English

Abstract

This paper presents an indoor vision-based system using a single camera for human localization. Without a priori knowledge of the operating environment, a map has to be built on-line to estimate the relative positions of the camera. When a model is a priori known, only the camera poses are computed. It results in distinctive algorithms which have both assets and drawbacks. Localization in an unknown environment is much more flexible but subject to drift while localization in a known environment is almost drift-less but suffer from recognition failures. We propose a new approach to localize a camera in an indoor environment. It combines both techniques described above benefiting from the knowledge of Georeferencing information to reduce the drift (comparatively to localization in unknown environment) while avoiding the user to be lost during long time intervals. Experimental results show the efficiency of our method.