Application of Linear Discriminant Analysis to Doppler Classification

In this work the author demonstrated a robust and efficient method for implementing Doppler classification through the use of Linear Discriminant Analysis (LDA). LDAs were used to reduce dramatically the data dimensionality and thereby eliminate redundancy and improve the efficiency of the classifier. The performance was assessed on a three-class problem of personnel, tracked and wheeled vehicles. Real radar data from a ground based system were used in the design and testing of the classifier. The classifier algorithm was optimised by choosing the best set of features that maximised the performance and the bootstrap method was used to measure the confidence interval. It was shown that only the first few LDA features were relevant. At the very least these were shown to contain information regarding the frequency extent of target Doppler sidebands. The classifier was shown to be robust to changes in target viewing geometry and speed. Overall, good classification was achieved for personnel with some misclassification between tracked and wheeled vehicles. 1.0 INTRODUCTION MTI (Moving Target Indication) radars can provide an all-weather, day/night, surveillance capability. Such radar systems provide very efficient location information on moving targets but traditionally have limited recognition capability. Automatic recognition algorithms developed for imaging radars, which exploit target spatial information, are not applicable for MTI systems because they operate in a low resolution mode. However, there is potential for classification based on target Doppler signatures. The Doppler signatures are shifted in frequency in proportion to the target radial velocity. Movement or rotation of structures on a target may induce additional frequency modulations on the returned radar signal and generate sidebands about the Doppler frequency shift of the target’s body. The signature characteristics of these Doppler sidebands provide a mechanism for classifying the target of interest. The Doppler classifier models each target class as a multivariate Gaussian mixture distribution (GMD). The parameters of the GMD model are estimated using labelled training data. The input feature vectors are generated from the radar Doppler spectra. It is assumed that each Doppler spectrum provides an independent feature vector. Training uses multiple Doppler spectra per target class. Recognition is performed using a single Doppler spectrum (feature vector). The size (and therefore the dimensionality) of the input feature vector depends upon the number of separate frequency bins in the Doppler spectra. Herein lies the limitation of a classification technique that uses the Doppler spectra directly for input feature vectors. Doppler spectra can comprise a large number of frequency bins (several tens, possibly hundreds) to cover sufficiently the full range of Doppler frequencies Paper presented at the RTO SET Symposium on “Target Identification and Recognition Using RF Systems”, held in Oslo, Norway, 11-13 October 2004, and published in RTO-MP-SET-080.