單一視角下以模組為基之三度空間工件的辨識

Model Based 3D Parts Recognition from a Single Image

簡正元(Cheng-Yuan Chien)

指導教授 ： 蔡篤銘

元智大學/工程學院/工業工程與管理學系/碩士(1993年)

摘要

本論文主要的目的是利用單一鏡頭及單一視角的機器視覺技術來辨識三度
空間的工件。之所以採用單一鏡頭的目的乃是為了節省設備成本及降低複
雜的座標轉換﹐加快辨識的速度。論文中所探討的工件乃是經由人工或機
器所製造出來的工件﹐其具有一般的幾何特性。首先我們假設已由使用圖
案投影法的方法來獲得工件的邊。並利用標記法將三度空間待測工件的各
個可見面由影像中萃取出來﹐並產生一個各面間毗鄰關係的毗鄰矩陣。由
於使用單一鏡頭及單一視角的關係﹐當工件表面正交向量和觀測物件的觀
測軸無法平行時﹐此時影像將產生扭曲﹐我們利用一種對投影不變的圖形
正規化方法對各個萃取出來的面做正規化﹐使其成為緊密的圖形而與視角
無關。之後我們對這些待測面與標準資料庫中的標準面做比對﹐並利用各
面間毗鄰的關係做全域性的限制﹐比對後將產生一個決策權重累加值﹐最
後根據決策權重累加值來做決策﹐判別該測試工件是屬於標準物資料庫中
的那一個工件。

關鍵字

三度空間工件辨識 ； 單一視角 ； 圖形正規化 ； 毗鄰矩陣 ； 決策權重累加值

並列摘要

In this research, a model-based 3D parts recognition from a
single view is presented. We use a single CCD camera at a
single view to recognize a 3D part from its corresponding 2D
image. The use of single camera and single view reduces the
cost of equipment and the complexity of the coordinate
transformation. The objects investigated in this research are
manmade workparts and have general geometric characteristics
with separatable surfaces. For each part under study,we assume
that the edges of part is obtained by the structured lighting
method. A labeling is followed to extract the visible surfaces
from image and genera- te an adjacent matrix that defines the
adjacency between surfa- ces. Under the single camera and
single view, the extracted surfaces may be skewed if the
normals of part surfaces are not parallel to the optical-axis
of the camera. Therefore,we employ a skew normalization
procedure to transform each part surface into its most compact
shape such that the surface becomes view-point invariant. This
model database is used to recognize the part in am image scene.
The match procedure starts by locally matching each visible
surface with the surfaces of a model part in the database. It
follows by a global match that verifies the correctness of
surface adjacency from the adjacent matrix. Finally, a weighted
accumulated value is generated. The model part with the
hightest score of accumulated value indica- te the evidence of
the scene object under test.

並列關鍵字

3D parts recognition ； Single view ； Skew normalization ； Adjacent