<title>3D reconstruction from monocular image sequences</title>

Abstract

Method for 3D reconstruction of known rigid objects from a single monocular image or a sequence of monocular images is presented. In the first part of the paper, a new computational approach to estimate pose and orientation of well-known objects in a 3D-scene from a single frame is discussed. The underlying theory is described in the context of a prototype matching problem and the existence of optimal solutions is proved. Furthermore, it is straightforward to extend the concept of prototype matching to the case of stereo or pseudo stereo applications and even more general setups. Hence, an improved 3D reconstruction with higher accuracy and increased stability can be achieved, for instance, by moving the camera along a linear sledge. The estimation of the ego-motion of a camera is covered as a special case by the introduced modeling via prototype matching. In the second part, experimental results for natural image sequences are analyzed. The derived accuracy and execution times of the described algorithms are illustrated. The last section deals with implementational details necessary to reduce the execution time under real-time restrictions.