Feature extraction for a neural network classifier

Abstract

Unattended ground sensor (UGS) networks are intended to detect and localize the presence of strategic relocatable targets in the theater of operation over several kilometers. Passive acoustic sensors, an integral part of UGS, have achieved a high level of maturity and will allow acoustic target classification for tracked and wheeled vehicles. Of primary importance in the classification problem is the selection of a robust feature extraction technique, tolerant of both the environment and the nonstationary nature of the acoustic signatures. Several feature extraction techniques were used with experimental acoustic data collected from a small baseline, circular array. Results will be presented of the classification for acoustic features using a backpropagation neural network with simple power spectrum, harmonic line association [J. A. Robertson, IIT Research Institute, in-house report], principal components [J. Mao and A. K. Jain, IEEE Trans. Neural Networks 6 (2) (1995)], and wavelet packet [K. Etemad and R. Chellappa, Proc. First Intl. Conf. on Image Processing (November 1994)] feature extraction techniques.

<title>Feature extraction and fusion of acoustic and seismic sensors for target identification</title>

“…After split-window peak-picking is performed on the PSE feature set, the algorithm finds the maximum peak P in the frequency set and assumes that this peak is some kth harmonic line of the fundamental frequency subject to the following soft constraint for fundamental frequency range, flund E{8,20} Hz, (1) and calculates total signal strength in this HLA set. The integer value k, which gives the maximum signal strength, is assumed to be the correct harmonic line number, and the hannonic lines of this particular set are retained as a feature vector.…”

“…5 Nonstationarity increases the difficulty in the feature-selection process, but methods have been reported to alleviate this difficulty;1'3 most notable is the use of the harmonic line association (HLA) algorithm with shape statistics. 6 The HLA algorithm takes advantage of spectral characteristics that are dominated by narrowband spectral peaks. In the past, the narrowband spectral peaks have been used for classification purposes, either in hierarchical clustering schemes or as direct inputs into an artificial neural network (ANN).…”

Efficient Maximum Likelihood Estimation for Multiple and Coupled Harmonics

“…So asymptotically, estimating multiple harmonic in colored noise completely decouples into estimating m single harmonics in white noise with variance equal to the local noise variance at that harmonic. (2,3,4,5,6,7,8,9,10,11,12) r and m = 11 in equation (6). The additive noise process Z in equation (7) consists of the cumulative effect of ambient noise such as wind, broadband and narrowband nontarget sources, and a portion of the target's acoustic signature not completely described by equation (6) figure 2, which shows the first 512 samples from the same data as in figure 1.…”

“…To simplify the analysis, I assume the variance is known, although all the results presented are equally valid in the unknown case. In this case, the MLE maximizes log(L(y,0)) = -^||y-s(ö)||^-n/21og(2^), (10) which is equivalent to minimizing the residual sum of squares…”

Classification of Vehicles Based on Audio Signals using Quadratic Discriminant Analysis and High Energy Feature Vectors

“…There is a number of feature extraction methods used in vehicle classification. Some of them produce too many features for a single input vector like estimation of Power Spectrum Density [6], [7], [8], and some are too complicated like Principal Component Analysis [9], [7]. Nevertheless, no feature extraction method fulfills all our requirements, and there is no systematic way to select best feature according to our criterion.…”