Radar-Based Road Surface Classification Using Range-Fast Fourier Transform Learning Models.

Area of Science:

• Engineering
• Computer Science
• Materials Science

Background:

• Traffic accidents from black ice pose significant public safety risks.
• Conventional detection systems struggle with low visibility and adverse weather.
• Millimeter-wave (mmWave) radar offers reliable road surface monitoring in challenging conditions.

Purpose of the Study:

• To evaluate the effectiveness of mmWave radar for recognizing diverse road surface conditions.
• To compare the performance of various machine learning models in classifying road states using mmWave data.

Main Methods:

• Experimentally simulated six road surface conditions (dry, wet, thin-ice, ice, snow, sludge) on asphalt and concrete.
• Collected mmWave radar data, processed using Range-Fast Fourier Transform (Range-FFT).
• Trained and evaluated classification models: XGBoost, LightGBM, CNN, and ViT.

Main Results:

• CNN and ViT models showed superior performance across all tested road conditions.
• Machine learning models had reduced accuracy on hazardous road states (ice, snow).
• CNN demonstrated consistent stability, while ViT offered competitive accuracy with enhanced pattern recognition.

Conclusions:

• mmWave radar is feasible for reliable road surface condition recognition.
• CNN and ViT show promise for advanced road safety systems.
• Future improvements may involve multimodal sensor fusion and time-series analysis.