本論文提供一個使用攝影機來辨識人體姿態的方法。我們結合了只使用單張影像的靜態辨識以及連續影像的動態資訊,並使用兩台不同角度的攝影機,來提高辨識的正確性。 在靜態辨識中,我們使用人的剪影(Silhouette)來做辨識。對於一組觀察到的影像,我們使用shape contexts來描述人的剪影,並從預先製作好的資料庫中比對出最接近的樣版。為了解決不同姿勢可能產生相同外觀,造成難以區別正確姿勢這種問題,我們使用上一個時間到目前時間影像上的變化來取得動態的資訊,並利用這個資訊來辨識目前人體的姿態。為了節省運算時間,我們只對靜態辨識的結果中是候選答案的樣版做動態資訊的辨識。 最後我們使用加權和(Weighted-sum)的方式來整合靜態與動態辨識的結果以獲得更準確的辨識結果。我們使用了人工合成及真實的影像資訊來做實驗,以驗證我們所提出的方法是迅速且有效的。
In this thesis, we describe a method which combines static recognition based on single image with motion information to recover human poses from video sequences with two cameras. Silhouettes extracted from two images are used for static recognition. For a pair of observed images, we match the obtained silhouettes with the examples generated from motion capture data using robust shape descriptors to find the solution candidates. In order to solve ambiguity problem which means that different poses may result in the same silhouette and save the computation complexity, the motion information is then introduced to evaluate the motion distance of solution candidates only. We use the intensity change as motion information, and thus error-prediction problem can be successfully overcome. Finally, a weighted-sum method is used for the combination of the results obtained with respect to two cameras. Several video sequences are used to validate the effectiveness of our proposed approach.