Date of Award

2011

Document Type

Thesis

First Advisor

Bikdash, M.

Abstract

We have completed the building of an extensive database of civilian vehicle sounds. The database consists of correlated acoustic and seismic signatures of a large number (exceeding 850) of civilian vehicles. Each acoustic signature is obtained through two high-quality microphones separated by 25 feet, and whose signals are exactly synchronized. In this work, spectral and tristimulus features of civilian vehicle sounds are computed and then submitted to further processing using principal component analysis. The “super” features, derived after principal component analysis is performed, are then used for classification. In this research effort, the performance of the quadratic classifier with that of the neural network classifier is compared. Results presented here show that the neural network classifier out-performs the quadratic classifier in distinguishing different and same branded vehicle sounds. The classification usually has small (at times 0%) classification errors.

Share

COinS