Visual Features for Long-Range Terrain Perception

A visual features based long-range terrain perception method for mobile robots is presented. This approach follows a near-to-far learning procedure. Firstly, the obtained image is segmented and the scale is normalized. Then, hue features for representing the color information and LBP (local binary pattern) features for representing the texture information are extracted. Next, the near field terrain samples are labelled obstacle and ground using stereo vision, and then a classifier is trained with these labelled samples to classify the rest unknown samples. Finally, confidence values are defined based on posterior probability to reclassify the low confident samples for further improving the classification accuracy. Experiment results show that the proposed approach can achieve long-range terrain perception accurately and stably.