由兩張影像計算相機主要運動參數

Abstract

在資訊科學領域中，影片的動態分析一直是個很重要的議題。其可協助影片分析與歸類，使得影片的瀏覽與檢索更為迅速。傳統上，運動分割的技術，多半建立在固定相機，或是已知相機參數的基礎上，找出相鄰影像的光流並將其分類，其缺點是無法實用於一般未知相機參數的多媒體影像上。近年來，基於對應點並搭配二維的運動模型之方法被廣泛的使用。但使用二維的運動模型處理三維的運動關係還是有其限制。針對此一問題，本論文主要是採用一三維的運動模型，提出改良後的霍氏轉換，解決三維運動模型參數估計上的困難，並搭配分支限界法解決傳統霍氏轉換速度上的問題。最後經由霍氏轉換估算影片中兩種以上的運動模型，以此做為影片前後景的分割，並且由背景的運動回朔相機的運動參數，借此幫助影片的分析與歸類。 改良後的霍氏轉換，不需在所有的參數空間取樣，藉由在第一台相機焦距以及旋轉上取樣，估算出第二台相機以及位移，藉此減少取樣之維度。並在旋轉的取樣搭配分支限界法，使我們只需取樣於原旋轉空間之子空間，提高取樣之效率。實驗結果顯示，本論文所提出的方法確實能有效的分辨一般大眾化影像之前後景，並且估算出的三維運動參數也具有一定的準確度，速度上也優於傳統的霍氏轉換。

Video motion analysis is a critical issue for multimedia signal processing. It can help video summarization and classification, which make the browsing and indexing more quickly. Clustering motions under fixed cameras or calibrated cameras has been developed recently. The drawback of these methods is it’s hard to realize on real video, which is usually un-calibrated. To deal with this problem, we propose a Hough transform based method which samples 3D motion parameters, including the focal length of the camera intrinsic matrix, the relative rotation, and translation. However, fully sampling all parameters is time consuming. We partially sample the focal length of the first camera and relative rotation. Then, we obtain the remaining parameters by solving a geometry optimization problem. The selection of multiple motion models is defined by a voting strategy. In addition, we combine a branch and bound algorithm with Hough transformation to solve high computation complexity of traditional Hough transform. The algorithm reduces sampling rotation space to its subspace by an efficient rotation space searching. The results of motion models are used to identify the dominant motion in videos.