Classification of road surfaces using convolutional neural network

In this paper a classifier of road surfaces, visible from the front of the car, is presented. It is intended to use in the driver assistance or the autonomous car systems. The classifier was prepared, tuned and tested using AlexNet/CaffeNet convolutional neural network. To perform experiments, an original database of about 500 surface images was prepared. Nine classes of road surfaces were recorded: asphalt, concrete, two types of concrete paver blocks, granite paver blocks, openwork surface, gravel, sand, and grass. Using this database, two groups of testing experiments were performed: recognition of a particular type of road surface and determination of the general condition of the surface. Classification results show accuracy over 85% in the first experiment and over 91% in the second experiment. The results are promising, especially that, e.g. the sensitivity of recognition of bad surfaces reaches 95%, what indicates the potential possibilities of usage of the proposed classifier in real cars.