Geometric descriptor for bag of visual words-based place recognition

Bayesian pattern recognition is a natural companion for graph-based simultaneous localization and mapping (SLAM) due to its ability to come up with high quality place matches based solely on image data, completely eschewing metric information. Recent SLAM-like approaches such as fast appearance-based mapping (FAB-MAP) [1] are very effective information filters, with the ability to provide place matching data in real-time to a very high degree of accuracy. In this study, the strong foundations of FAB-MAP and bag of visual words-based place recognition are refined to include geometric information directly into the image descriptor. To demonstrate the practicality of such an approach, the new descriptor type was incorporated into a full real-time graph-based SLAM stack. In experiments, places are recognized at a rate of 14% in a series of experiments performed in challenging, visually monotonous environments, compared to the 9% rate obtained by the direct application of OpenFABMAP to the same data set [2]. The system played an instrumental role in maintaining map integrity, without which SLAM would not have been possible.