weiwen deng, shuqing zeng, qingrong zhao and john daichenyanh/.downloads/ijvd_submission-0… ·...
TRANSCRIPT
Int. J. of xxxxxxxx, Vol. x, No. x, xxxx
Copyright © 2009 Inderscience Enterprises Ltd.
Modeling and Simulation of Sensor-Guided Autonomous Driving
Weiwen Deng, Shuqing Zeng, Qingrong Zhao and John Dai Global Research & Development Center
General Motors Company
Warren, MI 48307, USA
Abstract: This paper presents a modeling and simulation method for sensor-
guided autonomous driving. A generic model of range sensing and object
detection in 3D space is discussed first that represents their high-level functions.
The low-level physical characteristics of range sensing and object detection are
further investigated based on frequency modulated continuous wave (FMCW)
radar which is gaining wide popularity in automotive applications. These are the
enablers to modeling and simulation of vehicle interactions with one another
under traffic and with surrounding environment. A closed-loop adaptive cruise
control is then used as an example to demonstrate the vehicle limited
autonomous driving with the proposed model and method that has been shown
to be valid, effective and numerically efficient.
Keyword: Range sensing; Object detection; Modeling and simulation; Vehicle
active safety; Autonomous driving
Biographical notes:
Dr. Weiwen Deng has been working for General Motors R&D Center since
1996. He holds B.Sc. and M.Sc. degrees in Mechanical Engineering from China,
and M.Sc. in Mechanical Engineering and Ph.D. in Electrical Engineering from
U.S. He is currently a Staff Researcher at GM and the twice recipients of the
`Boss' Kettering Award, the most prestigious award in GM for technical
invention and innovation. He holds 17 US patents with another 16 pending, and
is the author of over 30 journal and conference publications. He also serves as
editor and associate editors for several international journals.
Dr. Shuqing Zeng received his PhD degree in Computer Science from Michigan
State University in 2004. He joined General Motors R&D Center in 2004 as a Sr.
Researcher. He was a member of Tartan Racing team who won the first place of
The Defense Advanced Research Projects Agency (DARPA) Urban Challenge
in 2007. He served as a reviewer to IEEE Transactions on Pattern Analysis and
Machine Intelligence and as a judge to Intelligent Ground Vehicle Competition
(IGVC). He served as the newsletter editor (2005-2007) of Autonomous Mental
Development TC, IEEE Computational Intelligence Society. He is currently an
associate editor of International Journal Humanoid Robotics (IJHR).
Dr. Qingrong Zhao is a Sr. Consultant to GM R&D Center from Wipro
Technologies (USA) since 2006 on various projects related to vehicle
autonomous driving and controls on vehicle autonomy. She holds B.Sc. and
M.Sc. degrees from Northwestern Polytechnical University of China and Ph.D.
from University of Cincinnati, all in Electrical Engineering. She is the author of
a dozen peer-reviewed papers of international journals and conferences, and co-
inventor of two filed US patents, and has received numerous honors and awards
from schools and government officials. Currently she serves as an associate
editor of International Journal of Vehicle Autonomous Systems.
Dr. John Dai received his Ph.D. degree in Systems Engineering from Oakland
University in 2002. His research interests include dynamic system modeling and
control methodology. He currently works in General Motors powertrain
engineering as a Sr. Engineer, developing cutting-edge engine control algorithm
and software.
1. Introduction
As new technologies emerge in recent years in the automotive industry, the
demands to extend the traditional human-driving capabilities have been greatly
increased with various active safety and driver-assistance features that enhance
driver comfort, convenience, and safety, which eventually will lead to full
sensor-guided autonomous driving. The trends have been largely driven not only
by the demands, but also by the technology progression and readiness of, in
particular, the low-cost sensors, such as radar, lidar and camera.
It is well known that physical on-road vehicle testing can be very costly, time-consuming, and sometimes even impossible. In addition, physical road
testing is typically unrepeatable and with limited testing scenarios and
conditions. To effectively develop, test and verify the sensor-guided autonomous
driving systems, it has become critically important to properly model and
simulate the system, its surrounding environment and traffic conditions with
adequate fidelity in sensing and object detection. This paper presents a modeling
and simulation approach for sensor-guided autonomous driving, in particular, the
object sensing and detection, to enable the research and development, early
testing and verification on features and functions of active safety and
autonomous driving under a virtual environment.
A sensor-guided autonomous driving system is a complex system that
includes a subject vehicle along with principle other vehicles (POV), road and
infrastructure of the road and traffic, surrounding environment including weather
and light conditions, and many others which may impact or impair autonomous
Modeling and Simulation of Sensor-Guided Autonomous Driving
driving. In addition, autonomous driving is not necessarily driverless, especially
at its early stage; human driver plays an important role. Since vehicle and driver modeling has been extensively discussed in the literature, and is relatively mature,
the focus of this paper is on object sensing and detection modeling, which is one
of the key enabling functions for simulating vehicle interaction with others and
its surrounding environment.
Although modeling object sensing and detection has long been researched in
the literature (Bacso and Bertolini, 1988 and Donohue, 1997), the progression in
senor technology has been greatly advanced during last few years. Thus, those
models with much of the focuses on the physics become less flexible, or even less useful for different sensors used today. In addition, physic model of a sensor
for object detection may not be always feasible practically, due to the complexity
and stochastic nature of the sensor, object, and often unstructured environment.
This paper presents a generic modeling approach with combined functional
and physic models of sensors to abstract range sensing and object detection: in a
higher level, the functions of object sensing and detection are to identify the
range, direction and speed of both moving and stationary objects. In a lower level,
such functions can include physical phenomena such as surface curvature,
material absorption and reflectivity, and be represented by capturing the
stochastic nature of the physics with physically interpretable parameters, such as
noise.
The higher-level geometric and functional representation of range sensor and
object detection in 3D space is briefly discussed in Section 2. More detail can be
found in a separate paper (Deng, et al, 2009). Section 3 presents an enhanced
modeling of range sensing and object detection with a radar model, as an
example, of a popular kind in automotive industry with frequency modulated
continuous wave (FMCW). This model takes the input of the scatters from the
geometric models discussed in Section 2, and generates object tracks with
superimposed noises from various sources. A closed-loop simulation of an
autonomous driving with range sensing and object detection, and control has
been conducted in Section 4. This further demonstrates that the proposed model
and algorithm are valid and effective in enabling the simulation of vehicle active
safety systems with multiple vehicle interactions under traffic. The paper
concludes in Section 5.
2. Functional Modeling of Range Sensing and Object Detection
To abstract the generic functions of range sensing and object detection, a geometric approach is first taken to represent the sensor beam to be a cone and the object to be a cuboid. Then the idealized sensing model is to generate the discretized scatters from the object(s). Noise will then be superimposed to the scatters and generates a more realistic object tracks with range, range rate and
z
x β
x α
y
R
sensor beam
x⋅tan(β/2)
y x⋅tan(α/2)
z
azimuth angle, to reflect stochastic nature of sensing physics.
A. Geometric Representation of Sensor Beam, Vehicle and Objects
A sensor beam is assumed to have attributes of maximal range R , azimuth
angle α and elevation angle β . The cone that simulates the sensor view volume
(SVV) is expressed in (1) where ( , , )x y z denote the coordinates of a 3D point.
Figure 1 shows the three orthographic projections of the sensor beam.
≤++
≤+
2222
22
2
22
2
1
)2
(tan)2
(tan
Rzyx
x
z
x
y
βα
(1)
Figure 1: Sensor beam (cone) in 3D space
An object is geometrically represented by a rectangular cuboid, as shown in
Figure 2, with 6 rectangles (faces), 12 edges, and 8 vertices. Its geometric
attributes are described by {L W H a b c}.
z3
y3 y3
x3
L
a
b
W
H c
Modeling and Simulation of Sensor-Guided Autonomous Driving
Z3
Y3
X3
OBJECT
Z2
Y2
X2
X1
Z1
Y1
Y0
X0
Z0
VEHICLE
SENSOR BEAM
WORLD
Figure 2: Geometric representation of object (a cuboid)
The axis and coordinate systems used in range sensing and object detection
are shown in Figure3 for world, vehicle, sensor, and object respectively. Under
world frame, denoted by o0x0y0z0, vehicle position and orientation (as yaw-pitch-
roll angles) can be defined as:
[ ]T
vzvyvx ppp=vp, [ ]T
vvv φθψ=vo
Similarly in vehicle frame, denoted by o1x1y1z1, sensor mounting position sp
and position error s∆p , and sensor mounting orientation so (as yaw-pitch-roll
angles) and orientation error s∆o are defined, respectively, as:
[ ]T
szsysx ppp=sp , [ ]T
szsysx ppp ∆∆∆=s∆p
[ ]T
sss φθψ=so , [ ]T
sss φθψ ∆∆∆=s∆o
Figure 3: Axis and coordinate systems for world (o0x0y0z0), vehicle (o1x1y1z1),
sensor, and object (o3x3y3z3)
z3
x3
y3
6
5
8
7
3
2
1
4
object cuboid
Under the world frame, the origin of the object coordinate system, its
orientation in yaw-pitch-roll sequences are defined, respectively, as:
[ ]T
ozoyox ppp=0p , [ ]T
ooo φθψ=0o
Range sensing is to compute the sensing attributes, such as range, range rate,
object azimuth angle and object elevation angle, while object detection is to
determine if the object is falling into the SVV.
In order to compute the sensing attributes, the position and orientation of the
object (a cuboid), under the world frame, need to be first converted to the sensor
frame, under which the sensor beams have explicit mathematical expression.
To transform the coordinates of a point from world frame to vehicle frame,
the consecutive rotations are Euler angles (Gillespie, 1992), or yaw, pitch, and
roll angles [ψ,θ,φ].
Thus the rotation matrix is given by φx,θy,ψz,
1
0 RRRR = :
−
=
100
0)cos()sin(
0)sin()cos(
ψψ
ψψ
ψz,R
−
=
)cos(0)sin(
010
)sin(0)cos(
θθ
θθ
θy,R
−=
)cos()sin(0
)sin()cos(0
001
φφ
φφφx,R
Consider both rotational and translational transformation, a homogeneous
transformation matrix from the world frame to the vehicle frame is defined
as 44×∈ R1
0H :
=
× 131
v
1
01
00
pRH
(2)
Similarly, the transformation from the vehicle frame to the sensor frame can
be represented by the homogeneous transformation matrix 2
1H :
Modeling and Simulation of Sensor-Guided Autonomous Driving
+=
× 131
ss
2
12
10
∆ppRH
wheressssss ∆φφx,∆θθy,∆ψψz,
2
1 RRRR +++= , and
∆+∆+
∆+−∆+
=+
100
0)cos()sin(
0)sin()cos(
ssss
ssss
ψψψψ
ψψψψ
ss ∆ψψz,R
∆+∆+−
∆+∆+
=+
)cos(0)sin(
010
)sin(0)cos(
ssss
ssss
θθθθ
θθθθ
ss ∆θθy,R
∆+∆+
∆+−∆+=+
)cos()sin(0
)sin()cos(0
001
ssss
ssss
φφφφ
φφφφss ∆φφx,
R
Further, we can define homogenous transformation matrix from the world
frame to the sensor frame as 2
0H , and from the sensor frame to the object frame
as 3
2H , respectively:
;1
2
02
0
=
×31
2
0
oRH
=
× 131
2
3
23
20
dRH
where
3
0
T1
0
T2
1
3
0
11
0
12
1
3
0
0
1
1
2
3
2 R)(R)(RR)(R)(RRRRR ==⋅⋅= −− (3)
2
1
1
0
2
0 RRR ⋅= , )o(p)(Rd 2o
T2
02 −= , )4,3:1(2
02 Ho =
where o2 is the origin of the sensor frame projected to the world frame, and d2 is
the origin of the target frame projected to the sensor frame.
Under the object frame, all the vertices of a cuboid can be expressed as simple
functions of its geometric size, and location with respect to the origin of the
cuboid:
8,,1),,,,,,( ⋯== iHWLcbaf ioi )(p
With properly defined coordinate transformation, the vertices of an object
(cuboid) can be defined now with respect to sensor frame as:
8,,1,11
3
2 ⋯=
=
i
oii )(pH
p
B. Abstracted Range Sensing and Object Detection
The problem of object detection can be abstracted geometrically as the
intersection(s) of a cone (sensor beam) with one or more cuboids (objects) in 3D
space, that is, if an object falls into sensor view volume (SVV), then the sensing attributes can be calculated accordingly. Furthermore, the intersection between a
cuboid and a cone can be regarded as the intersection of one or more faces of the
cuboid with the cone. Therefore, the problem becomes to detect which face(s)
and how they intersect with the cone, and then to determine the minimum
distance from the sensor origin to the face(s), and the associated azimuth and
elevation angles.
The method and algorithm described below are with respect to the sensor
frame (denoted as o2x2y2z2), into which the coordinates of the object vertices, planes, and their norms are all transformed.
Denote eight vertices of an object as821 p,,p,p ⋯ , along with their
coordinates defined under the sensor coordinate system as:
[ ] .8,,1, ⋯== izyxT
iiiip
Among six faces of a rectangular cuboid, the “visible” face(s) to the sensor
beam are those who are facing the sensor beam emission direction. To determine
if a face is “visible”, denote the four sequenced points of the face
as{ }4321 pppp . The positive normal vector of the surface can be calculated
and normalized as:
[ ]11)(
)(−∈
−×−
−×−=
=2312
2312
p(p)pp
p(p)ppn
z
y
x
n
n
n
The face is “visible” to sensor beam only if
0≤• nn2X
(4)
where T]001[=2xn denotes the unit vector of the X axis in the sensor coordinate
Modeling and Simulation of Sensor-Guided Autonomous Driving
system. Therefore, the criterion (4) can be alternatively expressed as .0≤xn
For each of the “visible” faces, the next step is to determine if it intersects with the cone, a solution of which is usually difficult to obtain directly, as the
face-and-cone intersection very often causes ill-shaped boundary shapes and
therefore complicates the analysis. In this paper, we propose an effective method
that simplifies the face-and-cone intersection problem into a problem with
intersection between a line and a cone. A “visible” face is first discretized into a
number of parallel lines, as illustrated in Table 1 and Figure 4. If any of the lines
on the face intersects the cone, then intersection between the face and the cone is detected. This way, the points of intersection can be calculated more easily and
accurately.
Figure 4: Illustration of discretization of a face into lines
Table 1. Discretization of a face plane into lines
Line
number 1
st vertex (q1) 2
nd vertex (q2)
L1 p1 p4
L2 1
1•
−
−+
N
12
1
ppp 1
1•
−
−+
N
43
4
ppp
L3 2
1•
−
−+
N
12
1
ppp 2
1•
−
−+
N
43
4
ppp
… … … LN p2 p3
Denote q1, q2 as the two vertices of a line. The intersection between the line
and the cone can be discussed in two cases as below.
Case A: both q1 and q2 are inside the cone
L2
L3
L1
LN
p4
O
Z
Y
X
p3
p1
p2
In this case, the face-and-cone intersection is given by the fact that both
vertices of the line are inside the cone. Then the minimum distance from the sensor origin to the line may exist at one of the two vertices, or a point (denoted
as q) in between. For the latter, the minimum distance must also be the shortest
distance. The point q can be calculated with the following fomula:
)( 121 qqqq −+= λ
where2
12
121
)q(qq
−
−•=λ and λ∈(0, 1) .
Case B: all other conditions
In this case, the line may either intersect the cone surface, or the spherical-
frontal surface, or both, or none of them. Any point q on the line from q1 to q2
can be expressed as
)( 121 qqqq −+= λ (5)
where λ∈[0, 1]
If q is an intersection point on the cone surface, it must satisfy the following
constraints
<<
≤++
=+
>
πβαβα ,0,1
)2
(tan)2
(tan
0
2222
22
2
22
2
Rzyx
x
z
x
y
x
(6)
If q is an intersection point on the spherical-frontal surface, it must satisfy the
following constraints
<<
=++
≤+
>
πβαβα ,0,1
)2
(tan)2
(tan
0
2222
22
2
22
2
Rzyx
x
z
x
y
x
(7)
Combining (5) & (6), or (5) & (7), we can calculate if there is a solution for q,
and then determine if the line intersects with the cone. If intersection occurs, we
can seek the point of minimum distance to the sensor origin from the line section
inside the cone. We can similarly find that point as Case A.
Modeling and Simulation of Sensor-Guided Autonomous Driving
With the above analysis, the minimum distance from the sensor origin to the
intersection face can be determined as the least of minimum distances from the sensor origin to each of the lines on the face. As defined by sensing attributes,
range is the least of minimum distances from the sensor origin to the object
“visible” face(s), while the range rate can be simply obtained by taking its
derivative. Given minimum distance point T
zyx ],,[=p in sensor frame, the
azimuth angle aξ and elevation angles
eη of the object, as shown in Figure 5, can
be computed respectively as:
( )0 0atan2( , ) 180 / , 90 90a a
y xξ π ξ°= • ∈ −
( )0 0atan2( , ) 180 / , 90 90e e
z xη π η°= • ∈ −
Numerical simulation under Matlab has demonstrated, as shown in Figure 6
and 7 with some of the results, that the proposed method and algorithm has taken into account all possible cases, and can effectively determine if a cuboid
intersects with a cone under full 3D space, or if an object is fallen into SVV
(sensor view volume).
Figure 5: Object azimuth angle ξa and elevation angle ηe
ξa
ηe
(x, 0, z)
(x, y, 0)
p (x, y, z)
z2
x2
y2
sensor beam
Figure 6: Object detection: cuboid surface is normal to the cone center axis
Figure 7: Object detection: the cuboid intersects the frontal surface of cone
3. Physical Modeling of Range Sensing and Object Detection
The physical mechanism of range sensing and object detection with a range sensing (or finder) device is in principle through the emission of a particular wave, such as electromagnetic, sound or light wave, that is reflected by the object and detected by a receiver followed by the recording of its echo. In addition to the reflection, the wave can also be absorbed or scattered during this process.
0
10
20
30
40
50 60
70-20
-10
0
10
20
-20 -15 -10
-5
0
5
10
15
20
Y axis
X axis
Z axis
-10
0 10
20
30
40
50
60
-20 -10
0 10
20
-20
-10
0
10
20
X axis
Y axis
Z axis
Modeling and Simulation of Sensor-Guided Autonomous Driving
In this section, a basic radar signal processing method is presented (Winkler,
and Schoor, 2007). Based on the method, a high-fidelity model is built that models the measurement process of the range, range rate, and azimuth angle of a
target for frequency-modulated continuous wave (FMCW) radars which have
gained significant popularity in automotive applications. Thus, without loss of
generality, FMCW radars are chosen for this study.
The key principle for FMCW radar is homodyne, i.e., detecting frequency-
modulated radiation by nonlinear mixing with radiation of a reference frequency.
Figure 8 shows the FMCW radar block diagram. The CW signal is modulated in
frequency to produce a linear chirp which is radiated toward a target through the TX antenna. The return radiation is collected by four receive channels with
identical distances between two adjacent receive antennas, forming a uniform
linear array (ULA). This array enables direction of arrival (DOA) estimation
using electronically scanned scheme or high-resolution digital beam-forming.
The return signals are mixed with the TX reference signal and fed into the low-
pass filter. The yielded beat signals are digitized by the analog-to-digital circuit
and fed into the signal processing module.
Figure 8: FMCW radar system diagram
3.1. Range and Doppler Processing
The basic idea in FMCW is to generate a linear frequency ramp (chirp). The
transmission frequency for one ramp with bandwidth B and duration T between
[0, ]T can be written as
( )T c
Bf t f t
T= + (8)
CW
Mu
ltiple p
ow
er splitter
TX
RX1
Low-pass
filter ADC
Signal
processing RX2
RX3
RX4
Chirp
where cf is the carrier frequency. We apply frequency modulation (FM)
principle, and after integration, the phase ( )T tϕ of the transmitted signal
cos( ( ))T tϕ becomes
0
2
02
1)(2)(2)( ϕπττπϕ −
+== ∫ t
T
Btfdft c
t
TT (9)
where 0ϕ is the initial phase. The phase of the down-converted signal (beat)
( )tϕ∆ from the mixed signal cos( ( ))cos( ( ))T Tt tϕ ϕ τ− is:
2( ) ( ) 2 ( )
2c
B Bt t f t
T Tϕ ϕ ϕ τ π τ τ τ∆ = − − = + − (10)
where τ is the delay between the transmitted and received signal of a target. For
ASDAS application, / Tτ is sufficiently small, and the last term in (10) can be
neglected. If the target at distance R with constant velocity v is assumed, this
leads to 2( )R vt
cτ
+= , where c is the speed of light in vacuum Ignoring the
high-order term, thus (10) becomes
2 2 2
( ) 2 c cf R f v BR
t tc c Tc
ϕ π
∆ = + +
(11)
The received beat signal B cos( ( ))s tϕ= ∆ is sampled with an interval
AT .
The samples are multiplied with a window function ( )w n , and zero padding is
performed, before a fast Fourier transform (FFT). A peak at frequency IFf can be
detected:
IF
2 2cf v BRf
c Tc= + (12)
Neglecting the Doppler shift term (the first term) in (12), it can be easily seen
that the range resolution is determined by the bandwidth B and the sweep
durationT :
2
A Z
TcR
BT N∆ =
where ZN is the size of the FFT window.
We note that (12) is ambiguous in a sense that the target's range and range rate cannot be determined simultaneously with one frequency ramp. Multiple ramps
are needed to resolve the range-Doppler ambiguity. Figure 9 shows L identical
end-to-end frequency ramps. The ramp repetition interval is denoted by RRIT .
The phase drift φ∆ across contiguous sweep cycles can be derived using the
Modeling and Simulation of Sensor-Guided Autonomous Driving
following 2D-FFT frequency analysis.
Figure 9: Ramp generation for 2D-FFT
Let B,2Ds denote the 2D temporal-spatial signal generated by the target (R,v) at
different chirp sweep cycles denoted by the variable l :
RRI RRIB,2D
22
RRI
2 ( ) 2 2 ( )exp 2
2 2 2exp 2
c
c c
f Ri
c cc
f R vT l f v B R vT ls A i t
c c Tc
vT f f v BRAe i l t
c c Tc
π
π
π
+ + = + +
= + +
(13)
where A denotes the amplitude, and the frequency increase can be neglected,
because the movement during the measurement is short compared to the distance
R. It can be easily seen from (13) that RRI2 cvT f
cφ∆ = where φ∆ is defined in
Figure 9.
The following 2D Fourier transformation is performed:
1 1
2D ,2
0 0
2 1 12RRI
0 0
( , ) ( , ) exp 2
2 2 2exp 2 exp 2
2exp
c
L N
B D
l n z Z
f R L Nic cc
A
l n Z
Z
lp nkS p k s l n i
L N
vT f f v BR nkAe i l i nT
c c Tc N
i lp
L
π
π
π π
π
− −
= =
− −
= =
= − +
= + −
−
∑∑
∑ ∑
where zL and
ZN are 2D FFT window size after zero-padding. We note a peak
appears at the following position:
φ∆ φ∆ φ∆ t
f
B T
L
TRRI
RRIRRI
2 2
2
cA Z
cD Z
f v BRk T N
c Tc
vf T Lp f T L
c
= +
= =
(14)
In order to fulfill the sampling theorem, the following constraint for the
maximum Doppler frequency D,maxf must be met:
D,max
RRI
1
2f
T<
Thus the maximum range rate is limited by max
RRI4 c
cv
T f< which can be used
to design the interval time of the sweep cycle.
It can be easily seen from (13) that the velocity resolution v∆ can be
determined by the overall measurement time, RRIT L :
RRI2
c
cv
f T L∆ =
3.2. Direction-of-Arrival Estimation
In the method outlined in Section 3.1, each ( , )R v cell corresponds to a
baseband sinusoid signal Bs and may be contributed by multiple targets. To
correctly distinguish the targets, we need to estimate direction of arrival (DOA).
The common method to estimate DOA is electronically scanned scheme. By
varying the signal phases of the RX antennas, the receiving antenna becomes
directional and is a function of the azimuth angle. Figure 10 shows that the beam
pattern is steered from the right to left at different time slices. In beam-forming,
both the amplitude and phase of each antenna are controlled. The combined
amplitude a kand phase shift
kθ is called a complex weight and is represented by
a complex constantkw , for the k-th antenna. As shown in Figure 11, a beam-
former applies the complex weight to the signals from RX antennas, then sums all the signals into one that has the desired directional pattern [Haynes,1998].
Modeling and Simulation of Sensor-Guided Autonomous Driving
Figure 10: Electronically scanned scheme
Figure 11: Receive digital beam-forming
One important drawback of the beam-former method shown in Figure 11 is
the lack of angular resolution due to the limited aperture size of the antenna, i.e.,
to resolve two or more closely spaced targets in the same (R,v) cell. However, the
DOA of the incident wavefront can be improved with much higher accuracy than
the beam width of the RX antenna. Figure 12 outlines the basic idea called
phase-comparison mono-pulse, or phase-interferometry. A wavefront from
direction θ arrives at antenna 1 first. Then, after travelling an additional path
distance sinL d θ∆ = , it arrives at antenna 2 where d is the distance between
two adjacent elements. Therefore, the refined DOA can be computed
as1sin
2 d
φλθ
π− ∆
=
where φ∆ is the measured phase shift.
B1sB2s B3s
B4s
Beam
-form
er
FMCW processing
Power Combiner
t-1 t-2
t-7
t
Figure 12: Phase-comparison mono-pulse (courtesy of wikipedia.com)
3.3. FMCW Radar Model
We have outlined the basic signal processing. In this section, we will outline
an efficient radar model.
High-fidelity radar model is difficult to implement due to considerable
number of variables and conditions to be taken into account. However, with the
new advances in modeling technology and computer speed, such a model is
feasible in Matlab-Simulink environment. For example, many off-the-shelf
blocks from the standard signal processing and communication blocksets can be
used to model the major components of a radar system.
To execute the simulation in real-time, only baseband subsystem of the radar
is modeled in this paper. We need a target model to represent the received signal
if the geometric shapes of targets and clutters are given. The simple way of
implementing is to represent the target model as a collection of point scatters.
Each scatter, therefore, will be characterized by distance from radar (and, thus, a
path loss associated with that distance), radial velocity, incident angle, and the
strength of reflection (i.e., RCS). Let K scatters be denoted by their distance,
radial velocity, azimuth angle, and RCS, i.e., { }( , , , ) | 1,...,k k k kR v k Kθ σ = . We
use the following baseband approximation for the arrival signal of the j-th RX antenna, holding for most practical cases:
22
RRI, ,
1
, 42
2 2 2exp 2
( ) 1
2(4 )
c k
j
f RK ik c c k kc
B j k j
k
it t k eff
k j
k
v T f f v BRs A e i l t
c c Tc
PG AA e
R
π
φ
π
σ θ
π
=
∆
= + +
= ⋅
∑ (15)
Modeling and Simulation of Sensor-Guided Autonomous Driving
In (15), the complex amplitude ,k jA is proportional to the receiving signal
power and its phase shifts a constant increment ( φ∆ ) depending on the incident
angle, i.e., sin
2d θ
φ πλ
∆ = . tP is the transmitted power (in the transmitted
pulse); R is the distance from the sensor to the object, tG is the directional gain
of the antenna measured in the direction of the object, and ( )eff
A θ is the
effective area of the sensor receiving antenna.
Figure 13: Block schematic of the FMCW model
Figure 13 illustrates the main building blocks of the FMCW radar model.
Based on the method described in Section 2, a target is defined geometrically in a
3D space, which is then converted into scatters, as shown above at the start of the
simulation. The remaining blocks are outlined as follows:
Baseband RX Signal: The baseband receiving signal is represented by (15),
given the K scatters.
Chirp Non-Linearity: The chirp slope B
bT
= is a constant chosen by the
design of the radar. However, if the chirp slope b is not linear, then the beat
frequency for a point target will not be constant, and the range accuracy will be
reduced. b can be modeled as a function of the distance, the deviation from the
Scatters Baseband
RX Signal
Chirp
Non-Linearity
Spread
Convolution Noise
Temperature
Ambient
Environment
Attenuation
Clutter
Noise
2D FFT
Beam-
forming &
DOA
Estimation
Phase Comparison
Mono-pulse
Peak Clustering-
Association
Tracking
AGWN
nominal constant will create a bias in range measurement. A lookup table is used
to model the non-linearity in the chirp slope in this block.
Spread Convolution: The two main noise sources covered in this block are
quantization noise at analog to digital (ADC) and phase noise by the non-ideal mixing process, frequency synthesizer, and other component. A 2D kernel
function was convoluted into the receiving signal. We note that the signal in (15)
can be represented as the sum of infinite spikes (Dirac function) in frequency
domain. The convolution blurs the spikes. In addition, an additive Gaussian white
noise (AGWN) is superposed on the center position of the kernel function.
Noise Simulation: Three additional sources of RF impairment are modeled
as: Noise Temperature, Ambient Environment Attenuation, and Clutter Noise.
The noise temperature block allows us to select effective radar system noise temperature. Three typical values can be chosen: 0 K (no noise), 40 K (very low
noise level) and 290 K (typical noise level). The ambient environment
attenuation block models the attenuation effects experienced by EM waves from,
such as atmospheric gases, rain, and fog. The clutter noise block superposes
noises contributed from clutter background (e.g., ground).
2D FFT: A Fourier transformation is performed on the 2D input signal. Then
a constant false-alarm rate (CFAR) threshold is applied to the yielded frequency
spectrum. The block outputs a list of range-Doppler bins whose signal-noise-ratio (SNR) is larger than the threshold.
Beam-forming & DOA Estimation: The block can resolve two or more
closely spaced targets in the same range-Doppler bin. We iterate through each
electronically steerable beam pattern. For each beam, we multiply the input
signal by the complex weight w and sum all weighted signals together. The
beam index of the amplitude of a summed signal that is larger than a threshold
indicates the DOA angle of a target. Therefore, the block outputs a list of range-
Doppler-beam bins, and each of those bins contains a target.
Phase Comparison Mono-pulse: We refine the DOA angle of a target
detected in the proceeding block using the phase-interferometry process outlined
in Section 3.2.
Peak Clustering-Association: The resulted bins are clustered and associated
with the track database (e.g., nearest neighbor).
Tracking: The tracks are updated by Kalman filtering. Those tracks
(including range, range rate, azimuth angle) are outputted as the sensor
measurements.
Modeling and Simulation of Sensor-Guided Autonomous Driving
3.4. Simulation Results
In this subsection, we show the simulation results of the proposed FMCW
model. As shown in Figure 13, the simulation results presented here are based on
the radar with the following key parameters: radar RF frequency is 24GHz; field-
of-view is 30°; the designed sensing range is [ 50,50]− m/s for range rate and
[0.1,150] m for range; the bandwidth B =1GHz; the ramp repetition interval
-56.2457 10RRI
T = × ; the number of range bins 1024ZN = ; the number of
range rate bins 256ZL = ; 32 receiving antenna elements are placed with the
space between two adjacent elements 5d = mm; 32 electronic steerable beams
are simulated; and transmit antenna power is 10 Watts. The simulated signal-to-
noise ratio is 6dB. The effective area of the sensor receiving antenna
1eff
A = 2m .
Three trihedral corner reflectors with radar cross section (RCS) of 1 2m are
simulated with following parameters:
• Target 1: range = 20 m, range rate = 9 m/s, azimuth angle = -15°
• Target 2: range = 32 m, range rate = -10 m/s, azimuth angle = 2°
• Target 3: range = 32m, range rate = -10 m/s, azimuth angle = -2°
We note that the Targets 1 and 2 can be easily separated by range-Doppler
bin. However, Targets 2 and 3 are in the same range and range rate and cannot be
separated by range-Doppler bin.
Figure 14 is the real part of the synthesized base band signal (c.f., (15)) for the
abovementioned three targets during a chirp sweep by a receiving antenna.
Figure 14: The synthesized baseband signal of a chirp sweep
Figure 15 is the power spectral density among the range-Doppler. The result is
derived by applying 2D FFT to the synthesized 2D temporal-spatial signal ,B js
defined in (15), for the given j-th receiving antenna. We note two detected peaks
that correspond to the simulated three targets. The left peak corresponds to
Target 1 while the right FFT peak corresponds to the lumped Targets 1 and 2,
which cannot be separated by range-Doppler bin. Given an FFT peak is detected
at bin ( , )n l , the corresponded range and range rate can be computed by
( 1) ( 1),
2 2 Z RRI c
n c l cR V
B L T f
− −= =
To handle negative range rate, we shift the zero component to the center of
spectrum.
Modeling and Simulation of Sensor-Guided Autonomous Driving
Figure 15: 2D FFT power spectrum among range-Doppler bins
Figure 16 illustrates the slices along the detected two peaks in 2D FFT power
spectrum shown in Figure 15. The upper plot is the spectrum density with
respect to the range, which can be derived using a 1D FFT on the signal ,B js in a
chirp sweep window (i.e., a row of ,B js ), giving the receiving antenna j. The
lower plot is the spectrum density with respect to the range rate, which can be
derived using a 1D FFT on a column of ,B js . One can easily verify that the two
peaks in the upper plot are the ranges 20 m and 32 m, respectively, and the two
peaks in the low plot are the range rates 9 m/s and -10 m/s, respectively.
Figures 15 and 16 illustrate that Targets 2 and 3 cannot be separated only by
range-Doppler bin. However, the electronically scanned digital beams can
discriminate the two targets. Figure 17 shows the response of the all the 32
beams for the two detected range-Doppler peaks. The green curve corresponding
to Target 1 shows there is a maximum peak at -15°, which indicates the azimuth
of the target. The blue curve corresponding to Targets 2 and 3 has two maximum
peaks at about -2° and 2°, respectively. This means the response of the digital
beams can separate targets with the same range and range rate bin.
Figure 16: 1D FFT power spectrum vs. range bins and range rate bins,
respectively.
Figure 17: The amplitude of the detected range-Doppler bins (clusters) as a
function of the steerable RX beams.
Modeling and Simulation of Sensor-Guided Autonomous Driving
4. Integrated Simulation with Sensing Model
To simulate vehicles on traffic, a traffic vehicle model is designed that is
much simplified to achieve efficient numerical computation, but with sufficient
fidelity to capture basic vehicle dynamics under both linear and nonlinear operating regions.
The traffic vehicle model is used to calculate the vehicle trajectory in terms of
its position ( , )x y and orientation in yaw plane ( Φ ), and its velocity in both
lateral and longitudinal directions, xv and yv , respectively. Denote the output
vector as [ ]Tyx vvYXy Φ=
Assume steering angle fδ . To simplify modeling on vehicle longitudinal
dynamics, a speed controller is used to represent vehicle powertrain and brake
systems. Assume a set of desired vehicle speed [ ]nx vvvv ˆˆˆˆ21 ⋯= is given,
which can be from driver’s input via set speed, but can be overridden by other
control strategies, such as adaptive cruise control based upon on-board sensors.
Figure 18 shows the block diagrams of the traffic vehicle model.
Figure 18: Simple traffic vehicle model
A simple PID control can be used for the speed controller as:
( )xxdxxixxpx vv
dt
dkdvvkvvkF −+−+−= ∫ ˆ)ˆ()ˆ( τ
which is subject to actuator saturation as shown in the figure.
A nonlinear bicycle model is defined as:
Vehicle
Dynamics
Model
fδ
xv̂
xv
xv∆ xF
Vehicle
Kinematics
Model
Φ
y
x
v
v
Y
X
Speed
Controller
Saturation
r
v
v
y
x
xF
yrfyffxz
yrfyffxxy
bFFFarI
FFFrvvm
−+=
++=+
)cossin(
cossin)(
δδ
δδ
ɺ
ɺ
where the front and rear tire lateral slip angles are defined respectively as below:
x
y
r
x
y
ffV
brV
V
arV −−=
+−= −− 11
tan,tan αδα
Thus, the tire lateral force at front or rear tire is a non-linear function of normal
force of front or rear axle, and slip angle of front or rear tire, as defined below
respectively:
ffzfffyf FfCF ααα ⋅== ),(
rrzrrryr FfCF ααα ⋅== ),(
The traffic vehicle model thus can be represented by a set of nonlinear
differential equations: ),,( tuxfx =ɺ , with state variables to be [ ]Ty rvx = , and
input variables to be [ ]TfxFu δ= .
With the simulated range sensing and object detection capability, the proposed
sensor model and algorithms can be applied to the simulation studies of many
xv̂
xv
xv∆Desired
speed
Distance
Controller
Desired
range
Range rate
Range
Desired
range rate
Radar
sensor
model
Vehicle
Longitudinal
Dynamics
Target
speed
xv
Speed
Controller
Figure 19: Simulation on an adaptive cruise control
Modeling and Simulation of Sensor-Guided Autonomous Driving
active safety, driver assistance, and autonomous driving features and functions
such as adaptive cruise control, collision warning and prevention, etc. As an application example, a closed-loop simulation of an adaptive cruise control
(ACC) system, shown in Figure 19, is presented in this section to demonstrate
that the proposed sensor model and algorithm enables the simulation of multiple
vehicles interacting with one another under traffic. The radar sensor model
simulates the in-lane target’s relative range and target’s speed. The distance
controller monitors the desired range from the driver’s input and the in-lane
target’s relative range rate, and outputs the desired range rate. The speed
controller monitors the error from the desired vehicle speed and the error from
the desired range rate, and generates either throttle or brake force command such
that the desired following distance between the vehicle and the target is
maintained.
Figure 20 shows an ACC simulation scheme created under iSim® and
Matlab/Simulink environment, with a scenario of a host vehicle (white color)
interacting with two target vehicles (yellow color) on a road. The animation is
shown in Figure 21. A range sensor, simulated with the proposed sensor model,
is mounted in front of the host vehicle with a forward-sensing range of 50 meters
to detect if any of the target vehicles is falling into its scanning zone. Based on
the sensor outputs such as range, range rate, detection flag, etc, the ACC module
generates either throttle or braking force command to adjust the host vehicle’s
speed such that the host vehicle maintains a desired following distance from the
target vehicle.
The simulation results are illustrated through the animation shown in
Figure 22. In the simulation, the host vehicle initially travels at a set speed
of 60kph, as shown in Figure 22(a). Within a close distance at the second
left lane, a target vehicle (marked as target vehicle 1) travels at a constant
speed of 50kph. Around time t=2 second, the target vehicle 1 changes its
lane and cuts in front of the host vehicle inside the sensing range, as
shown in Figure 22(b). The host vehicle quickly slowed down in order to
keep a pre-defined following distance around 15 meters. Shortly the target
vehicle 1 changes back to its previous lane and gets out of the sensing
zone, as shown in Figure 22(c), the host vehicle then accelerates back to
its pre-set speed of 60kph until another slower-traveling vehicle ahead
(marked as target vehicle 2) falls into its sensing range, then the host
vehicle slowed down again with adjusted speed to follow the target
vehicle 2 and maintain the desired following distance of 15 meters, which
lasts till the simulation ends. The simulation results further verify the
effectiveness of the range sensing model and object detection algorithms
proposed in this research.
Figure 20: Simulation on adaptive cruise control
(a) Host vehicle travels at set speed of 60 kph
Modeling and Simulation of Sensor-Guided Autonomous Driving
(b) Target vehicle 1 cuts in front of the host vehicle
(c) Target vehicle 1 gets out of sensing range
(d) Host vehicle accelerates to the set speed of 60 kph
(e) Target vehicle 2 falls into the sensing range
Figure 21: Animation for ACC Simulation
(a) ACC set speed and speed profiles of host vehicle and two target vehicles
(b) Desired range, actual range and detection flag
Modeling and Simulation of Sensor-Guided Autonomous Driving
(c) Range rate and detection flag
Figure 22: Simulation Results of ACC
5. Conclusions
This paper presents a generic approach in modeling range sensing and object
detection, both functionally and physically. This is to support the emerging
simulation of multiple vehicles interacting with one another under traffic and
with surrounding environment. The mathematical derivation behind the described
method is discussed in detail, along with algorithm for numerical implementation.
The proposed model and algorithm have been implemented and verified under Matlab with simulation under various scenarios. The results demonstrate that the
proposed models and object detection algorithm are valid, effective and
numerically efficient. A closed-loop simulation with an adaptive cruise control is
further presented to demonstrate an integrated modeling and simulation of
limited autonomous driving with the proposed sensing model and object
detection algorithm under iSim®, an integrated vehicle simulation environment.
REFERENCES
Bacso, J. et al, Project AMBER: Re-Engineering a Legacy Ground Radar
Modeling System into a Standard based Object Oriented Architecture, IEEE,
1988
Balanis, Constantine A., “Antenna Theory, Analysis and Design”, John Wiley &
Sons, Inc., 2nd ed. 1982
Bertolini, A. Long, J., “Radar target signal modeling and validation,” Simulation,
vol. 50, no. 4, pp. 137-142,1988
Cheng, David K: “Field and Wave Electromagnetics”, Addison-Wesley
Publishing Company Inc., Edition 2, 1998
Deng, W., Dai, J., and Zhao, Q., Litkouhi, B., Moshchuk, N. and Nisonger, R.,
“Modeling of Range Sensing and Object Detection for Vehicle Active Safety”, accepted by 12th International IEEE Conference on Intelligent Transportation
Systems, St. Louis, Missouri, U.S.A. October, 2009
Donohue, D. and Kuttler, J., “Modeling radar propagation over terrain,” Johns
Hopkins Applied Technical Digest, vol. 18, no. 2, 1997
Eberly, D., Intersection of a Triangle and a Cone,
http://www.geometrictools.com
Eberly, D., Intersection of Convex Objects: The Method of Separating Axes,
http://www.geometrictools.com
Gillespie, T.D., Fundamentals of Vehicle Dynamics, SAE International March
1992
Haynes, T., “A Primer on Digital Beamforming,” Spectrum Signal Processing,
March 1998
Laet, T., Schutter, J. and Bruyninckx, H., “Rigorously Bayesian range finder
sensor model for dynamic environment,” Proceedings of 2008 IEEE
International Conference on Robotics and Automation, Pasadena, CA, USA, May
19-23, 2008, pp. 2994 -3001
Schoor, M. and Yang, B., “High-Resolution Angle Estimation for an Automotive
FMCW Radar Sensor,” In Proceedings of the International Radar Symposium (IRS), Cologne, Germany, September 2007
Winkler, V., “Range Doppler Detection for Automotive FMCW Radars,” In
Proceedings of the 37th European Microwave Conference, pp.1445-1148,
Munich Germany, October 2007