Acoustic feature analysis for wet and dry road surface classification using two-stream CNN

• Main Authors: Bahrami, Siavash, Doraisamy, Shyamala, Azman, Azreen, Nasharuddin, Nurul Amelina, Shigang, Yue
• Format: Conference or Workshop Item
• Language: English
• Published: Association for Computing Machinery 2020
• Online Access: http://psasir.upm.edu.my/id/eprint/85375/1/Acoustic%20Feature%20Analysis%20for%20Wet%20and%20Dry%20Road%20Surface.pdf; http://psasir.upm.edu.my/id/eprint/85375/; https://dl.acm.org/doi/10.1145/3445815.3445847
• Institution: Universiti Putra Malaysia

Road surface wetness affects road safety and is one of the main reasons for weather-related accidents. Study on road surface classification is not only vital for future driverless vehicles but also important to the development of current vehicle active safety systems. In recent years, studies on road surface wetness classification using acoustic signals have been on the rise. Detection of road surface wetness from acoustic signals involve analysis of signal changes over time and frequency-domain caused by interaction of the tyre and the wet road surface to determine the suitable features. In this paper, two single stream CNN architectures have been investigated. The first architecture uses MFCCs and the other uses temporal and spectral features as the input for road surface wetness detection. A two-stream CNN architecture that merges the MFCCs and spectral feature sets by concatenating the outputs of the two streams is proposed for further improving classification performance of road surface wetness detection. Acoustic signals of wet and dry road surface conditions were recorded with two microphones instrumented on two different cars in a controlled environment. Experimentation and comparative performance evaluations against single stream architectures and the two-stream architecture were performed. Results shows that the accuracy performance of the proposed two-stream CNN architecture is significantly higher compared to single stream CNN for road surface wetness detection.