Vehicle Classification Using Acoustic Data Based on Biology Hearing Models and Multiscale Vector Quantization
D. A. Depireux, S. Varma, J. S. Baras, N. Srour and T. Pham
ARL Federated Laboratory 4th Annual Symposium, College Park, MD, March 21-23, 2000
The army is interested in using acoustic sensors in the battlefield to perform vehicle identification using passive microphone and seismic arrays. The main advantages of acoustic arrays are that they are non-line of sight, low cost, low power, can be made small and rugged and can provide 360 degrees coverage. Their capability includes target detection, bearing, tracking, classification and identification, and can provide wake-up and cueing for other sensors.
Acoustic arrays can be deployed in an expandable tracking system: the outputs of a network of acoustic arrays can detect, track, and identify ground targets at tactical range by triangulating the reports from several distributed arrays.
Here, we present a prototype of vehicle acoustic signal classification. To analyze the signature of the vehicle, we adopt biologically motivated feature extraction models. Several possible representations are used in a classification system. Different vector quantization (VQ) clustering algorithms are implemented and tested for real world vehicle acoustic signal, such as Learning VQ, Tree-Structured VQ and Parallel TSVQ. Experiments on the Acoustic-seismic Classification Identification Data Set (ACIDS) database show that both PTSVQ and LVQ achieve high classification rates.
The VQ schemes presented here have the advantages of not having to choose explicitly the features that distinguish the targets. The burden is shifted to having to choose the "best" representation for the classifier.