Comparative of signal processing techniques for Micro-Doppler

signature extraction with automotive radar systems

Berta Rodriguez-Hervas, Michael Maile, Benjamin C. Flores

Electrical and Computer Engineering Department, The University of Texas at El Paso,
El Paso, TX, 79968, USA
Autonomous Driving North America, Mercedes-Benz Research and Development North
America, Sunnyvale, CA, 94085, USA

ABSTRACT

In recent years, the automotive industry has experienced an evolution toward more powerful
driver assistance systems that provide enhanced vehicle safety. These systems typically operate
in the optical and microwave regions of the electromagnetic spectrum and have demonstrated
high efficiency in collision and risk avoidance. Microwave radar systems are particularly
relevant due to their operational robustness under adverse weather or illumination conditions.
Our objective is to study different signal processing techniques suitable for extraction of accurate
micro-Doppler signatures of slow moving objects in dense urban environments. Selection of the
appropriate signal processing technique is crucial for the extraction of accurate micro-Doppler
signatures that will lead to better results in a radar classifier system. For this purpose, we perform
simulations of typical radar detection responses in common driving situations and conduct the
analysis with several signal processing algorithms, including short time Fourier Transform,
continuous wavelet or Kernel based analysis methods. We take into account factors such as the
relative movement between the host vehicle and the target, and the non-stationary nature of the
target’s movement. A comparison of results reveals that short time Fourier Transform would be
the best approach for detection and tracking purposes, while the continuous wavelet would be the
best suited for classification purposes.

Keywords: Automotive systems, Micro-Doppler signature, non-stationary, Time-Frequency

analysis, FMCW Radar, Signal Processing techniques.

1. INTRODUCTION

Since the first introduction of radar systems into automotive research in the late 50’s, intensive
research in the field has led to increased integration of radar devices, operating alone or

Radar Sensor Technology XVIII, edited by Kenneth I. Ranney, Armin Doerry, Proc. of SPIE
Vol. 9077, 90771A · © 2014 SPIE · CCC code: 0277-786X/14/$18 · doi: 10.1117/12.2053323

Proc. of SPIE Vol. 9077 90771A-1

combined with other sensors. An important goal is to develop radar-based driver assistance
systems that provide increased safety to drivers and passengers. For instance, collision
avoidance/mitigation and occupant protection systems using radars have been developed [1, 2].

The commercialization of automotive radar systems began in the 1990’s, as sensing technologies
for driver assistance systems such as parking aid, collision warning and Adaptive Cruise Control
(ACC). Initially, collision warning systems emitted sound signals to assist the driver decision-
making process by providing information about the environment, giving the driver additional
time for responding to a threat and increasingly by providing the input for a control intervention
of the vehicle. Later on, research trends focused on the development of control brake systems,
capable of responding to a potential danger without driver intervention.

Automotive radar has shown its capabilities for collision avoidance and reducing the number and
severity of road accidents. Current research remains focused on increasing automotive safety,
and autonomous driving has become the main research trend for future advanced driver systems
assistance (ADAS) approaches [1, 2].

While several driver assistance systems have already been commercialized, none of them is
capable of distinguishing components of urban environments. Urban environments represent a
complex scenario due to the increased traffic density and diversity of traffic components,
including pedestrians and bicyclists, i.e. vulnerable road users. Pedestrians and bicyclists
represent an important percentage of traffic fatalities [3] and are the subject of increased research
in ADAS. Therefore, it is important to develop appropriate radar classification systems to allow
detecting vulnerable road users in order for the vehicle hosting the ADAS (HV) to react properly
to each of them. Object classification is primarily based on the difference in the micro-Doppler
spectrum of the different targets [4]. Classification of vulnerable road users will be particularly
relevant for the further development of current semi-autonomous driving systems [5] in order to
achieve full autonomy.

This article presents an analysis of signal processing methods used to most accurately extract
micro-Doppler signatures of traffic participants that will lead to a higher probability of correct
classification. Section 2 is an overview of Frequency Modulated Continuous Wave (FMCW)
radar, which is commonly used in the automobile industry, and includes the analysis of the
transmitted and received signals for short range environments. Section 3 describes car and
pedestrian movement patterns and Section 4 summarizes the results obtained by performing the
analysis with 3 different signal processing methods: Short Time Fourier Transform, Continuous
Wavelet and Wigner-Ville distribution. Through the use of those methods, we can capture the
transitions of the signal over time and obtain better time and/or frequency resolution when using
the correct transform parameters for our case of study. A comparison between methods,
constrains and results is presented on Section 5.

Proc. of SPIE Vol. 9077 90771A-2

 2. FMCW analysis for short range automotive radars

Frequency Modulated Continuous Wave (FMCW) radars are widely used in automotive
applications, commonly using a sequence chirp modulation. The block diagram of FMCW radar
is presented in Figure 1. The radar emits a continuous signal with linear frequency modulation,
and the received wave is mixed with a copy of the transmitted signal and downconverted and
filtered to eliminate high frequency components. The instantaneous range and velocity of the
target can be extracted from the phase of the received down-converted signal. To simplify the
analysis, we will consider the case of a system with only one receiving antenna for a single target
situation. Analysis can be extended for multiple antenna and target environments.

Transmitter
Antennna
Waveform
Transmitter
Generator

Receiver
Antenna
ADC LPF
O Receiver Front End

Figure 1. FMCW Radar processing squeme

Let us consider a FMCW system with frequency f0, bandwidth fsweep and a chirp duration tchirp.
As described in [6], the transmitted signal can be expressed as:

x(t ) = ATX cos(θ TX (t ) ) (1)

Where ATX is the amplitude of the signal and θTX is the instantaneous phase. For simplicity, we
will consider the amplitude maximum is normalized to 1. The instantaneous phase can be
formulated as:

Proc. of SPIE Vol. 9077 90771A-3

 ⎡ f sweep ⎤
θ TX = 2π ⎢ f 0t + t2⎥ (2)
⎢⎣ 2Tchirp ⎥⎦

The received echo is a copy of the transmitted signal, attenuated and delayed in time. For
simplicity, and since the focus of this paper is on frequency analysis, we will consider the
maximum amplitude of the received signal is normalized to 1. Thus, it can be expressed as:

r (t ) = ARX x(t − τ ) = ARX ATX cos(θ RX (t ) ) = cos(θ RX (t ) ) (3)

The instantaneous phase of the received signal can be formulated as:

⎡ f sweep ⎤
θ RX = 2π ⎢ f 0 (t − τ ) + (t − τ ) 2 ⎥ (4)
⎢⎣ 2Tchirp ⎥⎦

where τ is the delay of the received signal with respect to the transmitted one. This delay
depends on the distance of the target from the radar, since the travelling velocity of waves
through air is fixed. For moving targets, this will also depend on the movement. Since the
velocity would rarely be constant in normal road situations, an acceleration term must be
considered when performing the analysis. The delay τ can then be expressed as:

( R0 + v0 t + at 2 )
τ (t ) = 2 (5)
c

where R0 is the initial range, v0 is the initial velocity and a is the acceleration of the moving
target. The phase of the downconverted received signal can then be expressed as:

⎡⎛ R0 f sweep 4 R02 ⎞ ⎛ v0 f sweep 2 R0 f sweep 8R0 v0 ⎞ ⎤

⎜
⎢⎜ 2 − ⎟ + ⎜2 + − ⎟t ⎥
⎢⎝ λ 2Tchirp c ⎠ ⎜⎝ λ Tchirp c
2 ⎟
2Tchirp c ⎟⎠
2
⎥
θ d (t ) = 2π ⎢ ⎥ (6)
⎢ ⎛⎜ a 2 f sweep v0 f sweep 8R0 a f sweep 4v0 ⎞⎟ 2 f sweep 8v0 at f sweep 4a 2 t 4 ⎥
2 3

⎢+ ⎜ 2 λ + cT −
2T c 2
−
2T c 2 ⎟
t −
2T c 2
−
2Tchirp c 2 ⎥⎦
⎣ ⎝ chirp chirp chirp ⎠ chirp

And the beat frequency (fb), which is the frequency difference between transmitted signal and
received echo, can be written as:

Proc. of SPIE Vol. 9077 90771A-4

 1 d 2v f 2R f 4R 2 ⎛ a 2f v f 8R a f 4v 2 ⎞
f b (t ) = θ d (t ) = 0 + sweep 0 − sweep 20 + 2t ⎜⎜ 2 + sweep 0 − sweep 20 − sweep 20 ⎟⎟
2π dt λ Tchirp c 2Tchirp c ⎝ λ cTchirp 2Tchirp c 2Tchirp c ⎠
⎛ f sweep 8v0 a ⎞ f 2 4
− 3t 2 ⎜ ⎟ − 4t 3 sweep 4a t
⎜ 2T 2 ⎟
2Tchirp c 2
⎝ chirp c ⎠
(7)

For our specific case, we consider f0=76GHz, fsweep=500MHz, Tchirp= 20 to 200 µsec, 1m≤R0≤5m
and a sequence of 128 chirps. Substituting those values on the expression above to study the
values of each separate term, we identify negligible terms. Eliminating those terms, we can
rewrite the beat frequency as:

1 d 2v f 2R ⎛ a 2f v ⎞
f b (t ) = θ d (t ) = 0 + sweep 0 + 2t ⎜⎜ 2 + sweep 0 ⎟⎟ (8)
2π dt λ Tchirp c ⎝ λ cTchirp ⎠

This expression is valid to describe the beat frequency from a car target. In order to describe the
case of a pedestrian target, we need to consider that the velocity of the movement is different that
the punctual velocities of different parts of the body. Therefore, the expression should be
rewritten as:

2vm f sweep 2 R0 ⎛ a 2f v ⎞
f b (t ) = + + 2t ⎜ 2 + sweep 0 ⎟ (9)
λ Tchirp c ⎜ λ cTchirp ⎟⎠
⎝

Where vm is the individual velocity of the body part and v0 is the velocity of movement of the
pedestrian.

The first two terms of the left side of the expression are time independent. These terms are
composed of the Doppler term (fD) and the range frequency term (fR) from which the initial
range of the target may be extracted. For extremely short-range targets, those values will be
comparable as shown in Tables 1 and 2.

R0=1m R0=2m R0=3m R0=4m R0=5m

Tchirp =20µs 1.67*105Hz 3.33*105Hz 5.01*105Hz 6.67*105Hz 8.35*105Hz

Tchirp =200µs 1.67*104Hz 3.33*104Hz 5.01*104Hz 6.67*104Hz 8.35*104Hz

Table 1. Range Frequency for different chirp duration and range values

Proc. of SPIE Vol. 9077 90771A-5

 Velocity v0= 3m/s v0= 9m/s v0= 15m/s v0= 21m/s v0= 27m/s v0= 33m/s

Doppler shift 1540 Hz 4620 Hz 7700 Hz 10780 Hz 13860 Hz 16940 Hz

Table 2. Doppler Frequency values for different velocities

The time-dependent terms of the beat frequency represent the change of the frequency shift over
time due to the change in range of the target with respect to the radar due to its movement. The
combination of those two terms, whose expected values are shown in Tables 3 and 4, will cause
a progressive change on the beat frequency from the first to the last chirp of the sequence, due to
the change on range of the target caused by its movement.

v0=3m/s v0=9m/s v0=15m/s v0=21m/s v0=27m/s v0=33m/s

a=4m/s2 2.6013*103Hz 7.6013*103Hz 1.2601*104Hz 1.7601*104Hz 2.2601*104Hz 2.7601*104Hz

a=6m/s2 2.652*103 Hz 7.652*103 Hz 1.2652*104Hz 1.7652*104Hz 2.2652*104Hz 2.7652*104Hz

a=8m/s2 2.7027*103Hz 7.7027*103Hz 1.2703*104Hz 1.7703*104Hz 2.2703*104Hz 2.7703*104Hz

a=10m/s2 2.7533*103Hz 7.7533*103Hz 1.2753*104Hz 1.7753*104Hz 2.2753*104Hz 2.7753*104Hz

Table 3. Frequency shift increase/decrease over a 25 msec train due to movement and acceleration for Tchirp=200µsec.

v0=3m/s v0=9m/s v0=15m/s v0=21m/s v0=27m/s v0=33m/s

a=4m/s2 2.5101*104Hz 7.5101*104Hz 1.2510*105Hz 1.7510*105Hz 2.2510*105Hz 2.7510*105Hz

a=6m/s2 2.5152*104Hz 7.5152*104Hz 1.2515*105Hz 1.7515*105Hz 2.2515*105Hz 2.7515*105Hz

a=8m/s2 2.5203*104Hz 7.5203*104Hz 1.2520*105Hz 1.7520*105Hz 2.2520*105Hz 2.7520*105Hz

a=10m/s2 2.5253*104Hz 7.5253*104Hz 1.2525*105Hz 1.7525*105Hz 2.2525*105Hz 2.7525*105Hz

Table 4. Frequency shift increase/decrease over a 25 msec train due to movement for Tchirp=20µsec.

These results are in sharp contrast to the work of Smith and Baker [4] which assumed Doppler
Frequency shift within a chirp is negligible compared to range frequency shift for their target
range. Here the Doppler frequency shift for short range measurements is comparable to the range
frequency and the frequency shift varies over time from the first to the last chirp of the sequence
due to the range increase/decreased caused by the movement of the object. This variability of

Proc. of SPIE Vol. 9077 90771A-6

frequency over time establishes the need for a time-frequency analysis of the received signal
conducted over rather short-periods of time.

3. Environment Description

Urban environments are composed by a variety of objects, stationary or moving or both. For our
simulations we will focus on two types of moving objects: cars and pedestrians. We will analyze
the frequency shift of the backscatters for those two types of targets. We will model the radar as
and I/Q system with an antenna diameter of 60-100 mm [17], therefore all targets are located on
the far field or Fraunhofer region [18].

First, we consider a car target, initially located at R0=1m from the radar, and moving with an
initial velocity v0=21 m/s (75.6km/h) and experiencing a constant acceleration a=6m/s2 (77760
km/h2). Due to the frequency homogeneity of the point backscatters from this type of target, it is
expected that it will present a narrow Doppler spectrum [7, 8]. Next, we consider a pedestrian
target, with a displacement velocity (torso velocity) of 2m/s. Due to rotations and translations of
legs and arms superimposed to the displacement movement, the spectrum of the pedestrian
presents velocities ranging from zero to approximately 4 times higher than the torso mean
velocity, which yields an expanded micro Doppler spectrum [8]. For our simulations we will use
a typical model for description of pedestrian movement based on the six-point model depicted on
Figure 2. It divides the pedestrian in the superior half, known as passenger, which includes head,
arms and trunk (HAT) and the inferior part, that corresponds to the legs. The trunk or torso
moves with an oscillating velocity which mean corresponds to the velocity of the motion. The
legs have a cyclical trajectory and their velocity ranges between almost zero for the standing leg
to up to four times the torso velocity. Also the movement of the arms also contributes to the
spread of the Doppler spectrum, it is not included on the model because its swing is not
characteristic in the movement process [8].

Proc. of SPIE Vol. 9077 90771A-7

 Figure 2. The six reflection point model

Some approaches propose the processing of FMCW radar signals by performing a 2D FFT . In a
first step an FT is performed over N sampling points of each chirp, from which the range could
be obtained, assuming the velocity of the target is constant and the Doppler shift is negligible
compared to the range frequency. Applying the Fourier Transform over K chirps of the train, the
velocity of the target could be calculated. This processing must be performed individually for
every reflection. In the following sections we will study the results obtained for 3 different time-
frequency analysis approaches as applied to point target reflections at a very short range.

800
N=1 N=1
uu N=128 700 - N=128

800

700

600

500
E
400

300

200

100

-1.5 1 -0.5 0 0 5 1 1.5 2 1.5 0.5 0 0.5 1.5 2

freq (Hz)
x 105 freq (Hz)
x 1 05

Figure 3. Spectrum of reflection for first and last chirp of the train for a car (left) and a pedestrian(right)

Proc. of SPIE Vol. 9077 90771A-8

Figure 3 displays the Fourier Transforms for the first and last chirp of the signal train for a car
and a pedestrian. As we can see, we obtain a sinc function for every chirp, centered on the beat
frequency. This beat frequency evolves depending on the direction of movement of the target
and its velocity, stating the necessity of a time-frequency analysis.

4. Time-Frequency Analysis

In order obtain an accurate instantaneous frequency spectrum and an indicator of its evolution
over time, it is necessary to apply time-frequency analysis methods to the backscattered signal.
We present and analyze the response obtained for 3 of those methods: Short-Time Fourier
Transform (STFT), Continuous Wave Transform (CW) and Wigner-Ville Transform (WV).

4.1 Short-Time Fourier Transform Analysis

The idea of the Short Time Fourier Transform is to break the signal into smaller intervals, within
which we can consider the signal to be stationary. The length and type of window used to break
the signal determines the frequency resolution and the capability of capturing the non-stationary
nature of the signal [9]. The STFT expression is shown below, where x(t) is the signal under
study and h(t) is the function of the selected window.
∞
1
∫ x(t )h(t − τ )e
−iωτ
X (ω , t ) = dτ (10)
2π −∞

Where h(t) is a hamming window [14] with a length equivalent to a chirp duration:

⎛ 2πn ⎞
h(t ) = 0.54 − 0.46 cos⎜ ⎟ (11)
⎝ N −1⎠

Results obtained for car and pedestrian targets are displayed on Figure 3. We see how the
relevant features of this two signals were captured by the STFT. On the figure on the left we
observe how for the car the energy is concentrated around a single frequency while on the figure
on the right for the pedestrian the energy is spread over a range of frequencies for every
instantaneous time. We can also extract both range and Doppler information by using this
transform. From the slope of the time-frequency ramp we can derive the velocity of the target.

Proc. of SPIE Vol. 9077 90771A-9

Since the acceleration has a small impact over such a short time, it can be neglected, and the
slope of the frequency-time ramp would be directly proportional to the velocity of the target.The
absolute frequency value for each temporal instant gives us the frequency beat, which is the
addition of range frequency and Doppler frequency. Knowing the value of the target’s velocity,
the range calculation is straightforward. For longer observation times, the acceleration may have
an impact over the time-frequency distribution, causing the ramp to adopt a more parabolic
shape.

x 105 x 105

-1.5 -1.5

-1

7-7 -0.5

N
a 1 I: 1
r i* tY0F :' 1, 1: 1 : 1 C Mi -?
Lf 0.5

1
1

1.5
1.5

2
0 0.005 0.01 0.015 0.02 0.025 2
Time(s)
0 0.005 0.01 0.015 0.02 0.025
Time(s)

Figure 4. STFT of a train of chirps for a car experiencing constant acceleration (left) and a pedestrian target (right)

The most important issue when using the STFT is to select the appropiate length and type of
window, since an incorrect choice could led to loss of resolution [9,15].

4.2 Continuous Wave Transform Analysis

The Continous Wave Transform (CWT) transform provides an analysis of a given signal by
studying its similarity with a predefined wavelet on a scale ɑ and time b. Wavelets of different
scales are used to generate the transform coefficients. An appropiate wavelet form, depending on
the waveform used by our radar system, must be selected to extract accurate information. This is
a time-scale transform rather than a time-frequency transform (scale and frequency are inverserly
proportional) [9]. The Wavelet transform represent an alternative to STFT that overcomes the
resolution problem by providing a more flexible windowing approah, with the use of a scaled
wavelet. The transform is computed separately for different segments on the time-domain. [15]

Proc. of SPIE Vol. 9077 90771A-10

Results for CWT transform for both car and pedestrian targets, using a Morlet Wavelet [11] , are
displayed in Figure 4. The CWT expresion is shown below, where x(t) is the signal under study
and ψ(t) is the selected wave:

1 ⎛t −b⎞
X ωt (a, b) =
a
∫ x(t )ψ ⎜
⎝ a ⎠
⎟dt (12)

where a is the scale factor, b is the translation factor and ψ(t) is the Morlet Wavelet, described in
(11). The translation factor establish the center of the waveform and the scales determine its
amplitude and therefore its shape, since the area under the waveform must be constant for every
scale. Here fb is the wavelet bandwidth and fc is the wavelet center frequency. The translation
factor allows to focus on our region of interest, while the scaling parameter allow the use of
multiples scales to generate many wavelet transform coefficients. As result of this flexibility on
the basis functions, simultaneous analysis of high-frequency and low-frequency structures is
possible [9,12].

1
ψ (t ) = exp(2 jπf c t ) exp(−t 2 / f b ) (13)
πf b

0.005 0.01 0.015 0.02 0.025 0 00 0.0 0.015 0.02 0.025

Time Time

Figure 5. CWT of a train of chirps for a car (left) and a pedestrian (right)

For the car, the energy is concentrated around a small set of scale values, while for the pedestrian
it is distributed among a wider set of scale values. For example, for the time instant 0.002s, most
of the energy is concentrated between the scale range [3.1*10-5 , 3.7*10-5], while for the
pedestrian is spread in the scale range [4.3*105,6.3*10-5]. Also, the evolution of the energy
distribution within the scales over time provides an insigth on the movement of the target [12].

Proc. of SPIE Vol. 9077 90771A-11

An approximate relation between scale and frequency for an specific wavelet can be developed,
providing a time – pseudofrequency representation. [19]

4.3 Wigner-Ville Distribution (WVD)

The Wigner distribution is the classical example of a joint time-frequency distribution

quantitatively different from the spectrogram. The idea is to characterize time-frequency
functions by using an auxiliary function or kernel. The properties of a distribution are reflected
by simple constraints on the kernel. Several distributions have been widely used as kernel
functions. For our simulations we selected the Wigner-Ville distribution, which uses a kernel
function φ (θ ,τ ) = 1. This can be seen as the Fourier Transform of the instantaneous
autocorrelation function of the signal under study. By doing this, we obtain better time and
frequency resolution that with the STFT, and we don’t have to sacrifice one resolution in favor
of the other [9]. The expression of the Wigner-Ville distribution is displayed below, where x(t) is
the signal under study, and x*(t) its conjugate. [10]

1 ⎛ τ⎞ ⎛ τ⎞
∫ x⎜⎝ t + 2 ⎟⎠x * ⎜⎝ t − 2 ⎟⎠e
−iωτ
X (ω , t ) = dτ (14)
2π

10

. e'}'!'F1'i's.3f.,ÿ 4411..'e'l e: cS°.°,:..

o
0.005 0.01 0.015 0.02 0.025
o 0.005 0.01 0.015 0.02 0.025
Time(s)
Time(s)

Figure 6. Wigner-Ville distribution of a train of chirps for a car target (left) and a pedestrian target (right)

Proc. of SPIE Vol. 9077 90771A-12

Although the Wigner distribution allows us to capture the frequency-time variability of a signal
and to ascertain wheter a signal is multicomponent while providing good temporal and
frequential resolutions, it suffers from the artifacts derived from the emergence of cross-
frequency terms. Those cross terms complicate the interpretation of results. [16].

5. Conclusion

Dense urban environments, with their wide variety of stationary and moving objects, represent
complex scenarios for radar signal processing and analysis. In order to detect and classify
between differents components it is necessary to use time-frequency analysis methods that
provide us with information about the evolution of the movement through time and its nature.

For this purpose, the results for three time-frequency analysis methods are presented on this
article. Both STFT and CW constitute good alternatives for time-frequency analysis, although
their characteristics differ, making each of them better suited for different functions.

The STFT is able to capture the frequential variability of the signal over time, providing results
that are easier to interpret than in the case of the CW. This transform would be the better suited
for target detection and parameter estimation, since calculation of range and velocity of the
targets are straightforward. However, the frequenty resolution depends on the time resolution or
size of the window. Thefore we cannot simultaneously focus on high and low frequencies. This
could bring up some issues when studying targets at different very different ranges.[12]

The CW transform provides greater flexibility and, therefore, a better time-frequency resolution
than the STFT. However, calculation of parameters such as target range or target velocity is
harder, since the conversion scale to frequency is not straightforward. Therefore, this tool would
be better suited for performing target classification.[12]

In sharp contrast, the Wigner-Ville distribution, despite its good time and frequency resolution,
adds complexity to the analysis due to the introduction of cross terms. Thus, it is not the most
adequate signal processing tool since the emergence of frequential cross-terms adds more
complications to an already difficult scenario.

6. References

[1] Schneider, Martin. “Automotive Radar – Status and Trends”, Proceedings of the German
Microwave Conference 2005. University of Ulm, Ulm, Germany 05 Apr - 07 Apr 2005.

Proc. of SPIE Vol. 9077 90771A-13

[2] Meinel and Dickmann, “Automotive Radar: From Its Origins to Future Directions”,
MicroWave Journal, 2013. vol. 56(No.9):24-40.
[3] Pedestrian and bicyclist crash statistics. 2012. Pedestrian and Bicycle information center.
Web. http://www.pedbikeinfo.org/data/factsheet_crash.cfm.
[4] Smith, Woodbridge and Baker. “Micro-Doppler Signature Classification”. International
Conference on Radar. vol., no., pp.1,4, 16-19. Shangai, Oct. 2006
[5] Five Mind-Blowing Innovations from Mercedes-Benz. 2014. Autoblog. Web.
http://www.autoblog.com/2011/04/12/five-mind-blowing-innovations-from-mercedes-benz/
[6] Schroeder and Rohling. “X-Band FMCW Radar System with Variable Chirp Duration”,
IEEE Radar Conference vol., no., pp.1255,1259, Washington D.C., 2010.
[7] Rohling, Heuel, Ritter. “Pedestrian Detection Procedure integrated into a 24 GHz Automotive
Radar”. IEEE Radar Conference vol., no., pp.1229,1232. Washington D.C., 2010.
[8] Heuel, Rohling. “Two-Stage Pedestrian Classification in Automotive Radar Systems”.
Proceedings of International Radar Symposium (IRS) vol., no., pp.477,484, Leipzig, 2011.
[9] Bradley Lee Barnhart. “The Hilbert-Huang Transform: theory, applications, development”.
Diss. University of Iowa 2011. Web.
http://ir.uiowa.edu/cgi/viewcontent.cgi?article=2751&context=etd
[10] Leon Cohen. “Time-Frequency Analysis”. Prentice Hall Signal Processing Series. Print.
[11] Huailin, Jiren. “Instantaneous Feature Extraction of Non-stationary Signal Based on
Complex Morlet Wavelet”. IEEE International Conference on Electronic &Mechanical
Engineering and Information Technology vol.3, no., pp.1147,1150,. China, 2011.
[12] Chun-Lin, Liu “A Tutorial of the Wavelet Transform”. February 23, 2010. Web.
http://disp.ee.ntu.edu.tw/tutorial/WaveletTutorial.pdf
[13]Andres, Ishak, Menzel and Bloecher. “Extraction of Micro-Doppler Signatures Using
Automotive Radar Sensors”. Journal of RF-Engineering and Telecommunications, Volume 66,
Issue 11-12, Pages 371–377. December 2012.
[14] K. M. M. Prabhu . “Window Functions and Their Applications in Signal Processing”. CRC
Press. Print.
[15]Polikar. The Wavelet Tutorial. Rowan University. Web.
http://users.rowan.edu/~polikar/WAVELETS/WTpart3.html
[16] Auger, Frandlin, Gonçalvès, Lemoine. “ Time-frequency Toolbox. For use with Matlab”.
CNRS & Rice University,1995-1996.Web. http://tftb.nongnu.org/.
[17] Menzel and Moebius. “Antenna Concepts for Millimeter-Wave Automotive Radar Sensors”.
Proceedings of the IEEE Vol. 100, No. 7. July 2012.
[18] Balanis. “Antenna Theory: Analysis and Design”. John Wiley & Sons, 2012. Print.
[19] Scale to Frequency. Documentation Center, Mathworks. Web.
http://www.mathworks.com/help/wavelet/ref/scal2frq.html

Proc. of SPIE Vol. 9077 90771A-14