Abstract:
Road traffic accidents place a burden on the global economy. This impact is reduced by the
development of safer vehicles. Advanced Driver Assist Systems (ADAS) aim to reduce the frequency
and severity of accidents. ADASs are designed to operate in well-defined environments, such as first
world urban areas. However, 93% of fatal accidents occur in developing countries; areas often
without properly maintained roads. ADAS regularly fail to perform as intended in these challenging
environments. Terrain classification may improve the performance of ADAS. A lot of research has
been conducted on on-road terrain classification, but few studies focus on off-road terrain
classification. This study classifies several off-road terrains, based on road roughness using the
ISO8608:2016 standard, using a convolutional neural network (CNN). A database of images over
different terrains with known road roughness was created using forward and downward facing
cameras. Two different classification models were built: one is brand new and the other made use of
transfer learning on pretrained model. Terrain data was captured on several on-road and off-road
tracks. Results indicate that off-road terrain classification with cameras can be done with high
accuracy before a vehicle drives over a specific part of a road.