Robust terrain classification by introducing environmental sensors

This paper presents a vision-based off-road terrain classification method that is robust despite large environmental variations caused by seasonal or weather changes. In order to account for an overall image feature variation, we adopted environmental sensors, and to train a neural network based classifier, constructed a database according to environmental conditions. Robust classification could be achieved by selecting the training parameter set best suited for each environmental state. Also, we propose a hardware architecture that enables distributed parallel processing for real- time implementation of the present algorithm. Experimental results for real off-road images show that in spite of dissimilar conditions, degradation of classification performance could be minimized by replacing the nearest parameters.