視訊背景次像素平移量分析與分割

Abstract

在此篇論文中，我們提出了一種對於視訊背景次像素平移量的估計方法，此方法結合了影像空間以及頻率空間中的資訊。首先，我們利用影像空間的資訊作背景區域的比對，找出部分背景區域以及粗略估計背景平移量。接著利用行與列的一維傅立葉轉換，在頻率空間，利用影像轉換後的相位關係，分別分析背景區域的垂直及水平方向的次像素平移量。透過漸進式由粗而細的平移量分析，我們可以分割出背景區域。實驗結果顯示，利用上述演算法可以得到比像素為單位的平移量更為精準的次像素平移量。此外，本論文的作法也避免了利用頻率空間進行前景、背景分割時，無法處理前景物體形變的問題。

In this thesis, we propose an approach to sub-pixel background motion analysis in a video by both spatial and frequent information. First, rough translation of the background is obtained by partial background matching using image space information. Subsequently, more accurate sub-pixel motion analysis of background is carried out in one dimensional Fourier domain of image rows and columns using the phase information. Through the above coarse-to-fine procedure of motion analysis, major and significant background area in the video can be segmented Experimental results show that the sub-pixel motion analysis is indeed more accurate than pixel-based methods, while the problem of shape changing of foreground objects, often associated with segmentation in Fourier domain, is avoid.