Rugged terrain segmentation based on salient features

Abstract

In the last decade significant progress in computer vision based control of unmanned ground vehicles (UGV) has been achieved. However, until now textural information has been somewhat less effective than color or laser range information. In this paper we propose a computer vision based cross country segmentation system that is capable of distinguishing cross-country road, grass and trees during day-time and night times. For this purpose we extract Speeded-Up Robust Features (SURF) from the training image set and construct texture class models using two-layer feed-forward neural network. Using these constructed models and extracted features from the images captured by the CCD and IR cameras we estimate features' class membership values. These estimated values and features' spatial positions are then applied for image segmentation. A number of experiments are conducted with the lowest mean error segmentation rate of 16.78% and 20.60% for images in IR and visible spectrum correspondingly.