Professional Documents
Culture Documents
37
In the last few decades several car-following models have been proposed
and tested using mainly vehicle location data. The use of high-precision
Global Positioning System (GPS) data to test several car-following and
collision constraint models is reported, with a critical evaluation of these
models and proposal of a modified collision constraint formulation. GPS
receivers typically report time-stamped location or position fixes and
velocity. For a pair of leading and following vehicles, location and velocity data were used to examine estimates of acceleration, velocity, and
headway by Pipess; modified Pitts, or FRESIM; CARSIM; and INTELSIM car-following models. The important aspects of collecting accurate
GPS data are also highlighted.
During the last few decades, traffic flow models have been developed
to characterize the interaction between a following vehicle and a
leading vehicle when the spacing between the two is less than a few
hundred meters. For distances greater than this, vehicles are not considered to be following; that is, no vehicle interaction occurs. The
earliest studies on this subject, reported in the Highway Capacity
Manual (1), identified the relationship between velocity and spacing
to estimate the capacity of single-lane facilities. However since the
work by Pipes (2), one line of study has focused on determining the
acceleration or deceleration of following vehicles for a leading and
following vehicle pair. The response of a following vehicle is considered to be proportional to a stimulus, which has been characterized by researchers over the years. May (3) provides a good overview
of the history of the development of car-following models. Pipes proposed a model based on the California motor vehicle code that states
that for every 10 mph (16.1 km/h) in speed of the leading vehicle, the
following vehicle maintains a car-length gap. Other significant
efforts followed (4, 5). A series of experiments at the General Motors
Laboratory led to work on five generations of car-following models,
known as the GM models, reported over two decades (68).
Research efforts to develop and validate several car-following
models have used a few major data sets, including those from the GM
Test Track, New York Tunnel, Los Angeles, and Ohio.
The initial car-following experiments were conducted by researchers at General Motors on a 1-mi (1.6-km) test-track (9). An
instrumented follower vehicle had a reel and power unit mounted on
a small platform fastened to the front bumper, and few hundred
meters of wire was attached to the rear bumper of the leading vehicle.
A constant wire tension was maintained on the fine wire connecting
the follower vehicle and the leading vehicle. A direct-current generator tachometer provided a measure of the rate at which the wire was
wound or unwound, and therefore the velocity difference of the two
vehicles was estimated. A wheel attached to the follower vehicle
measured velocity, and an accelerometer mounted in the vehicle provided longitudinal acceleration. Twenty- to 30-min tests were conducted with speeds ranging from 10 to 80 mph (16.1 to 128.8 km/h),
including several braking actions by the lead vehicle.
The Ohio data (10) were collected using aerial photogrammetric
techniques on a four-lane, 3.5-mi (5.64-km) section of the I-70 freeway in Columbus, Ohio, with three on-ramps and three off-ramps
and a maximum 2 percent grade. Most photographs were taken from
an altitude of about 3,000 ft (914.4 m) at 1-s intervals. The reduced
data were reported to be accurate within 0.5 ft (15.24 cm) for
location and 1 mph (1.6 km/h) for velocity. The photographs provided headway, longitudinal positions, and velocity for 115 individual vehicles for about 4 min (between 7:45 and 7:50 a.m. on July
25, 1967). Four platoons of 5 to 15 vehicles were tracked under stopand-go conditions.
Several studies in the literature report results of applying carfollowing formulations to one of the data sets described above. For
example, Pipess model was applied to the GM data, the modified
Pitts model or FRESIM to the California data, and CARSIM and
INTELSIM to the Ohio data. In a more recent paper, Aycin and
Benekohal (11) report on the performance of the FRESIM, or modified Pitts model, CARSIM, and INTELSIM as applied to the Ohio
data. They also provide a good overview of all three models and
express all three formulations in the same consistent notation, thus
making it easier for the formulations to be compared. Their study (11)
focuses on comparing average velocity, headway, and density for
11 vehicles in a platoon.
In the current paper, a report is given on the first use of highprecision Global Positioning System (GPS) data to test these carfollowing models: modified Pitts, CARSIM, Pipess, and INTELSIM.
The estimates of velocity, headway, and acceleration by these models are compared with field data. This study examines the performance of each model, specifically for a leading and following vehicle
pair. GPS technology provides vehicle velocity every second, which
was used in the study, whereas most of the other studies provide location data every second. The important aspects of collecting and analyzing GPS data and applying car-following models to GPS data are
also highlighted.
38
CAR-FOLLOWING MODELS
Modified Pitts or FRESIM Model
Pitts car-following model has been modified and implemented in
the traffic simulation model CORSIM (12), extensively used for
research and design. In FRESIM, CORSIMs freeway simulation
module, the space headway maintained by a leader and a follower
is directly proportional to the relative velocity of the two and the
velocity of the follower, as follows:
x L , tn +1 x F , tn +1 = LL + m + k vF , tn +1 + bk (vL , tn +1 vF , tn +1 )
(1)
where
xL,tn, vL,tn = location and velocity, respectively, of leading
vehicle at time tn;
xF,tn+1, vF,tn+1 = location and velocity, respectively, of following
vehicle at time tn+1;
tn+1 = tn + t = time at beginning of (n + 1)th scanning interval,
where t is the scanning interval;
b = 0.1 if vF > vL (12, 13),
= 0.1 if vF vL 10 [Pitts model (13)],
= 0 otherwise;
m = 10 ft (3.048 m) (default in FRESIM);
k = driver sensitivity factor (default in FRESIM varies
from 0.6 to 1.5 for different drivers); and
LL = leading vehicle length.
In FRESIM, a 0.3-s reaction time is assumed, and therefore acceleration estimated by Equation 1 is applied for only 0.7 s in each time
step (1 s simulation time step minus 0.3 s reaction time). The following vehicle maintains a constant velocity during the 0.3 s of reaction
time; that is, zero acceleration is assumed.
Pipess Model
Pipess car-following model is based on the driving rule suggested
in the California motor vehicle code: A good rule for following
another vehicle at a safe distance is to allow at least the length of one
car between the following vehicle and the vehicle ahead for every
10 mph (16.1 km/h) of speed. The rule implies the following:
x L , tn +1 x F , tn +1 = cvF , tn + LL , tn
(2)
applied. Papers describing CARSIM (14, 15) suggest that the deceleration determined by the maximum of the two collision constraints
is compared with the comfortable deceleration and the deceleration
to desired speed. An examination of the source code (15) suggests
that deceleration based on the first of the two collision constraints
(A4) is never applied.
INTELSIM
For the car-following models presented above, the acceleration of
a following vehicle is determined every scanning interval on the
basis of the leaders action every time step. Recently, a significantly
different approach considers a following vehicles longer-term goal
of achieving a steady-state condition, that is, achieving the same
velocity as the leader while maintaining a minimum preferred time
headway (11, 16). In addition, a following vehicles acceleration is
considered to vary linearly.
The position of the leading vehicle at time tn + T is as follows:
x L , tn + T = x L , tn + vL , tn T + 0.5aL , tn T 2
(3)
where T is the time to reach steady state, and the velocity of the
follower at time tn + T is
vF , tn + T = vF , tn + aF , tn T + 0.5sF , tn T 2
( 4)
where
aF,tn = acceleration of following vehicle at time tn,
sF = slope of linear acceleration of the follower, and the third
term is included since linear acceleration is considered.
The two steady-state (at time tn + T ) conditions are that (a) the
velocities of the leader and follower are equal and (b) the minimum gap headway (dp) is achieved on the basis of the field-measured
preferred time headway. The following two conditions result:
vL , tn + aL , tn T = vF , tn + aF , tn T + 0.5sF , tn T 2
(5)
d p, tn + T = vF , tn + T t p = vL , tn + T t p = (vL , tn + aL , tn T )t p
(6)
vL , tn vF , tn ( aL , tn aF , tn )T
0.5T 2
( 7)
Khan et al.
eration for every time interval. However, this procedure does not
ensure that the two vehicles will not come closer than a critical distance for any action taken by the leading vehicle. The following vehicle may apply emergency control or an emergency deceleration in
response to a lead-vehicle action. This response may not leave sufficient or safe distance between the vehicles. The constraint applied to
avoid this situation is known as the collision constraint.
As shown in Figure 1, the most typical collision constraint examines a specific action of the leading vehicle by looking ahead. It
considers that a leading vehicle may abruptly apply a high deceleration (usually considered the maximum deceleration of the leading
vehicle) at the end of the current time step or scanning interval. As
a result, a following vehicle may apply its emergency deceleration.
The collision constraint checks if either a critical distance between
the vehicles is maintained or at least the deceleration required to
maintain a critical distance is applied.
a2 + a
+
( t c) ( t c)
2
2
2e
x L , tn +1 x F , tn vF , tn t LL cvF , tn (vF , tn vL , tn +1 )
0
( t c)2
2e
when
vL2 , tn +1 + e 2 c 2 ec vF , tn
( t c)
)>0
(8)
39
2( x L , tn +1 x F , tn vF , tn t LL )
( t c )2
when
vF , t n
< a <
( t c )
2( x L , tn +1
vF2 , tn
x F , tn LL )
vL2 , tn +1 + e 2 c 2 ec vF , tn )
( t c )
when
vF , t n
c)
( t
( 9)
(10)
(11)
2 dF
2dL
(12)
where vF,tn+1 = vF,tn + aF,tnt and dL, dF is the maximum deceleration rate
of the leading and following vehicles. The acceleration thus determined is referred to as the collision constraint, or A5 in the original
formulation presented (14, 15). As mentioned earlier, the collision
constraint applied in CARSIM estimates the following vehicles
acceleration taking reaction time into consideration. However, when
the following vehicles location and velocity are updated, the reaction
time is not considered (15).
INTELSIM Collision Constraint Model
FIGURE 1 Collision constraint showing leading and following
vehicles over time.
40
(13)
x F , tn + 1 +
vF2 , tn +1
2dL
x F , tn + (vF , tn c + 0.5aF , tn c )
0
2
(vF , t + aF , t c + aF , t t )
n
n
n+c
+ LL + K
+
2 dF
(15)
where
df = distance traveled by follower to stop to avoid collision
db = gap between vehicles after reaction time c,
= xL,tn+c xF,tn+c LL,
dl = distance for leader to stop = v2L,tn , and
dbuff = buffer gap between vehicles when stopped [value not specified, but typically considered 5 or 10 ft (1.524 or 3.048 m)
in other models]; in INTELSIM, actual buffer space data
from field were used (16).
Modifications to CARSIM
Collision Constraint Model
A more recent study (11) proposed a modification to the CARSIM
collision constraint model. This modification considers the distance
traveled by the following vehicle during reaction time c, where the
vehicles velocity at time t + t remains constant throughout the
reaction time:
x L , tn + 1 +
vL2 , tn +1
2dL
2
x F , tn + (vF , tn t + 0.5aF , tn t )
+ (vF , tn +1 + aF , tn +1 t )c
0
2
[vF , tn + aF , tn t ]
+ LL + K
+
2 dF
Point Positioning
GPS satellites provide the capability of determining location in terms
of longitude, latitude, and elevation by the simple resection process
using the distances measured to the satellites. In essence, the GPS
operates on the principle of trilateration, where the position of an
unknown point is determined by measuring the length of the sides of
a triangle between the unknown point and two or more known points,
the satellites.
(14)
Velocity Determination
Velocity may be calculated by extrapolating the first difference between positions in consecutive intervals. However, instantaneous
velocity of a moving vehicle is determined by using the Doppler
principle of radio signals. Because of the motion of the GPS satellites with respect to the moving vehicle, the frequency of a signal
broadcast by the satellite is shifted when received at the vehicle.
This measurable shift is proportional to the relative radial velocity.
The radial velocity of the satellite is known; therefore the radial
velocity of the moving vehicle can be deduced from the Doppler
observable. This method of velocity determination is very accurate;
the accuracy is 0.1 mph (528 ft/h) even for receivers with location
accuracy of 328 ft (100 m) and selective availability turned on. In
addition, GPS velocity reporting is instantaneous and independent
of position fixes.
Khan et al.
DATA COLLECTED
Car-following experiments were conducted on three days (February 27, 1999; March 12, 1999; and March 24, 1999) for several
hours on a 3-mi (4.83-km) section of the northbound I-25 freeway
in Denver, Colorado, between Evans and Logan Streets. GPS receivers were installed in the leading and following vehicles. Data collected during this period were later analyzed to ensure that only
car-following data were used. GPS data collected were differentially
corrected by postprocessing.
Data collection trips were planned to take advantage of optimal
satellite conditions based on forecast satellite availability and PDOP,
an indication of the current satellite geometry. GPS data were collected
only when at least five satellites were available and the minimum
PDOP value was 6, using a PDOP mask of 6.
A limited data set (six cases) from this large data collection effort
was used for the study. The duration of successive reporting varied
from 10 to 90 s, speed from 0.1 to 61.2 mph (0.16 to 98.53 km/h),
acceleration of 6.8 ft/s2 (2.07 m/s2) to +3.9 ft/s2 (1.19 m/s2)
[one outlier, 11.4 ft/s2 (3.47 m/s2)], and headway of 19.6 to 173.0 ft
(5.97 to 52.7 m).
41
EXAMINATION OF CAR-FOLLOWING
AND COLLISION CONSTRAINT MODELS
In this section the results of applying the models to the GPS data collected are presented. The first part of the analysis compares the performance of all the models on all six test cases. Graphs of model
estimates and field data were plotted for evaluation. For example,
Figure 2 shows the velocity plot of the CARSIM estimates and field
data for the following vehicle every second for all test cases. The
slope and R2 of a linear regression line through the origin and the
average absolute difference between the model estimated values and
the field values were used as performance measures.
For velocity estimates every second, these performance measures
are presented in Table 1 for all models. The high R2-values, with
slopes close to 1, and low average absolute difference for all models
indicate that velocity is tracked very well by all the models. It is
important to mention that FRESIM was tested on the basis of the best
values of the car-following sensitivity (k) values for each case as well
as on two fixed values of 0.6 and 1.7. The best k-values for Cases 1,
2, 3, 4, 5, and 6 were 1.7, 0.6, 0.6, 0.6, 1.2, and 1.1, respectively.
Table 2 presents the performance of the models for acceleration
estimates. From all the performance measures, FRESIM provided a
relatively better match between the field data and model-estimated
acceleration values. However, the car-following sensitivity factor has
FIGURE 2
42
TABLE 1
Comparison Between Models and Field Values of Velocity Estimated Every Second
a significant effect on the performance of the model. Estimated acceleration values for CARSIM and INTELSIM show very low correlation with field data. The slope of 0.60 for CARSIM also suggests
that the model underestimates acceleration.
Table 3 shows that FRESIM and Pipess model provide better estimates of headway. Once again, FRESIMs performance is dependent
on the selected car-following sensitivity factor (k). For a high sensitivity value of 1.7, the headway is overestimated, as indicated by the
slope of 1.36, and for a low sensitivity value of 0.6, the headway is
underestimated, as indicated by the slope of 0.68. The residuals were
high for both k = 1.7 and k = 0.6, yielding a negative R2. CARSIM
estimates also provided high residuals for a linear regression through
the origin.
In the following subsections, the characteristics of each model
are highlighted and therefore the best results of the models are not
necessarily presented.
Pipess Model
According to Equation 2 and assuming a vehicle length of 18 ft
(5.48 m), the gap maintained between vehicles is 0.8L for every
10 mph (16 km/h) of the following vehicles speed. However, the field
data suggest that vehicles maintain a gap between 0.33L and 1.9L.
A collision constraint is not specified in the Pipes formulation.
However, a check was made to ensure that the following vehicle
would not collide with the leading vehicle by the end of the simulation time step, and if such a collision was predicted, the calculated
acceleration rate of the following vehicle was revised to prevent
this collision. This constraint is similar to the A4 constraint of the
CARSIM model outlined earlier.
According to Equation 2, the gap maintained is a function of the
following vehicles velocity. For a leading vehicle traveling at a uniform velocity, a following vehicle accelerates at the rate of 80 percent of the velocity difference between the two vehicles till it
achieves an 0.8L gap and continues to maintain this gap. However,
since the leading vehicles velocity increases sharply, as shown at
10 s for Case 2 in Figure 4, the following vehicle falls behind 20 percent for every time interval, and this difference builds over time
and causes an increasing headway difference. Similar behavior is
observed for Case 3 (Figure 4); when the velocity of the leading vehicle sharply decreases at about 35 s, the difference in the predicted and
field headways starts to increase.
As is evident in both cases, the model predicts that the following
vehicles acceleration rate closely matches that of the leading vehicle.
CARSIM Model
As detailed earlier, the CARSIM model considers five different
acceleration rates in determining the response of the following vehicle. However, a review of the application of the model to Case 4 and
Case 6 data shows that for 70 of the 75 s the value is determined from
the collision constraint described (A5).
A comparison of the model estimates and field data for headway,
velocity, and acceleration is presented in Figure 5. As shown, the
model performs reasonably well in estimating the headway and the
velocity of the following vehicle.
Khan et al.
TABLE 3
Comparison Between Models and Field Values of Headway Estimated Every Second
There is a significant difference between the acceleration rates estimated by the model and those from the field data. Although the average of the acceleration rate predicted over a longer time period (5 s)
by the model is close to that of the field data, the individual estimated
values are either significantly greater or less than the field values.
INTELSIM Model
INTELSIM is based on the principle that a following vehicle tries to
reach the same velocity as the leading vehicle, a steady state, by
applying linear acceleration while maintaining a minimum preferred
headway. For two specific conditions (a steady-state time solution in
which the discriminant is either negative or results in double negative roots), a constant acceleration model is considered. As shown for
Case 4 in Figure 6, for a gradually increasing leading vehicle speed,
INTELSIM produces reasonable velocity and headway estimates.
For Cases 13, 5, and 6, the solution to the quadratic equation to
determine time to reach steady state could not be estimated on the
basis of the documentation available (16).
FIGURE 3
43
Comparison of Models
All the models examinedFRESIM, CARSIM, Pipess, and INTELSIMestimate the following vehicles velocity well. Acceleration
FIGURE 4
FIGURE 5
Khan et al.
FIGURE 6
45
INTELSIM: Case 4.
and headway estimates only are shown in Figure 7, since most differences in performance occur for these estimates. As shown, FRESIM
does well in estimating both headway and acceleration compared with
the other models. Pipess model does as well as FRESIM in estimating the followers acceleration; however, the headway difference
between the field and model data increases when the leading vehicles
CONCLUDING REMARKS
FIGURE 7
46
collection and use of high-accuracy GPS receivers, it is now feasible to collect data for car-following model analysis. In addition,
GPS velocity (accuracy 0.18 ft /s, or 5.48 cm /s) provides a more
accurate, first-difference estimate of acceleration instead of second-difference estimates from location data.
REFERENCES
1. Highway Capacity Manual. HRB, National Research Council, Washington, D.C., 1950.
2. Pipes, L. A. An Operational Analysis of Traffic Dynamics. Journal of
Applied Physics, Vol. 24, No. 3, 1953, pp. 274 281.
3. May, A. D. Traffic Flow Fundamentals. Prentice-Hall, Englewood
Cliffs, N.J., 1990.
4. Kometani, E., and T. Sasaki. On the Stability of Traffic Flow. Journal
of Operations Research, Vol. 2, No. 1, 1958, pp. 1126.
5. Forbes, T. W. Human Factor Considerations in Traffic Flow Theory. In
Highway Research Record 15, HRB, National Research Council, Washington, D.C., 1963, pp. 60 66.
6. Herman, R., et al. Traffic Dynamics: Analysis of Stability in CarFollowing. Operations Research, Vol. E17, 1958, pp. 86106.
7. Herman, R., and R. W. Rothery. Microscopic and Macroscopic Aspects
of Single Lane Traffic Flow. Operations Research, 1962, p. 74.
8. Herman, R., and R. W. Rothery. Car-Following and Steady-State Flow.
In Proc., 2nd International Symposium on the Theory of Traffic Flow
(J. Almond, ed.), OECD, Paris, 1965.
9. Chandler, R. E., R. Herman, and E.W. Montroll. Traffic Dynamics: Studies in Car Following. Operations Research, Vol. 6, No. 2, March 1958,
pp. 165184.
10. Treiterer, J. Investigation of Traffic Dynamics by Aerial Photogrammetry
Techniques. Ph.D. dissertation. Ohio State University, Columbus, 1975.
11. Aycin, M. F., and R. F. Benekohal. Comparison of Car-Following Models for Simulation. In Transportation Research Record: Journal of the
Transportation Research Board, No. 1678, TRB, National Research
Council, Washington, D.C., 1999, pp. 116127.
12. Halati, A., H. Lieu, and S. Walker. CORSIM: Corridor Traffic Simulation Model. Presented at the 76th Annual Meeting of the Transportation
Research Board, Washington, D.C., 1997.
13. CORSIM User Guide. FHWA, U.S. Department of Transportation,
McLean, Va., 1996.
14. Benekohal, R. F. Development and Validation of a Car Following Model
for Simulation of Traffic Flow and Traffic Wave Studies at Bottlenecks.
Ph.D. dissertation. Ohio State University, Columbia, 1986.
15. Benekohal, R. F., and J. Treiterer. CARSIM: Car-Following Model for
Simulation of Traffic in Normal and Stop-and-Go Conditions. In Transportation Research Record 1194, TRB, National Research Council,
Washington, D.C., 1988, pp. 99111.
16. Aycin, M. F., and R. F. Benekohal. Linear Acceleration Car-Following
Model Development and Validation. In Transportation Research Record
1644, TRB, National Research Council, Washington, D.C., 1998,
pp. 1019.
17. Hatipkarasulu, Y., B. Wolshon, and C. Quiroga. A GPS Approach for
the Analysis of Car Following Behavior. Presented at the 79th Annual
Meeting of the Transportation Research Board, Washington, D.C., 2000.
Publication of this paper sponsored by Committee on Traffic Flow Theory and
Characteristics.