Real-Time Terrain Condition Detection for Off-Road Driving Based on Transformer

Off-road driving is dangerous due to the deformation and irregularities of the road surface. The diverse nature of off-road surfaces makes it difficult to identify the safety of the driving surface. In this study, a Transformer-based neural network is proposed to estimate the drivability of various off-road surfaces, aiming to discern whether the terrain is safe to drive or potentially dangerous. The network only utilizes Controller Area Network (CAN)-bus signals from the vehicle, which makes it easy to implement on a readily available vehicle. To train the network, driving data was collected from a diverse range of off-road environments, from areas where novice drivers can drive safely to hazardous areas where expert drivers get stuck. We also propose a post-processing algorithm to filter out false estimations and limit frequent changes in estimation, as these can have detrimental effects on real-world systems. The performance of our algorithm was evaluated in real-time on various off-road surfaces showing high level of accuracy.