抄録
This paper proposes a method for estimating three-dimensional (3D) motion parameters of a rigid object, of which surface is covered with a texture, by three images from a single camera. The method estimates unit normal vectors of texels and the 3D shape of the rigid object in each image by using a Shape-from-Texture technique. The rotation parameters are obtained from average vectors of the normal vectors at the three images. The translation parameters are determined as the displacement between the centroids of successive shapes. Based on the 3D motion parameters, a correspondence between texels in successive images is established. The correspondence can provide a more exact shape of the object through stereo triangulation. The estimation of exact shape and 3D motion parameters should prove to be useful technique for the content-based image retrieval.
