Abstract:
The problem of target classification in the case of ground surveillance radar is very important because the availability of such information can enhance the situational awareness of the user and give a strategic advantage over an enemy. Information about the
target class can be extracted by applying digital signal processing and machine learning
techniques to the backscattered signal.
The purpose of this thesis is to develop an automatic target classifier for NR-V3 ground
surveillance radar. Three target classes considered for this research are pedestrians,
vehicles and "no target" classes. The proposed classifier exploits the micro-Doppler
features present in the target signatures to classify them into one of the three classes.
Field trials of NR-V3 radar were conducted in a controlled environment to collect signatures of different classes. Three different datasets, with signatures of 100 msec, 200
msec and 400 msec duration are created to test the performance of classifier for different
dwell times. The feature vector inputs for the classifier are obtained by preprocessing
of signatures. This includes clutter filtering, Fourier transform and Principal Component Analysis (PCA). The resulting feature vectors are classified using three types of
classifiers namely Linear Discriminant Analysis (LDA), Logistic Regression and Support Vector Machine (SVM) and their performance is compared. Classifiers are trained
using 80% of the total collected data to obtain the values of their parameters. During
the testing phase, remaining 20% of the data is used to test the performance of the classifiers. For 100 msec signal, the overall classification rate for LDA, Logistic Regression
and SVM classifiers is 86.81%, 91.37% and 96.48% respectively. The performances of
all the classifiers improve with increase in dwell time. The classification rate increases
to 95.58%, 96.96% and 97.24% for LDA, Logistic Regression, and SVM classifiers
respectively for 400 msec duration signal.
i
