A Robust Visual-Feature-Extraction Method for Simultaneous Localization and Mapping in Public Outdoor Environment

Abstract

We describe a new feature extraction method based on the geometric structure of matched local feature points that extracts robust features from an image sequence and performs satisfactorily in highly dynamic environments. Our proposed method is more accurate than other such methods in appearance-only simultaneous localization and mapping (SLAM). Compared to position-invariant robust features [1], it is also more suitable for low-cost, single lens cameras with narrow fields of view. Testing our method in an outdoor environment at Shibuya Station. We captured images using a conventional hand-held single-lens video camera. These environments of experiments are public environments without any planned landmarks. Results have shown that the proposed method accurately obtains matches for two visual-feature sets and that stable, accurate SLAM is achieved in dynamic public environments.