Since the numbers of features used in all views must be the same and sufficiently large in these methods, they are not applicable to the cases where not every feature point is visible in all views. On the other hand, the extracted features are often corrupted by image noise. Therefore, there is a need for a reconstruction algorithm that can handle missing feature points and image noise both in order to yield a robust result. We propose a method integrating a multi-view reconstruction and a least-squared-error estimation to produce a robust reconstruction result. Besides, the proposed method utilizes an auto-calibration scheme to convert the projective reconstruction to a Euclidean reconstruction. The method is tested on the synthetic and real data.