Semi-automatic geometric model construction for 3D traffic vision

This paper concerns the construction of 3D geometric object models from monocular intensity images. Two semi-automatic techniques are described. One is based on structure from motion, and the other on structure from symmetry. Both techniques fully exploit a priori knowledge such as the ground-plane constraint and the known topology of the object to be modelled. The techniques require human intervention in solving problems (such as feature detection and matching) which cannot be processed reliably and accurately by the computer. Both techniques have been implemented as interactive user application support tools. The tools have been used to generate tens of object models which are now routinely used in a 3D model-based traffic vision system