臺灣博碩士論文加值系統

在此篇論文中，我們提出一個方法，能夠正確的辨識受遮罩的目標物。為
了能夠把目標物從背景中抽取出來，我們使用二張序列影像來達到這個目
的。所謂二張序列影像乃是指包含相同背景和相同移動目標物的影像。首
先，我們做二張影像相減的動作以得到絕對差（absolute difference）
的影像，經由臨界值化（thresholding）後，我們得到合併後目標物的二
元影像（binary image）。由於我們需要的是一個完整合併後的目標物輪
廓，這樣才有利於特徵的抽取，因此，我們把滾動膨脹運算法（roll
dilation operation）用在合併後目標物的外部邊緣上，即可獲得所要的
輪廓。我們再將輪廓分成六個輪廓區段並且從這些區段抽取出特徵來，我
們選用一種不受物體位置和旋轉的區域化特徵表示法作為我們的特徵向量
。然後再用特徵比對的方法來辨識這些輪廓區段是屬於哪一類。我們用了
不同的方法來切割輪廓，這樣有助於我們使用投票法則來提高辨識效果。
經由實驗的結果顯示我們的方法可獲得不錯的效果。

Inhis thesis, we propose modified approach for occluded object
recognition.In order to extract the target from the background,
we use two image sequences that include the same background and
the moving target to achieve the objective. First, the two image
sequences are chosen to form an absolute difference image. This
provides the region of the moving target.After thresholding, we
obtain the binary image of the target. We use roll dilation
method to obtain the closed boundary of the combined target
which helps us to extract features from the target easily.To
recognize a partial occluded target, we divide the boundary into
six segments and extract features from them. We use a feature
matching method to classify the boundary segments. We also
divide the boundary of the combined target by different several
ways. These results help us to obtain a good recognition result
by using a voting rule. From recognition results, we find that
our proposed method can obtain a good performance for occluded
target recognition.