Radar target micro-Doppler signature classification
This thesis reports on research into the field of Micro-Doppler Signature (μ-DS) based radar Automatic Target Recognition (ATR) with additional contributions to general radar ATR methodology. The μ-DS based part of the research contributes to three distinct areas: time domain classification; frequen...
Main Author: | |
---|---|
Published: |
University College London (University of London)
2008
|
Subjects: | |
Online Access: | http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.564532 |
id |
ndltd-bl.uk-oai-ethos.bl.uk-564532 |
---|---|
record_format |
oai_dc |
collection |
NDLTD |
sources |
NDLTD |
topic |
621.3 |
spellingShingle |
621.3 Smith, G. E. Radar target micro-Doppler signature classification |
description |
This thesis reports on research into the field of Micro-Doppler Signature (μ-DS) based radar Automatic Target Recognition (ATR) with additional contributions to general radar ATR methodology. The μ-DS based part of the research contributes to three distinct areas: time domain classification; frequency domain classification; and multiperspective μ-DS classification that includes the development of a theory for the multistatic μ-DS. The contribution to general radar ATR is the proposal of a methodology to allow better evaluation of potential approaches and to allow comparison between different studies. The proposed methodology is based around a “black box” model of a radar ATR system that, critically, includes a threshold to detect inputs that are previously unknown to the system. From this model a set of five evaluation metrics are defined. The metrics increase the understanding of the classifier’s performance from the common probability of correct classification, that reports how often the classifier correctly identifies an input, to understanding how reliable it is, how capable it is of generalizing from the reference data, and how effective its unknown input detection is. Additionally, the significance of performance prediction is discussed and a preliminary method to estimate how well a classifier should perform is developed. The proposed methodology is then used to evaluate the μ-DS based radar ATR approaches considered. The time domain classification investigation is based around using Dynamic Time Warping (DTW) to identify radar targets based on their μ-DS. DTW is a speech processing technique that classifies data series by comparing them with a pre-classified reference dataset. This is comparable to the common k-Nearest Neighbour (k-NN) algorithm, so k-NN is used as a benchmark against which to evaluate DTW’s performance. The DTW approach is observed to work well. It achieved high probability of correct classification and reliability as well as being able to detect inputs of unknown class. However, the classifier’s ability to generalize from the reference data is less impressive and it performed only slightly better than a random selection from the possible output classes. Difficulties in classifying the μ-DS in the time domain are identified from the k-NN results prompting a change to the frequency domain. Processing the μ-DS in the frequency domain permitted the development of an advanced feature extraction routine to maximize the separation of the target classes and therefore reduce the effort required to classify them. The frequency domain also permitted the use of the performance prediction method developed as part of the radar ATR methodology and the introduction of a na¨ıve Bayesian approach to classification. The results for the DTW and k-NN classifiers in the frequency domain were comparable to the time domain, an unexpected result since it was anticipated that the μ-DS would be easier to classify in the frequency domain. However, the naıve Bayesian classifier produced excellent results that matched with the predicted performance suggesting it could not be bettered. With a successful classifier, that would be suitable for real-world use, developed attention turned to the possibilities offered by the multistatic μ-DS. Multiperspective radar ATR uses data collected from different target aspects simultaneously to improve classification rates. It has been demonstrated successful for some of the alternatives to μ-DS based ATR and it was therefore speculated that it might improve the performance of μ-DS ATR solutions. The multiple perspectives required for the classifier were gathered using a multistatic radar developed at University College London (UCL). The production of a dataset, and its subsequent analysis, resulted in the first reported findings in the novel field of the multistatic μ-DS theory. Unfortunately, the nature of the radar used resulted in limited micro-Doppler being observed in the collected data and this reduced its value for classification testing. An attempt to use DTW to perform multiperspective μ-DS ATR was made but the results were inconclusive. However, consideration of the improvements offered by multiperspective processing in alternative forms of ATR mean it is still expected that μ-DS based ATR would benefit from this processing. |
author |
Smith, G. E. |
author_facet |
Smith, G. E. |
author_sort |
Smith, G. E. |
title |
Radar target micro-Doppler signature classification |
title_short |
Radar target micro-Doppler signature classification |
title_full |
Radar target micro-Doppler signature classification |
title_fullStr |
Radar target micro-Doppler signature classification |
title_full_unstemmed |
Radar target micro-Doppler signature classification |
title_sort |
radar target micro-doppler signature classification |
publisher |
University College London (University of London) |
publishDate |
2008 |
url |
http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.564532 |
work_keys_str_mv |
AT smithge radartargetmicrodopplersignatureclassification |
_version_ |
1718140987582185472 |
spelling |
ndltd-bl.uk-oai-ethos.bl.uk-5645322015-12-03T03:27:07ZRadar target micro-Doppler signature classificationSmith, G. E.2008This thesis reports on research into the field of Micro-Doppler Signature (μ-DS) based radar Automatic Target Recognition (ATR) with additional contributions to general radar ATR methodology. The μ-DS based part of the research contributes to three distinct areas: time domain classification; frequency domain classification; and multiperspective μ-DS classification that includes the development of a theory for the multistatic μ-DS. The contribution to general radar ATR is the proposal of a methodology to allow better evaluation of potential approaches and to allow comparison between different studies. The proposed methodology is based around a “black box” model of a radar ATR system that, critically, includes a threshold to detect inputs that are previously unknown to the system. From this model a set of five evaluation metrics are defined. The metrics increase the understanding of the classifier’s performance from the common probability of correct classification, that reports how often the classifier correctly identifies an input, to understanding how reliable it is, how capable it is of generalizing from the reference data, and how effective its unknown input detection is. Additionally, the significance of performance prediction is discussed and a preliminary method to estimate how well a classifier should perform is developed. The proposed methodology is then used to evaluate the μ-DS based radar ATR approaches considered. The time domain classification investigation is based around using Dynamic Time Warping (DTW) to identify radar targets based on their μ-DS. DTW is a speech processing technique that classifies data series by comparing them with a pre-classified reference dataset. This is comparable to the common k-Nearest Neighbour (k-NN) algorithm, so k-NN is used as a benchmark against which to evaluate DTW’s performance. The DTW approach is observed to work well. It achieved high probability of correct classification and reliability as well as being able to detect inputs of unknown class. However, the classifier’s ability to generalize from the reference data is less impressive and it performed only slightly better than a random selection from the possible output classes. Difficulties in classifying the μ-DS in the time domain are identified from the k-NN results prompting a change to the frequency domain. Processing the μ-DS in the frequency domain permitted the development of an advanced feature extraction routine to maximize the separation of the target classes and therefore reduce the effort required to classify them. The frequency domain also permitted the use of the performance prediction method developed as part of the radar ATR methodology and the introduction of a na¨ıve Bayesian approach to classification. The results for the DTW and k-NN classifiers in the frequency domain were comparable to the time domain, an unexpected result since it was anticipated that the μ-DS would be easier to classify in the frequency domain. However, the naıve Bayesian classifier produced excellent results that matched with the predicted performance suggesting it could not be bettered. With a successful classifier, that would be suitable for real-world use, developed attention turned to the possibilities offered by the multistatic μ-DS. Multiperspective radar ATR uses data collected from different target aspects simultaneously to improve classification rates. It has been demonstrated successful for some of the alternatives to μ-DS based ATR and it was therefore speculated that it might improve the performance of μ-DS ATR solutions. The multiple perspectives required for the classifier were gathered using a multistatic radar developed at University College London (UCL). The production of a dataset, and its subsequent analysis, resulted in the first reported findings in the novel field of the multistatic μ-DS theory. Unfortunately, the nature of the radar used resulted in limited micro-Doppler being observed in the collected data and this reduced its value for classification testing. An attempt to use DTW to perform multiperspective μ-DS ATR was made but the results were inconclusive. However, consideration of the improvements offered by multiperspective processing in alternative forms of ATR mean it is still expected that μ-DS based ATR would benefit from this processing.621.3University College London (University of London)http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.564532http://discovery.ucl.ac.uk/18688/Electronic Thesis or Dissertation |