Abstract
We present a direct closed-form solution for estimating 3D depth and translational motion of a camera from image sequences where the scene illumination can vary from one frame to the next. The method is based on a dynamic image model that we proposed previously for the computation of optical flow. We use only selected moments of the image and its spatio-temporal derivatives in small image regions to calculate the solution. The direct solution for selected experiments with real images are given, and compared with that from the computed optical flow, to show that the former typically gives better estimates than the flow-based approach even though both techniques are based on the same dynamic image model. In particular, the direct solution is more robust with respect to motion sizes of more than one pixel per frame.