An Online Learning Method for Shadow Detection

Shadow detection is a critical issue for most applications of video surveillance. In this study, we present an object-wise online learning method to detect casting shadows without providing any priori scene information or threshold parameters. Hue, saturation, and intensity- difference histograms of moving objects are collected to learn a cumulative distribution separately. The accumulating strategy strengthens the impact of shadow parts but reduces the effects of non-shadow parts. In each cumulative distribution, the most significant peak is then fit as a Gaussian function by using a robust estimation method. The fitted Gaussian is treated as a shadow likelihood function. The integration of shadow likelihoods in hue, saturation and intensity are modeled as a data term into a Graph Cut model. The Graph Cut model also incorporates edge information of the current image as a spatial smoothing term. Therefore, the shadow pixels can thus be labeled by minimizing an energy function. Compared to a supervised thresholding method, experimental results reveal that the flexibility and adaptability of the proposed learning method on real surveillance scenarios.