CFAR Thesis - James Jen - PDF
Short Description
My Cal Poly Pomona Master's Thesis...
Description
A STUDY OF CFAR IMPLEMENTATION COST AND PERFORMANCE TRADEOFFS IN HETEROGENEOUS ENVIRONMENTS
A Thesis Presented to the Faculty of California State Polytechnic University, Pomona
In Partial Fulfillment Of the Requirements for the Degree Master of Science In Electrical and Computer Engineering
By James J. Jen 2011
SIGNATURE PAGE THESIS: A STUDY OF CFAR IMPLEMENTATION COST AND PERFORMATIVE TRADEOFFS IN HETEROGENEOUS ENVIRONMENTS AUTHOR: James J. Jen DATE SUBMITTED: Spring 2011 Electrical and Computer Engineering Dr. Zekeriya Aliyazicioglu _________________________________________ Thesis Committee Chair Electrical and Computer Engineering College of Engineering Dr. H.K Hwang _________________________________________ Electrical and Computer Engineering College of Engineering Dr. James Kang _________________________________________ Electrical and Computer Engineering College of Engineering
ii
ACKNOWLEDGEMENTS
I recall clearly last year when, at the end of a Signal Processing lecture, Profes‐
sor Hwang asked for interest in graduate research. My time at Cal Poly Pomona, up un‐ til that point, had been rocky and undistinguished. Tentative, I almost didn’t approach the professor that. I’m glad I did. The last year of graduate research with Professors Hwang and Zeki had meaningfully colored and enriched my graduate studies.
Much thanks to Professor H.K. Hwang and Zekeriya Aliyazicioglu: firstly, for the
opportunity for research and their many Friday mornings spent on us students, and secondly, for their invaluable guidance and encouragement throughout.
Appreciation too to Thales‐Raytheon for supporting the research, and particu‐
larly to Tom Nichols Walker Birrell for their expertise and input.
Thanks too to Nellie Qian, who, for most of the past year, had been my partner
in crime in MATLAB ventures from Minimax beamforming, to antenna array interfer‐ ence suppression, and finally to this CFAR thing that had become our thesis.
And thanks, most sincerely, to my mom and dad— Emily and Chin‐Ping Jen—
for their support, encouragement, and patience.
iii
ABSTRACT
CFAR— Constant False Alarm Rate— is a critical component in RADAR detec‐
tion. Through the judicial setting of detection threshold, CFAR algorithms allow RADAR systems to set detection thresholds and reliably differentiate between targets of inter‐ est and interfering noise or clutter.
In many operating conditions, noise and clutter distributions may be highly het‐
erogeneous— with sudden jumps in clutter power or with the presence of multiple tar‐ gets in close proximity. A good CFAR algorithm must reliably operate in these condi‐ tions but without prohibitively high implementation costs.
In this thesis, we investigate a number of CFAR algorithms— new and old— all
the while assessing their operational flexibility and cost of operation.
As balance of performance and implementation cost, two algorithms stood out
as desirable: Variability Index (VI) CFAR— a procedure that allowed for dynamically se‐ lection between the leading, lagging, or whole of the reference windows— and Switch‐ ing (S) CFAR— a test cell technique that allowed for the selection of representative subsets of the reference cell as compared to the cell under test. We conclude with de‐ veloping a new algorithm: Switching Variability‐Index (SVI) CFAR that combines the ad‐ vantages of both VI and S CFAR.
iv
TABLE OF CONTENTS
Signature Page ………………………………………………………………………………………………………… ii Acknowledgements ………………………………………………………………………………………………... iii Abstract ………………………………………………………………………………………………………………….. iv Table of Contents ……………………………………………………………………………………………………. v List of Tables …………………………………………………………………………………………………………... ix List of Figures ………………………………………………………………………………………………………….. x 1
Introduction ……………………………………………………………………………………………….. 1
1.1 Background …………………………………………………………………………………………… 1
1.2 Objective ………………………………………………………………………………………………. 1
1.3 Investigation Tool …………………………………………………………………………………. 2
2
Background …………………………………………………………………………………………………. 3
2.1 Introduction ………………………………………………………………………………………….. 3
2.2 The Radar System …………………………………………………………………………………. 3
2.3 Doppler Processing ……………………………………………………………………………….. 5
2.4 Radar Signals…………………………………………………………………………………………. 6
2.4.1 Target ……………………………………………………………….……………………. 6
2.4.2 Clutter ……………………………………………………………………………………. 8
2.4.3 Noise ……………………………………………………………………………………… 9
2.4.4 Square Law Detector …………………………………………………..…………. 9
2.5 Radar Signal Environment …………………………………………………………………….. 12
2.5.1 Homogeneous …………………………………………………………………………. 12
v
2.5.2 Multiple Targets ……………………………………………………………………… 13
2.5.3 Clutter Wall …………………………………………………………………………….. 14
2.6 About CFAR …………………………………………………………………………………………... 15
2.6.1 Probability of False Alarm ……………………………………………………….. 15
2.6.2 The CFAR Window …………………………………………………………………… 17
2.6.3 Probability of Detection ………………………………………………………….. 18
2.6.4 CFAR Loss ………………………………………………………………………………… 19
2.7 Monte Carlo Simulations ………………………………………………………………………. 20
3
Cell Averaging CFAR …………………………………………………………………………………. 22
3.1 Implementation ……………………………………………………………………………………. 22
3.2 Homogeneous Environment………………………………………………………………….. 24
3.2.1 Probability of False Alarm………………………………………………………… 24
3.2.2 Detection…………………………………………………………………………………. 25
3.3 Multiple Targets……………………………………………………………………………………… 27
3.4 Clutter Wall…………………………………………………………………………………………… 29
3.4.1 Clutter Wall False Alarm…………………………………………………………… 29
3.4.2 Clutter Wall Detection……………………………………………………………… 31
3.5 Summary: CA‐CFAR Advantages and Disadvantages……………………………….. 31
3.6 A Combat for Multiple Targets: Smallest‐Of Cell Averaging CFAR…...…….. 32
3.7 A Remedy for Clutter Walls: Greatest‐Of Cell Averaging CFAR ……………… 37
4
Variability‐Index CFAR …………………………………………………………………………… 42
4.1 Implementation ……………………………………………………………………………………. 42
vi
4.2 VI‐CFAR Decision Logic…………………………………………………………………………… 45
4.3 Homogeneous Environment…………………………………………………………………… 50
4.3.1 Probability of False Alarm………………………………………………………… 50
4.3.2 Probability of Detection…………………………………………………………… 51
4.4 Masking Targets……………………………………………………………………………………… 53
4.5 Clutter Wall…………………………………………………………………………………………….. 58
4.5.1 Probability of False Alarm………………………………………………………… 58
4.5.2 Probability of Detection…………………………………………………………… 59
5
Ordered Statistics CFAR ……………………………………………………………………………. 61
5.1 Implementation ……………………………………………………………………………………. 61
5.2 Homogeneous ………………………………………………………………………………………. 63
5.2.1 Probability of False Alarm ………………………………………………………. 63
5.2.2 Detection………………………………………………………………………………… 63
5.3 Multiple Targets……………………………………………………………………………………… 66
5.5 Clutter Wall ………………………………………………………………………………………….. 71
6
Ordered Statistics Greatest‐Of CFAR ………………………………………………………… 74
6.1 Implementation ……………………………………………………………………………………. 74
6.2 Homogeneous ………………………………………………………………………………………. 75
6.2.1 Probability of False Alarm ………………………………………………………. 76
6.2.2 Detection………………………………………………………………………………… 77
6.3 Multiple Targets……………………………………………………………………………………… 79
6.5 Clutter Wall ………………………………………………………………………………………….. 83
vii
7
Conclusions and Final Assessment………………………………………………………………. 86
References ……………………………………………………………………………………………………………… 89 Appendix A MATLAB Functions ………………………………………………………………………….. 90
A.1 Radar Return Generation …………………………………………………………………….. 90
A.2 Monte Carlo Simulations ……………………………………………………………………… 92
A.3 Neyman Pearson Detection …………………………………………………………………. 93
A.4 Cell Averaging CFAR ……………………………………………………………………………… 94
A.5 Greatest‐Of Cell Averaging CFAR ………………………………………………………….. 97
A.6 Smallest‐Of Cell Averaging CFAR …………………………………………………………… 100
A.7 Variability‐Index CFAR …………………………………………………………………………… 103
A.8 Ordered Statistics CFAR ………………………………………………………………………... 107
A.9 Ordered Statistics Greatest‐Of CFAR …………………………………………………….. 110
A.10 Switching CFAR ……………………………………………………………………………………. 113
A.11 Switching Variability‐Index CFAR …………………………………………………………. 116
Appendix B
MATLAB Scripts ……………………………………………………………………………….. 120
B.1 Probability of False Alarm in Homogeneous Environment……………………… 120
B.2 Probability of False Alarm in Clutter Wall Environment…………………………. 125
B.3 Probability of Detection in Homogeneous/ Multiple Target Environment. 129
B.4 Probability of Detection in Clutter Wall Environment…………………………….. 134
viii
LIST OF TABLES
Table 2.1 Swerling Targets ………………………………………………………………………………………. 6 Table 2.2 Clutter Types …………………………………………………………………………………………… 8 Table 4.1 VI‐CFAR Decision Logic……………………………………………………………………………… 44 Table 7.1 CFAR Comparison……………………………………………………………………………………… 86
ix
LIST OF FIGURES
Figure 2.1. Radar Block Diagram.…………………………………………………………………………….. 3 Figure 2.2. Doppler Processing………………………………………………………………………………… 5 Figure 2.3. Scattering for Different Swerling Targets……………………………………………….. 7 Figure 2.4. Diode Conductance Characteristic.…………………………………………………………. 11 Figure 2.5. Radar Return Statistics Before and After Square Law Detector ……………… 11 Figure 2.6. Homogeneous Environment…………………………………………………………………… 12 Figure 2.7. Multiple Targets. …………………………………………………………………………………… 13 Figure 2.8. Clutter Edge…………………………………………………………………………………………… 14 Figure 2.9. Exponential Distribution and Thresholding.……………………………………………. 16 Figure 2.10. CFAR Window………………………………………………………………………………………. 17 Figure 2.11. The Threshold. …………………………………………………………………………………….. 18 Figure 2.12. A Typical Probability of Detection Curve illustrating CFAR Loss…………….. 19 Figure 2.13. Monte Carlo with Exponential Distribution.…………………………………………. 20 Figure 2.14. Monte Carlo Weight Function for Exponential Distribution………………….. 21 Figure 3.1. Cell Averaging CFAR Block Diagram.…………………..…………………………………… 22 Figure 3.2. CA‐CFAR Threshold in Homogeneous Environment………………………………… 23 Figure 3.3. CA‐CFAR Theoretical v. Experimental PFA.……………………………………………… 24 Figure 3.4. CA CFAR Homogeneous Probability of Detection.…………………………………… 25 Figure 3.5. CA CFAR Homogeneous CFAR loss………………………………………………………….. 25 Figure 3.6. CA‐CFAR Threshold with 1 Masking Target…………………………………………….. 26 Figure 3.7. CA‐CFAR Probability of Detection with 1 to 3 masking targets……………….. 28
x
Figure 3.8. CA‐CFAR CFAR Loss with 1 to 3 masking targets……………………………………. 28 Figure 3.9. CA‐CFAR Threshold in Clutter Wall Transition………………………………………… 29 Figure 3.10. CA‐CFAR Clutter Wall Probability of False Alarm.…………………………………. 30 Figure 3.11. CA‐CFAR Clutter Wall Probability of Detection……………………………………… 31 Figure 3.12. SOCA‐CFAR Block Diagram……………………………………………………………………. 32 Figure 3.13. SOCA‐CFAR Experimental v. Theoretical PFA………………………………………… 33 Figure 3.14. SOCA‐CFAR Homogeneous PD.……………………………………………….…………….. 34 Figure 3.15. SOCA‐CFAR Homogeneous CFAR Loss ………………………………………………….. 34 Figure 3.16. SOCA‐CFAR 1 Masking Target Probability of Detection………………………… 35 Figure 3.17. SOCA‐CFAR 1 Masking Target CFAR Loss………………………………………………. 35 Figure 3.18. SOCA‐CFAR Clutter Wall Probability of False Alarm……………………………… 36 Figure 3.19. GOCA‐CFAR Block Diagram…………………………………………………………………… 37 Figure 3.20. GOCA‐CFAR Homogeneous Probability of False Alarm………………………… 38 Figure 3.21. GOCA‐CFAR Homogeneous Probability of Detection……………………………. 39 Figure 3.22. GOCA‐CFAR Homogeneous CFAR Loss………………………………………………….. 39 Figure 3.23. GOCA‐CFAR Probability of Detection with 1 Masking Target………………… 40 Figure 3.24. GOCA‐CFAR CFAR Loss with 1 Masking Target……………………………………… 40 Figure 3.35. GOCA‐CFAR Clutter Wall PFA...………………………………………………………………. 41 Figure 4.1. VI‐CFAR Block Diagram…………………………………………………………………………… 42 Figure 4.2. VI‐CFAR ‐ Case I……………………………………………………………………………………… 45 Figure 4.3. VI‐CFAR ‐ Case II……………………………………………………………………………………… 46 Figure 4.4. VI‐CFAR ‐ Case III……………………………………………………………………………………. 47
xi
Figure 4.5. VI‐CFAR ‐ Case IV……………………………………………………………………………………. 48 Figure 4.6. VI‐CFAR ‐ Case V…………………………………………………………………………………….. 49 Figure 4.7. VI‐CFAR Threshold in Homogeneous Environment…………………………………. 50 Figure 4.8. VI‐CFAR Theoretical v. Experimental Homogeneous PFA ………………………… 51 Figure 4.9. VI‐CFAR Probability of Detection Curve …………………………………………………. 52 Figure 4.10. VI‐CFAR CFAR Loss in Homogeneous Environment……………………………….. 52 Figure 4.11. VI‐CFAR Threshold 1 Masking Target……………………………………………………. 53 Figure 4.12. VI‐CFAR Probability of Detection with 1 Masking Target………………………. 54 Figure 4.13. VI‐CFAR PD Curve with varying Numbers of Masking Target…………………. 55 Figure 4.14. VI‐CFAR CFAR Loss Curve with varying Numbers of Masking Target……… 55 Figure 4.15. Three Targets……………………………………………………………………………………….. 56 Figure 4.16. VI‐CFAR Probability of Detection with Three Masking Targets……………… 57 Figure 4.17. VI‐CFAR Threshold in Clutter Wall Transition……………………………………… 58 Figure 4.18. VI‐CFAR Clutter Wall Probability of False Alarm..………………………………….. 59 Figure 4.19. VI‐CFAR Clutter Wall Probability of Detection..…………………………………….. 60 Figure 5.1. OS‐CFAR Block Diagram……………………………………..…………………………………… 61 Figure 5.2. OS‐CFAR Homogeneous PFA.…………………………………..………………………………. 63 Figure 5.3. OS‐CFAR Threshold in Homogeneous Environment.………………………………. 64 Figure 5.4. OS‐CFAR Homogeneous Probability of Detection…………………………………… 65 Figure 5.5. OS‐CFAR Homogeneous CFAR Loss ………………………………………………………… 65 Figure 5.6. OS‐CFAR Resistance to Multiple Targets …..…………………………………………… 66 Figure 5.7. OS‐CFAR Threshold with 1 Masking Target.……………………………………………. 67
xii
Figure 5.8. OS‐CFAR Probability of Detection with 1 Masking Target.………………………. 68 Figure 5.9. OS‐CFAR CFAR Loss with 1 Masking Target ……………………………………………. 69 Figure 5.10. OS‐CFAR Probability of Detection with 3 Masking Target.…………………… 69 Figure 5.11. OS‐CFAR CFAR Loss with 3 Masking Target ………………………………………… 70 Figure 5.12. OS‐CFAR in Clutter Wall Threshold.………………………………………………………. 71 Figure 5.13. OS‐CFAR Clutter Wall Probability of False Alarm…………………………………… 72 Figure 5.14. OS‐CFAR Clutter Wall Probability of Detection…………………………………….. 73 Figure 6.1. OSGO‐CFAR Block Diagram.……………………………………………………………………. 74 Figure 6.2. OSGO‐CFAR Threshold in Homogeneous Environment…………………………… 75 Figure 6.3. OSGO‐CFAR Homogeneous PFA.……………………………………………………………… 76 Figure 6.4. OSGO‐CFAR Homogeneous PD………………………………………………………………... 77 Figure 6.5. OSGO‐CFAR Homogeneous CFAR Loss……………………………………………………. 78 Figure 6.6. OSGO‐CFAR Resistance to Multiple Targets
79
Figure 6.7. OSGO‐CFAR Threshold with 1 Masking Target………………………………………… 79 Figure 6.8. OSGO‐CFAR PD with 1 Masking Target…………………………………………………… 80 Figure 6.9. OSGO‐CFAR CFAR loss with 1 Masking Target………………………………………… 81 Figure 6.10. OSGO‐CFAR PD with 3 Masking Target………………………………………………… 82 Figure 6.11. OSGO‐CFAR CFAR loss with 3 Masking Target……………………………………… 82 Figure 6.12. OSGO‐CFAR Threshold in Clutter Wall Transition…………………………………. 83 Figure 6.13. OSGO‐CFAR Clutter Wall PFA…………………………………………………………………. 84 Figure 6.16. OSGO‐CFAR Clutter Wall PD………………………………………………………………….. 85 Figure 7.1 CFAR Obstacle Course……………………………………..………………………………………. 88
xiii
CHAPTER 1 Introduction 1.1 Background
Radar return is often a mix of noise; clutter; and, if present, targets. A key com‐
ponent in RADAR processing is the setting of detection thresholds. These thresholds differentiate between targets of interest and unwanted radar returns. As operating en‐ vironments and conditions change, the amount and nature of noise and clutter also change. For accurate and reliable detection, the threshold must self‐adjust dynamically and intelligently.
CFAR, constant false alarm rate, represents a key technique in adaptively set‐
ting target detection threshold [1]. Employing a moving window, across range bins of data, CFAR algorithms look at neighborhoods of power returns in the estimate of the noise or clutter mean. By scaling the estimated mean with a pre‐calculated multiplier, the threshold is set as to limit false alarms to a desired rate.
CFAR algorithms are assessed for their abilities to maintain desired probabilities
of detections (PD) and their probabilities of false alarm (PFA). The probability of detec‐ tion describes the chances of successfully declaring a target, when a target is actually present. The probability of false alarm describes the odds of incorrectly declaring a tar‐ get, when the signal is, in actuality, noise or clutter.
Developed CFAR algorithms must be able to operate in a variety of canonical
environments: homogeneous, multiple targets, and clutter wall. In the homogeneous environment, a single target exists in a sea of clutter or noise of uniform noise. In the
1
multiple target exists, several targets exist in close proximity to one another. In the clutter wall environment, noise and/or clutter power experiences sudden, discontinu‐ ous increases or decreases. 1.2 Objective
The goal of this study is to investigate and assess the variety of existing CFAR
algorithms and judging their performance in homogeneous, multiple target, and clut‐ ter wall environments. Towards this, the Ordered Statistics Greatest‐Of CFAR (OSGO‐ CFAR), Variability Index CFAR (VI‐CFAR), and Switching CFAR (S‐CFAR) are studied. While Cell Averaging CFAR (CA‐CFAR) and Ordered Statistics CFAR (OS‐CFAR) are also looked at, they are used primarily as basis of comparisons for the algorithms of inter‐ est. We end with development of our own CFAR technique, Switching Variability Index CFAR (SVI‐CFAR) that represents a hybrid approach between VI‐CFAR and S‐CFAR. 1.3 Investigation Tool
Radar signals and CFAR algorithms are assessed through Monte Carlo, numeri‐
cal experiments. MATLAB was chosen as the investigation tool (Version 7.10.0, R2010a).
2
CHAPTER 2 Background 2.1 Introduction
Since their development during World War II, radar had grown to occupy criti‐
cal civilian and military roles. An introductory understanding of radar processing and signals is necessary before getting into the analysis and assessment of CFAR algo‐ rithms. 2.2 The Radar System
Originally acronymic as RADAR— for RAdio Detection And Ranging, common
usage had established the word in its own right, as radar— a free‐standing noun. The original word, though, helps remind of the technology’s dual purpose: To detect tar‐ gets of interests and to ascertain their distance.
Figure 2.1. Radar Block Diagram.
3
Figure 2.1 shows a basic block diagram of a RADAR system. The particular sys‐
tem is monostatic— where a single antenna is multiplexed for both transmission and reception [2].
The waveform generator forms a timing pulse to trigger the transmitter to send
an electromagnetic pulse at a designed frequency. As this pulse is sent it, it will en‐ counter different objects at varying distances, and these objects will scatter back some of these pulses back to the antenna.
The timing and frequency of the return pulse corresponds to the relatives
speeds and distance of the reflected target. The subsequent processing blocks— the A/D converter, the pulse compression, and the Doppler processor all helps extract the relevant information.
Pulse repetition interval (PRI) refers to the period of time before the successive
radar pulses. Pulse repetition frequency (PRF) is simply the inverse of the PRI. During the period between successive pulses, during a given pulse repetition interval, the ra‐ dar system waits for back‐echoed pulses. Detected echoes are detected, binned and stored.
The amount of time a sent pulse is reflected back to the radar antenna is di‐
rectly proportional to the distance of the reflector: ct R 0 2 R : range (distance) c 0 : speed of light t: time 4
Figure 2.2. Doppler Processing.
2.3 Doppler Processing
With doppler processing, a radar system can not only detect target range and
location, but also the radial velocity, the range rate.
A single target may be looked at several times over several pulse intervals . The
change in power over a given range bin, or distance, across many pulse intervals may give a sense of the doppler shift, and hence, the speed of the target (Figure 2.2).
When discrete Fourier transform (DFT), through fast fourier transform (FFT), is
performed on each range sample over its many range intervals, the pulse intervals are converted to doppler intervals (also called doppler bins).
5
2.4 Radar Signals
Radar signals are often composed of unwanted noise, undesired clutter returns,
and a target of interest. Over a short period of time, each signal elements possesses its own stochastic distribution with a given power. Statistical characteristics of the signals change as they pass through a square law detector. 2.4.1 Target
The power of a radar return depends on the distance and the effective radar
cross section (RCS) of the target. Over distance r, the power of the signal diminishes by a factor r4. Many factors of the radar, the target, and the operating environment come together to determine the target’s RCS: the illumination angle, the frequency, the po‐ larization of the transmitted wave, target motion, vibration, and kinematics of the ra‐ dar itself. Rather than hopelessly describing a target’s RCS deterministically, they’re characterizing stochastically.
Building upon Marcum’s pioneering studies, Peter Swerling developed four ca‐
nonical stochastic models: Swerling I, II, III, and IV. With these Swerling models, the re‐
Table 2.1. Swerling Targets Model Swerling I
Distribution Rayleigh
Swerling II Swerling III Swerling IV
p x
1
av
e
x
av
Notes
Scan‐to‐scan
Characterizes well targets like planes, with many independent Pulse‐to‐Pulse reflectors
Characterizes well targets like Chi‐Squared 2 x Scan‐to‐scan frigates, with one strong reflector (Four degrees p x 4 x e av Pulse‐to‐Pulse and many smaller, weaker inde‐ of freedom) 2 av pendent reflectors
Swerling V Non‐fluctuating
Also called Swerling 0 or Marcum’s Model 6
: Many scatter elements; each of approximately equal amplitudes.
(a)
: One scattering element producing an amplitude much larger than other elements.
(b)
Figure 2.3. Scattering for Different Swerling Targets. The Swerling models assumes each tar‐ get composed of a large number of independent scatters. (a) Swerling I/II. For Swerling targets I and II, none of the independent scattering elements dominates: their reflected powers is comparable to one another. (b) Swerling III/IV. For Swerling targets III and IV, there is a single dominant non‐fluctuating scatter together with other smaller scatters.
turn power of a target will fluctuate— randomly increase and decrease— with a char‐ acteristic probability density function (PDF) with either quickly or slowly (Table 2.1).
The Swerling I and II models both characterized with a Rayleigh distribution
and assumed to be composed of many independent small reflectors; and with none dominating, they each return radar signals of about the same power. Swerling I models fluctuate slowly— from scan‐to‐scan (Figure 2.3a). Each pulse return is highly corre‐ lated. Swerling II models, meanwhile, fluctuate quickly— varying from pulse‐to‐pulse. It is claimed that observed data on airborne aircraft targets agree with Swerling Models I and II.
Swerling III and IV models are described by the Chi‐squared distribution with
four degrees of freedom. They are conceptualized as many independent reflectors of roughly equal amplitude with one dominating reflector (Figure 2.3b). The Swerling III model varies slowly— from scan to scan while the Swerling IV model fluctuates 7
quickly— from pulse to pulse. The Swerling III/IV model is said to well‐characterize a frigate or corvette with one large target but with complex superstructures. 2.4.2 Clutter
Clutter is unwanted radar returns. They may be a echoes from the environ‐
ment— as from ground, mountain, sea— or objects of no importance— as a flock of geese. Designation of clutter changes with the aims and designs of the radar system. A doppler weather system, for instance, may regard radar returns from a storm cloud with prime importance while the same cloud may only be interference to a civilian air traffic control radar. The stochastic characteristic of clutter varies with the operating parameters of the radar as well as the nature of the clutter source (Table 2.2).
Older radar systems had longer pulse repetition intervals and larger range bins.
It was observed that ground clutter was well characterized with a Gaussian distribu‐ tion. However, as technology improved, and shorter pulse repetition intervals gave greater range resolution, large spikes in amplitude were observed in radar data. The
Table 2.2. Clutter Types Model
Application
Equation
Gaussian
Clouds
p x
Weibull Clutter
Land Clutter
p x
K‐distribution
Sea Clutter
p x
8
x 2 exp 2 2 2 2 1
4 h v 1 v x K v 1 2 hx v
cx bb
c 1
x c b
exp
Gaussian distribution no longer fitted.
Studies during the late 70’s and early 80’s found that ground clutter were more
accurately characterized with a Weibull distribution. Unlike the Gaussian distribution, the Weibull is characterized by two variables: a shaping coefficient and a mean coeffi‐ cient.
In contrast with ground clutter, sea clutter is highly correlated. The two‐
parameter k‐distribution well‐characterizes this. 2.4.3 Noise
Gaussian white noise results from random thermal motion in electronic sys‐
tems. Unlike target and clutter returns, noise is always present. As implied by its name, Gaussian noise is characterized by the bell‐shaped Gaussian distribution. 2.4.4 Square Law Detector
The stochastic characteristic of received radar signals may change at the diode
detector. The diode conductance characteristic is shown figure 2.4. Depending on the strength of the received signal, the detector may either operate in the conductance square or in the linear region. Each offers its advantages. Marcum showed that if the signal is low, lying below the conductance knee, the square law detector gave the best results. Linear detectors, meanwhile, offers the greatest dynamic range and is best for high signals. Mathematically, in analysis, the square law detector offers additional ad‐ vantages.
9
Both Rayleigh and Gaussian PDF’s, passing through the square law detector,
become exponential distributions (Figure 2.5). Given the relative simplicity of integra‐ tion of the exponential function, this fact comes in handy in analytical treatments of target detection and false alarm rates.
10
30 Square Law Linear Law 25
i
20
15
10
5
0
0
1
2
3
4 v
5
6
7
8
Figure 2.4. Diode Conductance Characteristic. Radar detection can either occur in the linear or the square region of the diode.
(a)
(b)
Figure 2.5. Radar Return Statistics Before and After Square Law Detector . The square law detector changes stochastic PDF of the signal.. (a) Gaussian Distribution. For Swerling targets I and II, none of the independent scattering elements dominates: their reflected powers is comparable to one another. (b) Exponential Distribution. For Swerling targets III and IV, there is a single dominant non‐fluctuating scatter together with other smaller scatters.
11
50
100 Target 1 Clutter
90
40
80
35
70
30
60
Magnitude
Magnitude
45
25 20
50 40
15
30
10
20
5
10
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
(a)
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
(b)
Figure 2.6. Homogeneous Environment (a) There is just one target of power 25dB at range gate 37. The noise power is uniform at 5dB. (b) MATLAB simulation of homogeneous environment.
2.5 Radar Signal Environment
A number of different simplified RADAR environments have been used to assess
and judge the efficacy and applicability of the various CFAR algorithm. Although no one model captures the full complexity and nuance of actual operating conditions, they each test important qualities of the CFAR algorithm: The homogeneous environment, the multiple target environment, and the clutter edge environment. 2.5.1 Homogeneous
The homogeneous environment is the most straight‐forward: It assumes a sin‐
gle target against an independent and identically distributed (iid) noise or clutter (Figure 2.6).
A homogeneous environment may arise when, for instance, the radar antenna
makes no detection, and noise of a uniform power dominates the range bins; or when 12
50
100 Target 1 Target 2 Clutter
45
80
35
70
30
60
Magnitude
Magnitude
40
90
25 20
50 40
15
30
10
20
5
10
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
0
0
10
20
30
(a)
40 50 60 Range Gates
70
80
90
100
(b)
Figure 2.7. Multiple Targets. (a) At range gates 37 and 42, there are targets of power 25dB. (b) MATLAB simulation of multiple targets with exponential distribution.
the clutter return over a great stretch of distance is of a constant environment (say, a forest, mountain, or ocean).
The homogeneous model assesses the ability of various CFAR algorithms as the
simplest test for properly assessing its ability to estimate the noise/clutter mean in set‐ ting the threshold of detection. 2.5.2 Multiple Targets
Targets of interest, however, don’t always exist in isolation along the range
gates of a given doppler bin. By chance or design, targets may be within close proximity of one another (Figure 2.6).
Some CFAR algorithms suffer serious performative decline in multiple target
cases. When the targets are within half a window length of one another, their high powers, improperly elevates the estimated mean of the background noise/clutter. 13
50
100 90
40
80
35
70
30
60
Magnitude
Magnitude
Clutter 45
25 20
50 40
15
30
10
20
5
10
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
(a)
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
(b)
Figure 2.8. Clutter Edge. (a) At range gate 41, there is a sudden jump in clutter power; from 5dB to 10dB. (b) MATLAB simulation of a clutter edge with exponential distribution.
When the resulting thresholds are set, they are higher than they lead to be, and result in losses in probability of detection (PD). 2.5.3 Clutter Wall
In addition to one‐target assumption, the iid assumption of the background
noise/clutter may also be violated. Radar environments often undergo abrupt changes in power— such as transitioning from clear land to a forest, or from clouds to the clear sky. These sudden jumps in mean power may throw off the CFAR algorithm, and intol‐ erably increase the probability of false alarm following the upward clutter wall transi‐ tion or decrease the probability of detection right before the clutter wall (Figure 2.8).
14
2.6 About CFAR
The goal of CFAR— constant false alarm rate— algorithms is to set thresholds
high enough to limit false alarms to a tolerable rate, but low enough to allow target detection. In the past, this differentiation between target and background noise and clutter were done by human radar operators upon a radar display screen. However, with the advent of computer technology, detection thresholding was able to be auto‐ mated, and thus bypassing much of the cost, inefficiencies, and limitations of a human operator. 2.6.1 Probability of False Alarm
The probability of false alarm, PFA, is the chance that spikes in noise or clutter is
mistaken by the CFAR algorithm as a target. As noise and clutter distributions are con‐ tinuous, extending from amplitudes very close to zero to amplitudes extending infi‐ nitely outwards. No matter how high thresholds are set, there will always be a finite chance of random noise or clutter exceeding that threshold. So rather than eliminating false alarms all together (this would be impossible), the goal of CFAR algorithms is to reliably estimate the mean noise, and scaling the estimated mean by a multiplier to obtain the threshold set high enough to limit false alarm rate to a tolerably small rate.
In many operating environments, the shape of the noise/clutter is known a
prior. From off‐line computation or from the experience of the engineering, the prob‐ ability density function (PDF) of the noise can be known ahead of time. For instance, land clutter is often‐characterized by the Weibull distribution and Clouds by the Gaus‐
15
Probability Distribution
Probability of False Alarm
Mean
Threshold
Noise Power
(a) Noise Mean
x
Threshold
Threshold Multiplier
(b) Figure 2.9. Exponential Distribution and Thresholding. (a) Exponential PDF. (b) To get the threshold, the mean is scaled by a threshold multiplier.
sian distribution.
Knowing the distribution, then, a measurement of the mean may be properly
scaled with a threshold multiplier for a threshold that yields the desired probability of false alarm (Figure 2.9)
So the question becomes: How to estimate the mean noise or clutter over sev‐
eral range bins? The answer: The CFAR window
16
Cell Under Test
Leading Window
Guard Cells
Lagging Window
Figure 2.10. CFAR Window
2.6.2 The CFAR Window
To estimate the mean noise/clutter present in a specific range bin, other local
bins are used. Towards this end, most CFAR algorithms utilizes a moving window. There are several components to the CFAR window (Figure 2.10).
The cell under test (CUT)— also called the test cell— is the range bin with which
the mean noise/clutter is estimated and with which the threshold is set.
To estimate the mean, references cells— cells local to the CUT— are used. Dif‐
ferent CFAR algorithms utilize different mathematical assessments and different deci‐ sions logical towards this end. Cell averaging CFAR, for instance, takes the mean of the reference cells while Ordered Statistics CFAR orders the reference cells from smallest to largest to take the kth largest as representative of the average noise.
Guard cells are optional and may be variable in length. A target, if present in
the CUT, may straddle consecutive; and this would lead to inaccurate estimates of the noise. Guard cells then, would guarantee then, guard against this. Typically, one guard cell is used to either side of the CUT.
17
30
30 Threshold Recieved Waveform
Threshold Recieved Waveform
25
25
Missed Detection
20 Magnitude (dB)
Magnitude (dB)
20
15
15
10
10
5
5
0
0
10
20
30
40 50 60 Range Gates
70
Detection!
80
90
0
100
0
10
20
30
(a)
40 50 60 Range Gates
70
80
90
100
(b)
Figure 2.11. The Threshold. (a) Missed detection. The target power falls underneath the threshold. (b) Successful detection. Target power greater than threshold.
2.6.3 Probability of Detection
The CFAR reference window is used to set the threshold (Figure 2.11). The
threshold helps differentiate between target a noise. If the signal power, for a given range cell, lies underneath the threshold, a target is deemed absent. If the same signal power exceeds the height of the threshold, then a target is declared present.
Since target powers are typically stochastic— with a Swerling I/II or Swerlign
III/IV distribution— the probability of detection is also stochastic. Around a mean value, the received target power would fluctuate to larger than the mean or smaller than the mean according to its characteristic PDF. Sometimes, it would flucturate as low as to miss the mean. So regardless to how well the threshold is set for a given probability of false alarm, there will always be the chance of missing a target, if it’s pre‐ sent. The probability of detection helps characterize the odds of detection over many trial runs or over many looks. 18
Probability of Detection - Homogeneous Environment
1 0.9
Probability of Detection
0.8 0.7 0.6
CFAR Loss
0.5 0.4 0.3 0.2
Neyman-Pearson CA-CFAR
0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
25
Figure 2.12. A Typical Probability of Detection Curve illustrating CFAR Loss
2.6.4 CFAR Loss
Although the PD is a good measure of CFAR performance, it is relative to operat‐
ing conditions like the signal to noise ratio (SNR). So rather than PD, CFAR loss is the usual ruler used to judge CFAR performance.
Figure 12.12 shows a typical probability of detection curve. As the SNR in‐
creases, the probability of detection also increases. The Neyman‐Pearson detector represents the theoretical best detection— if perfect knowledge of the mean noise is known a priori. It is against the Neyman‐Pearson detector that the CFAR loss is calcu‐ lated.
CFAR loss is the SNR difference, for a given PD between the Neyman‐Pearson
detector and the CFAR algorithm in interest. For instance, in Figure 12.12, for a prob‐ ability of detection of 0.4, the CFAR loss is close to 1dB of SNR. 19
f x
1
2
x exp 2
fm x
1
m
2
x exp 2 m
Q f x dx PFA 10 106 Th
Th
Figure 2.13. Monte Carlo with Exponential Distribution.
2.7 Monte Carlo Simulations
Frequently, desired probability of false alarms are very small— on the order of
10‐5 or 10‐6. Then, in MATLAB, during a CFAR numerical experiment, a great number of trial runs— on the order of 108 or 109— are necessary before a reliable experimental probability of false alarm can be ascertained. This often intolerably increases MATLAB run‐time or computer memory demands. Monte Carlo Importance Sampling repre‐ sents a class of numerical techniques used to reliably simulate very rare events [3].
Rather than using the original distribution to simulate events, with Monte Carlo
Importance Sampling, a similar function but with a different mean is used, but with the same threshold (Figure 2.13). Any successful hits, then, with the new similar function, rather than having a weight of one, is fractionally weight with a pre‐derived equation (Figure 2.14).
20
w x
m2 1 1 exp 2 2 x 2 m w x : The weighing function for successful Monte Carlo "hits" x : Random varaible with uniform distribution
2 : Variance of the estimated exponential distribution m 2 : Variance of the Monte Carlo exponential distribution Figure 2.14. Monte Carlo Weight Function for Exponential Distribution.
21
Chapter 3 Cell Averaging CFAR
Cell Averaging CFAR (CA‐CFAR) was developed early on, in 1947, by Howard
Finn [1]. Of all the CFAR algorithms, CA‐CFAR works best in a homogeneous environ‐ ment. With increasing window size, the probability of detection for CA‐CFAR ap‐ proaches that of the optimal Neyman‐Pearson detector.
However, when the homogeneous assumptions are violated, the performance
of CA‐CFAR rapidly deteriorates [4]. As a starting point in algorithmic development and as a basis of comparison, the algorithm is studied here. 3.1 Implementation
In Cell Averaging CFAR, every reference cell is added together and multiplied to
the threshold multiplier (Figure 3.1). For Gaussian noise and a square law detector, the
Σ
Σ
Σ x
PFA
Threshold
1 N
1
Threshold Multiplier
Figure 3.1. Cell Averaging CFAR Block Diagram.
22
threshold multiplier is given by:
PFA
1 N
1
Variable α is the threshold‐multiplier, N is the total number of reference cells, and PFA is the desired probability of false alarm.
Typically, PFA is designed to be very low, on the order of 10‐5 or 10‐6. The refer‐
ence window length, N, must be selected by the designer as a balance of performance in homogeneous and heterogeneous environments.
A cell averaging threshold with PFA = 10‐4 and N = 32 is simulated for a single‐
target in a homogeneous environment (Figure 3.2). A single target Swerling I/II target is in range bin 37 in homogeneous noise. SNR is set to 15dB. We see that the threshold,
Magnitude (dB)
20
15
10
5 CA-CFAR Recieved Waveform 0
0
10
20
30
40 50 60 Range Gates
70
80
90
Figure 3.2. CA‐CFAR Threshold in Homogeneous Environment.
23
100
across the noise‐only range bin, makes good clearance from the noise. There is a suc‐ cessful detection at the target. To either sides of the target in range bin 37 are elevated threshold plateau— these are artifacts from target presence. This is a characteristic of CA‐CFAR. Although it poses no problems in the single‐target homogeneous environ‐ ment case, with multiple targets, it negatively effects PD. 3.2 Homogeneous Environment 3.2.1 Probability of False Alarm
The primary function of any CFAR is to set the threshold to maintain the desired
probability of false alarm. To confirm the workability of the algorithm in a homogene‐ ous environment, we performed 10,000 trials for various designed PFA (Figure 3.2). The experimental PFA closely matched the expected theoretical results. 0
10
-1
10
Probability of False Alarm
-2
10
-3
10
-4
10
-5
10
-6
10
Theoretical PFA Experimental P FA
-7
10
2
4
6
8
10 12 14 Threshold Multiplier
16
18
Figure 3.3. CA‐CFAR Theoretical v. Experimental PFA.
24
20
22
3.2.2 Detection
Cell Averaging CFAR successfully sets thresholds to keep to desired false alarm
rates, but does this same threshold allow for good probability of detection? We first look at CA‐CFAR in a homogeneous environment— with a single target in a background of noise or clutter with uniform mean . The threshold was set for a single trial run in figure 3.1. In figure 3.4, we see the probability of detection profile over a range of SNR. For this data set, the CA‐CFAR window size was set to 32, and a million trial per SNR was performed. The CA‐CFAR PD curve closely matches to the contours of the Neyman‐ Pearson detector. In figure 3.5, we see that the CFAR loss is much less than 1dB.
25
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 Neyman-Pearson CA-CFAR
0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.4. CA‐CFAR Homogeneous Probability of Detection.
1.5 CA-CFAR
CFAR Loss (dB)
1
0.5
0
5
10
15 SNR (dB)
20
Figure 3.5. CA‐CFAR Homogeneous CFAR loss
26
25
25
Magnitude (dB)
20
15
10
5 CA-CFAR Recieved Waveform 0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 3.6. CA‐CFAR Threshold with 1 Masking Target
3.2.2 Multiple Targets
Typical to CA‐CFAR, as seen in figure 3.1, to either side of a detected targets are
elevated threshold plateaus. When another target falls within half a window length of it, each target’s probability of detection significantly decreases (Figure 3.6). This phe‐ nomena is referred to as “target masking.”
Detection performance diminishes drastically with numbers of interfering tar‐
gets (Figure 3.7). For CA‐CFAR with N=32, at SNR=20dB, the PD for a single target is 0.9. With one masking target, the PD drops to 0.7, and with three masking targets PD sinks to 0.4. The masking effect of multiple targets worsens with increasing signal to noise ratio (Figure 3.8). Whereas the CFAR loss for one interfering target is less than 2dB at
27
1 Neyman Pearson homogeneous one masking target two masking targets three masking targets
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.7. CA‐CFAR Probability of Detection with 1 to 3 masking targets
15 homogeneous one masking target two masking targets three masking targets
CFAR Loss (dB)
10
5
0
5
10
15 SNR (dB)
20
Figure 3.8. CA‐CFAR CFAR Loss with 1 to 3 masking targets
28
25
30
25
Magnitude (dB)
20
15
10
5 CA-CFAR Recieved Waveform 0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 3.9. CA‐CFAR Threshold in Clutter Wall Transition.
SNR=10dB, at SNR=20dB, the CFAR loss increases to 6B. 3.4 Clutter Wall
The clutter wall poses two different, but complementary challenges for CA‐
CFAR. Target masking occurs at the base of the clutter edge, and probability of false alarm rates increase drastically following a clutter wall edge. 3.4.1 Clutter Wall False Alarm
Figure 3.10 summarizes the false alarm rates for various distances before and
after the clutter wall. The pink dash line at 10‐4 indicates the designed false alarm rate.
29
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
-7
10
CA-CFAR
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 3.10. CA‐CFAR Clutter Wall Probability of False Alarm.
Simulated for N=32, the blue line at 16 demarcates when half the reference cells are immersed in clutter.
PFA decreases drastically in the cells before the clutter wall. Having a false alarm
rate smaller than the designed rate, at these regions, isn’t a concern. The problem oc‐ curs following the clutter wall, at reference cells 16 and outwards in Figure 3.10. At ref‐ erence cell 16, PFA reaches 10‐2; a false alarm rate nearly 100 times greater than the desired PFA at 10‐4! This degree of deviance from desired PFA is usually intolerable for a radar system’s designs; and other CFAR techniques were designed, in part, to remedy this deficiency.
30
1 CA-CFAR N/2 cells in clutter
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 3.11. CA‐CFAR Clutter Wall Probability of Detection
3.4.2 Clutter Wall Detection
The detection at clutter walls is complementary to its false alarm effects. With
CA‐CFAR, there is a gradual decrease of PD from before the wall to after it (Figure 3.11). Before the wall, from reference cells 1 to 16, the diminishment in PD can be attributed to a masking effect by the clutter region. The decrease in PD after the clutter wall can be accounted by the clutter to noise ratio (CNR) being smaller than the SNR. 3.5 Summary: CA‐CFAR Advantages and Disadvantages
Although CA‐CFAR performs very well in homogeneous environments, its detec‐
tion performances suffers drastically with closely spaced multiple targets. At the clutter wall, false alarm probabilities increase by as much as an order of 102. This renders the algorithm all but useless in clutter wall environments.
31
3.6 A combat for Multiple Targets: Smallest‐Of Cell Averaging CFAR
Smallest‐Of Cell Averaging (SOCA) CFAR was developed to remedy Cell‐
Averaging CFAR’s deficiencies in multiple target situations. If a secondary target intol‐ erably increases the average power of the leading or lagging window, then SO‐CFAR simply takes the smaller of the two. [5]. SOCA‐CFAR works by selecting the smaller of the sum of the leading and lagging windows (figure 3.12). This procedure requires a threshold multiplier different from CA‐CFAR:
1 PFA 2 SO 2 N /2
N /2
N2 1 N k SO 1 k 2 2 N /2 k 0 k
As there is no analytical solution for αSO, it is to be solved analytically from the above equation.
In the homogeneous case, the algorithm can successfully hold for the designed
probability of false alarm (Figure 3.13). Its Probability of Detection and CFAR loss is only slightly worse than CA‐CFAR (Figures 3.14, 3.15). The difference can be accounted
Σ
Σ
Min x
Threshold
Threshold Multiplier
Figure 3.12. SOCA‐CFAR Block Diagram
32
p
FA
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
-7
10
-8
10
0
Theoretical PFA Experimental PFA 5
10
15 20 Threshold Multiplier
25
30
Figure 3.13. SOCA‐CFAR Experimental v. Theoretical PFA
by SOCA‐CFAR using only half its reference windows (the smaller of the lead and lag) to estimate the noise or clutter mean.
In probability of detection, for the multiple‐target case, SOCA‐CFAR out‐
performs CA‐CFAR (Figure 3.16). SOCA‐CFAR follows the shape of the Neyman‐Pearson detector. Improvement over CA‐CFAR begins past SNR = 10dB for SOCA‐CFAR. In both PD and CFAR loss, SOCA‐CFAR is similar to Ordered Statistics CFAR (OS‐CFAR) in its ad‐ vantages over CA‐CFAR (Figure 3.17).
33
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR SOCA-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.14. SOCA‐CFAR Homogeneous PD. 1.5
CFAR Loss (dB)
1
0.5
CA-CFAR OS-CFAR SOCA-CFAR 0
5
10
15 SNR (dB)
20
Figure 3.15. SOCA‐CFAR Homogeneous CFAR Loss
34
25
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR SOCA-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.16. SOCA‐CFAR 1 Masking Target Probability of Detection
10 9
CA-CFAR OS-CFAR SOCA-CFAR
8
CFAR Loss (dB)
7 6 5 4 3 2 1 0
5
10
15 SNR (dB)
20
Figure 3.17. SOCA‐CFAR 1 Masking Target CFAR Loss
35
25
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
CA-CFAR OS-CFAR SOCA-CFAR
-7
10
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 3.18. SOCA‐CFAR Clutter Wall Probability of False Alarm
Whereas SOCA‐CFAR showed resistance to target masking, it is especially vul‐
nerable to the clutter wall environment (Figure 3.18). Near the clutter edge transition, the probability of false alarm is elevated by over 103 times over that of the design point! This is over 10 times worse than CA‐CFAR.
In practice, this clutter wall weakness of SOCA‐CFAR renders it unpractical.
However, as we will see in the next chapter, the philosophy of SOCA‐CFAR may be use‐ ful if utilized in a environment‐specific fashion.
36
3.7 A Remedy for Clutter Walls: Greatest‐Of Cell Averaging CFAR
Greatest‐Of Cell Averaging (GOCA) CFAR was a method designed to avoid the
elevated false alarm problem of clutter walls [6]. The algorithm works by summing each of the leading and lagging window and taking the larger of the two to be used with the threshold multiplier (Figure 3.19).
The threshold multiplier can be solved from this equation: PFA
GO 1 1 2 N /2
N /2
1 2 GO 2 N /2
N /2
N2 1 N k 2 1 k 2 GO N /2 k 0 k
Since an analytical equation doesn’t exist for αGO, it must be solved numerically.
In a homogeneous environment, the threshold multiplier was shown to be suc‐
cessfully maintaining the desired false alarm rate (Figure 3.20).
The probability of detection profile for GOCA‐CFAR in the homogeneous envi‐
ronment closely matches CA‐CFAR (Figure 3.21). The small additional CFAR loss of
Σ
Σ
Max x
Threshold
Threshold Multiplier
Figure 3.19. GOCA‐CFAR Block Diagram
37
FA
0
10
-1
10
Probability of False Alarm
-2
10
-3
10
-4
10
-5
10
-6
10
Theoretical PFA Experimental PFA
-7
10
2
4
6
8
10 12 14 Threshold Multiplier
16
18
20
Figure 3.20. GOCA‐CFAR Homogeneous Probability of False Alarm
GOCA‐CFAR can be accounted for by the fact that the CFAR only averages from half the available reference windows (the leading or lagging) as opposed to CA‐CFAR that makes use of the full set of reference windows.
In the masking target case, GOCA‐CFAR performs worse than CA‐CFAR (Figure
3.23). The algorithm suffers as much as a 2dB CFAR loss over CA‐CFAR (Figure 3.24). If a masking target exists, it will either elevate the average of the leading or lagging win‐ dow. Since GOCA‐CFAR takes the larger of the two windows, it will always take the win‐ dow artificially elevated by the masking target.
It is in the clutter wall environment that GOCA‐CFAR vast outperforms that of
CA‐CFAR (Figure 3.25). At the clutter wall, GOCA‐CFAR only suffers around a 10‐fold increase in PFA while CA‐CFAR suffers a 100‐fold increase. Over CA‐CFAR, this translates
38
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR GOCA-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.21 GOCA‐CFAR Homogeneous Probability of Detection
g
1.5
CFAR Loss (dB)
1
0.5
CA-CFAR OS-CFAR GOCA-CFAR 0
5
10
15 SNR (dB)
20
Figure 3.22 GOCA‐CFAR Homogeneous CFAR Loss
39
25
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR GOCA-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 3.23. GOCA‐CFAR Probability of Detection with 1 Masking Target
12
CFAR Loss (dB)
10
CA-CFAR OS-CFAR GOCA-CFAR
8
6
4
2
0
5
10
15 SNR (dB)
20
Figure 3.24. GOCA‐CFAR CFAR Loss with 1 Masking Target
40
25
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
CA-CFAR OS-CFAR GOCA-CFAR
-7
10
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 3.25. GOCA‐CFAR Clutter Wall PFA
to roughly a 10‐fold improvement for GOCA‐CFAR.
GOCA‐CFAR is, in many ways, the complement to SOCA‐CFAR. Whereas SOCA‐
CFAR improves over CA‐CFAR in multiple targets but suffers in clutter wall, GOCA‐CFAR suffers in multiple targets but improves in clutter wall.
41
Chapter 4 Variability‐Index CFAR
Variability‐Index CFAR (VI‐CFAR) dynamically switches between CA‐CFAR, SOCA‐
CFAR , and GOCA‐CFAR depending on the mean and the distribution of cells in the lead‐ ing and the lagging windows [7]. This approach captures particular advantages of each of the three CFAR approaches while avoiding their respective weaknesses. 4.1 Implementation
For each of the leading window and the lagging window, the mean and the vari‐
ability index is computed.
The variability index (VI) is a second‐order statistics that is closely related to an
estimate of the shape parameter. The greater the variability index, the greater the vari‐ ability in relatives magnitudes of the reference cells. When, for instance, a leading window contains the high powers of targets, the variability would differentially weigh
Σ
Σ
Σ(·)2
Σ(·)2 Logic
x
Threshold
Threshold Multiplier
Figure 4.1. VI‐CFAR Block Diagram
42
the target power over that of the lower powers of the noise and magnitude, resulting in a larger variability index.
The variability index may be calculated by this equation:
n
VI n
X
2
i
i 1
n Xi i 1
2
Here, Xi is the power of reference cell i. The summation is taken over all cells Xi of ei‐ ther the leading or the lagging window.
Once the VI is computed, the algorithm then makes a determination of whether
the leading and/or the lagging windows are variable. If the variability index is smaller than variability index constant KVI, then the window is declared not variable. If the vari‐ ability index is larger than KVI, then the window is declared variable:
VI K VI Not Variable VI K VI Variable
The Mean Ratio (MR) is a measure of how different the leading and lagging win‐
dow means are. It is computed as a ratio of the summations of the reference cells of each window:
X MR A XB
X X i A iB
i
i
If the MR falls within bounds defined by the mean ratio constant, KMR, then the leading and lagging window means are declared the same:
K MR 1 M R K MR Same M ean
43
Table 4.1. VI‐CFAR Decision Logic Leading Window Variable?
Lagging Window Variable?
Different Mean?
VI‐CFAR Adaptive Threshold
Equivalent CFAR Method
Illustrative Figure
No
No
No
αN ∙ ΣAB
CA CFAR
Figure 4.2 Case I
No
No
Yes
αN ∙ max(ΣA,ΣB)
GOCA CFAR
Figure 4.3 Case II
Yes
No
—
αN ∙ ΣB
CA CFAR
Figure 4.4 Case III
No
Yes
—
αN ∙ ΣA
CA CFAR
Figure 4.5 Case IV
Yes
Yes
—
αN ∙ min(ΣA,ΣB)
SOCA CFAR
Figure 4.6 Case V
If not, then they are declared as different means:
M R K MR 1 or M R K M R Different M ean
Once the variability and mean determinations had been made, then, according
to the logic of Table 4.1, CFAR is performed with either both the leading and lagging windows, one of them, the smaller of the two, or the larger of the two.
If only a half‐window is chosen— either only the leading or lagging window—
then the threshold multiplier becomes
N / 2 PFA
1 N /2
1
If both leading and lagging windows are chosen, as in case I, then the threshold multi‐ plier becomes:
N PFA
1 N
1
A little more can be explained of the different cases of Table. 4.1
44
Lagging Window: ‐ Not Variable ‐ Mean similar to Lagging
Leading Window: ‐ Not Variable ‐ Mean similar to Leading
Situation:
Homogeneous Environment CFAR to use:
100
100
90
90
80
80
70
70
Target 1 Clutter Cell Under Test
60
Power (linear)
Power (linear)
Cell Averaging CFAR with all reference cells
50 40
60 50 40
30
30
20
20
10
10
0
0
10
20
30
40 50 60 Range Gates
70
80
90
0
100
Signal Cell Under Test
0
10
20
30
40 50 60 Range Gates
70
80
90
Figure 4.2. VI‐CFAR ‐ Case I
4.2 VI‐CFAR Decision Logic Case I: Same means, both nonvaraible
When both the leading and lagging windows are nonvaraible, they aren’t likely
to contain targets; and when both windows have similar means, then the reference window probably doesn’t straddle a clutter edge. The cell under test, then, is most likely in a homogeneous environment. VI‐CFAR, thus, uses CA‐CFAR.
The threshold is set by summing all reference cells and using the CA‐CFAR
threshold multiplier for length N:
N PFA
1 N
1
45
100
Lagging Window: ‐ Not Variable ‐ Mean similar to Leading
Example Situation:
Leading Window: ‐ Variable ‐ Mean similar to Lagging
Interfering targets in the leading window. CFAR to use:
Cell Averaging CFAR with leading window only
150
200 180 160
Signal Cell Under Test
140 Power (linear)
Power (linear)
100 Target 1 Target 2 Target 3 Cell Under Test 50
120 100 80 60 40 20
0
0
10
20
30
40 50 60 Range Gates
70
80
90
0
100
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.3. VI‐CFAR ‐ Case II
Case II: Leading Window Variable, Lagging Window not Variable
A variable leading window is indicative of high spikes of power, and with a rea‐
sonable signal to noise ratio, these spikes are most likely the result of targets. The av‐ erage power, then, of the leading window wouldn’t be indicative of the noise/clutter contained. In this case, VI‐CFAR switches to using CA‐CFAR with the lagging window alone.
As the lagging window is nonvaraible, it is likely to contain noise/clutter only. To
set the threshold, all cells in the lagging window is summed and multiplied to the threshold multiplier for window length N/2: N / 2 PFA
1 N /2
1
46
Example Situation: Lagging Window: ‐ Not Variable ‐ Mean similar to Leading
Interfering targets in the lagging window.
Leading Window: ‐ Variable ‐ Mean similar to Lagging
CFAR to use:
Cell Averaging CFAR with lagging window only 150
200 180 160
Signal Cell Under Test
140 Power (linear)
Power (linear)
100 Target 1 Target 2 Target 3 Cell Under Test 50
120 100 80 60 40 20
0
0
10
20
30
40 50 60 Range Gates
70
80
90
0
100
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.4. VI‐CFAR ‐ Case III
Case III: Lagging Window Variable, Leading Variable Nonvariable
This is the converse situation of Case II. When only the lagging window is vari‐
able and not the leading, the lagging window is likely to contain targets and the leading noise/clutter only.
In this case, the threshold is set by summing only the reference cells of the
leading window and multiplied to a CA‐CFAR threshold multiplier for window length N/2: N / 2 PFA
1 N /2
1
47
Lagging Window: ‐ Variable ‐ Mean similar to Leading
Example Situation:
Leading Window: ‐ Variable ‐ Mean similar to Lagging
Interfering targets in both leading and lagging window CFAR to use:
Smallest of Cell Averaging 150
200 180 160
Signal Cell Under Test
140 Power (linear)
Power (linear)
100 Target 1 Target 2 Target 3 Cell Under Test 50
120 100 80 60 40 20
0
0
10
20
30
40 50 60 Range Gates
70
80
90
0
100
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.5. VI‐CFAR ‐ Case IV
Case IV: Both Windows Variable
When both the leading and lagging windows are variable, then both windows
very likely contains targets. Whichever window used is likely to result in target mask‐ ing. VI‐CFAR, then, seeks to minimize the masking by taking the smaller of the sums of the leading and the lagging windows. The threshold is set by multiplying this smaller sum to a CA‐CFAR threshold multiplier for window length N/2: N / 2 PFA
1 N /2
1
48
Leading Window: ‐ Not Variable ‐ Mean different from leading
Example Situation:
Leading Window: ‐ Not Variable ‐ Mean different from lagging
Near clutter wall. Can be on either edge of the clutter edge. Use:
Greatest of Cell Averaging 80
250 Noise/Clutter Cell Under test
70
Noise/Clutter Cell Under test 200
50
Power (linear)
Power (linear)
60
40 30
150
100
20 50 10 0
0
10
20
30
40 50 60 Range Gates
70
80
90
0
100
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.6. VI‐CFAR ‐ Case
Case V: Not Variable but Different Mean
When both the leading and lagging variables are nonvariable but different in
mean, then the cell under test is very likely near a clutter wall. The goal for VI‐CFAR , in this case, is to set the threshold by the higher edge of the clutter wall.
The threshold is set by taking the larger of the sums of the leading and lagging
window and multiplied to the CA‐CFAR threshold multiplier for window length of N/2: N / 2 PFA
1 N /2
1
49
Magnitude (dB)
20
15
10
5
0
CA-CFAR OS-CFAR VI-CFAR Recieved Waveform 0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.8. VI‐CFAR Threshold in Homogeneous Environment.
4.3 Homogeneous Environment 4.3.1 Probability of False Alarm
In a homogeneous environment, VI‐CFAR largely reverts to CA‐CFAR and uses
the CA‐CFAR threshold multiplier (Figure 4.8). In the threshold plot, we see that the threshold of VI‐CFAR almost entirely overlaps with that of CA‐CFAR. The discrepancy near and around the target— with CA‐CFAR displaying its distinctive elevated threshold plateaus near the target and VI‐CFAR with no such display— can be accounted for by the decision logic of VI‐CFAR. The CUTs N/2 cells in front of and N/2 cells behind the target will interpret the target as a masking target, and switch to using only the lagging and the leading windows respectively in the computation of CA‐CFAR.
50
FA
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
Theoretical PFA Experimental PFA
-7
10
2
4
6
8
10 12 14 Threshold Multiplier
16
18
20
22
Figure 4.9. VI‐CFAR Theoretical v. Experimental PFA in Homogeneous Environment
Not surprisingly, in the homogeneous environment, VI‐CFAR successfully holds
the probability of false alarm like CA‐CFAR (Figure 4.9). 4.3.2 Probability of Detection
Across various signal to noise ratios, we see that in the homogeneous environ‐
ment, CA‐CFAR consistently slightly outperforms VI‐CFAR (Figure 4.10). This slight dis‐ crepancy is to be expected since, due to the random distribution of the noise/clutter, VI‐CFAR may very rarely misinterpret the homogeneous environment as a Type II, III, IV, or V environment. In these cases, VI‐CFAR switches to using only its leading or lag‐ ging windows in the estimation of the noise.
In CFAR loss, this translates to about a 0.25 dB loss across all SNR’s.
51
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR VI-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
25
Figure 4.9. VI‐CFAR Probability of Detection Curve 1.5
CFAR Loss (dB)
1
0.5
CA-CFAR OS-CFAR VI-CFAR 0
5
10
15 SNR (dB)
20
Figure 4.10. VI‐CFAR CFAR Loss in Homogeneous Environment
52
25
Magnitude (dB)
20
15
10
CA-CFAR OS-CFAR VI-CFAR Recieved Waveform
5
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.11. VI‐CFAR Threshold 1 Masking Target.
4.4 Masking Targets
With the variability index, VI‐CFAR may reliably predict the presence of masking
targets in the leading window, the lagging window, or both windows and appropriately adjust to use of the lagging window or the leading window.
The VI‐CFAR threshold was simulated for a two‐target environment (Figure
4.11). It successfully detects both targets. In the graph, OS‐CFAR also allowed for de‐ tection of both targets. CA‐CFAR, in contrast, detects only the bin 42 target; setting the threshold too high for the bin 37 target.
VI‐CFAR’s resistance to target masking in two‐target environments is clear in its
probability of detection curve (Figure 4.12). For most signal to ratio values, the prob‐ ability of detection for VI‐CFAR is only slightly lower than that of OS‐CFAR. CA‐CFAR, 53
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson CA-CFAR OS-CFAR VI-CFAR
0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
25
Figure 4.12. VI‐CFAR Probability of Detection with 1 Masking Target
meanwhile, shows it usual loss in probability of detection wit target masking.
VI‐CFAR itself, however, begins to suffer masking effects when there are more
than one interfering target (Figure 4.13). Whereas the CFAR loss of either homogenous and one‐interfering VI‐CFAR is less than 1dB, the CFAR loss for two and three interfer‐ ing targets increase to as much as 10dB (Figure 4.14). How can we account for such dis‐ parate results?
When two masking targets are randomly placed around a CUT, part of the time,
there will be one in the leading window and one in the lagging window. In this case, no matter which window is selected, there will be a masking effect for the CUT (Figure 4.15). In this case, VI‐CFAR only mollifies the effect by switching to SOCA‐CFAR.
On the other hand, when both interfering targets are together, either in the
54
1 Neyman Pearson homogeneous one masking target two masking targets three masking targets
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
25
Figure 4.13. VI‐CFAR Probability of Detection Curve with varying Numbers of Masking Target
12 homogeneous one masking target two masking targets three masking targets
CFAR Loss (dB)
10
8
6
4
2
0
5
10
15 SNR (dB)
20
Figure 4.14. VI‐CFAR CFAR Loss Curve with varying Numbers of Masking Target
55
leading or the lagging window, VI‐CFAR can entirely avoid their masking effects by se‐ lecting the other window. The differences can be profound. In Figure 4.16, the prob‐ ability of detection is computed for different distributions of two masking targets.
56
150
Target 2: When the CUT reaches here, VI‐CFAR uses SOCA‐CFAR.
Power (linear)
100
Target 3:
Target 1:
0
When the CUT reaches here, VI‐CFAR takes only the leading win‐ dow for CA‐CFAR
When the CUT reaches here, VI‐CFAR takes only the lagging window for CA‐CFAR
50
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.15. Three Targets
1 0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 Neyman-Pearson Mean PD
0.2
Targets 1,3 Target 2
0.1 0
0
5
10 15 Signal to Noise Ratio (dB)
20
Figure 4.16. VI‐CFAR Probability of Detection with Three Masking Targets
57
25
Greater threshold clearance for VI‐CFAR
VI‐CFAR threshold increases earlier and more steeply than CA‐CFAR or OS‐CFAR
30
25
Magnitude (dB)
20
15
10 CA-CFAR OS-CFAR VI-CFAR Recieved Waveform
5
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 4.17. VI‐CFAR Threshold in Clutter Wall Transition.
4.5 Clutter Wall 4.5.1 Probability of False Alarm
When VI‐CFAR detects a clutter wall— a type V case— it switches to using to
GOCA‐CFAR. The algorithm uses the greater of the sum of the leading and lagging win‐ dow in its cell averaging CFAR computation.
In setting the threshold this way there are two consequences: Leading into the
clutter wall, the VI‐CFAR threshold slopes upwards earlier and more steeply than CA‐ CFAR, and after the clutter wall, the VI‐CFAR has already stopped sloping and achieved a sufficiently high threshold (Figure 4.17).
This effects translates to good maintenance of the probability of false alarm
(Figure 4.18). We see that VI‐CFAR suffers less than a 10‐fold increase in PFA.‐‐ from 10‐4
58
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
CA-CFAR OS-CFAR VI-CFAR
-7
10
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 4.18. VI‐CFAR Clutter Wall Probability of False Alarm.
to less than 10‐3. Cell averaging and Ordered Statistics CFAR, in contrast, suffers a 100‐ fold increase in PFA. 4.5.2 Probability of Detection
The drawback of VI‐CFAR’s clutter wall strategy is a greater drop probability of
detection than CA‐CFAR or OS‐CFAR (Figure 4.19). We see that for all cells under clut‐ ter, the probability of detection for VI‐CFAR is less than that of CA‐CFAR and OS‐CFAR. This is a result of two separate effects: threshold set too high before the clutter wall, and threshold set correctly after the clutter wall.
Due to VI‐CFAR’s GOCA‐CFAR clutter wall strategy, the threshold set before the
clutter wall is set by the greater of the leading and lagging window sums. This means
59
1 CA-CFAR OS-CFAR VI-CFAR N/2 cells in clutter
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 4.19. VI‐CFAR Clutter Wall Probability of Detection.
that the set threshold will be based, in part, on the larger powers of the clutter region. This elevates the threshold much higher than the level necessary to maintain the de‐ sired PFA of 10‐4.
After the clutter wall, when more than 16 cells are under clutter, all three
CFARs experience a low probability of detection. This is to be expected. In these re‐ gions of the clutter wall, instead of noise, the target is in a clutter background. Since the clutter to noise ratio (CNR) is much smaller than the signal to noise ratio (SNR), the probability of detection for any CFAR would be lower.
The detection of CA‐CFAR and OS‐CFAR is higher than VI‐CFAR after the clutter
wall because their thresholds were set too low— so low that the PFA is much greater than the designed 10‐4 point.
60
CHAPTER 5 ORDERED STATISTICS CFAR
Cell‐averaging CFAR was the first of the CFAR methods. Assuming homogene‐
ous Gaussian white noise, through a sliding window, CA‐CFAR estimates the noise level of the cell under test, CUT, through an arithmetic sum of N adjacent cells . In heteroge‐ neous environments, however, CA‐CFAR performance rapidly degrades.
Ordered Statistics CFAR, OS‐CFAR, had been developed to compensate for
much of CA‐CFAR’s pitfalls. Rather than taking the arithmetic mean of adjacent cells, OS‐CFAR ranks the N adjacent cells from smallest to largest and multiplies the kth larg‐ est cell to the appropriate threshold multiplier to set the threshold level. 5.1 Implementation
Each reference cell is sorted from smallest to biggest: x 1 x 2 x 3 x k x N
The kth‐largest cell is multiplied with the corresponding threshold multiplier to get the desired threshold:
Tˆ aOS x k
The kth‐ranked cell could be the median ranked value or even the largest value. As simulations will show, the larger x(k), the smaller the CFAR‐loss in homogeneous envi‐ ronment but the less the resistance to target masking and clutter‐edge transition.
61
Through a square‐law detector and in a background with Gaussian White Noise,
the threshold‐multiplier αOS may be found from this equation:
C /2 N OS N k ! PFA C /2 k N k ! OS M !
With the binomial given by:
N N! k k ! N k !
Here, N is the total number of reference cells; k is the rank of the representative cell used, PFA is the desired probability of false alarm, αOS is the threshold multiplier and Γ(∙) is the Gamma function.
The theoretical probability of detection is given by: D OS
OS 1
Sort
x1 x 2 xk xN
x
Threshold
Threshold Multiplier
Figure 5.1. OS‐CFAR Block Diagram
62
FA
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
Theoretical PFA Experimental PFA
-7
10
2
4
6
8
10 12 14 Threshold Multiplier
16
18
20
22
Figure 5.2. OS‐CFAR Homogeneous PFA.
5.2 Homogeneous 5.2.1 Probability of False Alarm
Using a window size of 32 and a kth rank of 22 (k chosen for 0.7∙N), thresholds
multipliers were used to set and simulated for probability of false alarms between 10‐1 and 10‐7 (Figure 4.2). The experimental probability of false alarm conformed with theo‐ retical expectation (Figure 5.2). 5.2.2 Detection
The ordered statistics threshold, in general, is higher than the cell averaging
threshold for the same designed probability of false‐alarm (Figure 5.2). This higher
63
Magnitude (dB)
20
15
10
5 CA-CFAR OS-CFAR Recieved Waveform 0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 5.3. OS‐CFAR Threshold in Homogeneous Environment.
threshold results from slightly smaller reliability of an ordered statistics estimations as versus a mean estimation of the average noise (Figure 5.3). In the probability of detec‐ tion curve, this translates to a slightly smaller PD (Figure 5.4).
Amongst ordered statistics curves of the same window length, In the CFAR loss
curve, the kth ranked cell chosen for the threshold multiplier also affects performance. In a homogeneous environment, in general, the higher the kth value used, the slightly better the probability of detection and the smaller the CFAR loss.
In the probability of detection curve, for a window size of 32, k=0.9N (or k=29)
gave a lower CFAR loss about half a dB lower than that of k=0.5N (Figure 5.5). The re‐ sults for 0.9N was also about 0.1 dB better than that of 0.7N. All results, however, had at least half a dB of CFAR loss as much as the cell averaging CFAR of the same window length. 64
1 0.9 0.8 0.7 0.6 0.5 0.4 0.3
Neyman-Pearson CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N
0.2 0.1 0
0
5
10
15
20
Figure 5.4. OS‐CFAR Homogeneous Probability of Detection
1.5
CFAR Loss (dB)
1
0.5 CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N 0
5
10
15 SNR (dB)
Figure 5.5. OS‐CFAR Homogeneous CFAR Loss
65
20
25
N reference cells smallest
largest
N‐k‐1 buffer cells kth cell Figure 5.6. OS‐CFAR Resistance to Multiple Targets
5.3 Multiple Targets
Ordered statistics CFAR offers natural resistance to target masking. With cell
averaging CFAR, any target within the reference cell gets added to be multiplied with the threshold multiplier. If the signal to noise ratio is 10dB, the estimated noise may be as much as 50% greater.
With ordered statistics, since only the ranked kth value is selected, interfering
targets don't’ directly affect the threshold multiplier. Since target powers are typically much greater than noise or clutter power, when ranked, they find themselves below the k‐value. The cells larger than the kth cell, in the reference then, act as a kind of buffer against target masking. If the number of buffer cells exceed the number of inter‐ fering targets, then OS‐CFAR can escape directly influence of masking targets. Also, even when there are less targets than buffer cells, the less targets there are less indi‐ rect influence the kth cell will feel from the target.
Figure 5.7 shows an example of a OS‐CFAR threshold in the multiple target case.
As is evident, OS‐CFAR draws a threshold low enough to detect both targets. CA‐CFAR,
66
25
Magnitude (dB)
20
15
10
5
0
CA-CFAR OS-CFAR Recieved Waveform 0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 5.7 OS‐CFAR Threshold with 1 Masking Target.
on the other hand, misses detection of one target.
With one masking target, the probability of detection curve of OS‐CFAR follows
the contour of the Neyman‐Pearson detector (Figure 5.8). The cell averaging detection begins to suffer major detection degradation beyond an SNR of about 10 dB. The selec‐ tion of different k values, 0.5N, 0.7N, or 0.9N (or k of 16, 22, and 29 respectively for a window size of N=32) led to no major differences in detection probability. The CFAR losses for each k value selection were no more than 0.5dB different from one another (Figure 5.9).
In contrast to the homogeneous case, k=0.7N produced better detection and
CFAR loss values than k=0.9N. The reason, most likely is that the buffer cells allowed by k=0.7N were more adequately insulating than that of k=0.9N. 67
1 0.9 0.8
CFAR Loss (dB)
0.7 0.6 0.5 0.4 0.3
Neyman-Pearson CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N
0.2 0.1 0
0
5
10 15 Probability of Detection
20
Figure 5.8. OS‐CFAR Probability of Detection with 1 Masking Target
3 CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N
2.5
CFAR Loss (dB)
2
1.5
1
0.5
0
5
10
15 SNR (dB)
20
Figure 5.9. OS‐CFAR CFAR Loss with 1 Masking Target
68
25
1 0.9 0.8
CFAR Loss (dB)
0.7 0.6 0.5 0.4 0.3
Neyman-Pearson CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N
0.2 0.1 0
0
5
10 15 Probability of Detection
20
25
Figure 5.10 OS‐CFAR Probability of Detection with 3 Masking Target
With three masking targets, the probability of detection for k=0.9N became
noticeably worse than that of k=0.9N while k=0.5N continued to give only a very slightly less detection curve than k=0.9N (Figure 5.10). In CFAR loss, k=0.9N got to as much as nearly 2dB worse for SNR of 25dB than k=0.7N. The CFAR loss difference be‐ tween 0.7N and 0.5N was about 0.25dB throughout the range of SNRs simulated.
69
5 CA-CFAR OS-CFAR k=0.5 N OS-CFAR k=0.7 N OS-CFAR k=0.9 N
4.5 4
CFAR Loss (dB)
3.5 3 2.5 2 1.5 1 0.5 0
5
10
15 SNR (dB)
20
Figure 5.11. OS‐CFAR CFAR Loss with 3 Masking Target
70
30
25
Magnitude (dB)
20
15
10
5
0
CA-CFAR OS-CFAR Recieved Waveform 0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 5.12. OS‐CFAR in Clutter Wall Threshold.
5.5 Clutter Wall
Ordered Statistics CFAR performs similarly to cell averaging CFAR at the clutter
wall. Both algorithms only gradually slope upwards their thresholds approaching the clutter wall, and past the clutter edge, the thresholds of both algorithms get as low as to frequently allow false alarms (Figure 5.12).
OS‐CFAR algorithms selected for different ranked kth values perform differently
at the clutter wall. In general, the greater the k, the less the probability of false alarm will increase at the clutter edge. Cell averaging CFAR for window length 32, at the clut‐ ter wall, suffers roughly a 100‐fold increase in false alarm rate: from the designed 10‐4 to nearly 10‐2 (Figure 5.13). Ordered Statistics CFAR for k=0.5*N (k=16) and k=0.7*N
71
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
-6
10
CA-CFAR OS-CFAR k=0.5*N OS-CFAR k=0.7*N OS-CFAR k=0.9*N
-7
10
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 5.13. OS‐CFAR Clutter Wall Probability of False Alarm
(k=22) actually perform worse than CA‐CFAR— suffering a false alarm rate that’s nearly a thousand times more than what was desired! OS‐CFAR for k=0.9*N, in contrast, im‐ proved upon CA‐CFAR, with a false alarm rate that’s only between 10 times and 100 times the designed rate of 10‐4.
Probability of detection performance at the clutter wall for the various k value
is reversed from that for probability of false alarm. The smaller the k value, the less the target masking effect before the clutter wall. K=0.5N gave better detection results than k=0.7N, and k=0.9N.
72
1 CA-CFAR OS-CFAR k=0.5*N OS-CFAR k=0.7*N OS-CFAR k=0.9*N
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 20 25 Number of Reference Cells in Clutter
Figure 5.14. OS‐CFAR Clutter Wall Probability of Detection
73
30
Chapter 6 Ordered Statistics Greatest‐Of CFAR
Although OS‐CFAR works well in avoiding the drawbacks of target masking, it
costs more than CA‐CFAR to implement and it suffers more false alarms than CA‐CFAR at the clutter wall. Ordered Statistics Greatest‐Of CFAR (OSGO‐CFAR) represents an at‐ tempt to kill both problems with one stone. 6.1 Implementation
OSGO‐CFAR operates similarly to OS‐CFAR in its sorting of reference cells. The
difference, though, is that in OSGO CFAR, the N reference cells of the leading window and the lagging window are sorted independently, into two N/2 sublists (Figure 6.1). The kth largest cells of the two sublists are compared, and the larger of the two are
Rank Order
Rank Order
x1 x 2 xk xN /2
y1 y 2 yk yN /2
Max
x
Threshold
Threshold Multiplier
Figure 6.1. OSGO‐CFAR Block Diagram
used with the threshold multiplier to set the threshold.
The threshold multiplier is given by this equation:
PFA
OSGO
N 2 2k 2 k
2 N N k k 2 2
j 0
N 2 k j i N N N j i 1 2 k 2 k 1 N N j i 1 i 0 i j i 2
As no analytical expression for threshold multiplier α exists, it must be solved numeri‐ cally.
Magnitude (dB)
20
15
10
5
0
CA-CFAR OS-CFAR OSGO-CFAR Recieved Waveform 0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 6.2. OSGO‐CFAR Threshold in Homogeneous Environment.
6.2 Homogeneous
In the homogeneous environment, OSGO‐CFAR sets its threshold very similarly
to OS‐CFAR. Neither method gives the elevated threshold plateaus around targets.
75
p
FA
0
10
-1
10
Probability of False Alarm
-2
10
-3
10
-4
10
-5
10
-6
10
-7
10
0
Theoretical PFA Experimental PFA 5
10 15 Threshold Multiplier
20
25
Figure 6.3. OSGO‐CFAR Homogeneous PFA.
6.2.1 Probability of False Alarm
To verify that the OSGO CFAR works in limiting noise/clutter to the desired
probability of false alarm, the algorithm was simulated to a number of desired PFA— from 10‐1 to about 10‐7. The measured probability of false alarm conformed well with expected results. 6.2.2 Detection
In probability of detections, OSGO‐CFAR performs comparatively with OS‐CFAR
(Figure 6.4). Like OS‐CFAR, OSGO‐CFAR gives a CFAR loss greater than that of OS‐CFAR. For CFAR algorithms of equal window length, OSGO‐CFAR always gave higher CFAR loss
76
1 Neyman-Pearson CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
0.9 0.8
CFAR Loss (dB)
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 Probability of Detection
20
25
Figure 6.4. OSGO‐CFAR Homogeneous PD
than OS‐CFAR. Also like OS‐CFAR, the greater the ranked kth value taken for the thresh‐ old multiplier, the less the CFAR loss (Figure 6.5). The CFAR loss for OSGO‐CFAR k=0.7*N/2 and k=0.9*N/2 were within 0.2 dB of the CFAR loss for k=0.7*N of OS‐CFAR.
77
5 CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
4.5 4
CFAR Loss (dB)
3.5 3 2.5 2 1.5 1 0.5 0
5
10
15 SNR (dB)
Figure 6.5. OSGO‐CFAR Homogeneous CFAR Loss
78
20
N/2 reference cells smallest
largest
N/2‐k‐1 buffer cells kth cell Figure 6.6. OSGO‐CFAR Resistance to Multiple Targets
25
Magnitude (dB)
20
15
10
CA-CFAR OS-CFAR OSGO-CFAR Recieved Waveform
5
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 6.7. OSGO‐CFAR Threshold with 1 Masking Target.
6.3 Multiple Targets
Using ordered statistics, OSGO‐CFAR, like OS‐CFAR, offers strong resistance
against target masking. However, due to the kth cell being taken from the larger of the sorted leading or lagging window, the amount of multiple target resistance is dimin‐ ished (Figure 6.6). Since OSGO‐CFAR only takes k from half a full window, there is only
79
1 Neyman-Pearson CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
0.9 0.8
CFAR Loss (dB)
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 Probability of Detection
20
25
Figure 6.8. OSGO‐CFAR PD with 1 Masking Target
N/2‐k‐1 buffer cells— as opposed to N‐k‐1 for OS‐CFAR.
The threshold set for OSGO‐CFAR looks a bit like that of OS‐CFAR (Figure 6.7).
Like OS‐CFAR, OSGO‐CFAR avoids much of the masking problem of CA‐CFAR.
The probability of detection across SNR=2 and 24 shows little difference be‐
tween OS‐CFAR and the test OSGO‐CFAR algorithms. In terms of CFAR loss, we see that the OS‐CFAR and OSGO‐CFAR algorithms, in the single interfering target environment, performs within 0.5dB of one another.
One thing to note, though, is that OSGO‐CFAR for k=0.9*N/2 begins to suffer
greater CFAR loss than OSGO‐CFAR for k=0.7*N/2 beyond an SNR of about 5dB. It is clear that for a window size of 32, the number of buffer cells offered by k=N/2*0.9 only offers enough masking resistance for one interfering target.
80
2.5
CFAR Loss (dB)
2
1.5
1 CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
0.5
0
5
10
15 SNR (dB)
20
Figure 6.9. OSGO‐CFAR CFAR loss with 1 Masking Target
At three masking targets, we see that the probability of detection for OSGO‐
CFAR of k=0.9*N/2 decreases precipitously (Figure 6.10). Though slightly lower than the OS‐CFAR curve, the detection curve for both OSGO k=0.7*N/2 and k=0.5*N/2 both followed the contour of the Neyman‐Pearson detector. The CFAR loss for these two curves were about 2 dB— about half a dB more than OS‐CFAR. (Figure 6.11).
81
1 Neyman-Pearson CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
0.9 0.8
CFAR Loss (dB)
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 Probability of Detection
20
Figure 6.10. OSGO‐CFAR PD with 3 Masking Target
5 4.5 4
CFAR Loss (dB)
3.5 3 2.5 2 1.5
CA-CFAR OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
1 0.5 0
5
10
15 SNR (dB)
20
Figure 6.11. OSGO‐CFAR CFAR loss with 3 Masking Target
82
25
30
25
Magnitude (dB)
20
15
10 CA-CFAR OS-CFAR OSGO-CFAR Recieved Waveform
5
0
0
10
20
30
40 50 60 Range Gates
70
80
90
100
Figure 6.12. OSGO‐CFAR Threshold in Clutter Wall Transition.
6.5 Clutter Wall
In previous chapters, we’ve seen how cell averaging CFAR and ordered statistics
CFAR both suffer from elevated false alarm rates near the clutter wall. Their PFA be‐ come elevated by as much as 100 times the set PFA. Can OSGO‐CFAR do better?
Figure 6.12 shows the OSGO‐CFAR threshold at a clutter wall. compared to CA‐
CFAR and OS‐CFAR, OSGO‐CFAR has a much better clearance at the clutter edge. This translates to significant improvements for keeping desired false alarm rates (Figure 6.12).
The probability of false alarm were generated over a million trials for OSGO‐
CFAR of window length N=32 and kth rank of N/2*0.5, N/2*0.7, and N/2*0.9. At worst,
83
0
10
-1
10
-2
Probability of False Alarm
10
-3
10
-4
10
-5
10
CA-CFAR OS-CFAR k=0.5*N OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
-6
10
-7
10
-8
10
0
5
10 15 20 25 Number of Reference Cells in Clutter
30
Figure 6.12. OSGO‐CFAR Clutter Wall PFA
OSGO‐CFAR only suffers a ten‐fold increase form the designed PFA of 10‐4. This is a ten‐ fold improve over cell‐averaging CFAR. These results are also comparable to those of the clutter wall false alarm performances of VI‐CFAR (chapter 4).
The clutter wall probability of detection performance for OSGO‐CFAR can be
found in figure 6.13.
84
1 CA-CFAR OS-CFAR k=0.5*N OS-CFAR k=0.7*N OSGO-CFAR k=0.5*N/2 OSGO-CFAR k=0.7*N/2 OSGO-CFAR k=0.9*N/2
0.9
Probability of Detection
0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0
0
5
10 15 20 25 Number of Reference Cells in Clutter
Figure 6.13. OSGO‐CFAR Clutter Wall PD
85
30
Chapter 7 Conclusions and Final Assessment
In the course of our research, we’ve studied a number of CFAR algorithms—
each approaching the same problem of noise/clutter mean estimation and threshold setting through different techniques. Cell Averaging CFAR (CA‐CFAR) though optimal in the homogeneous environment, suffers major performative drop‐off in the heteroge‐ neous environments: with multiple targets and the clutter wall environment.
Although Ordered Statistics CFAR (OS‐CFAR), Smallest‐Of Cell Averaging CFAR
(SO‐CFAR), and Greatest‐Of Cell Averaging CFAR (GOCA‐CFAR) each remedy one of the two heterogeneous conditions, they fail to fully remedy the other.
Two algorithms, Ordered Statistics Greatest‐Of CFAR (OSGO‐CFAR) and Variabil‐
ity Index CFAR (VI‐CFAR), balance performance in each environment— homogeneous, multiple targets, and clutter wall— to be practically feasible.
Figure 7.1 illustrates a practical CFAR environment. From it, we can see that
only OSGO‐CFAR and VI‐CFAR detect the homogeneous target, the multiple target, and make the clutter wall without a false alarm.
86
87
88
0
5
10
15
20
25
50
100
150
Figure 7.1. CFAR Obstacle Course
200
250
300
Cell Averaging Ordered Statistics Varability Index Ordered Statistics Greatest Of
REFERENCES [1] Finn, H.M., and Johnson, R.S.: ‘Adaptive detection mode with threshold control as a function of spatially sampled clutter estimates’, RCA Rev., 1968, vol.29, no.3, pp. 414–464 [2] Richards, M.A. Fundamentals of Radar Signal Processing. New York; McGraw‐Hill. 2005
[3] Kang, K.E. Radar System: Analysis, Design, and Simulation. Chapter 9: Monte Carlo Method and Function Integration. Pages 247‐260. Artech House 2008 [4] Gandhi, P.P.; Kassam, S.A.; , "Analysis of CFAR processors in homogeneous back‐ ground," Aerospace and Electronic Systems, IEEE Transactions on , vol.24, no.4, pp.427‐445, Jul 1988 [5] Trunk, G.V.; , "Range Resolution of Targets Using Automatic Detectors," Aerospace and Electronic Systems, IEEE Transactions on , vol.AES‐14, no.5, pp.750‐755, Sept. 1978 [6] Hansen, V.G.; Sawyers, J.H.; , "Detectability Loss Due to "Greatest Of" Selection in a Cell‐Averaging CFAR," Aerospace and Electronic Systems, IEEE Transactions on , vol.AES‐16, no.1, pp.115‐118, Jan. 1980 [7] Smith, M.E.; Varshney, P.K.; , "Intelligent CFAR processor based on data variability," Aerospace and Electronic Systems, IEEE Transactions on , vol.36, no.3, pp.837‐847, Jul 2000 [8] Rohling, H.; , "Radar CFAR Thresholding in Clutter and Multiple Target Situations," Aerospace and Electronic Systems, IEEE Transactions on , vol.AES‐19, no.4, pp.608‐ 621, July 1983 [9] Elias‐Fuste, A.R.; de Mercado, M.G.G.; de los Reyes Davo, E.; , "Analysis of some modified ordered statistic CFAR: OSGO and OSSO CFAR," Aerospace and Electronic Systems, IEEE Transactions on , vol.26, no.1, pp.197 [10]Cao, T.‐T.V.; , "Constant false‐alarm rate algorithm based on test cell information," Radar, Sonar & Navigation, IET , vol.2, no.3, pp.200‐213, June 2008
89
Appendix A MATLAB Functions A.1 Radar Return Generation function [P]=simRangeData(N,iN,S,iS) % [P]=simRangeData(N,iN,S,iS) produces a vector P that corresponds to % the power over range gates with complex gaussian white noise and % Swerling I/II targets. % %Vector N determines the power of the complex gaussian white noise across % Range Cells controlled by indices iN. % % Vector S determines the power of a Swerling I/II target in Range Cells % dictated by vector iS pwrSig=10.^(S/10); linear pwrNoise=10.^(N/10); linear
%Converting signal power from dB to
nRangeGate=iN(end);
%The total number of range gates
%%GENERATING NOISE voltNoise_I=zeros(nRangeGate,1); voltNoise_Q=zeros(nRangeGate,1);
%Pre-allocating for speed %Pre-allocating for speed
I=length(iN); different noise power iLead=1;
%Converinting noise power from dB to
%Number of range gate segments with
%This loop produces the voltage waveform for the gaussian white noise. %Every range gate segment between iLead and iLag will have its own power of %N(i) for i=1:I iLag=iN(i); voltNoise_I(iLead:iLag)=sqrt(pwrNoise(i)/2)*randn(iLag-iLead+1,1); % voltNoise_I Channel voltNoise_Q(iLead:iLag)=sqrt(pwrNoise(i)/2)*randn(iLag-iLead+1,1); % voltNoise_Q Channel iLead=iLag; end noise=voltNoise_I+j*voltNoise_Q; %Voltage waveform of the noise %%GENERATING TARGET RETURN signal=zeros(size(noise)); %Pre-allocating for speed I=length(iS); %The total number of signal returns for i=1:I %Power of the signal follows an exponential distribution pwrSwerling=exprnd(pwrSig(i),1,1); %Voltage waveform fo the signal signal(iS(i))=sqrt(pwrSwerling/2)+j*sqrt(pwrSwerling/2); end
90
%%GENERATING OVERALL SIGNAL data=signal+noise; P=abs(data).^2;
%Overall voltage waveform %Power of the overall return
91
A.2 Monte Carlo Simulations
function [expPfa]= monteCarloFalseAlarm(T,noisePwr) % This function performs a Monte Carlo Importance Sampling to determine % experimental probabilities of false alarm % % INPUTS % T - This is the threshold % % noisePwr - the noise power in linear scales % % OUTPUTS % expPfa: The experimental probability of false alram % % ASSUMPTIONS % Exponential distribution of noise/clutter % % AUTHORSHIP % James Jen % Cal Poly Pomona % % UPDATE HISTORY % May 28th, 2011 - created % nCFAR=size(T,2); nRangeGate=length(T); monteCarloPwr=noisePwr*20; iCFARwin=1:nRangeGate; iCFARwin=iCFARwin(T(:,1)~=0); nTrial=length(iCFARwin);
%The number of range gates %Monte Carlo mean 20 times greater %indicies for non-zero thresholds
%Exponential distribution with the Monte carlo power pwrMonteCarlo=exprnd(monteCarloPwr,nRangeGate,nCFAR); %The Monte carlo weight weight=monteCarloPwr/noisePwr*exp(-(1/noisePwr-1/monteCarloPwr) *pwrMonteCarlo); %with iCFAR win, we avoid the zero thresholsd at the ends of the range gate weight=weight(iCFARwin,:); T=T(iCFARwin,:); pwrMonteCarlo=pwrMonteCarlo(iCFARwin,:); %Logical indicies for when Monte Carlo power greater than threshold iFalseAlarm=T=iCUT pwrCUT=clutterPwr; else pwrCUT=noisePwr; end end
threshold(:,CA)=CA_CFAR1(Pfa,N,pwrSignal); threshold(:,OS)=OS_CFAR1(Pfa,N,k_OS,pwrSignal); threshold(:,SOCA)=SOCA_CFAR1(Pfa,N,pwrSignal); threshold(:,GOCA)=GOCA_CFAR1(Pfa,N,pwrSignal); threshold(:,OSGO)=OSGO_CFAR1(Pfa,N,k_OSGO,pwrSignal); threshold(:,VI)=VI_CFAR1(Pfa,N,pwrSignal); if ithreshold; nDetection=sum(iDetection,1); expPd(i,:)=nDetection/nTrial; end close
%close progress bar
% ############ % % Plotting % % ############ % cellLead=1:N/2+1; iLead=0:N/2; cellLag=N/2+2:N+2;
136
iLag=N/2:N; xMin=0; xMax=N; yMin=0; yMax=1; %PROBABILITY OF DETECTION CURVE for iCFAR=1:nCFAR figure if iCFAR==CA plot(iLead,expPd(cellLead,CA),'color',clr(CA,:),'linewidth',1.5) hold on elseif iCFAR==OS plot(iLead,expPd(cellLead,CA),'color',clr (CA,:),'linewidth',1.5,'linestyle','--') hold on plot(iLead,expPd(cellLead,OS),'color',clr(OS,:),'linewidth',1.5) else plot(iLead,expPd(cellLead,CA),'color',clr (CA,:),'linewidth',1.5,'linestyle','--'); hold on plot(iLead,expPd(cellLead,OS),'color',clr (OS,:),'linewidth',1.5,'linestyle','--') plot(iLead,expPd(cellLead,iCFAR),'color',clr (iCFAR,:),'linewidth',1.5) end %Line marking where the half-window point is line([N/2 N/2], [yMin yMax],'color','m','linestyle',':','linewidth',1.5); if iCFAR==CA legend(CFARtype([CA end]),'location','NorthEast'); plot(iLag,expPd(cellLag,CA),'color',clr(CA,:),'linewidth',1.5) elseif iCFAR==OS legend([CFARtype([CA OS end])],'location','NorthEast'); plot(iLag,expPd(cellLag,CA),'color',clr (CA,:),'linewidth',1.5,'linestyle','--') plot(iLag,expPd(cellLag,OS),'color',clr(OS,:),'linewidth',1.5) else legend([CFARtype([CA OS iCFAR end])],'location','NorthEast'); plot(iLag,expPd(cellLag,CA),'color',clr (CA,:),'linewidth',1.5,'linestyle','--'); plot(iLag,expPd(cellLag,OS),'color',clr (OS,:),'linewidth',1.5,'linestyle','--') plot(iLag,expPd(cellLag,iCFAR),'color',clr (iCFAR,:),'linewidth',1.5) end ylabel('Probability of Detection'); xlabel('Number of Reference Cells in Clutter'); axis([xMin xMax yMin yMax]); % grid on % set(gca,'Xcolor',[0.5 0.5 0.5]); % set(gca,'Ycolor',[0.5 0.5 0.5]); % Caxes = copyobj(gca,gcf); % set(Caxes, 'color', 'none', 'xcolor', 'k', 'xgrid', 'off', 'ycolor','k', 'ygrid','off'); end
137
View more...
Comments