Motion estimation using optical flow field

Over the last decade, many low-level vision algorithms have been devised for extracting depth from intensity images. Most of them are based on motion of the rigid observer. Translation and rotation are constants with respect to space coordinates. When multi-objects move and/or the objects change shape, the algorithms cannot be used. In this dissertation, we develop a new robust framework for the determination of dense 3-D position and motion fields from a stereo image sequence. The framework is based on unified optical flow field (UOFF). In the UOFF approach, a four frame mode is used to compute six dense 3-D position and velocity fields. Their accuracy depends on the accuracy of optical flow field computation. The approach can estimate rigid and/or nonrigid motion as well as observer and/or object(s) motion. Here, a novel approach to optical flow field computation is developed. The approach is named as correlation-feedback approach. It has three different features from any other existing approaches. They are feedback, rubber window, and special refinement. With those three features, error is reduced, boundary is conserved, subpixel estimation accuracy is increased, and the system is robust. Convergence of the algorithm is proved in general. Since the UOFF is based on each pixel, it is sensitive to noise or uncertainty at each pixel. In order to improve its performance, we applied two Kalman filters. Our analysis indicates that different image areas need different convergence rates, for instance. the areas along boundaries have faster convergence rate than an interior area. The first Kalman filter is developed to conserve moving boundary in optical How determination by applying needed nonhomogeneous iterations. The second Kalman filter is devised to compute 3-D motion and structure based on a stereo image sequence. Since multi-object motion is allowed, newly visible areas may be exposed in images. How to detect and handle the newly visible areas is addressed. The system and measurement noise covariance matrices, Q and R, in the two Kalman filters are analyzed in detail. Numerous experiments demonstrate the efficiency of our approach