針對視覺監控的應用，陰影偵測扮演著關鍵且重要的角色，本研究提出一種即時學習的陰影
偵測技術，該技術之特點在於無需使用者提供監控場景資訊或設定任何閥值，而是藉由統計的技
巧來自我學習該場景之陰影特性，進而達到陰影偵測之效果。該法首先利用背景相減法去除靜態
背景，並找出前景影像中包含陰影的移動物體區塊，然後將該移動物體區塊從紅綠藍色彩空間轉
換到色調飽和度及亮度色彩空間，並產生該移動物體之色彩特徵直方圖。隨著影片的播放，將出
現於該場景之移動物體之特徵直方圖加以持續累計，藉以強化陰影特徵的統計意義並弱化其他非
陰影之特徵，並產生一累計直方圖。此時，利用強健估測法在此累計直方圖上估測一陰影相似度
函數，利用其計算各像素之陰影信心程度，最後利用圖形分割演算法，即可正確地偵測出前景移
動物體區塊中的陰影部份。

Shadow detection is a critical issue for the applications in video surveillance. In this study, we present an object-wise unsupervised learning method to detect casting shadows without any priori scene information or threshold parameters. Color feature histograms of moving objects are collected for the estimation of shadow likelihood function based on the accumulative distribution data. The accumulating strategy will highlight the shadow parts and reduce the effects of other non-shadow parts. The most significant peak in the cumulative distribution is then fitted by a Gaussian by using robust estimation. The fitted Gaussian is treated as the shadow likelihood functions. The shadow likelihoods as a data term and edge information as a smoothing term are integrated into a Graph Cut model. Compared with a supervised thresholding method,
experimental results reveal the flexibility and adaptability of the proposed learning method on real surveillance scenarios.