立體視覺指向系統之準確性分析

Abstract

本論文中實作了一個即時的指向系統，不同於其他相關的研究中使用傳統三維重建的方法來找出指向點，我們沒有使用到相機的內部參數以及朝向等資訊，而是利用平面投影轉換以及一些三維空間內的平面關係作為系統實作的基礎，概念上較為直覺而且也相對地容易理解。而本系統與傳統使用三維重建的方法一樣都存在有誤差的問題，目前我們已經利用平滑過的指向點軌跡當作給予電腦的指令輸入，然而，如果能有更精確的系統，則我們將能研發出更多的應用。因此在建構出此指向系統後，我們亦針對系統的誤差進行分析，檢視誤差的來源並且使用適當的方法降低誤差。

In this thesis, we implement a real-time pointing system which, unlike ordinary 3D reconstruction approaches, uses homography and some known 3D reference planes to reconstruct the target point without using camera’s intrinsic parameters and orientation. The basic concept is quite intuitive and easy to realize. Despite the existence of reconstruction errors, like ordinary 3D reconstruction approaches, we have developed an application to use the smoothed trajectory of target point as a command to the computer. However, more applications can be considered if we could develop a more accurate system. To that end, we also try to identify the sources of the reconstruction errors and suggest possible ways to reduce them.