Three-dimensional camera motion estimation and foreground/background separation for stereoscopic image sequences
This work examines the problem of foreground/background extraction when stereoscopic image information is available. A two-stage algorithm is proposed, followed by postprocessing for the estimation of the global motion in the scene and the identification of moving objects. At the first stage, an estimate of the 2D global motion in the scene is formed which then serves as an initial estimate for the algorithm for a robust 3D global motion estimate. The 3D global motion estimation procedure is interleaved with a procedure for removal of the outlier vectors from the initial set of training vectors. Postprocessing of the foreground/background map produced results in an improved motion detection mask, which may then be used for 3D motion compensation.