Radar Signature Analysis of Indoor Clutter and stationary Human Target Classification

Open Access
Author:
Bufler, Travis Dale
Graduate Program:
Electrical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
March 03, 2016
Committee Members:
  • Ram Mohan Narayanan, Dissertation Advisor
  • Kultegin Aydin, Committee Member
  • Karl Martin Reichard, Committee Member
  • Randy Young, Committee Member
Keywords:
  • Radar
  • SVM
  • FDTD
  • RCS
  • Indoor Clutter
  • Human Scattering
  • Target Classification
Abstract:
Research in through-the-wall radar (TTWR) is a recent area of focus and as such many challenges have arisen related to target detection, location, and tracking. These types of radar systems are often plagued by harsh clutter scenarios caused by the variety and abundance of furniture elements within the antenna’s field-of-view which in turn degrade the radars ability to accurately detect targets. Furthermore, once an object is located, having the ability to identify the object under consideration is still an on ongoing research problem. This dissertation has two main focuses. The first analyzes and investigates the spectral signatures of indoor clutter elements through their radar cross section (RCS). The characterization of the wideband spectral properties for indoor clutter elements is accomplished using finite difference time domain (FDTD) techniques. Using FDTD, the spectral characteristics of clutter elements were obtained over a wide range of frequencies and aspect angles at different transmit-receive polarization combinations. The spectral properties are then compared and contrasted between the different objects as well as human biological models for analysis and use in TTWR. The RCS results obtained from the simulations are then compared to experimental data collected using a wideband network analyzer. The second main topic is the use of machine learning algorithms, specifically Support Vector Machines (SVMs) to construct models from the previously acquired spectral signatures to aid in target classification. Utilizing variables of frequency and polarization, a SVM classifier can be constructed to classify unknown targets as either human or clutter object. Additionally, the application of feature selection algorithms are applied to the spectral characteristics to reduce the model complexity and determine the SVM classification accuracy of a reduced dataset. Finally, the classification performance is assessed in the presence of additive white Gaussian noise (AWGN) and through various wall materials.