利用背景資訊及型態運算來快速自動分割視訊物件

Abstract

在視訊中擷取正在移動的物件，許多方法都針對物件本身來做處理，也就是前景的部份，這些方法通常綜合了空間上的分割法﹝如Canny邊緣判定法，分水嶺法等﹞和時間上的分割法﹝如運動預估法，改變偵測法等﹞。在這篇論文中，我們以改變偵測法為基礎，不同的是，我們不只以連續兩個視框的差異性做為分割時的參考，且利用了背景資訊的更新，來產生比前景可靠的背影記錄，再從背影記錄與視框的差異性獲得移動物件，最後應用型態學上的運算，消除camera產生的雜訊並使得物件更加平滑完整。實驗結果顯示我們提出的方法確實能自動且快速產生品質較佳的視訊物件。

The main subject is to extract out the moving object in video. There are many methods focus the object itself to process, the foreground part. These methods often are combined with spatial segmentation (Canny edge detection, the Watershed algorithm) and temporal segmentation (Motion estimation, change detection). In this thesis, our approach is based on change detection; unlike traditional change detection approach, we not only utilize the frame difference to be our reference resources of segmentation, but we use the up-to-date background information to generate a background record which is more reliable than the foreground one. Then, we use the difference between the background record and frames to obtain the initial moving object. At least, we use the morphological operations to remove noise from camera and get a new object with more smooth and complement. The experiment results show our approach indeed can make the VOPs with good quality fast and automatically.