カメラの回転運動を事前知識とした奥行き推定手法の提案

概要

論文の詳細を見る
To recover 3D-information using two images taken by a moving camera at different positions, we need two information, the location of cameras and relation of correspondences between images. When the camera motion is unknown, we must find all clues from only images. In this paper, we propose a method of depth estimation using the knowledge of camera rotation. We assume that the rotation of a camera is controllable and that approximate information of camera motion is available. The camera motion is described with six parameters. We set initial values of parameters from the information of camera motion. We calculate parameters that minimize the distance between epipolar lines and correspondent points. When an iteration is repeated, we update the values, which must keep a solution that is possible as real camera motion. In the simulation experiment, we confirmed that our method performed effectively when matching error was small, while there were many matching errors using knowledge only. We also examined in real environment and obtained good results. Epipolar lines recovered by our method passed through the correspondent points and lines recovered by knowledge only didnt. It proved that object shapes were generally reconstructed and the proposed method was useful and effective.
社団法人電気学会の論文