應用於無背景限制下三維建模之物體萃取演算法

Abstract

本論文提出一個可將任意背景影片中的前景物體萃取出來的演算法。由於這個演算法是針對三維物體模型重建系統所設計，因此利用重建系統中常用到的三維資訊重建演算法(camera tracker)做為前置處理器，先行重建出影片中的特徵點及相機位置等的三維資訊。萃取演算法首先將特徵點分為前景及後景兩類，再利用前景的特徵點產生前景物體的遮罩，利用遮罩即可將前景的物體萃取出來。基於這個萃取演算法，本論文同時提出一個三維物體模型重建的原型系統，用以測試萃取演算法之效能。與現存系統的最大差別在於，本系統原型可以處理任意背景的物體影片。由實驗結果得知，雖存在許多需要改進之處，但可驗證本論文提出之演算法及系統原型架構具有高度的可行性，亦可預期在進一步改良之後，本系統及演算法能發揮出之潛力。

This thesis proposed a novel algorithm for extracting foreground object in an image sequence of arbitrary background. The algorithm is mainly designed for the 3D object model reconstruction system. Hence, a camera tracker is adopted as a pre-processor to obtain the 3D structure information including feature point positions and camera poses of the image sequence. The proposed algorithm first separate the feature points into foreground and background, and then generates a mask from the foreground points to extract the object. Based on the proposed algorithm, a prototype of 3D object model reconstruction system is presented to verify the performance. The proposed system if capable of dealing with object image sequences of arbitrary background, which is not possible for the recent reconstruction systems. The experiment results show that, though it requires improvements in many aspects, the proposed algorithm and system prototype is functional in practice and should become more powerful after the improvements are done.