Clustering and probabilistic matching of arbitrarily shaped ceiling features for monocular vision-based SLAM

This paper presents improved extraction and matching methods for arbitrarily shaped (AS) ceiling features for monocular vision-based simultaneous localization and mapping. The feature descriptor, which is robust to illumination changes, comprises the vertex distribution, size, and orientation strength of the region of interest. However, to cope with the problem of vertices being detected at different positions in successive images, Bayes' rule is applied to preserve robust vertices and remove rarely observed vertices. Moreover, unstable features surrounded by similar features are clustered to create a robust feature by calculating their similarities to adjacent clusters. AS features from the proposed scheme are used as landmarks in the extended Kalman filter, and the effectiveness of the proposed scheme is verified through various experiments in real environments.