Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

116 Paper No.

99-1200 TRANSPORTATION RESEARCH RECORD 1678

Comparison of Car-Following
Models for Simulation
M. F. AYCIN AND R. F. BENEKOHAL

Presented in this paper are the car-following methods and algorithms of 3. Drivers’ perception thresholds are not considered in car
the NETSIM, INTRAS, FRESIM, CARSIM, and INTELSIM models. following.
Moreover, the car-following performance of these models is compared
with the field data. NETSIM, INTRAS, FRESIM, and CARSIM car-
In order to solve the foregoing problems with these simulation
following models first move the leader and then update the follower in
one simulation time step. Because of this approach, these car-following models, INTELSIM (INTELligent Vehicle SIMulator) was devel-
models cannot be used to command vehicles in real-time intelligent oped. INTELSIM moves follower and leader simultaneously and
transportation systems applications. Moreover, brake reaction times are provides solutions for a continuous time frame using a linear accel-
limited by the simulation time step because of this method of updating eration model. INTELSIM solutions are continuous and not limited
the vehicles. INTELSIM was developed to overcome these deficiencies. with the present time step. Therefore, reaction times of the drivers
INTELSIM moves vehicles simultaneously and produces solutions for
are not restricted by the simulation time steps.
a continuous time frame. INTELSIM produced the best agreement with
the field data and required the least amount of calibration effort. In this study, NETSIM, CARSIM, INTRAS, FRESIM, and
INTELSIM car-following methods are compared. Moreover, their
car-following performance is compared with the field data. Fur-
Car-following models form the basis of microscopic simulation thermore, a number of modifications to the FRESIM and INTRAS
models, and they explain the behavior of drivers in a platoon of models have been proposed in order to enhance the performance of
vehicles. Two approaches are used by most car-following simulation these models.
models: The derivation of each car-following algorithm is also presented
since the derivations of some of these models are not available from
1. Vehicles are advanced by considering the emergency braking the published papers. The authors believe that presenting car-
of their leaders, or following derivations of these models by using a common notation
2. Vehicles are spaced out behind the leader according to a certain comparing car-following approaches and the simulation results of
spacing equation. these models with each other will be a valuable reference for
researchers.
Car-following models such as NETSIM and CARSIM consider
emergency braking of leaders in their algorithms. However, the
INTRAS and FRESIM car-following algorithms use a combination
CAR-FOLLOWING MODELS
of the two approaches. INTRAS and FRESIM algorithms not only
place the vehicles at certain locations behind their leaders accord-
NETSIM
ing to a spacing equation but also consider emergency braking of
the leaders.
NETSIM is an urban street network simulation model. In NETSIM
These simulation models—NETSIM, CARSIM, INTRAS, and
car following, the leader is first brought to its new position when the
FRESIM—update the vehicles sequentially in the simulation. First,
simulation time is advanced by one time step. The follower is then
the leader is moved and then the follower is placed at a position
moved to a certain location such that if the leader decelerates to the
satisfying the design constraints of the model. That is, these car-
maximum deceleration limit, the follower will be able to stop with-
following models determine a vehicle’s speed and position after
out colliding with the leader. The main purpose of the car-following
updating its leader for the present time step. This process is
logic is to prevent collisions in any situation. The car-following algo-
described in Figure 1. The final distance (D) between vehicles is
rithm is developed by using the notation in Figure 2, in which
determined by the design constraints of each model. Generally, the
output of these models is the acceleration (+/−) of the following Vfi = speed of follower at beginning of time step;
vehicle. VL, Vf = speed of leader and follower at end of time step,
These car-following models, however, have some shortcomings: respectively;
XL = position of leader at end of time step;
1. Vehicles do not move sequentially in the real world. Therefore,
XLs, Xfs = stopped position of leader and follower, respectively;
these car-following models cannot be used to command vehicles in
X fi = position of follower at beginning of time step;
real-world intelligent transportation system (ITS) applications.
sL, sfT = distance to stop leader and follower, respectively;
T
2. Reaction times of the drivers are restricted by the simulation
∆sf = distance follower travels during time step;
time steps.
∆r = distance follower travels because of reaction time;
Newmark Civil Engineering Laboratory/MC-250, University of Illinois at L = length of leader; and
Urbana-Champaign, 205 North Mathews Avenue, Urbana, IL 61801-2352. T = time step.
Aycin and Benekohal Paper No. 99-1200 117

Similarly,

∆s f = Vf i + 0.5 ∗ a f (8)

and

∆r = Vf ∗ c = (Vfi + a f ) ∗ c ( 9)

s = X L − Xfi − L (10)

where c is the reaction time. If Equations 3 to 10 are inserted into


Equation 2,

VL2 1
XL + − X f − Vfi − a f − Vfi ∗ c − a f ∗ c
2 ∗ dl 2
FIGURE 1 Sequential updating of leader-follower pair.

(Vfi )2 −
Vfi ∗ a f
≥L (11)
2 ∗ df df
Conditions are as follows:
(V i ) 2
[s − Vfi ∗ (1 + c)] + 2 V∗Ldl − 2 ∗f df ≥ a f ∗  c + 12 + dff 
2 Vi
(12)
X Ls − Xfs ≥ L (1)

X L + sLT − Xf i − ∆sf − ∆r − s fT ≥ L (2) Therefore, af is obtained as

The distance to stop the leader is a f = F1 F2 (13)

VL2
sLT = (3) where
2 dl
VL2 ∗ df
F1 = 2 ∗ [ s − Vfi ∗ (1 + c)] ∗ df + − (Vfi )
2
The distance to stop the follower is
dl

s fT =
V f2
( 4) F2 = df ∗ (2c + 1) + 2 ∗ Vfi
2 df
The vehicles in NETSIM are updated by the following equations of
where dl and df are the deceleration rates of the leader and the
motion:
follower, respectively.
The simulation time step of NETSIM is 1 s, which is fixed.
Therefore, V f = V fi + a f ∗ T
(14)
Vf = Vfi + af (5) X f = Xf i + Vf i ∗ T + a f ∗ T 2

V f2 = ( )
2
V fi + 2 ∗ V fi ∗ af + a f2 (6 )

where af is the acceleration of the follower to be found. NETSIM FRESIM and INTRAS
neglects the term af2 and Equation 4 is rewritten as follows:
INTRAS is a microscopic freeway simulation model that was intro-
s fT =
(Vfi )2 + 2 ∗ Vfi ∗ a f ( 7)
duced in 1980s. In 1994, FRESIM was developed with enhancements
2 ∗ df to the INTRAS model. However, the car-following algorithm of

FIGURE 2 Notation used in development of car-following models.


118 Paper No. 99-1200 TRANSPORTATION RESEARCH RECORD 1678

INTRAS remained unchanged in FRESIM. Therefore, the following SIM incorporates a collision algorithm and a minimum separation
equations also apply to FRESIM. constraint for car following. The vehicles are moved sequentially in
INTRAS utilizes the PITT car-following model, which performs simulation as in NETSIM and INTRAS.
the car-following maneuver between simulated vehicles by main- The minimum separation constraint is given as
taining a space headway of h(t) = L + k p Vf + 10 + b p k p (VL − Vf)
between them. Here, L is the vehicle length of the leader, k is the [ ]
X L − X fi + V fi ∗ T + 0.5 ∗ ( a f ) ∗ T 2 ≥ L + K (22)
driver sensitivity factor, and b is a calibration constant defined as
where K is the buffer space between the vehicles and af is the fol-
VL = V f ≤ 10 lowers’ acceleration rate to satisfy the spacing constraint, found by
b = 
0.1
0 otherwise (15) using the equality in Equation 22.
The collision avoidance algorithm is given as follows:
In INTRAS, as in NETSIM, the leader is first advanced to its new
position at the end of the time step. The follower is then brought to
[ ]
X L − X fi + Vfi ∗ T + 0.5 ∗ ( a f ) ∗ T 2 − L − K ≥ Expression (23)
a space headway of h(t) behind the leader. An acceleration value to
  Vf2 V2 
bring the vehicle to this space headway in one time step is calculated Expression = max Vf ∗ c,  Vf ∗ c + − L 
  2 ∗ df 2 ∗ dl 
as follows by using the notation in Figure 2:
where Vf = Vf i + af p T.
X L − ( X fi + ∆ sf ) = L + 10 + k ∗ V f + b ∗ k ∗ (VL − V f )
2
(16)
The acceleration to be applied is the smallest positive value from
where Vf, VL are at the end of the time step. Therefore, Equation 22 or 23 or the acceleration to reach the desired speed or
the acceleration limit of the vehicle. The deceleration required is
V f = V fi + a f ∗ T (17) calculated either from Equation 23 or the minimum of Equation 22
or 23 compared with the comfortable deceleration and emergency
∆s f = V fi ∗ T + a f ∗ T 2 2 (18) deceleration rates. The resultant acceleration value (+ or −) is used
in Equation 14 to update the vehicles as in NETSIM.
where af is the required acceleration value to be found. Equations
17 and 18 are placed into Equation 16 to obtain
DEFICIENCIES OF CAR-FOLLOWING MODELS
X L − ( X fi + Vfi ∗ T + a f ∗ T 2 2) = L + 10 + k ∗ (Vfi + a f ∗ T )
The foregoing car-following models move the vehicles sequentially
+ b ∗ k ∗ (VL − Vfi )
2
(19) in simulation. This method of updating the vehicles prevents the
application of reaction times that are greater than the simulation
Note that af * T is neglected in Equation 17 to obtain (VL − Vfi )2 in time step. Moreover, vehicles do not move sequentially in the real
Equation 19. Equation 19 yields this car-following equation: world. Therefore, the NETSIM, FRESIM, INTRAS, and CARSIM
car-following algorithms cannot be used in controllers that com-
[
a f = 2 X L − X fi − L − 10 − Vfi ∗ ( k + T ) − b ∗ k mand vehicles in real time in the real world. Autonomous intelligent
cruise control (AICC) systems are examples of such applications.
∗ (VL − Vfi )
2
] (T 2
+ 2 ∗ k ∗ T) (20) Furthermore, since the position of the follower must be specified
exactly for each time step for models such as FRESIM and INTRAS,
Reaction time is introduced into the car-following equations as difficulty arises when one type of spacing equation is applied to dif-
follows: the speed and position of the leader are updated after the ferent conditions such as stop-and-go (congested) and noncongested
reaction time c: traffic. It is known that drivers do not car-follow in the same manner
in every situation, especially under congested and noncongested con-
V f = V fi + a f ∗ (T − c)
ditions. Therefore, a spacing equation that places the drivers at cer-
(21)
X f = X fi + V fi ∗ T + a f ∗ (T − c) 2 tain separations at each time step is not going to perform equally for
2

every situation. That is, the spacing equation calibrated for one con-
where c < T. dition does not perform well for another condition. Wicks (1) says
Since vehicles are updated sequentially in INTRAS, the reaction about validation of the INTRAS algorithm, “The behavior of the pla-
times must be smaller than the time step, which is 1 s. Moreover, toon was simulated for 50 s. As a result of examining position trajec-
INTRAS uses fixed reaction times of 0.3 s for deceleration and 0.2 s tories in detail, it was noted that even for this comparatively short
for acceleration for all vehicles. period of time, individual drivers tended to change their type of
INTRAS has emergency constraint equations, which check the car-following behavior, that is, their desired car-following distance.”
acceleration found in Equation 20 for emergency deceleration of the NETSIM and CARSIM perform car-following by considering
leader to stop. The emergency constraint equations are given in emergency braking of the leader. It can be argued that drivers do not
detail by Wicks and Andrews (1). consider emergency braking of their leaders and do not have the
information about the deceleration capability of their leaders.
Although CARSIM has a spacing constraint (Equation 22), this
CARSIM equation only ensures that the vehicles keep the minimum distance
(L + K) during car-following.
CARSIM was developed by Benekohal and Treiterer (2) to solve the The NETSIM and CARSIM algorithms anticipate that drivers are
problems that INTRAS has in simulating stop-and-go traffic. CAR- going to lose some amount of time equal to their reaction time if the
Aycin and Benekohal Paper No. 99-1200 119

leader goes through emergency deceleration. However, reaction where


times are not applied in updating the vehicles in Equation 14. The
t = time to reach steady state,
vehicles are updated as if there is no reaction time. Because of the way
tp = preferred time headway,
the reaction time is used, drivers with high reaction times follow the
( ∆V )t 0 +c = (vf − vL)t0+c, and
vehicles at greater spacings than the drivers with low reaction times
in NETSIM and CARSIM. ( ∆a)t0 + c = (af − aL)t +c.
0

FRESIM and INTRAS apply the reactions correctly as seen in The required slope value is found from
Equation 21. However, when Equation 21 is examined, the acceler-
ations of vehicles during the reaction time period c are neglected
and assumed to be zero. Moreover, reaction times are limited by the
[
s f = −2 ∗ ( ∆V )t0 + c + ( ∆a)t0 + c ∗ t t 2] (26)

simulation time step, and fixed reaction times of 0.3 and 0.2 s are
This slope is applied after the reaction time at time t = t0 + c.
used for all vehicles.
As a check to the car-following algorithm, the time required to
The foregoing car-following models do not consider drivers’ per-
reach the steady-state condition is compared with the time the
ception thresholds. It is known that drivers cannot perceive very small
leader needs to come to a stop at his present rate of deceleration.
changes in speed and acceleration difference and headway deviation.
The derivation of the car-following algorithm is explained in detail
Because of these deficiencies of car-following models, actual
elsewhere (9).
driver behavior in car-following cannot be simulated realisti-
Equation 27 is used to update the vehicles. A slope term is added
cally. INTELSIM was developed to solve these car-following
to the equations of motion used by NETSIM and CARSIM:
model problems.
( a )t 1 = ( a )t 0 + s ∗ T
t1
INTELSIM
∫ (a)t dT = (v)t 1
= (v)t0 + ( a)t0 ∗ T + 0.5 ∗ s ∗ T 2
t0
INTELSIM was developed to simulate the driver car-following behav-
t1
ior as realistically as possible and to simulate vehicles using AICC (3).
INTELSIM was formulated by combining the information about ∫ (v)t dT = ( x )t 1
= ( x )t0 + (v)t0 ∗ T + 0.5 ∗ ( a)t0 ∗ T 2
t0
drivers’ car-following behavior from the literature. The principles + 0.167 ∗ s ∗ T 3 (27)
used in the INTELSIM car-following model are as follows:
Note that Equation 27 is used if the acceleration and slope remain
1. A driver reacts to a decrease in the speed of its leader by trying constant between the previous time step and the current time step.
to equalize his speed with that of the leader’s in order to maintain his If a new acceleration value is to be applied during this time period,
spacing (4). Equation 28 is used:
2. The preferred time headway (tp) is the time headway of a driver
during steady-state car-following (5). (a)t1 = (a)t0 + ∆c + st0 + ∆c ∗ (T − ∆c)
3. The separation at the steady-state is termed the desired space
and can be expressed as follows: (v)t1 = (v)t0 + (a)t0 ∗ ∆c + 0.5 ∗ (s)t0 ∗ ( ∆c) 2 + (a)t0 + ∆c

∗ (T − ∆c) + 0.5 ∗ ( s) t0 + ∆c ∗ (T − ∆c)


2
Desired space = speed ∗ preferred time headway (24)

4. A linear acceleration-deceleration model can represent drivers’


( x )t1 = ( x )t0 + (v)t0 ∗ ∆c + 0.5 ∗ (a)t0 ∗ ( ∆c) 2 + 0.167
coming to a stop or accelerating to desired speeds (6). Similarly, ∗ ( s) t0 ∗ ( ∆c)3 + (v) t0 + ∆c ∗ (T − ∆c) + 0.5 ∗ ( a) t0 + ∆c
drivers use the same principles to reach the steady state with their
leaders during car following (7). ∗ (T − ∆c) + 0.167 ∗ ( s) t0 + ∆c ∗ (T − ∆c)
2 3
(28)
5. Drivers have perception thresholds to changes in the speed
and acceleration of the leader and to the headway deviation (8). where
∆c = time difference between previous update time and time
INTELSIM provides continuous solutions using a linear acceler-
to apply reaction,
ation model. The model moves vehicles simultaneously, and because
(a)t0 + ∆c = new acceleration to be applied after reaction time,
of continuity, simulation time steps do not restrict the reactions of
vehicles. Moreover, INTELSIM can be used in the controllers that (s)t0 + ∆c = new slope to be applied after reaction time, and
move the vehicles in real time in the real world. (v)t0 + ∆c = speed at end of reaction time, given by
INTELSIM’s car-following approach can be described as fol-
lows: assume that a leader and follower pair is initially separated by (v)t0 + ∆c = (v)t0 + (a)t0 ∗ ∆c + 0.5 ∗ (s)t0 ∗ ∆c (29)
a distance ∆ x at time t = t0 and that they have different speeds. The
follower will reach the steady state with the leader at time t = t1 at In Equation 28, a reaction that was scheduled in a past time step is to
its desired spacing. A slope of acceleration is calculated that will be applied at a time that falls in between the present and the previous
bring the follower to steady state. The car-following equation is simulation time steps.
A follower reaches the steady state with his leader at a certain
0.166 ∗ ( ∆a) t 0 + c ∗ t 2 + [0.666 ∗ ( ∆V ) t 0 + c + a L ∗ t p ] ∗ t time in the future. Therefore, the spacings are not specified at each
time step. This approach reduces the need to precisely specify the
− ( ∆ x ) t 0 + c + vl ∗ t p = 0 (25) vehicle spacings and reduces the model calibration effort.
120 Paper No. 99-1200 TRANSPORTATION RESEARCH RECORD 1678

COMPARISON OF CAR-FOLLOWING stopped separation. Therefore, the effective vehicle length used in
MODEL PERFORMANCE the simulations is 5.5 m (18 ft) (1).
Three sets of INTRAS simulation runs were performed with k-
Software values of 1, 1.45, and 1.9 from the range of 1 to 1.9 for all the vehi-
cles in the simulation. Plots of average speed and density versus
Separate simulation programs were written for each car-following time (Figure 3) show that k = 1 provided the closest speed and den-
model to be able to compare the performance of each model with the sity approximation to the field data. Since k = 1 is the smallest value
field data. In the simulations, 13 vehicles were all given starting posi- from the range of 1 to 1.9, a new range of k-values was needed.
tions, speeds, and accelerations as in the field data. These vehicles FRESIM uses a new set of k-values ranging from 0.6 to 1.5. Bet-
followed the field leader, whose position and speed were input from ter results were obtained when the simulation was run with ran-
the data, under the command of one of the car-following models. The domly generated k-values from this range. The best results of 10
same vehicle characteristics were used in all simulations. random simulation runs are given in Figure 4. Although FRESIM
provided better results than INTRAS because of the k-range it uses,
there is still a need for some improvement. As can be seen from Fig-
Data Set ure 4, the average k of 0.9 provided the closest results to the field
data for the congested condition when the vehicles were coming to
Data collected by Treiterer (10) were utilized in the study. The vehi- a stop. On the other hand, k = 1.06 yielded density and volume fig-
cles in the data set remained in the car-following mode for 135 s and ures identical to the field data for noncongested conditions. Conse-
experienced stop-and-go conditions. This data set provided positions, quently, FRESIM did not provide good results for congested and
speeds, and accelerations of vehicles for 1-s time intervals. noncongested conditions at the same time. The density and volumes
The preferred time headways of drivers in the data set were found could be improved for congested conditions at the expense of den-
by averaging the time headways of a vehicle at the same speeds as sity and volumes for noncongested conditions. The reason is that
those of his leader. The preferred time headways of the drivers in the FRESIM and INTRAS use one type of equation that cannot be cal-
field data were found to be in the range 1.1 to 1.9 s, with an average ibrated for congested and noncongested conditions separately. In
of 1.47 s. order to improve the simulation results, a new set of equations that
The reaction times of the drivers are not reported in the field data. can simulate stop-and-go conditions without affecting freeway
However, the reaction times of the drivers were estimated to be in performance will be presented.
the range of 0.88 to 1.51 s. Aycin and Benekohal (9) explain the Another problem with FRESIM and INTRAS is that it is difficult
calculation of preferred time headways and reaction times in detail. to come up with the right k-values. The k-parameter is known to
affect the throughput or capacity outcomes; however, it is not clear
what the k (driver sensitivity) parameter refers to at an individual
Methodology vehicle level. Next, a method to improve k-parameter selection in
FRESIM and INTRAS will be proposed.
Macroscopic measures of average speed, density, and volume versus
time were obtained to compare the car-following models with field
data. Only plots of average speed and density versus time are included Proposed Method
in this paper since volume is a product of speed and density.
In the density-versus-time plots, density values are calculated by Aycin and Benekohal (9) used preferred time headways of drivers
in order to simulate their car-following on the freeways. The de-
D = 5, 280 ∗ [ N ( Xfirst − Xlast + L )] (30) sired spacing equation was given as Equation 24. The vehicles in
INTELSIM try to reach their desired spacings in steady state with
where their leaders. INTRAS, on the other hand, places the vehicles at the
following space headway, h(t), at every simulation time step:
D = density (veh/km);
N = number of vehicles in platoon, 13 in this case; h(t ) = L + 10 + k ∗ Vf + b ∗ k ∗ (VL − Vf )
2
(31)
L = car length of last vehicle (m), and
Xfirst, Xlast = positions (m) of first and last vehicles in platoon at If the adjustment factor b * k * (VL − Vf) is ignored, Equation 31 is
any given time. similar to Equation 24 where the preferred time headway is the
time headway between the front bumper of the follower and the
rear bumper of the leader. Therefore, the k-parameter in INTRAS
FRESIM and INTRAS and FRESIM is indeed a time headway expression for drivers. An
alternative representation of the INTRAS spacing equation is
In FRESIM and INTRAS, drivers are generated with driver sensi-
h(t ) = L + t p ∗ Vf + b ∗ t p ∗ (VL − Vf )
2
tivity factors, or k-values. The mean k-value is the determinant (32)
factor for density and volume outcomes. Therefore, calibration of
FRESIM and INTRAS to yield reasonable density and volume fig- where tp is the preferred time headway.
ures comparable with field data involves selecting a suitable range Note that 3 m (10 ft) of extra spacing has been dropped and the
of k-values. actual vehicle length, 4.5 m (15 ft), is used. Moreover, the emer-
INTRAS default values of k are from 1 to 1.9 in increments of gency constraint of INTRAS was also modified as follows. If the
0.1, each corresponding to a particular type of driver (11). More- leader decelerates at the maximum deceleration rate to stop, the fol-
over, INTRAS increases the vehicle length by 0.9 m (3 ft) for lower must be able to stop behind the leader at his buffer space
Aycin and Benekohal Paper No. 99-1200 121

apart. The distance that the follower travels because of his reaction V f = V fi + afi ∗ c + a f ∗ (T − c)
time, af, is also included to satisfy the following equation:
X f = Xfi + V fi ∗ c + afi ∗ c 2 + (V f ) c ∗ (T − c)
1
 2
V2
[
X L + L −  Xfi + V fi ∗ T + 0.5 ∗ a f ∗ T 2 + V fi + ( a f ) ∗ T
2 ∗ dl 
]
+ 0.5 ∗ a f ∗ (T − c)
2
(35)
[
V fi + ( a f ) ∗ T ]
2

∗c + + L + K = 0 (33) where afi is the acceleration present at the beginning of the time step
2 ∗ df  and af is the acceleration calculated for the present time step.
The simulation runs using the proposed method yielded very good
Note that this equation is similar to the CARSIM car-following agreement with the field data (Figure 4). FRESIM and INTRAS spec-
equation. ify only 10 different types of drivers. Instead, a suitable range for the
If the leader has already stopped, the follower must come to a preferred time headways of drivers can be specified to represent the
stop at his buffer spacing behind the leader. Note here that Equation driver population.
34 may not bring the vehicle to a stop at the end of the simulation
time step:
NETSIM
a f = − 0.5 ∗ (Vf i ) ( XLs − Xfi − L − K)
2
(34)
Although NETSIM is an urban street simulation program, it is
As mentioned earlier, Equation 21 does not consider any acceler- included in this study because the field data have a stop-and-go
ation that is present during the reaction time. The authors recommend portion that NETSIM can simulate and to find out why NETSIM is
that it be changed as follows: particularly used for urban street simulations.

FIGURE 3 INTRAS plots of average speed and density versus time for k-values of 1, 1.45,
and 1.9.
122 Paper No. 99-1200 TRANSPORTATION RESEARCH RECORD 1678

FIGURE 4 FRESIM plots of average speed and density versus time for random k-values in range of
0.6 to 1.5 and proposed FRESIM model results.

Unlike INTRAS, the NETSIM car-following equation does not follow at farther distances than drivers with low reaction times in
have parameters to be calibrated. However, the maximum deceler- NETSIM.
ation rates of the following and leading vehicles and reaction times The effect of inserting actual buffer spaces of the vehicles rang-
of the drivers could be adjusted to simulate the field data. First, sim- ing from 1.5 to 5 m (5 to 17 ft) instead of the 0.9-m (3-ft) buffer with
ulations were performed with default values. The maximum decel- the field reactions produced lower densities. Since NETSIM includes
eration rate for both leader and follower, 3.6 m/s2 (12 ft/s2), was these buffers to vehicle length to find the “effective vehicle length,”
selected as the default, and the brake reaction time or “lag” was using larger effective lengths affected the overall performance of the
taken as 1 s for all the vehicles. NETSIM also adds a 0.9-m (3-ft) simulation instead of affecting only the stop-and-go portion.
buffer space to the vehicle length as the buffer space in stopped The best results obtained came from the use of 4.9-m/s2 (16-ft/s2)
separation. deceleration rates for the leaders and 3.6-m/s2 (12-ft/s2) rates for the
Plots of density and volume versus time show that NETSIM followers with default values.
default values produce much higher densities and volumes than the NETSIM, in general, did not space out the vehicles as speeds
field data (Figure 5). Utilizing the field reaction times with an aver- increased and provided too close car-following for high-speed con-
age of 1.15 s instead of 1 s reaction lag produced better results. This ditions. The use of different deceleration rates for followers and
result is due to the fact that drivers with high reaction times car- leaders was found to be the best strategy to calibrate NETSIM for
Aycin and Benekohal Paper No. 99-1200 123

FIGURE 5 NETSIM plots of average speed and density versus time (1ft = 0.3 m).

freeway data. On the other hand, NETSIM can easily be calibrated the density is less than 98 veh/km (60 vpm), traffic is considered
for congested conditions on freeways, which very much resemble noncongested, and “surprised” brake reaction times with maximum
urban street situations. deceleration rates of 4.9 m/s2 (16 ft/s2) for both leaders and follow-
ers are utilized. For congested conditions, maximum deceleration
CARSIM rates of 3.9 m/s2 (13 ft/s2) and 4.9 m/s2 (16 ft/s2) are used for the fol-
lowers and leaders, respectively. Moreover, a buffer spacing of 3 m
CARSIM uses different reaction times and buffer spaces to show (10 ft) was used for noncongested conditions as opposed to 1.5 to
dual traffic behavior for congested and noncongested conditions. If 2 m (5 to 7 ft) of buffer spacing for congested conditions.
124 Paper No. 99-1200 TRANSPORTATION RESEARCH RECORD 1678

First, simulation runs were performed by using the reaction times In order to see the effect of various reaction times on car-following,
of the drivers from the field. Density and average speed plots for this simulation runs were performed with randomly generated reac-
run provided good agreement with the field data (Figure 6). Since tion times from a cumulative distribution based on Johansson and
CARSIM utilizes buffer spaces of the vehicles, 1.5 to 5 m (5 to Rumer’s data (2). Different reaction times shifted the density plots,
17 ft) of buffer space from the field was substituted for the default as expected. In order to show this point clearly, runs were generated
buffer space for the second runs. The results did not differ from for drivers with high reaction times, 1.2 s on average, and for drivers
those of the first run because these buffer spaces are used if speeds with low reaction times, 0.8 s on average. The results are presented
are less than 2.2 m/s (7.3 ft/s) (2). In both simulation runs, the vehi- in Figure 6.
cles were following each other at closer spacings than the vehicles These results show that reaction times can be used to calibrate
in the field after the stop-and-go region. CARSIM. Higher reaction times space out the vehicles and decrease

FIGURE 6 CARSIM plots of average speed and density versus time.


Aycin and Benekohal Paper No. 99-1200 125

densities. On the other hand, shorter reaction times cause closer car STATISTICAL ANALYSES
following in simulation and increase densities. This is, however, not AND DISCUSSION OF RESULTS
necessarily the case in the real world because drivers’ car-following
distances may not depend on their reaction times. Since the vehicles in the simulations were following the leader from
the field data, plots of average speed versus time did not show much
variation between models and runs for a particular model. There-
INTELSIM fore, in most of the cases, the average density variations between
field data and simulation results determined the success of each run.
INTELSIM was run with drivers’ preferred time headways, reac- Regression analyses were performed between simulation results
tion times, and buffer spaces obtained from the field data. INTEL- and the field data. Correlation coefficients (R2) of simulation results
SIM gave the closest results to the field data among the versus field data for all the simulation runs are given in Table 1.
car-following models compared (Figure 7). Moreover, INTELSIM Regression analysis results show that there are almost no differ-
was able to use the information from the field data directly, which ences between models in terms of their R2 values. This finding is
reduced calibration time considerably. These results can be attrib- rather unexpected because the plots of density versus time in Fig-
uted to the fact that INTELSIM uses a different car-following ures 3 through 7 show that there are differences between models
approach than the other models and aims to simulate actual driver and even that the same model performs differently depending on
car-following behavior. the calibrations performed.

FIGURE 7 INTELSIM plots of average speed and density versus time.


126 Paper No. 99-1200 TRANSPORTATION RESEARCH RECORD 1678

TABLE 1 Correlation Coefficients of Simulated Versus Actual Densities of All Simulation Runs
for Each Model

The results of the regression analysis suggest that there might be CONCLUSIONS
problems in the data set or that regression analysis may not be appro-
priate. Regression analysis assumes that estimation errors are un- The similarities and differences of well-known car-following mod-
correlated. However, the speed and density values are correlated with els have been presented and their performance compared with the
the next and previous speed and density values. Moreover, residual field data. NETSIM, CARSIM, FRESIM, and INTRAS sequen-
plots of fitted regression lines suggest that there is nonconstant vari- tially update vehicles. This approach prevents application of these
ance in the regression. Therefore, regression analysis results are not models to situations in which the vehicles are commanded by real-
meaningful for this type of data. Therefore, the simulation results time controllers. Moreover, brake reaction times cannot be greater
will be compared with the field data visually. than the simulation time step because of this approach.
Since calibrations performed on a model play a very important role Although NETSIM is developed for urban street simulations, its
in model fitting to field data, they can determine how good a model car-following model yielded good agreement with the field data for
is compared with another model. Therefore, car-following models stop-and-go conditions. However, it could not perform well in non-
can be evaluated for their performance by considering congested conditions. NETSIM did not space out the vehicles as
speeds increased because the NETSIM algorithm only aims to pre-
1. How well a calibrated model can simulate field data; compar- vent collisions. Therefore, it provides too close car-following for
ing microscopic and macroscopic plots required for this task; high-speed conditions.
2. The amount of calibration effort required to make a model fit FRESIM provided results that were similar to those from field data
to the field data; for either noncongested or congested conditions but not for both at
3. How well the calibrations applied to a particular model reflect the same time. FRESIM and INTRAS require proper selection of a
the field data; and k-parameter, which determines the throughput from the simulation.
4. The ease with which a model can utilize the available infor- FRESIM utilizes a k-parameter range that yields better agreement
mation from the field. with the field data than the INTRAS model does; however, the selec-
tion of a mean k-value is not straightforward. It was shown that the
According to the above criteria, INTELSIM produced the best fit k-parameter is a time headway assignment for the drivers. Therefore,
to the field data among the group of models compared. Moreover, using tp (preferred time headway) instead of the k-parameter in the
INTELSIM used the information from the field and consequently FRESIM and INTRAS algorithms makes them easier to calibrate.
required the least amount of calibration. The proposed method for The proposed modifications to FRESIM and INTRAS car-following
FRESIM comes second because it produced good results by using models yielded very good agreement with the field data.
the information from the field data. CARSIM’s performance was CARSIM is a more detailed model than NETSIM, INTRAS, or
better for the stop-and-go region than FRESIM’s. However, FRESIM FRESIM. It tries to simulate dual traffic behavior in congested and
showed better performance for noncongested conditions than in noncongested conditions. Therefore, it is easy to calibrate for
CARSIM. FRESIM performed better than INTRAS because of the stop-and-go conditions and yields good results.
k-parameter range it employed. NETSIM was developed for urban INTELSIM required the least amount of calibration effort since it
streets; therefore, it required some tricky calibrations to make it was able to use driver characteristics (preferred time headway, buffer
work for the field data. spaces) from the field. INTELSIM updates the vehicles simultane-
Aycin and Benekohal Paper No. 99-1200 127

ously. Therefore, it can be used to simulate vehicles commanded by 3. Aycin, M. F., and R. F. Benekohal. A Linear Acceleration Car-Follow-
controllers such as AICC. Moreover, reaction times of drivers are not ing Algorithm for Autonomous Intelligent Cruise Control Systems.
Proc., Fifth International Conference on Applications of Advanced
restricted by the simulation time steps. Furthermore, INTELSIM Technologies in Transportation Engineering, 1998.
produced the best agreement with the field data. 4. Michaels, R., and D. Solomon. Effect of Speed Change Information on
Regression analysis between field data and simulation results did Spacing between Vehicles. Public Roads, Vol. 31, No. 12, 1962.
not provide meaningful results because of the correlation that exists 5. Winsum, W. V., and A. Heino. Choice of Time-Headway in Car-
between data points. Following and the Role of Time to Collision Information in Braking.
Ergonomics, Vol. 39, 1996, pp. 579–592.
This comparison was based on one data set for freeway traffic. 6. Lee, C., and T. Rioux. The TEXAS Model for Intersection Traffic:
Therefore, more field data should be collected for further validation Development. Research Report. University of Texas at Austin, Dec.
and comparison of car-following models and to understand driver 1977.
car-following behavior. Moreover, field data are required to deter- 7. Spurr, R. T. Subjective Aspects of Braking. Automotive Engineer, Feb.
1969.
mine the distribution of the drivers’ preferred time headways. In addi- 8. Hattori, Y., K. Asano, N. Iwama, and T. Shigematsu. Analysis of Driv-
tion, the relationship between the drivers’ preferred time headways er’s Decelerating Strategy in a Car-following Situation. Vehicle System
and their reaction times must be examined. Dynamics, Vol. 24, 1995, pp. 299–311.
9. Aycin, M. F., and R. F. Benekohal. A Linear Acceleration Car-Following
Model Development and Validation. In Transportation Research Re-
cord 1644, TRB, National Research Council, Washington, D.C., 1998,
REFERENCES pp. 10–19.
10. Treiterer, J. Investigation of Traffic Dynamics by Aerial Photogramme-
1. Wicks, D. A., and B. J. Andrews. Development and Testing of try Techniques. Final Report EES 278. Transportation Research Center,
INTRAS, a Microscopic Freeway Simulation Model, Vol. 1: Program Department of Civil Engineering, Ohio State University, Feb. 1975.
Design, Parameter Calibration and Freeway Dynamics Component 11. Cheu, R., W. Recker, and S. Ritchie. Calibration of INTRAS for Simu-
Development. Report FHWA-RD-80-106. FHWA, U.S. Department of lation of 30-s Loop Detector Output. In Transportation Research
Transportation, 1980. Record 1457, TRB, National Research Council, Washington, D.C.,
2. Benekohal, R. F., and J. Treiterer. CARSIM: Car-following Model for 1994, pp. 208–215.
Simulation of Traffic in Normal and Stop-and-Go Conditions. In Trans-
portation Research Record 1194, TRB, National Research Council, Publication of this paper sponsored by Committee on Traffic Flow Theory
Washington, D.C., 1988, pp. 99–111. and Characteristics.

You might also like