適用於靜態背景視訊並以型態學處理物件邊緣之即時視訊切割

Abstract

在本論文中，我們設計並實現一個在個人電腦上的視訊影像切割系統。此系統可以應用於靜態背景的視訊電話及視訊會議。 此系統的基本概念是利用物件的邊緣銳化來精準得到移動的物件。首先，我們使用一個兩級的雜訊估計方法來估計攝影機的雜訊，並且把此結果拿來當做往後參數調整的參考。為了取得一個較佳的物件遮罩，我們觀察六張連續畫面的變化情形來取得一個初步的前景以及利用連續兩張畫面的差異選擇一適合的臨界值將移動的物件萃取出來。接著，我們利用Canny edge detector偵測出整張畫面的邊界資訊，利用影像變動偵測所得之遮罩及前一張畫面所求得之物件遮罩可以將屬於背景的邊界移除，然後使用收縮的技巧來取得一個粗略的物件遮罩，並利用此資訊來取得移動物件最外圍之邊界資訊。為了得到更精確的物件遮罩，我們使用Dijkstra所提之最短路徑搜尋演算法，將物件的邊緣連接並利用此封閉的物件輪廓將移動的物件給萃取出來。最後的模擬結果顯示我們所提的方法，切割出來的物體邊界是相當精確的，且在經過我們的加速之後，整個系統能快速且精準的切割出移動的物件。 在1.733-GHz CPU 及1024-MB RAM的個人電腦上且攝影機不移動時，目前的執行速度是每秒約四張CIF Frame及每秒約二十張QCIF Frame，在格式影片的應用上，我們有提出一簡化的方法，則每秒約十二張。

In the thesis, we consider the design and implementation of video segmentation on a personal computer. The system can be applied on video conference and videophone with stationary background. The basic idea of the system is a graph-based edge linking technique. At first, we adopt a two staged method for camera noise estimation and those thresholds are adjusted based on the estimated camera noise. To get the change detection mask, we consider six consecutive frames as time of observation to estimate the background and select a suitable threshold to estimate the foreground by two consecutive frames. Next, we use Canny edge detector to get the edge information of entire frame. We use the change detection mask and the moving object mask of previous frame to remove the edges of background. Then, we shrink the object mask to edge map. To refine the object mask, we use Dijkstra’s shortest path search algorithm to link the boundary of moving object and extract the object by the closed contour. Simulation results show that our method can give correct segmentation results. After optimization, the proposed segmentation system can get the moving object accurately and quickly. With a Intel Pentium M 1.733 GHz CPU and 1024-MB RAM, the system can achieve 20 QCIF frames per second and 4 CIF frames per second. We also propose a simple method and it can achieve 12 CIF frames per second.