Combining features for adaptive terrain classification based on ART neural network

Terrain classification focuses on determining a safe region for robot to traverse while labeling obstacles. This paper study terrain classification based on scene imageries of the natural environment and attempts to combine color, texture and geometry moment features to train an ARTMAP neural network. After learning the relationship between the combined features and the traversability of terrains, the neural network can be used to assess the front terrain. In this paper, we used the combined features to carve up the environment and enable the classification more lighting independent, season adaptable and more efficient in the whole. Thanks to the adaptability of the ARTMAP neural network classifier and the strategy of features combination, the results show that the presented method can classify the terrain more accurately across different scenarios with fairly high classification efficiency.