行人切割在許多視覺影像處理的問題中,扮演著相當重要的腳色,也可以做許多的應用,例如:行為分析、肢體切割、行人計數器…等。在連串影像中,可以依據前後的影像差異的資訊套用不同的高斯混合模組將特定的行人切割出來,然而在動態攝影之下,這樣的方法無法順利得到完整的行人肢體,因此,本篇論文將針對單一影像進行行人切割並且不需事前建立行人模組的資料庫。 本系統在區塊的基礎下進行切割處理,將一張影像先做兩種低階切割處理-三角化及分水嶺法,使整張影像切成許多的小區塊,再將兩個低階切割法的結果做一次合併重切割,以得到更精細的區塊結果。再依據每個小區塊的顏色與相關位置的資訊,分別套入不同的核心強度估算器中去計算每個區塊屬於前景或是背景的機率。最後,根據比對後的結果,重新更新每個區塊屬於前景或背景的身分。反覆執行數次以得到最後的結果。 由實驗結果得知,我們的系統可以適用於許多不同條件底下,並且有很好的表現。
This paper proposes a novel approach for pedestrian segmentation in region-based procedure based on deformable triangulation. The traditional algorithm to segment pedestrian from a video is using Gaussian Mixture Models. To extract precise body of a pedestrian from a query image, KDE-EM is combined with a shape template-based detection for specific object segmentation. This thesis proposes kernel density estimation which applies individual kernel to estimate the probabilities of each region belonging to foreground and background. According to information of color and position after doing pedestrian detection, apply color histogram to determine its similarity in color kernel, for spatial kernel, it is directly represented by the geometric distance. After estimating each region similarity to foreground and background, redefine its class label and do with some iteration to get the best result. Experimental results reveal the performances in several different conditions.