Segmentation, matching and estimation of structure and motion of textured piecewise planar surfaces

We present an algorithm which segments and matches the regions and then estimates non-iteratively 3D motion and structure of a moving piecewise planar textured surface from two perspective views. The algorithm has two major steps. The first step is coarse. Here, the local planar nature of the surface is used to obtain polynomial expressions for image plane displacements of features. Using regions as moving features, the image is segmented using Hough transform such that the regions in each segment have the same polynomial coefficients. The values of these coefficients and region properties (e.g., area) are then used to identify region correspondences. In the second step, for each planar surface, the region correspondences are used to compute the corresponding motion parameters and surface orientation in closed form. The second step uses a finer model of motion than the first step. Experimental results are presented for three image pairs.