This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 1

An IMM/EKF Approach for Enhanced

Multi-Target State Estimation for Application to
Integrated Risk Management System
Beomjun Kim, Kyongsu Yi, Hyun-Jae Yoo, Hyok-Jin Chong and Bongchul Ko

 system (AEBS), etc., have been developed and commercialized

Abstract—For the generic assessment and the total for the driver’s safety and handling enhancement. And
management of collision risks with multi-traffics in complex state-of-the-art ADAS researches focus on the integration of
driving situations, it is essential to estimate and represent the existing safety functions [1]. In the same context, integrated
target vehicles’ overall behaviors such as heading, yaw rate,
risk management system (IRMS) has been proposed which is
absolute velocity and acceleration as well as relative position and
relative velocity which are the state of the target vehicle with the approach to integrate vehicle safety functions and deal with
respect to the host vehicle. To achieve this, this paper describes an complex driving scenes in a generic way using an integrated
interacting multiple model (IMM) approach using extended perception layer, a unified decision layer and an optimally
Kalman filters (EKF) to improve multi-target states estimation coordinated action layer [1], [2].
performance with utilization of automotive radars. Automotive The effectiveness of safety function of ADAS depends on the
radar is the best fitted vehicular surround sensing technology with
interpretation of the information arriving from sensors, which
respect to functionality, robustness, reliability, dependence on
weather conditions etc. In an application of the automotive radar, provide details of the multi-surrounding vehicles, the road
the most important issue is to handle an uncertain measurement environment, and the driver-assisted vehicle itself [3].
model problem that is wandering on the target’s physical Especially, for the generic assessment and the total
boundary. To cope with this problem, multi-models are management of collision risks with multi-target vehicles in
formulated and a new multi-target tracking algorithm is complex driving situations, it is essential to estimate and
developed based on IMM approach, global nearest neighbor
represent the target vehicle’s overall behaviors such as heading,
(GNN) based data association and EKF method with elaborated
modeling of automotive radar. The performance of the proposed yaw rate, absolute velocity and acceleration as well as relative
multi-target tracking algorithm is verified via vehicle tests in real position and relative velocities which are the state of the target
roads. And performance comparison with model-switching vehicle with respect to the host vehicle.
algorithm which is a simple approach to handle the multi-model We consider the multi-target state estimation problem with
problem has been conducted. It is shown that the target vehicle’s automotive radar which is the best fitted vehicular surround
overall behavior can be estimated by proposed elaborated models
sensing technology with respect to functionality, robustness,
and the estimation performance can be significantly enhanced by
the proposed IMM based algorithm reliability, dependence on weather conditions, etc. Techniques
for tracking maneuvering multi-targets are already well
Index Terms—Interacting Multiple Model, Multi-target state established topics. The switching multiple model approach has
estimation been found to provide quite better performance than using a
single model. A number of multiple model algorithms to track a
maneuvering target have been proposed; the multiple-model
I. INTRODUCTION algorithms, the IMM algorithm, the adaptive IMM, the fuzzy
IMM, state dependent IMM, and others [3]-[10]. In these
R ECENTLY, numerous advanced driver assistant systems
(ADAS), such as adaptive cruise control (ACC), lane
keeping assistance system (LKAS), advanced lane change
previous works, considerable attention is limited to multiple
process models.
assistance (ALCA) system, and advanced emergency braking However, in an application of automotive radars, one of the
most significant problems is an uncertain measurement model
This work was jointly supported by the Hyundai Motor Company, problem which is wandering on the target’s physical boundary
SNU-IAMD, the BK21 program, a Korea Research Foundation grant funded by as can be seen in Fig.4-5. Therefore, multiple model approach
the Korean government (MEST) (no. KRF-2009-200-D00003), and a National has to be extended to the problem of measurement modeling.
Research Foundation of Korea grant funded by the Korean government (no.
2011-0001277). To cope with these problems, the uncertain wandering
B. Kim and K. Yi. Authors are with Department of Mechanical and measurement point problem is formulated as multiple
Aerospace Engineering, Seoul National University, Seoul, Korea. measurement models and a new multi-target state estimation
(corresponding author to provide phone: 02-880-1941; fax: 02-888-7194;
e-mail: [email protected]).
algorithm is developed based on the IMM approach, the GNN
H. Yoo, H. Chong and B. Ko Authors are with the Intelligent Vehicle Safety based data association technique and EKF method with
System Development Team, Hyundai Motor Company, Seoul, Korea. (e-mail: enhanced modeling of automotive radar.
[email protected]; [email protected]).

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 2

The contributions of this paper are as follows. A standard follows: 1) to find the set of validated measurement for each
process model which can describe multiple modes of behavior target and 2) to find the minimum variance estimated state and
of target vehicle has been derived and discretized. Second, the associated error covariance matrix for each target.
through an analysis of the automotive radar characteristics, a
standard measurement model is derived and twelve specific
B. Coordinate Systems
measurement models are selected so that can cover all possible
measurement patterns. Finally, an IMM algorithm using In this subsection, coordinate systems used to derive
extended Kalman filters (IMM/EKF) for multi-target state standard process and measurement models are presented. These
estimation is provided with proper parameter design and coordinate systems are described in Fig.1. Frame {0} is the
validated via vehicle tests in real road. ground based fixed coordinate system of which the position is
same with the host vehicle, frame {1} is the host vehicle’s
II. PROBLOEM FORMULATION body-fixed moving coordinate system, frame {2} is the
sensor’s body-fixed moving coordinate system, and frame {3}
In this section, a stochastic hybrid system is formulated in
is the target vehicle’s body-fixed moving coordinate system.
the form of an IMM/EKF algorithm for tracking multiple target
And superscript means a corresponding time step. The standard
vehicles. A standard process model to describe behaviors of
process model describes the predicted state of the target
target vehicle is derived and discretized. And a standard
vehicle’s mass center on the frame of {0}k+1 based on the
measurement model is defined to describe automotive radar
current state of the target vehicle’s mass center and system
characteristics.
input which are defined on the frame of {0}k. And the standard
A. Stochastic hybrid system measurement model describes the state of measured point such
Assume that there are total Ntarget targets and the system of as position and longitudinal velocity on the sensor based
moving frame {2}.
each target can be modeled as one of the N mode hypothesized
models where the target set is denoted as T  1, 2, , Nta rget  and
the model set is denoted as M  1, 2, , Nmode  . From the work of
Li and Bar-shalom [8], a general description for a stochastic
hybrid system of target n Τ during the sampling period
 tk 1 , tk  can be given as:
xnj  k   fnj  xnj  k  1 , u k  1  wnj k  1 (1)
(a) at time k
z  k   h  x  k  , u  k   v  k 
j
n
j
n
j
n
j
n (2)
where xnj  is the state vector of target n when model
j M  is in effect, u  is the input vector(same for all targets),
fnj   and hnj  are the nonlinear and time invariant system
structure. The process noise w nj  k  and the measurement noise
v nj  k  are mutually uncorrelated zero-mean white Gaussian with
covariance Wnj  k  and Vnj  k  respectively. And a (possibly
(b) at time k+1
state dependent) Markovian transition probability of the system Fig.1. Four coordinate systems used to derive standard process model and
mode index is given as follows. standard measurement models
i, j  Μ
(3) C. Standard Process Model
Pr mn  k  1  j mn  k   i  Φi , j  xn  k 
Various driving patterns of a vehicle including straight line,
where mn  k  is the system mode index of target n which is curve, cut-in/out, U-turn and interchange can be represented by
in effect at time step k . a combination of a constant velocity rectilinear motion, a
The following notations and definitions are used regarding constant acceleration rectilinear motion, a constant angular
the measurement. In general, at any time k , some velocity curvilinear motion and a constant angular acceleration
measurements may be due to clutter and some due to the target, curvilinear motion [1].
i.e. there can be more than a single measurement [6]. The To describe all these motions, the state vector and input
measurement set (not yet validated or associated) is denoted as vector of target n are defined as follows:
Yk  y1  k  , y 2  k  , , y N  k  where Nmeasure is the number of xn   pn, x pn, y n  n an, x  n 
T
measure
vn, x (4)
measurements. The set of validated measurement of target n at u   vx  
T
(5)
time k is denoted by
where subscript x and y denote x-axis and y-axis of each
Zk  z1  k  z 2  k  , , z N  k  (4)
target
frame, subscript n denotes “of target n “, p denotes the
Therefore, the goals of the algorithm can be summarized as relative position  denotes the relative yaw angle, v denotes

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 3

the velocity,  denotes the yaw rate, a denotes the f3 

n
T 2   n T  n f4  
an ka
T 2  an T  vn
acceleration, and  denotes the yaw acceleration. And variable 2 2
k  n
without subscript n means “variable of the host vehicle”. f5  
2
T 2   n T   n f6 
an
2
 ka 2T 2  2ka T  2
Therefore vx denotes the longitudinal velocity of host vehicle
n
and  denotes the yaw rate of host vehicle. Note that pn, x , pn, y f7 
2
 k 2T 2  2k T  2
and  n are defined on the host vehicle’s body-fixed moving w  k  ~  0, W  k 

frame, {1}, and other elements are defined on the ground-based W  k   E  w  wT 

fixed frame, {0}. With no slip assumption ( vn,y  0 ), the  T 2    T
 E   T  I   a  xn , u  q    
continuous process model can be formed as:    2 x n   
xn  a  xn , u   q
 T 2  
 a1 a 2 a7   q
T  T  I   a  x n , u   E q  qT   T
a3 a 4 a5 a6
 2 x n 
a1  vn , x cosi  vx  pn , y   a 2  vn , x sin i  pn , x   (6)
 T 2  
a3   n   a 4  an , x a5   n a 6  ka a7  k  T  I   a  x n , u   Q  T
 2 x n  xn  xn  k 

q ~  0, Q 
u  u k 

where a is the nonlinear process vector equation, k a is the u

d
dt
 T

 vx     a x  
T

decay-rate of the longitudinal acceleration, k is the decay-rate where u is time-derivative of input vector where the
of the yaw-acceleration, and q is the continuous white process components’ physical meanings are host vehicle’s longitudinal
noise with covariance Q . Above continuous process model can acceleration and yaw-acceleration. As the result, xn  k  1
be discretized via Taylor methods as follows [11]. describes the predicted state of the target vehicle’s mass center
x n  k  1  x n  t  T  on the frame of {0}k 1 . Note that various driving patterns of a
d 1 d2
 xn  t   x n  T   2 x n  T 2  h.o.t. vehicle can be represented by the above standard discretized
dt 2 dt
process model by adjusting a and Q of continuous process
 x n  t   a  x n , u   q  T   a  x n , u   q  T 2
1 d
2 dt model. The details about specific modes of process model will
 x n  t   a  x n , u   q  T   be discussed in following subsection, III-A.
 
 1     2 D. Standard Measurement Model

 2  x a  x , u   x  a  x , u   u  q    T 
u
n n n
  n   In an automotive target tracking, changes in the target aspect
 x n  t   a  x n , u   q  T   with respect to the radar can cause the apparent point of radar
 
     reflections (relative position seen by the antenna) to wander
  1  x a  x n , u    n
a  x , u   q 
  (7) significantly [12]. To represent these characteristics, the
  n
  T 
2

2     measurement model can be elaborated by introducing new

  a  xn , u   u  q  
  u  parameters to specify the sensor position and the measured
  x n  t   a  x n , u   T    point. These are depicted in Fig.2. As the result, the
  
 1     
2 
measurement vector of target n is defined as follows.
   a  x , u   a  x , u   a  x , u   u    T 
    z n  k   h  x n  k  , u  k   v n  k 
n n n
   n
2 x u   
    
  T  I  T   a  x n , u   q   h n1 h n 2 hn3   v n  k 
2 T
 (8)
   2 x n   
v n  k  ~  0, Vn  k 
 f  x n  k  , u  k   w  k 
where
 f1 f 2 f7   w  k 
T
f3 f 4 f5 f6
h n1  pn , x  sx  bn , x cos n  bn , y sin  n
where h n 2  pn , y  s y  bn , x sin  n  bn , y cos n
  host  pn , x  host  vn sin n   
  h n 3  vn , x cos n  vx  pn , y     bn , x sin  n  bn , y cos n     n 
 2 
 a cos   v sin    pn , y  host  vhost  where h is nonlinear measurement equation which describe
f1       T 2   T  pn , x
 v cos  
n n n n n

 2 2   n n 
the state of measured point on the sensor based moving frame
 a p   {2}. s is sensor position vector defined on the host vehicle’s
  host  n , y host 
 2 2  body-fixed moving frame {1} and b n is measured point vector
 
 an sin n   n vn cos n  pn , x  host  of target n defined on the target vehicle’s body-fixed moving
   
2 2 2  2  pn , x  host 
f 2   T   T  pn , y frame {3}. The first order approximation of the measurement
  host  pn , y  host  vhost  vn cos n     v sin  
   n n  error covariance has been presented in previous works as
 2  follows [13].

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 4

  r 2  rn 2  2 bn  cos 2 n sin 2 n   A. Set of Multiple Process Model

  sin 2 0 
Vn   2  n b  cos 2 n  21  (9) Various driving patterns considered in this section are
  v 2  depicted in Fig.3. To describe these various driving patterns of
 012
a vehicle, four specific modes of process model will be
where
discussed in this section. As aforementioned, four motions
 2  r 2 2  hn 2 
bn  r 2 n 2  2 rn  h n12  h n 2 2 n  tan 1   include the constant velocity rectilinear motion, the constant
 r  rn    hn1  acceleration rectilinear motion, the constant angular velocity
where rn and  n are the range and azimuth measurements, curvilinear motion and the constant angular acceleration
respectively.  r ,   and  v are the standard deviations of the curvilinear motion are considered. Assuming that accelerations
in the steady state are quite small, linear accelerations or
range, the azimuth and the relative velocity, respectively. The
decelerations can be reasonably well covered by process noises
approximation is validated to be useful for practical parameters
with the constant velocity model. With these assumptions, the
[13]. The various measurement patterns due to the target aspect
Constant Velocity Rectilinear (CVR) motion can be modeled
with respect to radar can be represented by the above standard
by:
measurement model by adjusting the measured point vector b n .
x n  a  xn , u   q
The details about specific modes of measurement model will be
 a1 a 2 0 0 0 0  q
T
discussed more fully in subsection, III-B a3,CVR (10)
q ~  0, QCVR 
where
a3,CVR  
QCVR  diag 0 0 0 Q44 0 0 0 
where Q is the appropriate covariance which can represent
the zero-mean Gaussian white noise assumed accelerations. In
Fig.2. Parameters to derive the standard form of the refined measurement model the same way, the Constant Acceleration Rectilinear (CAR)
motion can be formulated as follows:
xn  a1 a 2 a3,CVR 0 0 0  q
T
a 4,CAR
III. SELECTION OF MULTIPLE MODEL SET AND (11)
PARAMETER DESIGN q ~  0, QCAR 

To obtain the best estimate, the IMM algorithm has to be where

properly designed and tuned to meet the special requirements of a4,CAR  an, x
the automotive application. And these requirements are QCAR  diag 0 0 0 0 0 Q66 0 
fulfilled by means of 1) selection of vehicle motion models for The constant angular velocity curvilinear motion is a turning
all modes of driving, 2) selection of measurement models for all motion with a constant yaw rate along a road of constant radius
cases of measurements, and 3) determination of the parameters of curvature. With a zero-mean Gaussian shite noise to the
of the underlying Markov chain, that is, the transition derivative of yaw rate, the Constant Velocity Curvilinear (CVC)
probabilities [8]. Therefore, in this section, after analyzing the motion can be modeled by:
driving patterns and automotive radar characteristics, four
x n  a1 a 2 0 0 0 0  q
T
a3, CVC
specific modes of process model, twelve specific modes of (12)
measurement model, and event dependent transition probability q ~  0, QCVC 
matrix are derived and introduced. where
a3,CVC   n  
QCVC  diag 0 0 0 0 Q55 0 0 
With some modifications to (13), the Constant Acceleration
Curvilinear (CAC) motion can be derived by:
x n  a1 a 2 a3,CVC 0 0  q
T
0 a5,CAC
(13)
Fig.3. Various driving patterns of multi target vehicles. q ~  0, QCAC 
where
a5,CAC   n
QCAC  diag 0 0 0 0 0 0 Q77 
Evidently, the discretized process model derived from
(10)-(13) are special forms of (7) and can be summarized by (1).
In addition, it is reasonable to assume that the transition
Fig.4. Various measurement patterns due to the target aspect with respect to the
host vehicle’s radar sensor. between the driving modes of a target vehicle has the

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 5

Markovian probability governed by (3). Consequently, the A transition probability matrix of the proposed algorithm can
process model can be suitable described in the framework of be decomposed into two sub-matrices; one is the sub-transition
the stochastic hybrid systems. probability matrix related to multiple process models and the
other matrix is about multiple measurement models. Relation of
B. Set of Multiple Measurement Model
three matrices is given as:
As shown in Fig.4, by means of relative position and  Φ1process  Φmeasure Φ1process  Φ measure Φ1process  Φ measure 
1 2 1
orientation between host and target vehicle, the configuration  process measure 
 Φ21  Φ Φ Φ Φ2process process  Φ
process measure measure
22  N mode 
of measurement can be altered. Ideally, the detecting point Φ  (14)
should be located on the four sides or four edges of the target  
Φ process  Φ measure Φ process
Φ measure
Φ N process N process  Φ
process measure 
vehicle. As an example, for the case of the cutting in vehicle, a  Nmode
process
1 process
N mode 2 mode mode 
rear right edge is the most likely measurement model while lane where Φ process is the matrix defining a transition probability
changing. However, after cutting-in, the most likely measured between process mode indices, N is the number of process process
mode

point is transited to a center of rear side of the target vehicle. model set, Φmeasure defines a transition probability between
As aforementioned, to describe these various measurement measurement mode indices. Φ process can be expressed in a single
patterns, several specific modes can be derived by adjusting b n matrix, but this is impossible in the case of Φmeasure as
in standard measurement model. From the viewpoint of aforementioned. To solve this problem, different measurement
detecting position defined on the target-body-fixed frame, a mode transition matrices are used depending on the target
specific mode set depicted in Fig.5 can be utilized to describe aspect event. All eight possible events are defined in Fig.6.
the variations of the measurement patterns. We can see that the From the work of Zhang and Bar-shalom [14], an approach
proposed mode set include twelve points around the boundary using state-dependent mode transition probabilities has been
of the target vehicle and all case examples in Fig.4 can be presented to track move-stop-move targets. And in our research,
represented well by the proposed mode set. Consequently, this approach is extended to eight stages and applied with some
derived set of measurement models can be suitable described modifications. Target index is omitted unless otherwise
by (2) of the stochastic hybrid systems. required for clarity. First, measurement mode transition
probability matrix with no event (physically not allowed) is
designed in Gaussian distribution manner as follows:
i, j  Μ
1  1  d 2  (15)
Φimeasure ,0
 exp    i , j  
2 d  2   d  
,j

where di , j is the distance travelled from i to j along the

vehicle’s body line. And with this raw transition probability, a
transition probability conditioned on a specific event E is
given by:
Φmeasure,E  Φmeasure,0  ΘE  Φimeasure
,j
,0
  ΘE (16)
where Θ E is an event dependent diagonal matrix of which
elements are defines as follows:
1 i  j  M valid
measure , E


Θ E
i, j  0 i  j  M invalid
measure , E
(17)
Fig.5. Twelve modes of detected point of measurement model to describe 0
 i j
various measurement patterns.
where Mvalid
measure , E
is the subset of valid measurement modes
C. Event Dependent Transition Probability Matrix
and Minvalid
measure , E
is the subset of invalid measurement modes on an
In a conventional IMM estimator the Markov chain of the
mode states is described by a single mode transition matrix. event E . Consequently, the conditional transition probability
However, unlike the conventional system, mode transition matrix has zero values as entries of columns of invalid modes,
process of an automotive radar system cannot be modeled by a which means that under that given event, transitions to invalid
single mode transition probability matrix. As can be seen in modes are not allowed.
Fig.6, transitions to invalid modes are infeasible since the radar Finally, the actual transition probability matrix of the system
cannot penetrate the vehicle’s body. And a subset of invalid at time step k is calculated as a combined form which can be
modes can be different for different situations. As an example, written as follows:
i  E
mode transition from the front-right edge to rear-left edge is
(18)
impossible in event.1 while it is possible in event.8. Therefore,
i

Φmeasure   Pr E  k   i z 0 , 
, z  k   Φmeasure,,i
we can conclude that the mode transition probabilities of the
automotive radar system are target aspect dependent and multi Where E  i i  1, ,8 is a set of eight event indexes and
transition matrices should be derived to describe all cases of the 
Pr E  k   i z 0 , , z k   is an event probability which means the
mode jump process.
probability of the target being in a specific event i at time step

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 6

Fig.6. Set of measurement model, their two subsets classified by validity (of each target vehicle in multi traffics), and eight events of target aspect.

k conditioned on the cumulative set of measurements up to x host  A host  x host  q host

time k . The event probability can be calculated directly as: 0 1 0 
 0 0 0   vx    q host
T
i  E ax (20)
(19)

Pr E  k   i z 0 , 
, z  k     f s z0,
A.i ,z k   s k  ds 0 0 0 
q host ~  0, Q host 
where s denotes a sensor position vector transformed in
target-body-fixed coordinate, fs z0, ,zk   is a conditional z host  k   H host x host  k   v host
1 0 0 
probability density function associated with s conditioned on
 0 1 0  vx  k  ax  k    k   v host
T
(21)
the cumulative set of measurements, and A.i is an associated
0 0 1 
area with event i which is depicted in Fig.7.
v host ~  0, Vhost 
The given process model can be discretized as [11].
x host  k   x host  k  1  t  q host
 Fhost  xhost  k  1  w host
(22)
w host ~  0, t 2  Q host 
~  0, Whost 
Fig.7. Associated area of possible target aspect event on target-centered where t is the sampling time (0.1s in this study). With
coordinate
above process and measurement model, vehicle states are
recursively estimated by using the Kalman filter which is a
sequence of time and measurement update steps as following
IV. IMM/EKF MULTI TARGET STATE ESTIMATION
specific equations.
In this section, an IMM/EKF based filtering algorithm for
tracking multiple target vehicles is proposed. The overall 1) Time update
structure of the algorithm is shown in Fig.8. The host vehicle xhost  k   Fhost  xˆ host  k  1
filter estimates the host vehicle’s longitudinal speed and yaw (23)
Phost  k   Fhost  Pˆ host  k  1  Fhost T  Whost
rate which take roll of inputs as can be seen in (5). And based
on IMM approach, the organized measurement set is filtered 2) Measurement update
K host  k   Phost  k   H host T   H host  Phost  k   H host T  Vhost 
1
and the best estimates of multi target vehicles are obtained.
Each component of the algorithm is presented in following Pˆ host  k    I  K host  k  H host   Phost  k 
sections. In the part of track management, the raw measurement (24)
xˆ host  k   xhost  k   K host  k    z host  k   H host  xhost  k 
cloud is validated and associated for each target by using most
uˆ  k    xˆ host  k 1  xˆ  k 
T
likely validation law. 
host 2 

A. Host Vehicle Filter B. IMM/EKF based Filtering

The Kalman filter is used to estimate host vehicle states such For the each track, after the validated measurement is
as longitudinal velocity and yaw rate. These values can be received, the multiple EKFs are used as the each mode-matched
estimated from the vehicle sensor signals such as wheel speed, local filter and combined through the IMM method to obtain a
acceleration, and yaw rate of which each time derivative is refined state estimate. A complete cycle of the IMM/EKF based
considered as the Gaussian white noise. The state of the host filtering algorithm can be summarized as follows. Time index
vehicle filter and its time invariant linear system structures can is omitted unless otherwise required for clarity. Note that
be defined as follows. 1) Interaction ( i, j  M / n  T ):
Starting with the N mode weights ni  k  1 , the N mode means
xˆ in  k  1 and the N mode associated covariances Pˆ ni  k  1 , compute

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 7

Measurement update:

Fig.8. Architecture of IMM/EKF multi target state estimation algorithm

the mixed initial condition for filter matched to mode state j . xˆ nj  k   xnj  k   K nj  k rnj  k  (36)
Note that, when the modes used have different dimension state ˆ j  k    I  K j  k   H j  k  P j  k 
P (37)
n n n n
vectors, an augmentation with zeros the lower dimension state
3) Mode Probability Update:
estimate may lead to a bias toward zero for the state
The N mode weights  ni are updated from the innovations of
components of the larger state vector. In the previous work, a
simple procedure to avoid this “biasing” is presented, together the N mode extended Kalman filters.
with a suitable augmentation of the covariance of the smaller Likelihood function:
 1 
 exp    rnj  k    Snj    rnj  k 
state that yields an unbiased and consistent mixing [15].  nj  2 Snj
1/2 T 1
(38)
Predicted mode probability:  2 
nj Pr mn  k   j z n  k  1 Mode probability:
(25) nj   nj
  Φij  k  1 ni  k  1 nj  (39)
i  ni  in
i
Mixing probability:
4) Combination:
ni j Pr mn  k  1  i mn  k   j , z n  k  1
for output purpose only, xˆ n  k  and Pˆ n  k  are computed
Φij  k  1 ni  k  1 (26)
 according to
 j
n
Combined condition:
Mixed condition: xˆ n  k    xˆ nj  k  nj (40)
xˆ 0n j  k  1   xˆ in  k  1 ni j (27) j

 
Pˆ n  k    Pˆ nj  k    xˆ nj  k   xˆ n  k   xˆ njj  k   xˆ n  k  nj
i T
(41)
 xˆ n  k  1  xˆ n  k  1 
  i 0j 
 i j j
Pˆ n0 j  k  1   Pˆ ni  k  1   T
 (28)
  xˆ n  k  1  xˆ n  k  1 
  i 0j  n
i
  C. Track Management
2) Extended Kalman Filtering: In track management, there are three processes required
Each of the N mode pairs weights xˆ 0n j  k  1 , Pˆ n0 j k  1 is used as during each time step: track update, track deletion and track
input to an extended Kalman filter matched to mode state j . creation. In this study, each track has a rewarding counter that is
Time update: rewarded when the track is updated with validated
fnj measurement and decreased when the measurement is missed.
Fnj  k  1  (29) In the part of track update, the measurements are associated
x xˆ 0n j  k 1,uˆ  k 1
with the existing tracks. Following (28), the most likely
xnj  k   fnj  xˆ 0n j  k  1 , uˆ k  1 (30)
candidate measurement for target n at time k can be defined as
P  k   F  k  1  P
ˆ   k  1 F k  1  W k  1 T
n
j
n
j 0j
n n
j
n
j
(31) follows:
 k   min  y  zˆ n  k    Sn  k    y  zˆ n  k 
T 1
Filter gain: z candi
n (42)
yY
h nj
k

H nj  k   (32) where
x xnj  k ,uˆ  k 
zˆ nj  k   hnj  xnj  k  , uˆ k 
Snj  k   Hnj  k   Pnj k   Hnj k   Vnj k 
T
(33)
zˆ n  k    zˆ nj  k  nj
K nj  k   Pnj  k   Hnj k   Snj k 
T 1
(34) j

Innovation: 
Sn  k    Snj  k    zn  k   znj k   zn k   znj k 
T
 n
j

r  k   z n  k   hnj  xnj  k  , uˆ  k 
j
j
(35)
 k  is validated as z n  k  if
n
The candidate measurement z candi
n

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 8

and only if situation are shown in Fig.12. We can see that, from 0-12s, the
z  k   zˆ n  k   Sn k    z k   zˆ n k   rth
1
candi T
(43)
candi target vehicle is detected by rear radar and left side of the target
n n
is in effect during that period. A period from 12-14s is blind
where rth is the gate threshold corresponding to the gate
area where there is no measurement for the target vehicle by
probability PG . Note that all targets share a common threshold. any radar. From 14-20s, the target vehicle is detected by front
If the track has no validated measurement, the rewarding radar and rear side of the target is in effect during that period.
counter is decreased and the time-prediction is performed as the
current estimate (which means a zero-gain EKF). If a rewarding
counter drops below a certain threshold, the track is deleted.
And newly arriving measurements generate new tracks. The
initial state of new track is decided by a model selection
strategy depicted in Fig.15. And the new track is assumed to
have the same speed with the host vehicle, zero yaw-rate, zero
acceleration, and zero yaw acceleration. Additionally, some
large value of error covariance is given for the newly generated
target.

V. VEHICLE TESTS RESULTS

The developed IMM/EKF based multi-target tracking
Fig.9. Configurations of the host vehicle and the target vehicle
algorithm is implemented and verified via vehicle tests in real
road. The test scenario is an overtaken situation by a target
vehicle to validate all-around target states estimation B. Implementation and Evaluation
performance. To investigate the performance enhancement, As aforementioned, in target tracking research area,
comparison with model-switching algorithm which is a simple considerable IMM approaches are limited to various modes of
approach to handle the multi-model problem has been target motion. And the main novelties of this paper are that
conducted. Additionally, the experimental results dealing with process model and measurement model have been elaborated,
multi-target scenarios on the real road are presented to explore and multiple measurement models have been derived as shown
the multi-target tracking performance of the proposed in Fig.5. Therefore, twelve measurement model set has been
algorithm. implemented and investigated for the proposed IMM/EKF
As aforementioned, in target tracking research area, algorithm (with single standard process model). Target states
considerable IMM approaches are limited to various modes of estimation results in the given situation are summarized in
target motion. And the main novelties of this paper are that Fig.10-13.
process model and measurement model have been elaborated, Fig.10 gives three case examples of mode probability update
and multiple measurement models have been derived as shown and target state estimation. The host vehicle is depicted by
in Fig.5. Therefore, twelve measurement model set has been green vehicle, the actual target vehicle by red vehicle, the
implemented and investigated for the proposed IMM/EKF estimated target vehicle by gray vehicle and the associated
algorithm (with single standard process model). measurement by red square. As can be seen in the figure, in all
cases, mode estimation results are quite reasonable and
A. Configuration of Vehicle Tests estimated target positions maintain good performance
compared to the actual values even in no measurement situation
Fig.9 shows a host vehicle and a target vehicle which are
(case.2).
used in this study. The host vehicle is equipped with a
Fig.13 shows the results of mode probability adaptation
front-radar, two rear-side-radars, wheel speed sensors, and a
along the time axis. The update results represent the actual
6DOF sensor, etc. And a RT-Range is used for an accurate
mode transition which can be inferred from Fig.11. From 0-10
measurement of the relative motion between the host and target
sec, which is a period that the target vehicle is detected by rear
vehicle. The RT-Range is a precision tool for evaluating and
radar and front-left edge of the target is mainly in effect, we can
verifying the accuracy of radar and tracking algorithm [16].
see that the probability of mode.6 ranked highest. In a period
Since this paper focuses on the target tracking algorithm,
from 10-12 sec which is blind area, although there are no
additional details about the test vehicles’ system configurations
measurements, the most likely mode has been changed from
have been omitted.
mode.6 to mode.2 along the target vehicle’s left side by the
The test scenario is an overtaken situation by a target vehicle
interaction/mixing step of the algorithm. This transition in the
as shown in Fig.11. The host vehicle (denoted by green vehicle)
blind area is quite reasonable in the view of changes of the
drives in a straight lane with an initial speed of 50 km/h and a
target aspect. And finally, from 12-20 sec, which is the period
slight deceleration and the target vehicle (denoted by red
that the target vehicle is detected by front radar, mode.1 (center
vehicle) started its overtaking at 7s and finished at 16s while
of rear side), mode.12 (rear-right edge), and mode.2 (rear-left
driving with an initial speed of 55kph with a slight acceleration.
edge) are evaluated as the highly likely modes during the period.
Typical characteristics of the automotive radar in the given

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 9

Moreover, as the target vehicle becomes the in-lane preceding

vehicle, we can see that mode.1 gets stronger mode probability
which denoted by red-solid line.
Fig.14 presents the estimation results for the target states and
their error distribution. From the figure, we can see that the
overall trends of each estimated state are similar to actual
values and their error distributions can be well approximated by
normal distribution. The statistics of estimation error of
IMM/EKF based filtering algorithm are given in Table 1.

Fig.10. Three case examples of mode probability update

Fig.11. Test scenario: overtaking target vehicle

Fig.14. Estimated target states and the distributions of the estimation errors

C. Comparison with Model-switching/EKF

To examine the performance enhancement of the proposed
IMM/EKF algorithm, performance comparison with a
model-switching/EKF algorithm has been conducted.
Model-switching/EKF is the simple approach to handle the
multi-model problem by selecting one model from the
multi-model set according to some predefined model selection
criterion. Architecture and model selection strategy of
model-switching/EKF algorithm is summarized in Fig.15. The
strategy of model selection is grounded on the assumption that
radar gives the information of the closest point on the boundary
Fig.12. Test data: actual position (from RT-Range) and the associated radar of the target vehicle. As an example, if the measurement is
signals of target vehicle
detected from the front-right side, the second measurement
model is selected as the model in effect and implemented to the
algorithm.
Statistical analysis of estimation errors of
model-switching/EKF has been conducted following the same
procedure under the same situation of B. The statistics are given
in Table 1. Over all, the proposed algorithm performed
considerably better in all state estimations than the
Fig.13. Mode probability update results
model-switching algorithm. As the standard deviation is the
accuracy specifications of each algorithm, the proposed

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 10

TABLE 1
STATISTICS OF ESTIMATION ERRORS (N=400)
Model-switching/EKF IMM/EKF
Standard (Ratio to Standard (Ratio to
RMS (Ratio to IMM) (Ratio to IMM) RMS
Deviation Model-Switch) Deviation Model-Switch)

(m) 3.8412 5.1326 3.7492 5.0768 0.7484 0.1948 0.7385 0.1970

(m). 1.9908 3.0961 1.9113 3.0843 0.6430 0.3230 0.6197 0.3242
(deg) 5.2473 2.7931 5.1047 2.8513 1.8787 0.3580 1.7903 0.3507
(m/s) 4.0688 5.9465 4.0777 6.2087 0.6842 0.1682 0.6568 0.1611
(deg/s) 5.2730 4.4813 5.2158 4.6778 1.1767 0.2232 1.1150 0.2138
2
(m/s ) 0.7476 4.8278 0.6380 4.1492 0.1549 0.2071 0.1538 0.2410
2
(deg/s ) 2.9483 12.2687 2.9285 12.6176 0.2403 0.0815 0.2321 0.0793

algorithm is approximately three times more accurate than the The selected test set is a complex driving scene where the
model-switching/EKF. targets may enter and leave the field of view, showing the
Finally, computing complexity of these two approaches have capability for birth and death of targets.
been compared. An average consumption time of IMM/EKF Fig.16 shows a frontal image of the vision camera, and the
algorithm is about 4.9806ms and mode-switching/EKF is about multi-target state estimation results in complex driving
1.6787ms. When considering that IMM uses twelve filters in situation. The radar measurements are depicted by blue circles
parallel, this three times more complexity is relatively good with inner star. And based on the lane information, a region of
performance. Moreover, although IMM based algorithm has interest is defined and the outside is represented by red area. As
high complexity compared to mode-switching method, as the can be seen in the figure, it is evident that the proposed
sampling rate of the overall algorithm is 100ms, computing algorithm manages to follow the correct true targets. A true
complexity of IMM/EKF algorithm is not in question. target is most generally defined to be an object that will persist
in the tracking volume for at least several scans. False alarms
(or false targets) refer to erroneous detection events (such as
those caused by random noise or clutter) that do not persist over
several scan. In the figure, we can show that those events
(caused by road struchers such as guardrail in red area) cannot
generate new track. At t=8sec, the proposed algorithm tracks 5
targets. And at t=13sec, one target enters the field of view and
track#19 is newly generated. And at the same time track#10
starts leaving the sensing range. At t=15.5sec, track#10 leaves
the sensing range completely, and zero-gain EKF is applied.
The track#14 starts its lane changing and cuts-in front of the
host vehicle. Due to this lane changing, a measurement of
hidden vehicle is newly arrived. And at t=20sec, track#12 is
newly generated from this measurement while tracks other
existing 4 targets. In summary, these results indicate that GNN
based track management can be utilized with satisfactory. The
GNN based track management approach, which considers the
single most likely hypothesis for track update and new track
initiation, works well because of widely spaced targets,
Fig.15. Architecture of model-switching/EKF Algorithm relatively accurate measurements in the given experiment
situation. However, the problem of closely-spaced unresolved
targets can occur and this can be solved by MHT which forms
VI. EXPERIMENTAL RESULTS WITH MULTI-TARGET SITUATION
multiple hypotheses and manages these hypotheses [17].
In this section, the experimental results dealing with Fig.17 presents the target state estimation results for
multi-target scenarios are presented. Real-road driving data track#14. Each graph shows each state variables, rewarding
were collected from a highway and an urban area. The camera function, and mode probability update result as a function of
is used to analyze the driving situation and to acquire the lane time. As can be inferred from the vision image, the vehicle of
information. Especially, one test set for validation is extracted track#14 changes the lane and cuts into the host vehicle’s lane.
from the collected real-driving data to explore the multi-target From the figure, we can see that the overall trends of each
tracking performance and target state estimation performance. estimated state can represent actual target motion with

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 11

Fig.16. Multi-target state estimation in complex driving situation

satisfactorily while the rewarding function of track#14 driving in the left lane during that period. And a transition to
maintains the maximum value (100 in this paper). In the case of mode.1 (center of rear side) from mode.12 has occurred in a
mode probability update results, from 0-18 sec, mode.12 period of lane changing. Finally, the probability of mode.1
(rear-right edge) is evaluated as the most highly likely mode. ranked highest after the target vehicle becomes the in-lane
This can be explained by the fact that the target vehicle was preceding vehicle. Overall, the mode evaluation results are

Fig.17. Target state estimation results for Track 14 (a target cutting in front of Fig.18. Target state estimation results for Track 10 (a target vanishing to
the host vehicle) rear-right-blind area)

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 12

quite reasonable, in the view of target aspect change. REFERENCES

Fig.18 shows the results for track#10. Track#10 drives with [1] C. Laugier, I. E. Paromtchik, M. Perrollaz, M. Y. Yong, J. D. Yoder, C.
lower speed than the host vehicle in the right lane and leaves the Tay, K. Mekhnacha, and A. Negre, “Probabilistic Analysis of Dynamic
Scenes and Collision Risks Assessment to Improve Driving Safety,”
field of view at about 13sec. From the figure, each state is
Intelligent Transportation Systems Magazine, IEEE , vol.3, no.4, pp.4-19,
estimated reasonably so that can represent actual motion. The winter 2011
rewarding function of track#10 shows a typical track [2] R. Kastner, T. Michalke, J. Adamy, J. Fritsch, and C. Goerick,
management process. It keeps increasing until 10sec, maintains “Task-Based Environment Interpretation and System Architecture for
Next Generation ADAS,” Intelligent Transportation Systems Magazine,
the maximum value from 10sec to 13sec, and drops after the IEEE , vol.3, no.4, pp.20,33, winter 2011
target disappears from the scope of the radar. And in the case of [3] Y. S. Kim, and K. S. Hong, “An IMM algorithm for tracking maneuvering
mode probability update results, from 0-11sec, mode. 2 vehicles in an adaptive cruise control environment,” International Journal
of Control Automation and Systems, vol.2, no.3, pp.310-318, 2004
(rear-left edge) is evaluated as the most highly likely mode. [4] D. S. Caveney, “Multiple Target Tracking in the Adaptive Cruise Control
This can be explained by the fact that the target vehicle was Environment Using Multiple Models and Probabilistic Data Association,”
driving in the right lane during that period. And a transition to M. S. Thesis, California Univ., Berkeley, USA., 1999
[5] M. Chadli, A. Akhenak, J. Ragot, and D. Maquin, “State and unknown
mode.5 (front of the left side) has occurred at 11sec. This input estimation for discrete time multiple model,” Journal of the Franklin
drastic mode change can be explained from the change of radar Institute, vol.346, no.6, pp.593-610, Aug. 2009,
measurement at that moment. In the figure, the radar [6] C. Bing, and K. Tugnait, “Tracking of multiple maneuvering targets in
clutter using IMM/PDA filtering and fixed-lag smoothing,” Automatica,
measurement is depicted as a red circle line and we can show vol.37, no.2, pp.239-249, Feb. 2001,
that longitudinal measurement changes drastically at 11sec. [7] K. H. Kim; J. G. Lee and C. G. Park, “Adaptive Two-Stage Extended
This reflects the fact that the mode adaptation of the proposed Kalman Filter for a Fault-Tolerant INS-GPS Loosely Coupled System,”
Aerospace and Electronic Systems, IEEE Transactions on , vol.45, no.1,
IMM algorithm is quite reasonable.
pp.125-137, Jan. 2009
[8] Li, X. Rong, Y. Bar-Shalom, “Design of an interacting multiple model
algorithm for air traffic control tracking,” Control Systems Technology,
VII. CONCLUSION IEEE Transactions on , vol.1, no.3, pp.186-194, Sep. 1993
[9] E. Mazor, A. Averbuch, Y. Bar-Shalom, and J. Dayan, “Interacting
An interacting multiple model (IMM) approach using multiple model methods in target tracking: a survey,” Aerospace and
extended Kalman filters (EKF) to improve multi-target states Electronic Systems, IEEE Transactions on , vol.34, no.1, pp.103-123, Jan
1998
estimation performance for application to integrated risk [10] B. Pattipati, C. Sankavaram, K. Pattipati, Y. Zhang, M. Howell, and M.
management system (IRMS) has been developed. To cope with Salman, “Multiple model moving horizon estimation approach to
an uncertain measurement model problem of automotive radar, prognostics in coupled systems,” AUTOTESTCON, 2011 IEEE ,
pp.149-157, 12-15 Sept. 2011
the problem is formulated as a multiple model approach and a [11] N. Kazantzis, C. Kravaris, “Time-discretization of nonlinear control
new multi-target state estimation algorithm is developed by an systems via Taylor methods,” Computers & chemical engineering, vol.23,
application of the IMM approach, the GNN based data no.6, pp.763-784, 1999
[12] M. I. Skolnick, Introduction to Radar Systems, 2001 :McGraw-Hill
association technique, and EKF method with enhanced [13] D. Lerro, Y. Bar-Shalom, “Tracking with debiased consistent converted
modeling of automotive radar characteristics. measurements versus EKF,” Aerospace and Electronic Systems, IEEE
The proposed IMM/EKF based multi-target state estimation Transactions on , vol.29, no.3, pp.1015,1022, Jul. 1993
[14] S. Zhang and Y. Bar-Shalom, “Tracking Move-Stop-Move Targets with
algorithm is implemented and verified via vehicle tests in real State-Dependent Mode Transition Probabilities,” Aerospace and
road. From the test results, it is found that the mode probability Electronic Systems, IEEE Transactions on , vol.47, no.3, pp.2037,2054,
adaptations are quite reasonable and the proposed algorithm Jul. 2011
[15] T. Yuan, Y. Bar-Shalom, P. Willett, E. Mozeson, S. Pollak, and D.
can provide good estimates of the target vehicle’s overall
Hardiman, "A multiple IMM estimation approach with unbiased mixing
behaviors. To investigate the performance enhancement, for thrusting projectiles," Aerospace and Electronic Systems, IEEE
comparison with model-switching algorithm which is a simple Transactions on, vol. 48, pp. 3250-3267, 2012.
approach to handle the multiple-model problem has been [16] RT-RANGE User Manual, [Online]. Available:
http://www.oxts.com/products/rt-range/
conducted. It is shown that the target state estimation [17] S. S. Blackman, "Multiple hypothesis tracking for multiple target
performance can be significantly enhanced by the proposed tracking," Aerospace and Electronic Systems Magazine, IEEE, vol. 19, pp.
algorithm with approximately three times more accurate results 5-18, 2004.
while maintain real-time performance with reasonable
computation complexity. Beomjun Kim received his B.S. degrees in
The proposed algorithm can be utilized as a general-purpose mechanical and aerospace engineering
perception module for various advanced driver assistance from Seoul National University, Korea, in
systems (ADAS) such as emergency driving support (EDS) 2011. He is a Ph.D. student in mechanical
system, advanced emergency braking system (AEBS), and aerospace engineering at the Seoul
side-crash prevention (SCP) system, advanced lane change National University, Korea.
assistance (ALCA) system. And furthermore, it is expected to His research interests are control systems,
make possible the generic assessment and the total driver assistant systems, autonomous
management of collision risks with multi-target vehicles in vehicle control and active safety systems of a ground vehicle.
complex driving situations for integrated risk management
system (IRMS).

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/TVT.2014.2329497, IEEE Transactions on Vehicular Technology
VT-2013-01321.R2 13

Kyongsu Yi received the B.S. and M.S.

degrees in mechanical engineering from
Seoul National University, Korea, in 1985
and 1987, respectively, and the Ph.D.
degree in mechanical engineering from the
University of California, Berkeley, in 1992.
He is a Professor at the School of
Mechanical and Aerospace Engineering at
Seoul National University, Korea. He currently serves as a
member of the editorial boards of the KSME, IJAT and ICROS
journals. His research interests are control systems, driver
assistant systems, and active safety systems of a ground
vehicle.

Hyun-Jae Yoo received Engineering and

M.S. degrees from the Universite de
Technologie de Compiegne, France, in
2007 and Ph.D. degree from MINES
ParisTech, France, in 2011. He is a senior
research engineer at Hyundai Motor
Company, Korea.
His research interests are autonomous
vehicle control and intelligent vehicle
safety systems.

Hyok-Jin Chong received B.S. and M.S.

degrees in Mechatronics Engineering from
Chungnam National University, Korea, in
2000 and 2002, respectively. He is a senior
research engineer at Hyundai Motor
Company, Korea.
His research interests are autonomous
vehicle control and intelligent vehicle
safety systems.

Bongchul Ko received the B.S., M.S. and

Ph.D. degrees in electronic and electrical
engineering from Ulsan University, Korea,
in 1998, 2000 and 2004 respectively. He is
a senior researcher at Hyundai-motor
Company, Korea.
His research interests are control systems,
autonomous vehicle control, driver
assistant systems, recognition technology
for car and road environment and active safety systems of a
ground vehicle.

0018-9545 (c) 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.