On accuracy of 3D localization obtained by aligning 3D model with observed 2D occluding edges

摘要

In this paper accuracy of the 3D localization obtained by aligning a 3D model with 2D observed occluding edges is discussed on. Aiming at accurate localization of a robot in a nuclear power plant, a method for aligning a 3D environmental model with an image observed by a camera mounted on the robot was proposed [1]. By effectively using the two-type predicted views which are calculated by a graphics system (e.g. OpenGL etc), the method succeeded in robust alignment even though the scene consists of complicated occluding edges of pipes. However accuracy of the 3D localization obtained by the alignment has not been yet enough examined. In this paper some factors affecting the error in the 3D localization are investigated. The experimental results using both synthetic and actual data make it clear that the error in the focal length of the camera model causes relatively large translation error in the view direction. For the case the camera parameter is not known precisely, we propose to utilize two cameras to decrease the errors and show its effect.