利用結構光源對基本物件作三維位置及外形參數測量的計算機視覺

Abstract

本論文提出一利用主動感測式計算機視覺系統，對基本物件作三維位置測 量及其外形參數計算的方法。對於人造物體依統計大約有８５％可利用平 面、圓柱體、圓錐體、及球體等基本物件組合而成，因此對於基本物件位 置及外形參數的測量，是自動度量三維物體幾何構形中一個重要的基礎。 我們將在論文中提出演算法，對上述基本物件計算其三維位置以及它們的 外形參數。本系統利用主動式結構光源的格子編碼技術將兩組格子線投射 至待測物體表面，利用其在物體表面形成的條紋之變形，由變形前後條紋 圖形之間的某種三度幾何關係去推導出上述所欲獲得的資訊。一個重要的 步驟是我們將一個對複雜的曲線形條紋的計算問題，轉換成一個對直線形 條紋的計算問題，這使得測量計算上得以簡化及有效。它主要的方法是在 曲面形狀的物體上，利用一些幾何特性開發出一個虛構的平面，並推導出 此一平面應該存在的直線條紋圖形及其位置。由此我們得以去計算出各類 基本物件的三維位置及外形參數。同時計算上較其它方法更為簡單及正確 。為使本方法能實用化，我們必須考慮影像中含有雜訊及計算的敏感度的 問題。我們開發出多種的規律去規範計算中的每一步驟，而可以降低因影 像處理所引入的雜訊或計算誤差，這可以增加計算結果的精確度及強健性 。大量實驗結果證明，由單張經格子編碼的影像，利用虛構平面及直線形 條紋去計算基本幾何物件的三維位置及外形參數，十分的精確及快速。在 靈敏度分析中，我們也藉不同的幾何模型及雜訊分佈，對計算結果作各項 的分析，也同時證明我們的方法是有效及強健的。 In this dissertation, the algorithms for estimating the 3-D location and shape parameters of primitive objects using a structured light vision system are proposed. Complex surface/ volume structures can be constructed as hierarchies of simpler components. It has also been observed that 85% of man-made objects can be perfectly represented or can be well approximated by a small number of primitive patches, namely, planar, cylindrical, conic, and spherical patches. Hence, the determination of 3-D location and shape parameters of primitive objects is of fundamental imprtance in 3-D object reconstruction. In our vision system an expanded laser beam passes through a code plane marked with two sets of parallel and equally spaced lines that are perpendicular to each other and the resultant grid light impinges on an object surface to create a spatial-encoded image for analysis. The projected curved stripe pattern appearing on the object surface is related to the position and structure of the imaged object surface. A crucial step in our method is to convert the estimation problem with the complex curved stripe patterns to an equivalent, but simpler estimation problem with linear stripe patterns. Based on the obserable or constructible geometric entities that are formed on the object surface, we can estimate the location and shape parameters of primitive objects. Using the various physical constraints or properties derived from the structured light geometry, the object surface geometry and image formation principle, we can improve or refine the estimation result significantly.