Pothole Sensing and Analysis with Doppler Audio Signal Spectrum Employing Machine Learning Algorithms

• Published in: IEEE Symposium on Wireless Technology and Applications, ISWTA
• Main Author: Zainuddin S.; Ismail M.A.D.; Nasir H.M.; Rashid N.E.A.; Shariff K.K.M.; Abidin I.Z.
• Format: Conference paper
• Language: English
• Published: IEEE Computer Society 2024
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85203798235&doi=10.1109%2fISWTA62130.2024.10652168&partnerID=40&md5=19b5cf77b3a062a5f6fb37273ec39924

Potholes are imperfections, whether minor divots or massive craters, on the surface of roads, streets, or pavements caused by depressions or holes that are dangerous for cars and pedestrians. Several approaches have been investigated to raise pothole detection accuracy; each has benefits and drawbacks. This research proposes a pothole sensing approach by applying machine learning (ML) algorithms and the power spectrum density (PSD) of Doppler audio signals. Although potholes are notable to be localised based on their depth, continuous waveform (CW) is known for its capability to identify based on targets’ movement. Furthermore, neighbouring harmonics are likely to conceal the target's reflected signal. A distinct viewpoint is provided by placing the radar on a moving vehicle where the radar is moving despite the subject of interest. In this research, machine learning (ML) algorithms are compared, and the potential of the Doppler audio signal for pothole identification is analysed. Waveform audio file (.wav) datasets for potholes and non-potholes were obtained using a commercially available KLC2 Doppler radar. In order to extract the spectral features, audio Doppler signals were transformed to the power spectral density (PSD). Various ML classification methods were applied to the extracted features. The outcome indicates that the Cubic Support Vector Machine (SVM) produced the highest performance of 97.22% accuracy, 100% precision, 95.2% recall and 97.56% F1-Score, for the 80:20 distribution, in comparison to other models presented. © 2024 IEEE.