Terrain Classification Using Inside-Wheel Cameras Based on Wheel-Terrain Interaction Characteristics.

The wheeled rover is widely used in off-road planetary exploration tasks nowadays. However, complex operating environment like soft deformable terrains on the planet surface demands the high performance of rover. In order to reduce danger, environmental perception technologies need to be used on the rover. Terrain classification is a critical method in environmental perception. In this research, a wheel-vision-based terrain classification method is proposed. This method uses wheel-terrain interaction parameters obtained from inside-wheel cameras which are strongly related to the terrain as input. We use single-wheel testbed to collect and process data to form a dataset. The data is obtained under different slip ratios and different terrains. Then, we train four terrain classification models using machine learning algorithms and analyze their performance. From the results, all performance criteria of the models achieve over 95%, with DT being the optimal method. The feasibility of our method in the case of fewer training samples and the rationality of dataset selection are also verified through the results of different training set sizes and the ablation experiment.