Off-road Place Recognition using Fused Image Features

This paper presents a method of image-based localization in rough off-road terrain. It assumes an underlying topological map of the environment modeling places that are relevant for navigation as nodes in a graph. Hence, localization is reduced to the coarse recognition of previously visited places. While many approaches try to identify objects that characterize a specific location, the presented algorithm creates and matches as many redundant representations of a place as possible. It therefore only uses features of different nature that are already available in camera images and can be easily extracted. Then, it fuses the matching result of each single representation into one estimation, compensating for possible weaknesses in the individual feature extractors. The presented algorithm was tested in a forest environment and the experimental results show that the fusion of redundant information is sufficient for the characterization of places as natural landmarks.