Vehicle Classification Based on Sound

Abstract

Vehicle classi cation or vehicle type identi cation is one of the most impor- tant components in tra c management. Due to the high implementation cost of the current techniques and technologies, implementation of such systems in larger scale is quite challengeable for the developing countries like Sri Lanka. So, this paper is focus on sound signature based vehicle classi cation, which is a research on implementing a cost bene cial vehicle classi cation system. As this research focuses on cost bene cial aspect, most of the equipment and techniques used are cost bene cial rather than advanced equipments currently available. Sound capturing and enhancement unit which enhance the captured vehicle sound and a classi cation system which evaluate features of the signal and di erentiate vehicles based on them would be the basic components of such a system. Simple electronic circuit for ltering and ampli cation used for the recording and cone shaped architecture used to provide directional behavior to the microphone. Most of the features selected for the classi cation are commonly used in the audio processing and music classi cation domains. Low level signal properties such as zero crossings rate, root mean square value, signal energy and spectral roll-o and gamma tone lter banks which mimics the human hearing behavior are among the decided feature set. The main categories of the vehicles that are considered in this project are cars, vans, buses, motor bikes, three wheelers and trains. K-means clustering algorithm used for the classi cation and evaluates the performance of the classi cation in each combination of features and vehicle type domains. Usually the classi cation is better when there is less number of classes and it can be increased with the strength of the features chosen. Combination of several features may be not always helpful to increases the performance.