Doppler radar-based pothole sensing using spectral features in k-nearest neighbors

Abstract

Potholes, resulting from wear, weather, and traffic, pose a substantial road safety concern, driving up maintenance costs and government liabilities. Numerous studies have explored pothole detection systems, however, there is a limited focus on radar-based approaches. This study investigates the use of Doppler radar mounted on moving vehicles to collect asphalt road surface data, with the aim to leverage this unique perspective point. Spectral features from power spectral density (PSD) are extracted and explored by incorporating Doppler signal PSD features into a k-nearest neighbors (KNN) within a machine learning framework for road condition classification. Six KNN algorithms are applied, and results indicate that potholes exhibit distinct spectral differences characterized by higher variability, with fine KNN performing the best, achieving an accuracy rate of 95.38% on the test dataset. In summary, this research underscores the effectiveness of Doppler radar-based pothole sensing and emphasizes the significance of algorithm and feature selection for achieving accurate results, proposing the viability of radar systems and machine learning.