Wavelet-Based Neural Networks Applied to Automatic Detection of Road Surface Conditions Using Tire Noise from Vehicles
W. Kongrattanaprasert - wuttiwat@ew3.ee.uec.ac.jp
H. Nomura, T. Kamakura
Dept. of Electronics, The University of Electro- Communications
1-5-1 Chofugaoka, Chofu-shi, Tokyo 182-8585
K. Ueda
Nagoya Electric Works Co., LTD
19-1 Mentoku, Miwa-cho, Ama-gun, Aichi 490-1294
Popular version of paper 4pSP5
Presented Thursday afternoon, May 21, 2009
157th ASA Meeting, Portland, OR
The detection of road surface conditions is an important process for efficient road management. In particular, in snowy seasons, prior information about the road conditions, such as an icy state, helps road users or automobile drivers to avoid serious traffic accidents.
Tire noise from moving vehicles was recorded at the two different locations. They both were on the sides of four-lane national roads near Sapporo city. The roads are asphalted. Vehicles passed by at 60 km/h to 80 km/h on average. Tire noises emitted from moving vehicles on the road surface was recorded with a microphone. As everybody notices, the tonal qualities of these noises can change, depending on various factors like road surface conditions. We can, in principle, detect the surface conditions using appropriate signal processing techniques.
Noise signals were recorded continuously for more than three days. As a rule of thumb, we try to classify the road surface conditions into four categories using only tire noise. The four categories are snowy (LISTEN), slushy (LISTEN), wet (LISTEN) and dry (LISTEN). We focused on the classification into the snowy condition because serious traffic accidents often occur in snowy seasons.
Noise signals are first converted into digital signals. After that, we pay attention to the frequency spectrum of the noise because the tone qualities are directly related to the spectrum. We use the wavelet transform to obtain the power spectrum and propose several classification features so that the classification of the road conditions may be readily implemented. Furthermore, the usefulness of the proposed classification method is examined in the sets of multiple neural networks. And a final decision about the road surface conditions is made by combining the outcomes of the networks using the decision-making scheme.
Several practical methods for predicting road surface conditions have been proposed so far. For example, some researchers proposed a promising method using weather forecast data and field data. Different views and treatments were adopted in the past by other researchers, who used images of road surfaces and traffic noise from vehicles to classify road surface conditions without weather data. Unfortunately, their approach suffered from systematic problems of high cost and unstable accuracy. For cost reduction and simplicity in system structure, we propose the present detection method that is realized using only tire noises and is able to attain the prediction of the surface conditions with high accuracy. Of course, we have some plans that put our system into practice on actual roads.