Model based techniques for image registration and three-dimensional scene analysis from image sequences

In order to achieve accurate registration of images of a scene taken from different views, and to obtain accurate estimates of three-dimensional scene structures, it is necessary to obtain accurate displacement fields. However, the accurate estimation of displacement fields requires the use of a-priori constraints on the structure of the displacement field. The paper focuses on techniques that integrate the two problems into a single estimation process, wherein camera motion and projection models are used to constrain the displacement field estimation process, as well as the recovery of three-dimensional structure. A family of progressively complex image displacement models including the 3D affine camera motion and full perspective projection model, and its various approximations and special cases is described. A family of mutually similar techniques suitable for the different models is also described. The discussion emphasizes practical issues with respect to the applicability of the different models and the accuracy and stability of estimation. Results of applying these techniques to different real image scenes is shown