Visual place recognition for multi-robots maps merging

Abstract

This paper presents a new approach allowing to a group of robots team to merge their individual maps without any a priori knowledge about their relative positions. While moving, each robot acquires laser data and a video stream. The first data are used to create individual maps (SLAM), while images sequences are used to derive invariant visual descriptions of the visited places. The later are then exchanged between robots to determine a probability of being close to each other and sharing a common place in order to initiate a map merging process. In such a way, ambiguous situations that usually occur when a single sensor is used are reduced. We present the solution we developed in order to extract a compact visual description that remains constant regardless to the actual pose of the robots. Two or more robots having such description can verify if they are or not present at the same place, before merging their respective individual maps. This scheme, e.g. the fusion of laser-range data and visual information enhances and accelerates the construction of the global map and consequently disambiguate individual maps. The validation of the approach has been performed on various indoor environments mainly, office-like spaces and houses.