Abstract: Based on conditional random fields(CRFs),an online,adaptive,and near-to-far long-range terrain perception approach is proposed.First,the current image is segmented into superpixels,and feature vectors and terrain categories of near-field superpixels are incorporated into terrain database as learning samples.Second,superpixel features and spatial relationships between far-field superpixels are modeled by using conditional random fields and terrain database.Finally,terrain categories of far-field superpixels are inferred based on the on-line learned model parameters.Experimental results show that the proposed approach outperforms other existing methods in terms of accuracy,robustness and adaptability to dynamic environments.