基於視覺的智慧型指向系統

Abstract

由二維的影像資訊重建出三維的場景模型，在電腦視覺領域已被研究多年，這項技術若應用在指向系統上，便可以幫助我們進行場景的重建，以及計算使用者所指向的位置。目前有關於指向系統的研究，大部分都是分析手勢的影像，來計算所指的方向，但這類的方法精確度不高。本論文以一套現有的電腦視覺軟體系統，首先進行場景特徵點的三維重建，保留這個步驟所計算出來的基本矩陣，之後每次重建，不須再重算基本矩陣，直接用第一次重建時所保留的基本矩陣去重建指向物體，這個方法的優點是計算時間短且有較高的精確度。實驗方面，分別探討指向物體不同，基本矩陣不同，解析度不同以及指向方向不同的情況下，所重建出來的三維場景及指向位置，並列出幾點可能影響指向精確度的因素，以供日後改進做參考。

Computer vision techniques for 3D scene reconstruction from 2D images have been studied for decades. These technologies can be used to establish pointer systems to obtain the position users point at through 3D reconstruction. Most recent studies on pointer systems calculate the pointing direction and position by analyzing the images of gestures, which is not very accurate. In this research, we develop a computer vision system which carries out the 3D reconstruction of the feature points in the scene and calculates the fundamental matrix. The obtained fundamental matrix is then used to perform the reconstruction. The advantage of this method is that only one calculation of the fundamental matrix is needed and with high accuracy. Discussions of the accuracy of our 3D scene reconstruction and the obtained target position using different pointers, fundamental matrix, image resolution and pointing direction are given. Several possible methods that can improve the accuracy of our pointer system are also listed in this thesis.