本篇論文提出一種新穎的車輛偵測方法。對於彩色影像中之車輛，利用色彩與邊緣資訊加以偵測並確認。大多數過去的方法皆採用移動資訊(Motion features)，即假設車輛為影像中之移動物，但此法對於靜止之車輛完全失效。本文提出一種新穎的色彩空間轉換方法，如同人臉偵測時會先尋找膚色區域一般，快速地找出影像中屬於車輛顏色之像素點。由於車輛具有各種不同之顏色，同時戶外環境伴隨著季節與天候不同，複雜的光線因素導致極少的論文採用色彩資訊作車輛偵測。而本文所提之色彩空間轉換方式則為一強有力之工具，足以在不同的光線條件下區分車輛與背景之像素點。 在找出可能屬於車輛之像素點後，本文結合三種有效之特徵，分別為角點（corners）、邊緣資訊(edge maps)與小波轉換之係數，用以建構一連串且多重維度(multi-channel)之車輛分類器。此分類器對輸入影像中可能之車輛像素點作有效之確認。由於先前已利用色彩資訊濾除大量無關的背景像素點，故此確認步驟將可快速且有效的執行。實驗結果證明結合整體色彩資訊與局部邊緣資訊之車輛偵測方式是強而有效的。平均偵測率達到94.5％ In this thesis, a novel approach for detecting vehicles using color and edge information from static images is presented. Different from traditional methods which use motion features to detect vehicles, the proposed method introduces a new color transform model to find important “vehicle color” for the quick finding of possible vehicle candidates. Since vehicles have various colors under different weather and lighting conditions, seldom works were proposed for the detection of vehicles using colors. The proposed new color transform model has extremely excellent capabilities in identifying vehicle pixels from background ones even though the pixels are under varying illuminations. After finding possible vehicle candidates, three important features including corners, edge maps, and coefficients of wavelet transform are used for constructing a cascade and multi-channel classifier. According to this classifier, an effective scanning is performed to verify all possible candidates. The scanning can be quickly achieved because most background pixels are eliminated by the color feature. Experimental results show that the integration of global color feature and local edge feature is powerful in the detection of vehicles. The average accuracy rate of vehicle detection is 94.5%.
