Method for modeling cast shadows in videos

A method for modeling a scene, comprising the steps of:

• acquiring a video of a scene with a camera, and for each frame of the video:
• in a computer:
• updating, for each pixel, a set of background models;
• representing the pixels in each background model and each shadow model with a plurality of layers, each layer including a plurality of Gaussian distributions, each Gaussian distribution including a mean and a covariance, the covariance being an inverse Wishart distribution, and the updating is according to a recursive Bayesian estimation process;
• updating, for each pixel, a set of shadow models;
• updating, for each background color, a set of shadow flow vectors from the set of shadow models;
• determining, for each pixel, a confidence score for each model;
• comparing, for each pixel, a most confident shadow model with a most confident background model to determine a disparity vector;
• projecting the disparity vectors into a quantized color space including bins, each bin representing a color value in a background of the scene;
• and
• aggregating, for each bin, the projected disparity vectors to locate the shadow flow vector in each bin;
• and
• updating, for each pixel, a background shadow map using the set of shadow flow vectors to detect shadows in the video.