基於立體視覺之即時指向系統之誤差分析

Abstract

在本論文中，我們基於立體視覺，利用兩台攝影機實作一套即時指向系統。此系統避開以往複雜的三維重建方式，改以影像分析技術，平面投影轉換及三維幾何等數學理論，作為系統實作的基礎。而與傳統三維重建方法一樣，本系統亦存有誤差的問題，故本論文在介紹系統運作流程後，先檢視本系統的所有誤差來源，了解所擷取影像特徵的誤差範圍。而後，再由已知影像上的誤差範圍，探討影像誤差與最終重建誤差之間的關係，也探討了在最大影像誤差的情況下，重建點的誤差上限會如何分佈。最後則介紹我們嘗試在系統中降低誤差的方法，並探討這些方法對靜態與動態指向點重建精確與穩定度之改進。由實驗結果可知，我們將能根據誤差分析之結果，來設計使系統能精確且穩定運作的使用者介面。

In this thesis, we use two cameras to implement a real-time vision-based pointing system. Without complex 3D reconstruction processes, the implementation of the pointing system is based on image analysis, homography, and 3D geometry. Similar to 3D reconstruction approaches, the reconstruction errors also exist in the proposed pointing system. Based on the procedure of the system, we first study the error sources, and the range of extraction errors of image features. Then, we investigate the relation between image errors and the ultimate reconstruction errors, and try to find the worst-case error range in the latter with respect to a fixed magnitude of the former. Finally, we develop some methods to reduce the error range for both static and dynamic pointing actions. Based on the derived characteristic of error range of reconstruction, as observable from the experimental results, pointing systems with more accurate and stable user interface can be developed.