遮蔽輪郭線を用いた物体モデルに基づく3次元物体位置姿勢推定

概要

論文の詳細を見る
This paper describes a method for model-based 3D pose estimation. The object model includes a triangular surface mesh, model points, and model geometrical features. Model points and model geometrical features are generated using contour generators, which are estimated by the occluding contours of the images of the triangular surface mesh projected along multiple viewing directions, and are stored depending on the viewing direction. This modeling allows easy parallel processing. Multiple hypotheses for an approximate model pose are generated by comparing the model geometrical features and geometrical features extracted from pose data. These hypotheses are limited by the viewing directions used to generate model geometrical features. Each hypothesis is verified and improved by using model points and 3D boundaries, which are reconstructed by segment-based stereovision. In addition, each hypothesis is improved by using the triangular surface mesh and 3D boundaries. Experimental results obtained for objects with various shapes show the effectiveness of the proposed method.