This content is not included in your SAE MOBILUS subscription, or you are not logged in.

Supervised Terrain Classification with Adaptive Unsupervised Terrain Assessment

Journal Article
ISSN: 2641-9637, e-ISSN: 2641-9645
Published April 06, 2021 by SAE International in United States
Supervised Terrain Classification with Adaptive Unsupervised Terrain Assessment
Citation: Kurup, A., Kysar, S., Bos, J., Jayakumar, P. et al., "Supervised Terrain Classification with Adaptive Unsupervised Terrain Assessment," SAE Int. J. Adv. & Curr. Prac. in Mobility 3(5):2337-2344, 2021,
Language: English


Off road navigation demands ground robots to traverse complex and often changing terrain. Classification and assessment of terrain can improve path planning strategies by reducing travel time and energy consumption. In this paper we introduce a terrain classification and assessment framework that relies on both exteroceptive and proprioceptive sensor modalities. The robot captures an image of the terrain it is about to traverse and records corresponding vibration data during traversal. These images are manually labelled and used to train a support vector machine (SVM) in an offline training phase. Images have been captured under different lighting conditions and across multiple locations to achieve diversity and robustness to the model. Acceleration data is used to calculate statistical features that capture the roughness of the terrain whereas angular velocities are used to calculate roll and pitch angles experienced by the robot. These features are used to train a k-means clustering classifier, where k is the number of terrain types the robot is expected to encounter. Human intervention is required only to set k. Cluster indices are auto labelled during training. In the operating phase, the SVM predicts the terrain that the robot will encounter and the corresponding cluster center predicts the vibration features associated with that terrain type. These features can be used to assess traversability. Vibration features are measured and the clusters are updated as the robot traverses, thus adapting to change in terrain over time. An accuracy of 92% has been achieved using visual data to classify terrain and 76% to predict the physical properties across different terrain classes.