Abstract

The use of random finite sets (RFSs) in simultaneous localization and mapping (SLAM) has many advantages over the traditional random vector based approaches. These include the consideration of detection and clutter statistics and the circumvention of data association and map management heuristics in the estimation stage. To take full advantage of RFS based estimators in feature based SLAM, the feature detector's detection and false alarm statistics should be modelled and used in each SLAM estimation update stage. This paper presents principled techniques to obtain these statistics for semantic features extracted from laser range data, and focusses on the example of the extraction of circular cross-sectioned features, such as trees, pillars and lampposts, in outdoor environments. Comparisons of an RFS based SLAM algorithm ¿ Rao-Blackwellized, Probability Hypothesis Density (RB-PHD)-SLAM, which utilizes the derived, variable feature probabilities of detection, and the same SLAM algorithm based on the typically assumed constant feature detection probabilities, within the sensor field of view, are provided. The results demonstrate the advantages of explicitly modelling feature detection statistics.