Globally Consistent 3-D Reconstruction by Utilizing Loops in Camera Movement

A common approach to 3-D reconstruction from image sequences is to track point features through the images, followed by an estimation of camera parameters and scene geometry. For long sequences, the latter is done by applying a factorization method followed by an image-by-image calibration. In this contribution we propose to integrate the tracking and calibration steps and to feed back already known camera parameters to both tracking and calibration. For loop-like camera motion, reconstruction can thus be optimized by using loop-closing algorithms known from robot navigation.