An Unsupervised Shadow Detection Algorithm

Abstract

Shadow detection is a critical issue for most applications of video surveillance. In this study, we proposed an unsupervised method to detect shadow caused by moving objects without setting any priori parameters or threshold values. The color histograms of moving objects’ image patch are collected to learn an accumulated distribution in each color domain. The accumulating strategy strengthens the impact of shadow parts in the image patch but reduces the effects of other non-shadow parts. By taking the advantage of robust estimation, the most significant peak, i.e. the shadows, in the accumulated distribution is fitted by a Gaussian function. As a result, if a pixel’s feature value is within the 2σ range of the Gaussian function,, it can thus be determined as shadows.