標題: 動態攝影機下的前景背景分割
Separation of Foreground and Background for Dynamic Cameras
作者: 邱永昌
Chiu,Yung-Chang
莊仁輝
Chuang, Jen-Hui
資訊科學與工程研究所
關鍵字: 動態式攝影機;前景背景分割;Dynamic camera;foreground and background separation
公開日期: 2008
摘要: 在本篇論文中,我們提出一個方法,能在動態式攝影機拍攝影像中偵測前景物體。本方法一開始會需要一張或數張單純只有背景物的影像以建立背景高斯混合模型,之後再利用此高斯混合模型所提供的背景影像建立以顏色為基礎的像素群組、以邊線(edge)所切割出的區塊以及用角點偵測器所獲得之角點。當有新進的影像須要做前景物偵測時,以同樣方式對該新進影像建立對應的群組、區塊以及角點,接著藉由配對上述三種背景影像與新進影像的對應特徵,便可以估算兩張影像間的移動向量。接著根據此移動向量,將背景高斯混合模型移動到正確的位置,意即我們對齊背景高斯混合模型與新進影像,然後才進行前景背景的分割,之後再進一步地消除雜訊以及攝影機些微旋轉的影響,以得到最後的前背景分割結果。從實驗結果中可以看到,在背景景深差異不是太大的前提下,即使兩張影像間的移動量並不是十分的小,我們的方法仍可以成功地切割出前景物。
In this thesis, we propose an algorithm to detect foreground objects in an image sequence obtained with dynamic cameras. The algorithm needs one or several initial images containing pure background environment to establish its Gaussian mixture model. Then, we establish color-based pixel clusters, edge-secluded image blocks, and corners of image regions from the background image supported by such a Gaussian mixture model. In order to detect moving objects inside an upcoming input image, the above three types of image features are extracted. By matching these features between the background image and the input image, the motion vector between the background image and the input image can be estimated. Such a process corresponds to an alignment between the Gaussian mixture model and the input image, and the moving objects can be detected using aligned Gaussian mixture model. After removing some noise and partly eliminating the effect due to small camera rotation, the final foreground separation result can be obtained. According to the experiment results, even when the displacement between two consecutive images is not very small, the proposed algorithm can still detect moving objects if the background has a small range of depth.
URI: http://140.113.39.130/cdrfb3/record/nctu/#GT079655505
http://hdl.handle.net/11536/43307
Appears in Collections:Thesis


Files in This Item:

  1. 550501.pdf

If it is a zip file, please download the file and unzip it, then open index.html in a browser to view the full text content.