環物影像的自動相機校正與模型重建

Abstract

在這篇論文中，我們提出了一套方法夠為一系列的影像進行自動的相機校正，並重建物體的三維模型。相機的校正包含兩個步驟，第一步是參考自近年來從運動恢復結構的研究，從影像中抽取一些特徵點，並利用特徵點之間的對應關係來估計相機的方向和相對位置。第二步則是利用物體的輪廓資訊來提升校正的準確度，藉由調整相機參數使重建的視覺外型在相機的成像能盡可能符合物體原有的輪廓。取得每張影像的校正資訊以後，我們的模型重建是利用圖形分割演算法，以影像色彩的一致性作為考量，對視覺外型進行刻劃取得凹面的細節，同時藉由輪廓資訊標示出一些位於物體表面的空間點，作為尋找最佳分割面的限制條件，保留物體表面的突出部份。 我們提出的方法在實際影像的相機校正上有很好的表現，模型重建結果也令人滿意。由於並不需要假設相機已事先經過校正，我們的方法非常適用於博物館擴增實境的文物展示以及藝術品的數位典藏。

In this paper, we propose a method for automatic camera calibration and model reconstruction from image sequences. Our approach for camera calibration consists of two steps. The first step is based on the recent work of structure from motion that recovers camera orientations and positions from a sets of corresponding features points extracted from images. The second step exploits the concept of silhouette coherence to refine camera parameters in the way that the projections of the reconstructed visual hull can be consistent with the original silhouettes as much as possible. Once camera parameters have been recovered, our multi-view stereo reconstruction use a graph-cuts algorithm to carved the visual hull by optimizing the photo-consistency of the surface in a global manner to pursue concavities. A set of surface points identified from silhouettes are also imposed to the optimization as hard constraints to preserve protrusions of the surface. The proposed method is shown to perform well on camera calibration for real data, and the results of model reconstruction are also satisfactory. Since it does not require images being calibrated in advance, our work is suitable for applying augmented reality on the museum exhibition and the digital archives of museum artifacts.