Abstract

Planetary rovers face mobility hazards associated with various classes of terrains they traverse, and hence it is desirable to enable remote prediction of terrain trafficability (ability to traverse) properties. For that reason, the development of algorithms for assessing terrain type and mobility properties, as well as for coupling these data in an online learning framework, represent important capabilities for next-generation rovers. This work focuses mainly on 3-way terrain classification (classifying as one of the types: Sand, Bedrock and Gravel) as well as on the correlation of terrain types and their mobility properties in a framework that enables online learning. For terrain classification, visual descriptors are developed, which are primarily based on visual texture and are captured in form of histograms of edge filter responses at various scales and orientations. The descriptors investigated in this work are HOG (Histogram of Oriented Gradients), GIST, MR8 (Maximum Response) Textons and the classification techniques implemented here are nearest and k-nearest neighbors. Further, monochrome image intensity is used as an additional feature to further distinguish bedrock from the other terrain types. No major differences in performance are observed between the three descriptors, leading to the adoption of the HOG approach due to its lower computational complexity (over 3 orders of magnitude difference in complexity between HOG and Textons) and thus higher applicability to planetary missions. Tests demonstrate an accuracy between 70% and 93% (81% average) for the classification using the HOG descriptor, on images taken by NASA’s Mars rovers.
To predict terrain trafficability ahead of the rover, exteroceptive data namely terrain type and slope, are correlated with the trafficability metrics namely slip, sinkage and roughness, in a learning framework. A queue based data structure has been implemented for the correlation, which keeps discarding the older data so as to avoid diminishing the effect of newer data samples, when there is a large amount of data. This also ensures that the rover will be able to adapt to changing terrains responses and predict the risk level (low, medium or high) accordingly.
Finally, all the algorithms developed in this work were tested and verified in a field test demo at the CSA (Canadian Space Agency) mars yard. The risk metric in combination with the queue based data structure, can achieve stable predictions in consistent terrains, while also being responsive to sudden changes in terrain trafficability.