Robust extraction of arbitrary-shaped features in ceiling for upward-looking camera-based SLAM

This paper proposes a novel scheme of extracting the features in the ceiling using an upwardlooking camera for SLAM in indoor environments. The conventional approaches based on corner or line features have difficulties in associating adjacent indistinguishable features since their descriptors do not have enough information to distinguish them. Therefore, we used various properties of the objects in the ceiling, such as the distribution of nodes, and size and orientation strength of the region of interest. Also, the similarities among adjacent features are calculated to distinguish between unique and non-unique features. The non-unique features are discarded before they are registered to the feature map. The robustness of the proposed scheme is verified by several experiments using a mobile robot. The extended Kalman filter is used to estimate both robot pose and feature position and orientation, and the experimental results show that the proposed scheme successfully works in various indoor environments.