Importance sampling forests for location invariant proprioceptive terrain classification

The ability for ground vehicles to classify the terrain they are traversing or have previously traversed is extremely important for manoeuvrability. This is also beneficial for remote sensing as this information can be used to enhance existing soil maps and geographic information system prediction accuracy. However, existing proprioceptive terrain classification methods require additional hardware and sometimes dedicated sensors to classify terrain, making the classification process complex and costly to implement. This work investigates offline classification of terrain using simple wheel slip estimations, enabling the implementation of inexpensive terrain classification. Experimental results show that slip-based classifiers struggle to classify the terrain surfaces using wheel slip estimates alone. This paper proposes a new classification method based on importance sampling, which uses position estimates to address these limitations, while still allowing for location independent terrain analysis. The proposed method is based on the use of an ensemble of decision tree classifiers trained using position information and terrain class predictions sampled from weak, slip-based terrain classifiers.