基於動作偵測之戶外靜態場景影片分層研究

Abstract

大多數的video matting 技術皆為image matting技術的延伸，由於缺乏時域變化的資訊，在面對前景與背景顏色極度相像的影片時，並無法利用現有的顏色資訊將前景分離出來。 本篇論文中提出一套可用於戶外複雜背景的video matting系統，利用偵測物體移動來幫助分離前景及背景。針對兩個相鄰的影像以及前景-影像-背景的立體架構，分別建立matting Laplacian，並利用這些像素間的關係，計算出畫面中正在移動的區域。最後將不同影像的結果互相傳遞，達到時域變化一致的前景。實驗結果顯示我們的系統可應用於戶外光影多變的場景，並且只需要極少的使用者提示，即可得到相當正確的前景。

Most of existing video matting systems utilize the extensions of image matting techniques and do not consider much about the use of temporal information. They often experience difficulty in videos which exhibit foreground and background color ambiguity. In this thesis, a motion aware video matting system that can be applied to outdoor stationary background videos is proposed. Based on the matting Laplacian matrices constructed between two adjacent frames and the affinity established over a 3-layered, foreground-current frame-background, the system generates a smooth motion index with accurate boundaries and uses it as a feature in addition to color to resolve color ambiguity problems. Then the system propagates alpha results between frames to generate a temporally more coherent result. The proposed approach utilizes temporal information, in addition to colors, in deriving stable motion indices and insights into variations in object-background boundaries among successive frames in the implementation of video matting system, which is thus able to handle severe kinds of color ambiguity in an outdoor scene based on only a few user scribbled key-frames.