Robot vision: model synthesis for 3D objects

This paper presents the automated model synthesis component of an integrated passive 3D vision system. The synthesized models can be used by the object recognition and pose determination components. The model synthesis obtains the 3D object model from images acquired by a CCD camera posed at various known positions. This paper presents developments and discusses automatic model synthesis. The tasks include robust 2D feature detection; 2D feature post-processing for eliminating noise and recovering missing features; 2D feature grouping of structurally related 2D features; stereo triangulation with a new form of epipolar line constraint; projective inversion of ellipses; synthesis for circular shape in 3D space from its projective views based on the ellipse pose hypothesis; and incremental model synthesis of model from multiple views based on the vertex triangulation. To demonstrate full automation, we use a single CCD camera mounted on the last link of a robot arm. The integrated robot system is able to move the CCD camera around the object and capture images at various vantage points and furnish the camera pose corresponding to each image acquired. The intelligent system then synthesizes the extracted features from each image to obtain a 3D model of the object. Such an approach, though more difficult than the direct use of range data through range sensors, is of great importance for space and industrial automation where cost and flexibility are of concern.