Road Surface Classification Using Frequency- Modulated Continuous Wave Radar for Autonomous Driving Robots

Autonomous driving robots (ADRs) are being deployed in increasingly diverse environments. Consequently, the systematic classification of road surfaces is critical for safe and adaptive control. This paper proposes a lightweight, real-time road surface classification (RSC) system implemented on a field-programmable gate array (FPGA) for ADRs. The RSC system utilizes a 60 GHz frequency-modulated continuous wave (FMCW) radar to acquire road surface data, which is particularly effective in adverse conditions where visual sensors are unreliable. Road surface data are converted into range-Doppler (RD) maps that contain a combined representation of range and velocity. To enhance classification accuracy and predictive capability, 4-channel input data were generated from near- and far-range profiles extracted from the RD maps. We employ a binarized neural network (BNN) for classification to achieve significant model lightweighting and computational efficiency, making it suitable for resource-constrained ADRs. The proposed RSC system classifies the target road surface types with 93.84% accuracy in our experimental setup, while achieving 99.59 mW power consumption and 25 μs inference times.