Abstract
電腦視覺系統已廣泛應用於工業檢驗，但由於一般視覺系統所取得的是平面式的影像，故多只能作二度空間的影像分析。立體視覺系統以兩台照相機同時對物體取得影像，藉由檢視物體上特徵在兩張影像中的位置偏移量，便可利用三角函數求得此特徵的三度空間座標，故立體視覺系統最重要的工作即解決特徵的對應問題。本研究之目的即發展一套以曲線為特徵，應用模糊鬆弛理論之立體視覺對應演算法 ( Curve-based MatchingAlgorithm by Relaxation, CMAR )，運用一些常見的影像處理方法，自影像中擷取出曲線，每條曲線均賦與定義之屬性，經由衡量屬性的相似程度可以找出可能的對應曲線。但此時仍存在許多模稜兩可之對應情況，所以可應用模糊鬆弛理論來解決曲線的對應問題，藉著比較各影像中曲線間的相對位置關係(包括相對距離與相對方位)，可決定正確的曲線對應，如此可結合屬性相似程度及相對位置關係兩項考量因素之優點。又本研究改進模糊鬆弛的程序，故可以較少的反覆次數仍得到正確的曲線對應。大部份的特徵對應演算法只能解決特徵之對應問題， CMAR 法利用對應兩影像之掃瞄線可達到特徵中細部點對應的效果。將 CMAR 法驗証於簡單的幾何物體、複雜形狀物體及不完整物體影像，其對應結果良好且計算快速，相較其它類似對應方法，CMAR可得到較佳之結果。論文內容第一章為緒論，第二章為文獻探討，第三章詳述 CMAR 的方法，第四章則是系統構建與實驗分析，第五章是結論。Computer Vision has been widely applied in automatedinspection, but most vision techniques can only handle withplanar objects. Stereo vision systems provide the capabilityof determining 3D coordinates of objects by examining thedisparity existed between a pair of images. For stereovision systems, the correspondences (matching) problem hasto be resolved before employing the triangulationprinciple.Most segment-based stereo techniques just match the definedfeatures, but the point correspondences problemremains unresolved. In this research, a curve-basedmatching algorithm by relaxation (CMAR) has been developed.Curve segments extracted from two images arecharacterized by some pre-defined attributes. By comparingthe similarity measures of attributes, possible matchescan be identified. However, ambiguities induced bynoises may influence the matching results. Therefore,fuzzy relaxation method is applied to match curves byverifying the geometrical relationships of curves (bothrelative distances and orientations). In CMAR, therelaxation process is improved so that the number ofprocessing iterations required will be reduced. Furthermore,the information acquired from curves matching is extended forepipolar lines matching. Thus, a detailed matchingcondition (point to point) of two images can be obtained.