Long-range terrain perception using convolutional neural networks

Autonomous robot navigation in wild environments is still an open problem and relies heavily on accurate terrain perception. Traditional machine learning techniques have achieved good performance for terrain perception; however, most of them require manually designed classifiers, meaning they have a poor generalization ability for learning new unknown environments. In this work, we integrate a deep convolutional neural network (CNN) model with a near-to-far learning strategy to improve the accuracy of terrain segmentation and make it more robust against wild environments. The proposed deep CNN model consists of an encoder and a decoder, which perform downsampling and upsampling for terrain feature extraction, respectively. The near-field terrain information obtained directly from the stereo disparity maps is fed into the CNNs as reference to aid in learning the far-field terrain information. Experimental results on a benchmark dataset demonstrate the effectiveness of the proposed terrain perception method.