紅外線影像中之前景物偵測

Abstract

在本篇論文中，我們針對紅外線影像，提出一個新的前景物偵測方法。此方法推廣傳統的高斯混合模型，加入了位置變數以對於整張影像而非每一像素，建立數個「區域高斯模型」，因此所建立的高斯模型數量會比傳統的高斯混合模型少很多。建立起始的背景模型後，之後影像中每一像素是以一個5 5的鄰近區塊，來對前一張影像做區域高斯模型的比對，再以最符合的區域高斯模型做更新。實驗結果會看到在攝影機移動不大的情況下，利用本論文之方法在區分紅外線影像的前景跟背景確實能獲得較佳的結果。

In this thesis, we propose a novel method of foreground object detection for infrared images. We generalize the Gaussian Mixture Model (GMM) to construct a new Regional Gaussian Mixture Model (RGMM), by adding two random variables of image coordinates. Since the models are built for the whole image, not for every image pixel, the number of RGMM is much smaller than that of GMM for common videos. After an initial background construction, the RGMMs are updated by examining the existence of previous RGMMs in a 5 5 neighborhood for each image pixel, followed by the identification of the best-fit model which is then used in the update process. Experimental results show that better separation of foreground object from background can be achieved by using RGMM for infrared images obtained by a camera with small movements.