An acoustic signature based neural network model for type recognition of two-wheelers

Vehicles of a given type, in different working conditions, generate dissimilar sound patterns. Each sound pattern is viewed as acoustic signature. Sounds of moving vehicles provide clues of their traits such as makes, possible faults, performances of sub systems and the like. Different work conditions mean vehicles running at different speeds, under different road conditions, different accelerations and the like. In such situations tracking of faults manually becomes difficult and automatic acoustic surveillance enables easy monitoring of certain conditions of the vehicles and future consequences. These could be accidents, over speeding of the vehicles, compliance with traffic rules and regulations etc. In this paper, we have proposed an acoustic signature based neural network model for recognizing different types of two-wheelers. We have used simple time-domain features such as Average Zero Crossing rate(ZCR), Root Mean Square(RMS), and Short Time Energy(STE), and frequency-domain features such as Mean and Standard Deviation of Spectrum Centroid (CMEAN and CSD). Two-wheelers of three major Indian makes, namely Hero Honda, Bajaj and TVS, are considered in the work. The vehicles are classified into Bikes and Scooters. It is observed from the results that classification accuracy depends on different factors such as their usage, maintenance, environmental and road conditions. We have considered age of the vehicle as a factor in choosing the samples. The recognition results show 73.33% accuracy.