Please use this identifier to cite or link to this item:
標題: Kernelized Orthogonal Subspace Projection Techniques for Target Detection
作者: 陶金旭
關鍵字: 目標偵測;target detection;電信工程;紋理特徵抽取;目標定位;應用研究;texture feature extraction;target localization
To cover larger surveillance area, the unmanned air vehicle (UAV) can be used to monitor the military objects (such as warships on the sea). The goal of this project is to develop a system to automatically detect the desired objects in a wide range of area so that the payload operator can locate and track the desired objects efficiently. Due to the fact that the characteristics of the images from the payload (CCD camera) are often influenced a lot by the environments, the automatic target detection is a complicated task. In the machine learning research area, using kernel functions to construct a nonlinear version of the classifier has become popular these years. In this project, the orthogonal subspace projection techniques, the idea of signature subspace classifier, and the kernelization-based learning algorithm will be combined to develop Kernelized SSC (KSSC). To avoid the computation in the high dimensional space, the computation of the classifier will be done in terms of the kernel function. The color and texture feature together with KSSC will be adopted to analyze the images captured from the UAV so that objects in the image can be detected automatically and the payload operator can locate and track the desired objects efficiently.

在執行大範圍的軍事目標偵測及監控等任務(如監控海上船艦等)時,我們可以使用無人飛行載具(UAV)在高空飛行,以涵蓋較大區域。本計畫的目的在於如何於大範圍的區域內,快速自動偵測出該區域內的目標,使酬載操控人員能迅速選擇有興趣的目標進行定位或追蹤。由於酬載影像(CCD影像)受外在影像擷取環境影響,所拍攝影像的特性會有很大的差異,因此「自動目標偵測」是一困難且值得研究的主題。在機器學習的領域中,利用核函數(kernel function)方法來建構線性演算法的非線性版本,是近年來的趨勢。本計畫將結合正交次空間投影技術中,特徵子空間分類器(SSC)的概念,以核化(kernelization)為基礎的學習理論,來定義Kernelized SSC (KSSC)。為了避免在高維度特徵空間作計算,分類器將以核心函數的形式來進行核化。同時配合適當的彩色及紋理特徵來分析無人飛行載具所擷取的影像、快速自動偵測出影像區域內的目標,使酬載操控人員能迅速於大範圍區域內對目標進行定位或追蹤。
其他識別: NSC96-2623-7005-003-D
Appears in Collections:電機工程學系所

Show full item record

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.