利用三維資訊做人臉辨識

Abstract

不同於以往二維人臉辨識的方法，本論文是研究利用兩張人臉影像所具有的三維資訊，從三維空間的角度來探討人臉辨識的問題。其主要方法是從人臉上挑選取三個特徵點以形成一參考平面，並求出其他各特徵點相對於此平面的關聯仿射結構。因為此種結構具有投影幾何的不變性，所以即使相機內外參數不同，或人臉以不同的角度取像，同一個人所求出對應於此一結構之特徵值並不會改變，而不同的人所求出之結果則恆有誤差。基於這樣的特性，所提出的方法便可達到辨識人臉的目的。

Different from many existing face recognition approaches, this thesis considered using the 3D information obtained from two face images to solve the problem of face recognition. The proposed approach first chooses three feature points from a face to establish a reference plane, then computes the relative affine structure（RAS）with respect to this plane using the rest of feature points. Such a structure is an invariant under projective geometry for the same person, even under the variations of intrinsic or extrinsic camera parameters, but is different for different persons. A numerical measure is thus established to assess the difference between two RAS’s and to facilitate face recognition.