Distributed acoustic sensor data processing for target classification

Abstract

The Acoustics Signal Processing Branch at the U.S. Army Research Laboratory has been investigating tracking and classification of military and civilian vehicles using acoustic sensors. Currently the target signals are modeled as a sum of harmonics and then they are classified using multivariate Gaussian classifier at each individual node. When multiple targets are present in the scene overall classification of the targets deteriorates as the signals from several targets are mixed together and determinations of individual target harmonics become difficult. This is true particularly for civilian vehicles. In order to improve the overall probability of correct classification a distributed classifier will be implemented. In a distributed processing each sensor node would broadcast the classification information, that is, probability of detection of various targets, to all the sensors within its vicinity. At each sensor node a distributed Bayesian classifier is used to determine the overall classification of each target. The distributed processing is robust to failures in sensor nodes unlike the centralized processing. Although the technique is known, it had been tested using only simulated data. In this paper we present the results of the algorithm on real data that was collected using several acoustic sensors using a mixture of military and civilian vehicles. This would identify how well the distributed processing works or its limitations in classifying multiple targets using acoustic data.