利用立體影像和形隨影的複合方法作三維物體的辨識

Abstract

本篇論文提出一個利用影像系統辨識三維物體的複合方法。由於影像系統具有類似人類'看'的能力，一直是作為辨識的主要工具。而三維物體的辨識無論在電腦視覺或機器人學上都是很重要的研究課題。因此應用影像來作三維物體的辨識很早就開始發展。但一般傳統辨識方法主要有兩項缺點。(1)每種方法都只能辨識某些特定的三維特徵(3-D fature)，對物體的辨識不夠完整。(2)大多數的方法都是被動的(passive)，只能處理給定的影像。若所給的影像不夠就無法對物體作完整的辨識。 我們提出的複合方法就是針對這兩個缺點加以改進。首先我們結合立體影像(stereo imaging)和形隨影(shape-from-shading)這兩種辨識方法。由於立體影像和形隨影的資料可以互相結合我們可以得到相當完整的辨識資料包括物體的位置大小和表面結構。但若單從某個觀察方向(viewing drection)去辨識三維物體是無法窺得物體全貌的。為了達到完整的辨識我們將攝影機架在機械手上並提出一套三維辨識的程序主動地去取得足夠的影像來作辨識。然後整合不同觀察方向的辨識資料就可完整地將三維物體辨識出來。

A hybrid method for three-dimensional (3-D) object recognition is proposed ill this thesis. There are two drawbacks in the traditional recognition methods. (1) All methods extract only part of 3-D features. No method can extract all the features. (2) The traditional recognition methods are passive, they extract information from given images. If there are not enough images, complete information can not be obtained To overcome these two problems, the hybrid method we proposed contains two operations. The first operation is to extract the most useful 3-D features (positions, sizes and surface structures) of objects by integrating the information from stereo imaging and from shape-from-shading. The second operation is the active recognition procedure which will obtain complete 3-D information by moving the camera around the objects to obtain sufficient images. The verification of the hybrid method is illustrated by the results of two simulations and two experiments.