You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/245563370

From Existing Accident-Free Car-Following Models to Colliding Vehicles

Article in Transportation Research Record Journal of the Transportation Research Board · December 2008
DOI: 10.3141/2088-06

CITATIONS READS
70 606

2 authors:

Samer H. Hamdar Hani S. Mahmassani


George Washington University Northwestern University
50 PUBLICATIONS 1,223 CITATIONS 551 PUBLICATIONS 20,111 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Hani S. Mahmassani on 02 February 2014.

The user has requested enhancement of the downloaded file.


From Existing Accident-Free Car-Following
Models to Colliding Vehicles
Exploration and Assessment

Samer H. Hamdar and Hani S. Mahmassani

The study explores the specifications of microscopic traffic models that ity to describe driver behavior under extreme and incident conditions
could capture congestion dynamics and model accident-prone behaviors is limited.
on a highway section in greater realism than existing models currently This paper explores specifications of microscopic traffic models
used in practice (commercial software). A comparative assessment of that could capture congestion dynamics and model accident-prone
several major acceleration models is conducted, especially for congestion behaviors on a highway section in greater realism than existing
formation and incident modeling. On the basis of this assessment, alter- models currently used in practice (commercial software). The focus
native specifications for car-following and lane-changing models are is on the following models:
developed and implemented in a microscopic simulation framework. The
models are calibrated and compared for resulting vehicle trajectories 1. Gazis–Herman–Potts (GHP) model (4),
and macroscopic flow–density relationships. Experiments are conducted 2. Gipps model (5),
with the models under different degrees of relaxation of the safety con- 3. Cellular automaton (CA) model (6),
straints typically applied in conjunction with simulation codes used in 4. SK model (7 ),
practice. The ability of the proposed specifications to capture traffic 5. Intelligent driver model (IDM) (3),
behavior in extreme situations is examined. The results suggest that 6. Intelligent driver model with memory (IDMM) (8), and
these specifications offer an improved basis for microscopic traffic sim- 7. Wiedemann model (9).
ulation for situations that do not require an accident-free environment.
As such, the same basic behavior model structure could accommodate The third section of the paper presents the results obtained for indi-
both extreme situations (evacuation scenarios, oversaturated networks) vidual vehicle trajectories, macroscopic flow–density relationship, and
as well as normal daily traffic conditions. the ability to model accidents when relaxing the safety constraints. On
the basis of this analysis, a modified car-following model with a sim-
plified lane-changing framework is constructed. The model with its
Acceleration models are at the core of traffic operational behaviors. new specification is shown to capture certain traffic characteristics dur-
These models include car-following models, which capture the inter- ing breakdown situations. Moreover, the results suggest that these spec-
action between a lead and a following vehicle. Fundamentally, car- ifications offer an improved basis for microscopic traffic simulation
following models describe the trajectory of the nth vehicle in a traffic for situations that do not require an accident-free environment.
lane given the trajectory of the (n − 1)th vehicle in the same lane.
Accordingly, the main assumption in these models is that a relation-
ship exists between a leader and a follower traveling on the same lane SELECTED ACCELERATION MODELS
when intervehicle spacing is within a given range, typically between
In previous continuous-time, single-lane car-following models, the
0 and 125 m (1). According to Boer (2), more recent acceleration
main response to a given stimulus was performed through accelera-
models are structured to account for several factors, such as task sched-
tion or deceleration (stimulus–response or General Motors models).
uling and attention management and the use of perception rather than
The stimulus consisted of the velocity of the driver, the relative veloc-
Newtonian variables.
ity between a vehicle and the front vehicle, and the corresponding
Although several studies have sought a connection between micro-
space gap (10). A limitation of these models is that they are not applic-
scopic behavior of drivers and the macroscopic properties of the traf-
able under very low traffic densities. Another is that in dense traffic,
fic system, phenomena associated with congestion dynamics such as
small gaps will not induce braking reactions if the front vehicle is
flow breakdown on freeways and hysteresis remain to be fully eluci-
traveling at the same velocity (zero relative velocity).
dated for microscopic determinants (3). Moreover, since all existing
Newell addressed this problem by introducing the concept of the
car-following models are built in a crashfree environment, their abil-
velocity depending adiabatically on the gap (11). As with all previous
car-following models, the Newell model is collision free. Moreover,
Department of Civil and Environmental Engineering, Northwestern University, since there is an immediate dependence of the velocity on the den-
Transportation Center, Chambers Hall, 600 Foster Street, Evanston, IL 60208. sity (gap), very high and unrealistic accelerations can be produced.
Corresponding author: H. S. Mahmassani, masmah@northwestern.edu. To overcome this limitation, Bando et al. modified Newell’s model
by controlling the change in velocity by a relaxation time (12), result-
Transportation Research Record: Journal of the Transportation Research Board,
No. 2088, Transportation Research Board of the National Academies, Washington,
ing in the optimal velocity model (OVM). The model is known to pro-
D.C., 2008, pp. 45–56. duce possibly unrealistic accelerations when the relaxation time is less
DOI: 10.3141/2088-06 than 0.9 s.

45
46 Transportation Research Record 2088

In 1998 (13), the so-called generalized force model offered a gener- in the absence of congestion, most of the existing microsimulation
alized optimal velocity function that incorporates reaction to velocity models can imitate real traffic conditions reasonably well (15, 16). To
differences and different rules for acceleration and braking. However, increase the confidence in the comparison, three additional simulation
although this model was able to produce time-dependent gaps and runs were used to test the effect of randomness on the obtained results.
velocities, unrealistic small accelerations and decelerations were The same range of output values and the same patterns of behavior
produced as well. Although the Newell model, the OVM, and the were observed. Finally, the common random number method is used
generalized force model offer important insight to the car-following when generating the vehicle characteristics and their interarrival times.
logic, they are not be included in the comparative tests presented in
the next section, in part because of the known issues they face and
also because the character of the behavior they produce is subsumed Data Analysis
in other models.
To compare the seven models of interest consistently, the parameters
found suitable in the calibration study are used during the data analy-
NUMERICAL ASSESSMENT sis. The comparison is performed for flow–density relationships, indi-
vidual vehicle trajectories, and incident formation when relaxing the
To compare the performance of the seven models, their respective safety constraints.
logics were implemented by using Visual C++ language. The seven
models were then calibrated.
Fundamental Diagrams

Model Calibration In this section, the mean interarrival time was modified so that the
flow–density data points can cover most of the fundamental diagram
The data used for calibration are provided through the FHWA Next space. Seven fundamental diagrams were thus obtained, one for
Generation Simulation (NGSIM) project. The data set includes tra- each of the seven models that use the parameters calibrated in the
jectory data for 4,733 vehicles for 0.5 h (2:35 to 3:05 p.m.) observed previous subsections. These fundamental diagrams are illustrated
on Interstate 80 in Emeryville, California, on December 3, 2003 (14). in Figure 1.
The study area is a straight 2,950-ft freeway section consisting of six It can be seen that the GHP model (Figure 1a), the original Gipps
lanes (Lanes 1 through 6) with an on ramp (Lane 7) at the beginning model (Figure 1b), and the SK model (Figure 1d) did not capture either
of the section and an off ramp (Lane 8) at the end. The x and y coor- the metastable congested state or the instability encountered during or
dinate location is captured every 1/15 s. These data are also processed at the beginning of the traffic breakdown. This problem has been men-
so aggregate traffic measures such as flows and space–mean speeds tioned in several publications in the literature in the case of the GHP
are calculated over the period of the study. Table 1 shows the aggre- or Gipps model. However, although the SK model is a simplified ver-
gate traffic state variables extracted from the data for the period of sion of the Gipps model, previous studies indicated that the SK model
interest. The calibrated models were subsequently implemented and offers the advantage of capturing congested traffic behavior because
tested on a single-lane straight freeway section 10 km long. The vehi- of the deceleration randomization inherited form the CA model.
cles are generated and injected into the network following an expo- In its turn, the CA model showed the beginning of a traffic break-
nential interarrival time; 3,600 vehicles are generated in a period of down when reaching a flow capacity of 1,800 vehicles per hour
2 h before ending the simulation. (vph); see Figure 1c. However, as mentioned earlier, the CA lacks
The calibration is performed so that acceptable parameters values the cognitive logic behind it, making the model “mechanical” and
are determined for each model to perform the comparative analysis sometimes unrealistic. This is because the model is controlled heavily
described in the next subsections. The parameter values that corre- by the constant deceleration rate attributed to the drivers. Although
spond to the traffic characteristics shown in Table 1 are presented in still more improvement is needed on the cognitive side of the model,
Table 2. Average flows, densities, speeds, and headways for the seven the IDM has improved on the CA model in that respect. Both IDM and
models are collected by placing virtual detectors at the end of each IDMM showed realistic fundamental diagrams with a stable region
1-km section of the freeway (10 detectors), collecting average data for and an unstable region (Figures 1e and 1f, respectively).
each 30 min. Thus, 40 data points are obtained in each simulation run. Finally, the Wiedemann model showed the same congestion insta-
The results are posted in Table 3. The similarity between the average bility (Figure 1g) exhibited by the IDM and the IDMM models. More-
flow values and average speed values is seen when compared to the over, its complexity, although a disadvantage for some researchers,
results obtained in the NGSIM data. This is expected since previous allows a more realistic and complete view of the different factors
research, especially that performed by Brockfeld et al., found that encountered in the driving task.

TABLE 1 Aggregate Results, NGSIM Data

Lane

Measure 1 2 3 4 5 6 Average

Flow (vph) 1,744 1,764 1,406 1,540 1,506 1,506 1,578


Space–mean speed 29.97134 24.95984 24.5949 24.75938 24.5692 25.02666 25.62804
Hamdar and Mahmassani 47

TABLE 2 Calibrated Parameter Values for Various Models

Parameter Calibrated Value Parameter Calibrated Value

GHP Model IDMM Model


Constant parameter c 12.1 m/s (40.2 mph) Desired velocity v0 31 m/s
Reaction time (lag) T 1s Netto time gap T0 0.85 s
Gipps Model Maximum acceleration a 0.8 m/s2
Desired velocity Vn 29 m/s Comfortable deceleration b 1.8 m/s2
Mean reaction time τn 0.66 s Adaptation factor βT = Tjam /T0 1.8 s
Mean acceleration an 0.73 m/s2 Adaptation time τ 600 s
Deceleration bn −(2* an) m/s2 Minimum distance s0 1.6 m
Mean vehicle length sn 6.5 m Vehicle length l = 1/ρmax 6m
Continuous CA Model Wiedemann Model

Maximum velocity vmax 28 m/s AXadd 1.25


Maximum acceleration amax 1.37 m/s2 AXmult 2.5
Maximum deceleration σ 1.73 m/s2 BXadd 2.0
Vehicle length 4m BXmult 1.0
EXadd 1.5
SK Model
EXmult 0.55
Maximum velocity vmax 25.7 m/s
OPDVadd 1.5
Maximum acceleration b 1.37 m/s2
OPDVmult 1.5
Maximum deceleration b 0.73 m/s2
CX 50
Model parameter  0.4
BNullmult 0.1
Vehicle length 4m
BMAX 3.5 − (3.5/40). V(I)
IDM Model
BMIN −20 + (1.5/60). V(I)
Desired velocity v0 31 m/s VDES 33 m/s
Safe time headway T 1.6 s VMAX 44 m/s
Maximum acceleration a 0.73 m/s2
Modified Gipps Model
Desired deceleration b 1.67 m/s2
Desired velocity Vn 33 m/s (one distribution adopted)
Acceleration exponent δ 4
Mean reaction time τn 0.66 s
Jam distance s0 2m
Mean acceleration an 1.7 m/s2
Jam distance s1 0m
Mean deceleration bn −3.4 m/s2
1
Vehicle length l = 5m Critical space separation s0 2m
ρmax
Safe space separation ss 50 m
Mean vehicle length sn 5m
Original risk factor D0n 0m

TABLE 3 Results Obtained by Each Model After Calibration Trajectories


Average Average Average After an examination of the fundamental diagrams, a closer micro-
Travel Average Flowrate Density scopic look is presented in this subsection. Sample vehicle trajec-
Model Time (s) Speed (m/s) (vph) (vpk)
tories at the same time (minute 55 to 56) in the same freeway
section (second kilometer of the freeway section) are constructed.
GHP-calibrated 29.3 25.6 1,533.1 16.7
Figure 2 shows the results for the seven implemented models. The
Gipps-calibrated 40.0 25.0 1,532.0 17.1 figures show how the GHP models allow vehicles to follow each
IDMM-calibrated 40.0 25.2 1,525.5 17.0 other at high speeds with extremely small space headway. Another
IDM-calibrated 39.9 25.3 1,423.1 15.9 unrealistic behavior is observed in Figure 2c (CA model); as sug-
SK-calibrated 39.4 25.4 1,355.4 14.9 gested in the literature, a vehicle can follow a leader with 0 m
CA-calibrated 40.8 25.6 1,503.0 17.1 separating it from the lead vehicle: it is forced to stop at that loca-
tion by using an unrealistic deceleration rate (safety constraint).
Wiedemann- 41.2 24.2 1,553.5 17.8
calibrated The only two models that allowed more uniform and larger space
headways are the same models that captured traffic instability
Modified Gipps- 34.3 24.1 1,457.6 17.9
calibrated during congestion: the IDM model, the IDMM model, and the
Wiedemann model.
48 Transportation Research Record 2088

2000 2000
1800 1800
1600 1600
1400
Flow (veh/hr)

1400

Flow (veh/hr)
1200 1200
1000 1000
800 800
600 600
400 400
200 200
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Density (veh/km) Density (veh/km)
(a) (b)

2500 2500

2000 2000

Flow (veh/hr)
Flow (veh/hr)

1500 1500

1000 1000

500 500

0 0
0 10 20 30 40 50 0 5 10 15 20 25 30
Density (veh/km) Density (veh/km)
(c) (d)

1800 1600
1600 1400
1400 1200
Flow (veh/hr)

Flow (veh/hr)

1200
1000
1000
800
800
600
600
400 400
200 200
0 0
0 5 10 15 20 25 0 5 10 15 20 25
Density (veh/km) Density (veh/km)
(e) (f)

1800 2000
1600 1800
1400 1600
1400
Flow (veh/hr)

Flow (veh/hr)

1200
1200
1000
1000
800
800
600 600
400 400
200 200
0 0
0 5 10 15 20 0 20 40 60 80 100
Density (veh/km) Density (veh/km)
(g) (h)

FIGURE 1 Flow–density relationship for (a) GHP model, (b) Gipps model, (c) continuous CA model, (d) SK model, (e) IDM model,
( f ) IDMM model, (g) Wiedemann model, and (h) modified Gipps model.
Hamdar and Mahmassani 49

(a) (b)

(c) (d)

(e) (f)

(g) (h)

FIGURE 2 Sample vehicle trajectories for (a) GHP model, (b) Gipps model, (c) continuous CA model, (d) SK model, (e) IDM model,
( f ) IDMM model, (g) Wiedemann model, and (h) modified Gipps model.
50 Transportation Research Record 2088

Accident Modeling Equation 1) with a standard deviation of 0.5 and range of [[ẋn (t) −
ẋn+1 (t)] −1, [ẋn (t) − ẋn+1 (t)] + 1].
Safety constraints imposed in each model are relaxed when the vehi- Although the relaxation is on the order of 0.1, a complete break-
cles traverse the first kilometer of the highway stretch. The parameters down with a chain-type accident is immediately produced (561 acci-
presented in Tables 1 and 2 are kept the same. When an accident is gen- dents). Such unrealistic scenario is due to the little elasticity the
erated (space separation between two consecutive vehicles is less GHP model offers between the particles. The vehicles are not even
than 0), the vehicles involved is the accident decelerate at a maximum able to stop once an accident is generated. Once traffic is disrupted
deceleration rate of 6 m/s2 until they come to a complete stop. The (accident, shockwave), the GHP model is not feasible.
separation between the two consecutive vehicles is then set to zero.
Gipps Model The safety constraint is the Gipps model is presented
GHP Model The main factors that allow an accident-free envi- in the expression x′n−1 − sn−1 > x′n. It indicates that when a driver starts
ronment in the GHP model is the form of the sensitivity term λ and decelerating so that the vehicle will stop at a given location x′n−1, the
the exact relationship between acceleration and relative speed. The following vehicle will decelerate and come to rest at x′n before hit-
relation that governs the GHP models is ting the rear of the preceding vehicle. To relax this condition, the
risk term Dn is subtracted from x′n−1 − sn−1. In this case, even if x′n−1 −
xn +1 ( t + T ) = λ [ x n ( t ) − x n +1 ( t )] (1)
sn−1 − Dn > x′n, the distance between two vehicles can be negative
where and an accident may be generated.
If the safety conditions are kept, the following relations are obtained:
T = reaction time,
.
xn (t) = position of a vehicle n at time t, vn −1 ( t )
2

x n′ −1 = x n −1 ( t ) −
..
x n+1 (t) = position of a vehicle n + 1 following vehicle n at (2)
time t, and 2bn −1
λ = sensitivity term.
τ n vn ( t + τ n )
2

The dots represent differentiation with respect to time t. x n′ = x n ( t ) + [ vn ( t ) + vn ( t + τ n )] − (33)


The sensitivity term λ received the most input in earlier research 2 2bn
(10). It took different functional forms:
and
1. λ = a, that is, a constant.
if s ≤ scritical x n′ −1 − sn −1 ≥ x n′ (4)
⎧⎪a
2. λ = ⎨
⎪⎩b if s > scritical After introducing Dn, Equation 3 will be transformed to
that is, a step function. s is the spacing (xn+1 − xn − ln+1) between two
vn −1 ( t )
2
vehicles. ln is the length of vehicle n. scritical is a threshold specified by τ
the modeler. x n −1 ( t ) − − sn −1 − Dn ≥ x n ( t ) + [ vn ( t ) + vn ( t + τ n )]
2bn −1 2
c
3. λ = vn ( t + τ )
2

s + vn ( t + τ ) θ − (5)
2bn
This form is adopted in the GHP model and is called reciprocal
spacing. c is a constant.
The final expression for the velocity of vehicle n at time t + τn is
c i x n +1
4. λ =
s2
⎧⎪ v (t ) ⎞ ⎛ v (t ) ⎞
12

adopted in Edie’s model. vn ( t + τ n ) = min ⎨ vn ( t ) + 2.5an τ n ⎜ 1 − n ⎟ ⎜ 0.0225 + n ⎟ ;
⎩⎪ ⎝ Vn ⎠ ⎝ Vn ⎠
c
5. λ =
s2 ⎡ 2 ( x n −1 ( t ) − sn −1 − x n ( t )) ⎤ ⎫
⎛τ ⎞ bn2 τ 2n
− bn ⎢⎢ ⎥ ⎪ (6)
bn ⎜ n ⎟ + v (t )
2
leads to the famous macroscopic Greenshield’s flow–density ⎬
⎝ 2⎠ 4 − vn ( t ) τ n − n −1 + Dn ⎥ ⎪
relationship. ⎢⎣ bn −1 ⎥⎦ ⎭
c
It can be seen that having the spacing s in the denominator λ = instead of
s
will reduce the acceleration response tremendously for smaller head-
⎪⎧ v (t ) ⎞ ⎛ v (t ) ⎞
12

ways. Moreover, assuming that the driver will be able to observe and vn ( t + τ n ) = min ⎨ vn ( t ) + 2.5an τ ⎜ 1 − n ⎟ ⎜ 0.0255 + n ⎟ ;
measure exactly the relative speed term, the vehicle will travel at the ⎩⎪ ⎝ Vn ⎠ ⎝ Vn ⎠
same speed of the leader.
To relax this safety condition, the λ term is treated as a random ⎡ 2 ( x n −1 ( t ) − sn −1 − x n ( t ))⎤ ⎫
bn τ + b τ − bn ⎢⎢ ⎥⎪
v (t ) ⎥ ⎬
2 2 2
variable with a normal distribution, a mean of (7)
− vn ( t ) τ − n −1
n

⎢⎣ ⎪
λ mean =
c bn −1 ⎦⎥ ⎭
s
a standard deviation of λstd = 0.1, and a range of λmean − 0.1 ≤ λ ≤ Treating the risk factor Dn as a normally distributed random vari-
λmean + 0.1. This modification did not cause the formation of any able with a mean of 0.1, a standard deviation of 0.1, and a range of
accidents. The accidents were created when treating the relative 0 < Dn < 0.2, 42 accidents were created. Although many fewer acci-
speed stimulus as a normally distributed random variable (see dents were obtained than in the GHP model, chain-type accidents
Hamdar and Mahmassani 51

can still be seen with a relatively high number compared to real-life Accordingly, accidents are obtained by removing this safety buffer.
situations. The logic is still not suited for incident scenarios. A complete traffic breakdown with 1,211 accidents for IDM and
674 accidents for IDMM is observed. Trying to decrease the desired
CA Model The continuous limit of the CA model (17) is defined minimum gap s*(v, Δv) by a value up to 1 did not produce any
as follows: accidents.

vdes = min ⎡⎣ v ( t ) + amax , vmax , sgap ( t ) ⎤⎦ , Wiedemann Model The main mode that prevents the formation
of accidents in Wiedemann model is the emergency braking mode.
v ( t + 1) = max [ 0, vdes − σnran ,0 ,1 ] , It needs to be replaced by a normal mode of deceleration. Moreover,
it was essential to remove a safety buffer term BXadd from the desired
x ( t + 1) = x ( t ) + v ( t + 1) (8)
spacing threshold BX; the desired space threshold ABX that needs
to be followed is dominated by the AX = ABX − BX.
where With all these changes, 17 chain-type accidents are observed.
sgap(t) = free space to the vehicle ahead, Although more resistant to the traffic breakdowns compared to other
amax = maximum acceleration, models, 17 accidents is unrealistically high for such relaxation.
nran,0,1 = random number in the interval (0, 1), and Some of the discussed models are more stable (Wiedemann, Gipps,
σ = maximum deceleration due to the noise. CA) than others (IDM, IDMM, SK, GHP) when relaxing safety con-
straints. However, even when using calibrated parameters for normal
The safety condition is imposed by the term sgap(t). Setting the max- traffic conditions and the smallest degree of relaxation possible,
imum velocity equal to the space gap between two successive vehi- no model allows the observation of one accident where the follow-
cles irrespectively of the required deceleration needs to be changed. ing drivers stop smoothly. One of the aims of the modified model
By allowing the velocity to be equal to sgap(t) + 0.1 m (increase of presented in the next section is to reach this goal.
100 cm), 29 accidents are produced. This relatively low number of
accidents compared to the GHP model and CA Gipps model is due
to the absence of constraints on the deceleration (braking) rate that IMPROVED MODEL FORMULATION
can be applied: unrealistic behavior can still be seen.
The advantage of the Gipps model is in its ability to model driving
SK Model In the SK model, there is a safe-velocity term vsafe that is behavior following some cognitive thinking that may be adopted by
increased by a value of 0.27 m/s (∼1 km/h). However, no accidents the driver. This led to an explicit incorporation of the reaction time
were generated until increasing vsafe by 0.45 m/s (1.62 km/h). Although τn and an asymmetric application of acceleration versus decelera-
it resisted any creation of accidents at the beginning, 2,013 chain-type tion. Moreover, the Gipps model showed an acceptable degree of
accidents are obtained, occupying most of the 10-km highway stretch. stability (relatively low number of accidents) when relaxing its safety
constraints. Motivated by these properties, the objective is to follow
IDM and IDMM Models The IDM and IDMM models assume that
the same logic applied in the Gipps model while modifying it so that
interactions between drivers during high-density situations can be
the acceleration is a continuous function of the velocity va, the gap sa
captured. This includes speed fluctuations causing the creation of
and the velocity difference Δva:
multiple-velocity waves in the traffic flow. In the speed–density rela-
tionship, this will be indicated by observing a metastable state and a
⎡ 2

⎛ v ⎞ ⎛ s* ( vα , Δvα ) ⎞ ⎥
δ
sort of traffic breakdown in the fundamental diagram.
(α ) ⎢ ⎜ ⎟
vα = a − α
⎢ ⎜⎜ (α ) ⎟⎟ ⎜−
⎟ ⎥⎥
1 (9) Moreover, further specifications and a simplified lane-changing
⎢ ⎝ v0 ⎠ ⎝ sα ⎠ ⎦ logic are added so an acceptable incident modeling framework can

be offered. The conditions in which incidents are created are called
This expression can be seen as the integration of two tendencies. “extreme conditions” for the rest of the paper.
The first tendency is to accelerate with
Acceleration Model
⎡ ⎛ v ⎞δ ⎤
a f ( vα ) = a (α ) ⎢
1− α ⎥
⎢ ⎜⎝ v0(α ) ⎟⎠ ⎥
(10) The Gipps model is modified by (a) relaxing some constraints in the
⎣ ⎦ model, such as a safety threshold at the individual driver level (which
may then give rise to accidents or other types of incidents); (b) alter-
As for the second case, when vehicle α comes too close to the leading ing the structure of the equations in the model (by either completely
vehicle, drivers tend to brake with a deceleration of changing the shape of the equations or adding or removing variables);
2
and (c) changing the values of the input variables of the model, as a
⎛ * v , Δv ⎞
bint ( sα , vα , Δvα ) = − a
( α α )⎟
(α ) ⎜ s
(11)
way of representing new traffic situations in different locations.
⎜ ⎟ The model aims to capture driver behavior under incident or
⎝ sα ⎠ extreme conditions and congested situations. The modifications that
can be made on the variables included in the simulation model
The desired gap s* is set by the following equation: (the rest already described in the original Gipps model) as well as
the description of the initial risk factor D0n are as follows:
v vΔv
s* ( v, Δv ) = s0(α ) + s1(α ) + T (α ) v + (12)
v0(α ) 2 a(α ) b(α ) 1. an (m/s2). Under extreme conditions, drivers typically can be
willing to apply higher acceleration rates than under normal condi-
In these two models, the last term of Equation 12 is responsible of pre- tions, causing irregularities and possible instabilities in traffic flow
venting crashes although the safe time headway is already included. patterns (18). This variable is drawn from a truncated Gaussian-
52 Transportation Research Record 2088

shaped (normal) distribution with a given mean and variance. The The new relationship dominating the Gipps logic is
truncation is performed through a range variable and is based on the
⎧⎪ v (t ) ⎞ ⎛ v (t ) ⎞
12
value of the mean chosen during the sensitivity analysis. The main goal ⎛
for this truncation is to deal with negative values. Although drivers vn ( t + τ n ) = min ⎨ vn ( t ) + 2.5an τ n ⎜ 1 − n ⎟ ⎜ 0.0225 + n ⎟ ;
⎩⎪ ⎝ Vn ⎠ ⎝ Vn ⎠
may act chaotically under extreme conditions, there may still be a
⎡ 2 ( x n −1 ( t ) − sn −1 − x n ( t )) ⎤ ⎫
distribution describing the variation of this behavior across drivers.
2. bn (m/s2). The value of bn can increase in absolute value. The ⎛τ ⎞ bn2 τ 2n
− bn ⎢⎢ ⎥ ⎪ (14)
bn ⎜ n ⎟ + v (t )
2

hypothesis is that under extreme conditions, drivers tend to have higher ⎝ 2⎠ 4 − vn ( t ) τ n − n −1 + Dn ⎥ ⎪
braking rates or increased use of emergency braking (19). This value ⎢⎣ bn −1 ⎥⎦ ⎭
is also drawn from a truncated normal distribution with a given mean,
variance, and range at the beginning of the simulation.
3. Vn (m/s) is the speed at which the driver of vehicle n wishes to Lane-Changing Decision Logic
travel. In extreme conditions, the value can be randomly chosen from
a probabilistic mixture of two normal distributions. For the first dis- In addition to the car-following model, Gipps offers another model
tribution, the mean is higher than the suggested mean in the Gipps that explains the structure of lane-changing decisions. Although
model. For the second distribution, the mean is lower than the sug- well detailed, his model discusses complex objectives behind lane-
gested Gipps mean. This choice is consistent with an illustration by changing behavior that do not apply to the basic situation of this study
Daganzo (20) of the disruptions and the irregularities in traffic flow (21). Moreover, lane changing is based on the gaps offered by traffic
resulting from velocity differentials (idealized as two classes of in the adjacent lanes. Accordingly, accepting these gaps will be related
drivers, so-called slugs versus rabbits). It is also consistent with the to the relative speed and acceleration of both the leading and the lag-
description of human psychology when facing a threat: drivers become ging vehicle in the adjacent lane. Gipps’ model takes into account only
either lost (slugs) or more aggressive (rabbits), causing velocity the properties of the leading vehicle.
fluctuations (20). It should be noted that the relative composition of In this study, a driver’s logic for changing lanes is based on the
the driver population into each of the two types is itself a parameter answers to the three following questions:
reflecting a particular incident situation and is a worthwhile sub-
ject of investigation. Furthermore, the two-class representation is • Is it desirable to change lanes?
only a simplified representation of a richer population mix with many • Is it possible to change lanes?
underlying classes. • Is it necessary to change lanes?
4. D0n (m). An initial risk factor, it represents the distance a
driver is willing to travel beyond the safety threshold. The safety Lane-changing decisions are strongly related to the speed at which
threshold indicates the distance between the driver and the leading a driver wishes to travel. A driver traveling at a speed less than his
vehicle at which the driver would start decelerating so that the vehi- desirable speed will seek to increase his speed in the same lane. If
cle can come to a complete stop before hitting the preceding vehi- another vehicle is in the way (space headway between the two vehi-
cle. This value is added to the model to allow potential accidents to cles is less than 5 m, which is the average length of a car), the follow-
be generated. It reflects the willingness of a driver to take a risk. The ing driver will consider changing lanes. However, the driver must
value of Dn for each vehicle n is initially drawn from a truncated nor- check first if this maneuver is possible with the gaps offered in the
mal distribution. When this value is positive, the driver is willing to adjacent lane. Checking these gaps is a procedure to be specified as
take a risk, and this may increase the probability of causing an acci- part of the lane-changing model.
dent. If this value is negative, the driver prefers to stay within the Conversely, it was found that the average lead or lag times for all
safety margin so that he or she can come to a stop without hitting the traffic conditions are almost equal (22). Accordingly, it may be sug-
vehicle in front. gested that neither the lead nor the lag dominates the gap-acceptance
5. sn (m) is size of vehicle n instead of its effective size. This can decision in lane changing. Therefore, both the leading and the lagging
be drawn from a normal distribution or “mu,” which consists of its vehicles in an adjacent lane are objects of interest in this study.
physical length plus a margin (headway) into which the following The theoretical estimate of the minimum safe lead value based an
vehicle is not willing to intrude, even when at rest. assumed desirable deceleration rate and an average braking perception
6. n (s) is the reaction time corresponding to the driver of vehi- and reaction time is given by the following equation:
cle n. It is not constant for all vehicles as it was indicated in the
original Gipps model. vn ( t + τ n ) v (t + τ n )
2 2

L1 = vn ( t ) τ n + − m (15)
2 bn 2 bm
By using the same logic adopted for the Gipps model to relax the
safety constraint, the relationship of x′n−1 − sn−1 − Dn > x′n will allow
where
the distance between two vehicles to be negative, and an accident
may be generated. Moreover, another safety factor illustrated by the m =
subscript for a leading vehicle in destination lane,
safety margin vn (t + τ)θ in Equation 5 (θ is normally equal to τ/2) is L1 =
safe lead distance for lane changing (m),
removed. After introducing Dn (function of D0n), Equation 5 will be vn(t + τn) =
speed of lane-changing vehicle n (m/s),
transformed to vm(t + τm) =
speed of leading vehicle m in destination lane (m/s),
bn =
deceleration rate vehicle n can sustain (m/s2),
vn −1 ( t )
2
τ bm =
deceleration rate vehicle m can sustain (m/s2),
x n −1 ( t ) − − sn −1 − Dn ≥ x n ( t ) + [ vn ( t ) + vn ( t + τ n )] n τn =
apparent reaction time for vehicle n (braking perception
2bn −1 2
and reaction time, s), and
vn ( t + τ n )
2
τm = apparent reaction time for vehicle m (braking percep-
− (13)
2bn tion and reaction time, s).
Hamdar and Mahmassani 53

With the same logic, the theoretical estimate of the safe lag value is immediate surrounding: deceleration (braking) rate bn and risk fac-
tor Dn are considered functions of the vehicle’s separation (distance
vm+1 ( t + τ m+1 ) v ( t + τ m+1 )
2 2
separating the front end of a lagging vehicle to the rear end of the
L2 = vm+1 ( t ) τ m+1 + − n (16) leading one). Moreover, alert distance Rn depends on the initial
2 bm+1 2 bn
risk factor D0n. Since literature offers general and not detailed
descriptions of how these characteristics may change, the functional
where relations are assumed to be piecewise linear; they are presented
m+1 = subscript for a lagging vehicle in destination lane, in Figure 3.
L2 = safe lag distance for lane changing (m), In Figures 3a and 3b, Critical_Separation is equal to the aver-
vm+1(t + τm+1) = speed of lagging vehicle m + 1 in destination lane age vehicle length, which is 4 m. It is the space separation below
(m/s), which drivers will tend to react exaggeratedly to any stimulus (9),
bm+1 = deceleration rate vehicle m can sustain (m/s2), and including use of the maximum deceleration rate a vehicle can sup-
τm+1 = apparent reaction time for vehicle m + 1(braking port (bn = −6 m/s2). Below this separation, drivers are not willing
perception and reaction time, s). to take any risk (Dn = 0 if the original risk factor D0n was greater
than zero), but this may be too late for a collision. However, if the
However, it is suggested that both lag and lead distances are over- original risk factor was greater than zero (D0n < 0: drivers already
estimated (22). The use of different parameters during extreme con- conservative), the risk factor always will be equal to the original risk
ditions will help deal with this subject. First, the higher deceleration factor and will not depend on the space separation. Safe_Separation
rates in absolute value will decrease the safe leads and lags to be is equal to 10 average vehicle lengths (40 m), and it is seen as the
accepted. This is expected during panic behavior, especially on the distance separation above which drivers tend not to take extra pre-
part of aggressive drivers, since their patience is limited and they tend cautions for preventing collisions: the initial deceleration rate bn
to accept shorter gaps. and the initial risk factor D0n already drawn for normal distributions
are used.
In Figure 3c, the alert distance (alertness of the drivers) is seen
Incident Modeling negatively correlated to the initial risk factor (aggressiveness). The
functional relationship is also assumed to be linear.
The preceding model faces several limitations in its ability to repre- The preceding model is rich in behavioral driving parameters
sent car-following behavior during incidents, especially regarding offering flexibility and model dynamics to imitate real-life cognitive
uncontrollable chain-type accidents. Note every time the headway driver behaviors.
between two vehicles is less than zero, the speed of both vehicles
will decrease at a rate of −6 m/s2, the maximum deceleration a vehi-
cle can sustain. The vehicles will come to a rest at the end. The head- ANALYSIS OF RESULTS
way of the two vehicles will be reset to zero governed by the location
of the front vehicle. Incident-Free Environment
The accidents still will be possible in this lane-changing model
because of the duration required for the lane-changing maneuver. NGSIM data described earlier are used to calibrate the acceleration
The general idea is that if the lane-changing maneuver is seen as model presented in the previous section (improved model formula-
possible, the respective locations of vehicles n, m, and m + 1 are com- tion). For comparison, the same simulation environment is created
puted in the target lane after a given lane-changing time. This time here. Calibrated parameters are given in Table 2. The macroscopic
is also drawn from a normal distribution varying from one driver to results of this calibration are presented in Table 3.
another. Moreover, to capture sudden lane changing (19) to escape an
incident, lane-changing time is reduced to have a mean of 2 s, a value
that is found feasible by sensitivity analysis. If xm(t) − xm+1(t) − sm Flow–Density Relationship
is less than or equal to sn, the respective velocities of the three vehi-
cles are set to be equal to zero in the same manner as for car fol- The fundamental diagram corresponding to the modified Gipps
lowing. This indicates the occurrence of an accident in that lane. It model is presented in Figure 1h. Congestion patterns were captured
may be suggested that accidents due to lane changing may block both when the flows reach a value of 1,600 vph. At that time, two states
lanes of travel. are seen: the first state is a slow traffic movement state (metastable
To prevent traffic breakdown quickly, once an accident occurs state) until hitting the flow capacity bar of 1,900 vph. The other
another vehicle-specific variable enters into the equation. This vari- state is the traffic breakdown state, located under the slow-moving
able is called “alert distance” (Rn). If the vehicle is within an alert vehicle line. This kind of shape has some aspects in common with
distance from the incident location, the risk of this vehicle will be reset the hysteresis loop.
to zero. In other words, drivers will tend to be more attentive and may
slow down when they see an accident and for some time after they
pass it (18); even if on the side of the road, accidents have an effect Trajectory Data
on traffic conditions.
As in the section on numerical assessment, sample vehicle trajecto-
ries at the time duration (minute 55–minute 56) in the freeway sec-
Parameters Change over Time tion (kilometer 1–kilometer 2) are constructed. The model was able
to capture different velocity waves and irregularities in the interaction
The change in some driver characteristics over time is performed between drivers. This conclusion is reached in comparing Figure 2h
through linking them to dynamic traffic properties in the driver’s to the rest of the trajectories.
54 Transportation Research Record 2088

Risk or Dn (m)

Original Risk:
If Original Risk or Dn 0 ≥ 0 Dn0 (m)

Separation (m)
Safe_Separation
Critical_Separation

Risk or Dn (m)

If Original Risk or Dn 0 ≥ 0

Separation (m)
Original Risk:
Dn0 (m)
(a)

Deceleration or bn
(m/s2)
Safe_Separation
Critical_Separation

Original Separation (m)


Deceleration:
2
bn0 (m/s )

Maximum
Deceleration: B
(m/s2)
(b)

Alert Distance or Rn

1500

15 Original Risk
or Dn0 (m)
(c)

FIGURE 3 Parameters change over time with (a) D n  f 1 (separation), (b) b n  f 2 (separation), and (c) R n  f 5 (D n0 ).

Extreme Conditions sive drivers mentioned earlier. The others are the slower slug drivers.
Moreover, in Figure 4b, the risk factor is no longer zero. Accordingly,
Since there are no field data to calibrate or validate this model the horizontal lines the last until the end of the simulation indicate that
with the presence of incidents, the model relies on sensitivity analy- an accident has occurred. Some of the vehicles will be stuck behind
sis to examine the feasibility of the logic used. The initial input these accidents before they can change lanes. Accordingly, chain-
parameters and the different scenarios performed are presented in type accidents can be avoided by just waiting behind the vehicles
Tables 4 and 5. involved in the crash.
At the beginning, the trajectories of 50 sample vehicles were dressed After talking about the microscopic trajectory data, the macro-
for one simulation run using Scenarios 1 and 3. Figure 4 shows these scopic data of average travel times, average speeds, total number of
trajectories in the form of time–space diagrams. lane changes, total number of vehicles crashed, average headway,
The first set of trajectories shows two families of drivers; those with average flows, and average densities (more than 10 simulation runs)
steeper velocity lines possess higher velocities and thus are the aggres- are shown in Table 6.
Hamdar and Mahmassani 55

TABLE 4 Main Input Parameters Used in Simulation TABLE 5 Scenarios Adopted in Sensitivity Analysis
of Extreme Events
Mean
Input Variable Value # of Interarrival Risk
Scenario Description Scenario # Vehicles Time (s) (m)
Time T (s) 7,200
Road length L (m) 10,000 Free-flow risk free 1 1,200 6 0
Mean reaction time (s) 1 Free-flow minor risk 2 1,200 6 1
Mean Vd1 (m/s): mean desired velocity for slugs 20 Free-flow major risk 3 1,200 6 5
Mean Vd2 (m/s): mean desired velocity for rabbits 35 Free-flow break down 4 1,200 6 10
Percent Vd1(%): percent of slug drivers 40 Congested risk free 5 3,600 2 0
Mean LCT (s): mean of the lane changing time 2.5 Congested minor risk 6 3,600 2 1
Mean acceleration (m/s2) 2 Congested major risk 7 3,600 2 5
Mean deceleration (m/s2) −3 Congested break down 8 3,600 2 10

It was found that the variance of the speed will increase with the risk CONCLUSION
factor; this is expected since accidents will bring an increasing num-
ber of traffic fluctuations. In terms of the means, with higher risks, acci- This paper presented a detailed analysis of existing car-following
dents, travel times, and lane changes will increase. However, speeds models with an assessment of their qualities and limitations. The
and flows decrease. This special rule is sometimes violated in the focus was on seven models: GHP, Gipps, CA, SK, IDM, IDMM,
traffic breakdown situation, because most vehicles are trapped at the and Wiedemann. These models were implemented, calibrated, and
beginning of the road length because of the accident. Few vehicles will tested for vehicle trajectories, flow–density relationships, and ability
escape and can thus travel at higher speeds and with higher headways. to model driver behavior during incident situations.
This lower number of vehicles is reflected in the low density value for This kind of analysis allowed the formulation of an improved car-
Scenario 8 compared to Scenario 7 (Table 6). following model that shows instability during congestion without the
Table 7 focuses on Scenario 3 and how vehicles escape and travel necessity of an accident-free environment. The modified model could
between accidents, avoiding the unrealistic chain effect observed in capture some congested flow dynamics (hysteresis effect, beginning
and discussed earlier. of traffic breakdown). Moreover, when incorporating the space risk

12,000

10,000
Position (m)

8,000

6,000

4,000

2,000

0
0 2,000 4,000 6,000 8,000 10,000
Time (Multiple of 0.1 s)
(a)

12,000

10,000
Position (m)

8,000

6,000

4,000

2,000

0
0 2,000 4,000 6,000 8,000 10,000 12,000
Time (Multiple of 0.1 s)
(b)

FIGURE 4 Time–space diagram for one run for (a) Scenario 1 and (b) Scenario 3.
56 Transportation Research Record 2088

TABLE 6 Output Data Corresponding to Scenarios Adopted in Sensitivity

Performance Measure Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Scenario 7 Scenario 8

Average travel time (s) 54.9 46.7 48.9 50.9 60.1 59.9 74.0 49.2
Average speed (m/s) 18.3 27.4 25.4 22.8 16.7 27.8 23.7 22.4
Number of cars crashed 0.0 14.0 24.0 161.0 0.0 14.0 39.0 108.0
Number of lane changes 74.0 8,733.0 11,753.0 3,247.0 268.0 37,162.0 50,542.0 995.0
Average flow rate (vph) 291.3 290.0 287.4 43.7 874.9 822.8 842.4 15.9
Average density [veh/(lane km)] 4.5 4.7 5.2 8.1 14.6 15.2 19.6 6.1

TABLE 7 Avoidance of Chain-Type Accidents: 4. Gazis, D. C., R. Herman, and R. Potts. Car-Following Theory of Steady
Escape by Lane Changing (Scenario 3) State Traffic Flow. Operations Research, Vol. 7, 1959, pp. 499–505.
5. Gipps, P. G. A Behavioral Car-Following Model for Computer Simulation.
Number of Transportation Research B, Vol. 15, 1981, pp. 101–115.
Vehicles Crashed 6. Nagel, K., and M. Shreckenberg. A Cellular Automaton Model for
Road Segment Number of Freeway Traffic. Journal of Physics I, Vol. 2, 1992, pp. 2221–2229.
Index (1 km) Lane 1 Lane 2 Lane Changes 7. Krawss, S., and P. Wagner. Metastable States in a Microscopic Model of
Traffic Flow. Physical Review E, Vol. 55, No. 5, 1997, pp. 5597–5602.
1 0 2 1,023 8. Treiber, M., and D. Helbing. Memory Effect of Microscopic Traffic
2 2 2 1,575 Models and Wide Scattering in Flow-Density Data. Physical Review E,
Vol. 68, 2003, PDF 046119.
3 0 2 1,086 9. Wiedemann, R., and U. Reiter. Microscopic Traffic Simulation, the
4 2 0 1,135 Simulation System Mission. PhD dissertation. University of Karlsruhe,
5 2 0 1,022 Germany, 1991.
10. Gazis, D., R. Herman, and R. Rothery. Nonlinear Follow-the-Leader
6 0 2 1,193 Models of Traffic Flow. Operations Research, Vol. 9, 1961, pp. 545–567.
7 0 2 1,036 11. Newell, G. F. Nonlinear Effects in the Dynamics of Car Following.
8 2 2 1,363 Operations Research, Vol. 9, No. 2, 1961, pp. 209–229.
12. Bando, M., K. Hasebe, A. Nakayama, A. Shibata, and Y. Sugiyama.
9 2 0 1,128
Dynamical Model of Traffic Congestion and Numerical Simulation.
10 0 2 1,192 Physical Review E, Vol. 51, 1995, pp. 1035–1042.
13. Tilch, B., and D. Helbing. Generalized Force Model of Traffic Dynamics.
Physical Review E, Vol. 58, No. 133, 1998.
14. Cambridge Systematics. NGSIM Task E.1-1: Core Algorithms Assessment.
factor, drivers showed an acceptable degree of maneuverability either FHWA, U.S. Department of Transportation, 2004.
by standing behind accidents without being involved in a rear-end col- 15. Brockfeld, E., R. D. Kuhne, and P. Wagner. Calibration and Validation
lision or by changing lanes. Long and unrealistic chain-type accidents of Microscopic Traffic Flow Models. In Transportation Research Record:
causing a total traffic breakdown could be avoided. Journal of the Transportation Research Board, No.1876, Transportation
Research Board of the National Academies, Washington, D.C., 2004,
The results suggest that the specifications presented in this paper
pp. 62–70.
offer an improved basis for microscopic traffic simulation for situ- 16. Brockfeld, E., R. D. Kuhne, and P. Wagner. Calibration and Validation
ations that do not require an accident-free environment. As such, the of Microscopic Models of Traffic Flow. In Transportation Research
same basic behavior model structure could accommodate both extreme Record: Journal of the Transportation Research Board, No.1934,
situations (evacuation scenarios, oversaturated networks) as well as Transportation Research Board of the National Academies, Washington,
D.C., 2005, pp. 179–187.
normal daily traffic conditions.
17. Krauss S., P. Wagner, and C. Gawron. Continuous Limit of Nagel-
Shreckenberg Model. Physical Review E, Vol. 54, No. 4, 1996,
pp. 3707–3712.
ACKNOWLEDGMENT 18. Todosiev, E. P. The Action-Point Model of the Driver-Vehicle-System.
Ohio State University, Columbus, 1963.
This study is based in part on research funded by the National Science 19. Querejeta-Iraola, A., and U. Reiter. Calibration, Validation and Testing
Foundation’s Human and Social Dynamic Systems. of Multi-Lane Simulation Model. EC DRIVE Project ICARUS (V-1052),
Brussels, Belgium, 1991.
20. Daganzo, C. F. A Behavioral Theory of Multi-Lane Traffic Flow, Part I:
Long Homogeneous Freeway Sections. Institute of Transportation Studies,
REFERENCES University of California, Berkeley, 1999.
21. Gipps, P. G. A Model for the Structure of Lane Changing Decisions.
1. Rothery, R. W. Traffic Flow Theory: A State-of-the-Art Report: Revised Transportation Research B, Vol. 20, 1986, pp. 403–414.
Monograph on Traffic Flow Theory. Transportation Research Board, 22. Lane Changing on Multi-Lane Highways. FHWA, U.S. Department of
National Research Council, Washington, D.C., 1999. Transportation, 1969.
2. Boer, E. R. Car Following from the Driver’s Perspective. Transportation
Research F, Vol. 2, No. 4, 1999, pp. 201–206. The authors are responsible for the contents of this paper.
3. Treiber, M., K. Hennecke, and D. Helbing. Congested Traffic States in
Empirical Observations and Microscopic Simulations. Physical Review E, The Traffic Flow Theory and Characteristics Committee sponsored publication of
Vol. 2, No. 2, 2000, pp. 1805–1824. this paper.

View publication stats

You might also like