Does the Intelligent Driver Model Adequately Represent Human

Drivers?

Zeyu Mu, Fatemeh Jahedinia and B. Brian Park

Link Lab and Department of Engineering Systems and Environment, University of Virginia,

Keywords:

Intelligent Driver Model, Human Drivers, Microscopic Simulation, Mixed Trafﬁc Evaluation.

Abstract:

The Intelligent Driver Model (IDM) is one of the widely used car-following models to represent human drivers

in mixed trafﬁc simulations. However, the standard IDM performs too well in energy efﬁciency and comfort

(acceleration) compared with real-world human drivers. In addition, many studies assessed the performance

of automated vehicles interacting with human-driven vehicles (HVs) in mixed trafﬁc where IDM serves as

HVs based on the assumption that the IDM represents an intelligent human driver that performs not better

than automated vehicles (AVs). When a commercially available control system of AVs, Adaptive Cruise

Control (ACC), is compared with the standard IDM, it is found that the standard IDM generally outperforms

ACC in fuel efﬁciency and comfort, which is not logical in an evaluation of any advanced control logic with

mixed trafﬁc. To ensure the IDM reasonably mimics human drivers, a dynamic safe time headway concept is

proposed and evaluated. A real-world NGSIM data set is utilized as the human drivers for simulation-based

comparisons. The results indicate that the performance of the IDM with dynamic time headway is much closer

to human drivers and worse than the ACC system as expected.

1 INTRODUCTION

A human driver model is a key component in rep-

resenting driver characteristics in microscopic trafﬁc

simulations. This is because the assessment of vehi-

cle safety, energy efﬁciency, and trafﬁc characteristics

is highly associated with the reliability of the human

driver models. Generally, the car-following model of

human drivers estimates the velocity/acceleration ac-

cording to the velocity/acceleration of its preceding

vehicle, and the driver characteristics is represented

by standard parameters or calibrated parameters

based on real-world data. The commonly used car-

following models are Gazis-Herman-Rothery (GHR)

model (Gazis et al., 1959), Optimal Vehicle Model

(OVM) (Bando et al., 1995) and Intelligent Driver

Model (IDM) (Treiber et al., 2000). These models can

effectively capture drivers’ basic behaviors without

requiring much load on the mathematical framework.

Due to the collision-free characteristics and mathe-

matical efﬁciency of the IDM, many studies have em-

ployed the IDM to represent the car-following behav-

iors of human drivers in microscopic trafﬁc simula-

tion, especially mixed trafﬁc simulation. To utilize

the IDM as human drivers, the IDM is supposed to

describe the microscopic dynamics of the individual

drivers as well as macroscopic aspects of trafﬁc ﬂow.

However, the complexity and uncertainty of human

drivers present a great challenge for IDM to capture

their characteristics.

The logic behind the IDM is to model the reac-

tions of a human driver to his/her preceding vehi-

cle’s motions (e.g., speed and distance) correspond-

ing to the desire to achieve the desired speed, mean-

while, keeping a safe gap from the preceding vehi-

cle and collision precaution of human drivers in re-

ality. The characteristics of drivers are deﬁned by

parameters, e.g., maximum acceleration/deceleration,

desired speed, safe time headway, and minimum gap,

which can be calibrated with empirical data or real-

world data. Many studies (Jiang et al., 2017; Bhat-

tacharyya et al., 2020; Sharma et al., 2021; Mu et al.,

2022) employ the IDM modeling homogeneous or

heterogeneous human drivers to evaluate the micro-

scopic or macroscopic characteristics of automated

vehicles (AVs) in the mixed trafﬁc environment, e.g.,

fuel efﬁciency, driving comfort, trafﬁc capacity and

stability. Generally, the IDM is utilized to model tra-

ditional human-driven vehicles interacting with con-

nected automated vehicles and connected human-

Mu, Z., Jahedinia, F. and Park, B.

Does the Intelligent Driver Model Adequately Represent Human Drivers?.

DOI: 10.5220/0011996800003479

In Proceedings of the 9th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2023), pages 113-121

ISBN: 978-989-758-652-1; ISSN: 2184-495X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

113

driven vehicles. In terms of the interaction with au-

tomated vehicles in trafﬁc, (Rahman and Abdel-Aty,

2018; Wang et al., 2019; Sun et al., 2020; Guo and Jia,

2021; Ding et al., 2022) evaluated the performance of

trafﬁc to show the impact of penetration of connected

automated vehicles (CAVs) in energy efﬁciency in

mixed trafﬁc. The IDM parameters refer to the cal-

ibrated parameters of real-world human drivers. For

interactions with connected human-driven vehicles,

(Rahman and Abdel-Aty, 2018; Sharma et al., 2019;

Jiang et al., 2019; Zhang et al., 2020) evaluated the

safety and mobility of connected vehicles in mixed

trafﬁc with connected human-driven vehicles. The

IDM was used to represent human driving behaviors

of connected vehicles, which can receive other con-

nected vehicles’ motion information. However, the

limitation of these studies is that neither of them eval-

uated or compared the performance of each control

mode under the mixed trafﬁc of human drivers and

automated vehicles. Even though the capability of

IDM to represent the stochastic human drivers’ car-

following is not the focus of their research, it is impor-

tant for researchers to use a model that can represent

the human drivers for reliable assessment.

The most common method of employing IDM

serving as human drivers is calibrating parameters

with real-world human-driven vehicle data. Various

calibration methods have been investigated, e.g., least

squared errors or maximum likelihood, to ﬁnd the

range or the distribution of these parameters to sim-

ulate different drivers’ characteristics (Kesting and

Treiber, 2008; Stern, ; Ro et al., 2018; Hegde et al.,

2021). However, these studies focus on identifying

the most ﬁtting parameters in a long time driving and

do not consider that each driver’s characteristics could

change over time. To overcome these limitations, on-

line estimation methods are utilized to capture the dy-

namics of individual drivers by utilizing real-time in-

formation to update those model parameters. (Bhat-

tacharyya et al., 2020). However, the online esti-

mation method is more about an estimation of hu-

man driver behaviors online instead of modeling hu-

man driver behaviors in trafﬁc simulation, consider-

ing the estimated parameters cannot be used to regen-

erate human behaviors for new simulations since the

estimated parameters are highly associated with the

preceding vehicles’ behaviors, which could change in

different simulations.

To improve IDM model applications, many re-

searchers (Bhattacharyya et al., 2020; Kesting et al.,

2010; Eggert et al., 2015; Yi et al., 2020) seek to in-

corporate realistic human driver features. The En-

hanced IDM (Kesting et al., 2010) presents an im-

proved IDM application for safety by preventing the

model from over-reactions even when the driver of

the preceding vehicle suddenly brakes with the max-

imum possible deceleration. The Foresighted Driver

Model (FDM) (Eggert et al., 2015) extends the IDM

by assuming that a driver balances the risk of possi-

ble collisions with travel time and the smoothness of

the ride. The models mentioned above focus on im-

proving the utilization of the IDM considering safety

when human drivers control the vehicles, while the

limitations are also their emphasis on the safety of

IDM aiming at generating smooth acceleration. How-

ever, those models neglect the fact that human drivers

actually could have aggressive behaviors. It is appar-

ent that few research investigated the uncertainties of

IDM parameters and their impacts on human driving

behaviors.

The objective of this research is to assess whether

the IDM adequately represents human driving behav-

iors. To achieve this objective, the performance be-

tween the IDM with calibrated parameters and real-

world human drivers is to be compared. In addition,

since ACC is one of the most commonly used au-

tomated control for AVs, which should have better

driving comfort and fuel economy than IDM for hu-

man drivers, the calibrated IDM is to be compared

with the ACC system. If the IDM does not repre-

sent human driving behaviors, this research is to con-

duct experimental design-based evaluations to iden-

tify key parameters affecting the performance of the

IDM and explore the feasibility of adjusting IDM pa-

rameters to adequately represent human driving be-

haviors. The rest of the paper is organized as follows.

Section 2 presents the car following models, includ-

ing the IDM and ACC, and how they are to be im-

plemented in this paper. Section 3 discussed the real-

world human driving data and the efforts given to cal-

ibrate the IDM and the comparison results among the

human drivers, IDM and ACC. In section 4, the im-

pacts of the IDM parameters are evaluated through an

experimental design covering all IDM parameters, an

attempt is made to choose an IDM parameter that can

best represent human driving behaviors, and the pro-

posed IDM model is validated using a NGSIM data

that is not used in the calibration. Finally, section 5

summarizes the ﬁndings of this research and discusses

future research.

2 CAR-FOLLOWING MODELS

2.1 Intelligent Driver Model

The standard Intelligent Driver Model (IDM) is a

deterministic car-following model that describes the

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

114

dynamics of a human-driven subject vehicle by es-

timating its acceleration with respect to its speed,

speed difference, and the gap from the vehicle ahead.

The speed difference introduces additional caution to

make IDM crash-free. Five parameters, including

safe time headway, desired speed, maximum accel-

eration, maximum deceleration, and minimum dis-

tance, represent the car-following characteristic of the

driver, which can be calibrated with empirical or real-

world data. The estimated acceleration of the subject

vehicle is calculated by Eq. (1) and Eq. (2).

(t) = a



1 −(

(t)

)

−(

△ d

s,d

,△ v

,t)

△ d

(t)

)



(1)

△ d

s,d

= s

+ max



(t) ∗T +

(t)· △ v

(t)

√

, 0



(2)

Where △ d

(t) = d

(t) −d

(t) is the gap between

the subject vehicle’s (s) and its preceding vehicle (p)

at time t. △ v

(t) = v

(t) −v

(t) is the relative speed

difference of the subject vehicle and its preceding ve-

hicle at time t. d

(t), v

(t), a

(t) are the states rep-

resenting the position, speed, and acceleration of the

subject vehicle at time t, respectively. △ d

s,d

is the

desired bumper to bumper gap and △ d

(t) is the ac-

tual bumper to bumper gap at time t. The ﬁve pa-

rameters are interpreted as follows. v

is the desired

speed of the driver driving in free ﬂow. a is the max-

imum acceleration. b represents a comfortable decel-

eration. s

is the minimum inter-vehicle gap that the

driver prefers to maintain at the stop. T is the safe

time headway.

2.2 Adaptive Cruise Control

Adaptive Cruise Control (ACC) (Vahidi and Eskan-

darian, 2003) is a commercially available advanced

driver assistance system of longitudinal control that

is designed for autonomous vehicles and aims at im-

proving safety, driving comfort, energy economy, and

trafﬁc ﬂow(Marsden et al., 2001). The ACC systems

utilize the measured motion of the preceding vehicle

and control the subject vehicle to maintain a safe gap.

A proportional-derivative (PD) controller is utilized

for the ACC systems in this study. The control input

(t) can be written with respect to the spacing error

e(t):

u(t) = k

e(t) + k

˙e(t) (3)

where k

, and k

are proportional and derivative gains

of the controller, respectively. The spacing error

e(t) = d

(t) −d

(t) is the difference between the de-

sired gap of the subject vehicle from the immediately

preceding vehicle d

and the actual gap d(t) at time

t. The desired gap d

(t) = v

(t) ·T

+ d

is calcu-

lated by the desired time headway T

, current speed

and standstill distance d

. The low-level controller

is modeled by a ﬁrst-order lag τ to the acceleration

command u(t) and vehicle acceleration a(t) :

˙a(t) = −

a(t)

u(t)

(4)

In this study, we adopted system parameters, k

0.7,k

= 0.5, and τ = 0.3.

3 HUMAN DRIVERS DATA AND

IDM CALIBRATION ISSUE

3.1 NGSIM Dataset and Pre-Processing

To understand microscopic car following behaviors

of human drivers, the freeway US-101 data set from

the Next Generation Simulation (NGSIM) (Admin-

istration, 2017) was utilized. The data set consists

of about 2000 vehicles’ trajectories on ﬁve lanes ob-

served within a short distance (roughly 640 m) and for

the ﬁrst 15 minutes of the dataset, and it reﬂects dense

highway ﬂow, the transition between uncongested and

congested conditions, as well as full congestion. In

this study, we only utilized the data from uncongested

conditions to analyze car-following behaviors.

This NGSIM trajectory data includes the position

and speed proﬁle of vehicles at a 0.1-second time in-

terval. Due to the propagation of the measurement er-

ror in the speed proﬁles, considerable noise (i.e., unre-

alistic jerks) in acceleration could be generated when

derived from the speed proﬁles of vehicles. There-

fore, in this study, the locally weighted scatterplot

smoothing (LOWESS) is applied to the speed proﬁles

of vehicles, and the size of the sliding window is cho-

sen as 2s. After smoothing, the speed proﬁles of ve-

hicles are less noisy, and the jerks are always below

15m/s

, which is more mechanically realistic (Punzo

et al., 2011).

3.2 Intelligent Driver Model

Calibration Issue

Several IDM calibration methods (Chen et al., 2010;

Ciuffo et al., 2014; Bhattacharyya et al., 2020; Al-

hariqi et al., 2022) to estimate the IDM parameters

were used in many studies where the parameter es-

timation employed optimization techniques to mini-

mize the error between the simulated and measured

output. Most studies focused on ﬁnding the constant

parameters of IDM to minimize the speed and gap er-

rors from human drivers. One limitation is that most

Does the Intelligent Driver Model Adequately Represent Human Drivers?

115

studies showed the error or how much improvement

the error is with the proposed method. In contrast,

limited studies show the actual performance compar-

ison of human drivers and IDM. Besides, the com-

plexity and uncertainty of human drivers’ character-

istics are neglected. To make a fair comparison with

other studies, the cost function which is the sum of the

square error of the simulated speed proﬁles and time

headway proﬁles of IDM from that of the actual hu-

man drivers is minimized during the IDM parameters

estimation.

The cost function which is related to the simulated

speed v

sim

, actual speed from data v

data

, the simu-

lated time headway T

sim

, actual time headway from

data T

data

, is deﬁned in Eq.5

f (v

sim

) =

∑

t=0

[(v

sim

(t) −v

data

(t))

+(T

sim

(t) −T

data

(t))

]

(5)

min

sim

f (v

sim

)

s.t. T

min

≤ T ≤ T

max

dmin

≤ v

dmax

0min

≤ s

0max

min

≤ a ≤ a

max

min

≤ b ≤ b

max

(6)

Each parameter of the IDM is constrained by its max-

imum value and minimum value. The safe time head-

way, T , is in range of (0.8, 2) m; the desired speed,

, is in range of (20, 28) m/s; the maximum accel-

eration/deceleration, a/b, is in range of (1, 5) m/s

and the minimum distance, s

, is within (1, 5) m. The

optimization problem, Eqs. 5 and 6, is solved by ap-

plying a trust-region reﬂective least squares algorithm

with constraints (Coleman and Li, 1996). The algo-

rithm is simple yet powerful and specially designed

to solve nonlinear equations and is efﬁcient for non-

convex problems with constraints.

Twenty pairs of human car-following trajectories

from the NGSIM data set were selected and utilized

as the preceding vehicles of the subject vehicles mod-

eled by IDM and evaluated in terms of acceleration

Figure 1: Comparison of speed, gap, time headway and ac-

celeration between real-world human drivers and IDM.

and fuel consumption performance. It is noted that

the safe time headway of IDM and ACC are set as

the same for fair comparisons. Seven metrics, speed

mean, speed standard deviation (Std), gap mean, gap

Std, acceleration (Accl.) mean, acceleration Std and

fuel consumption that is estimated by Virginia-Tech

fuel consumption model(Rakha et al., 2011) are eval-

uated for performance comparisons.

Table 1 shows the IDM performance with cali-

brated parameters and comparisons to real-world hu-

man drivers and the ACC systems. The calibrated pa-

rameters are shown in Table 2. From Table 1, it could

be seen that the speed and gap of the IDM are similar

to those of human drivers, while the acceleration and

fuel consumption of the IDM have a signiﬁcant differ-

ence to real-world human drivers, as shown in Figure

Table 1: Performance comparison of human drivers, calibrated IDM, and ACC.

Model

Speed Mean

[m/s]

Speed. Std

Gap Mean

[m]

Gap. Std

Accl. Mean

[m/s

]

Accl. Std

Fuel

[ml]

Fuel Std

Human

Drivers

12.95 2.97 22.37 5.34 0.76 1.07 64.92 7.92

IDM 12.91 2.75 21.73 5.17 0.42 0.49 45.42 7.32

ACC 12.96 2.90 20.07 4.25 0.47 0.55 48.64 7.72

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

116

1. More importantly, the IDM performs even better

than the ACC systems, which is unlikely and makes

the IDM not suitable to serve as a human driver model

in mixed trafﬁc, including automated control systems.

Table 2: Calibrated parameters of IDM (20 pairs).

Parameter Name (Unit) Mean Std

Safe time headway (T,s) 1.12 0.34

Maximum acceleration (a

max

,m/s

) 2.45 1.30

Maximum deceleration (b

max

,m/s

) 4.28 1.17

Desired speed (v

,m/s) 24.89 2.02

Minimum distance (s

,m) 2.23 1.32

4 IMPACT ASSESSMENT OF IDM

PARAMETERS

4.1 Identifying Key IDM Parameters

In this study, an analysis using the Monte Carlo tech-

nique is implemented to assess how the parameters

and states of the IDM inﬂuence the optimization cost

function. This approach uses a representative set of

samples to explore the design space. The ﬁve param-

eters of the IDM, safe time headway T , maximum

acceleration a

max

and deceleration b

max

, and desired

speed v

and minimum distance s

are evaluated for

20 pairs of trajectories from NGSIM data, and their

ranges are set as (0.8, 2) s , (1, 5) m/s

, (1, 5) m/s

(20, 26) m/s and (1, 5) m respectively. The distribu-

tion of randomly generated parameters is assumed to

be a uniform distribution. The reason why the uni-

form distribution is set is that the restricted interval is

small so that any value within such range is equally

likely. 1000 samples for each parameter from the fol-

lowing distribution are generated for analysis.

The design requirements for this analysis are to

match the IDM’s time headway and speed trajectories

with those of real-world human drivers. As the rela-

tionship between ﬁve parameters and the cost func-

tion is complex, three different statistical analysis,

Rank correlation(Rutherford, 2005), Kendall corre-

lation (Kendall and Gibbons, 1990), and Rank stan-

dardized Regression (Greenland et al., 1991) are con-

ducted to analyze the impacts of parameters from

IDM. The rank correlation, referred to as Spearman

analysis, is to measure the degree of similarity be-

tween two variable rankings based on the assumption

that a nonlinear monotonic relation between the pa-

rameters and the cost function. The Kendall corre-

lation measures the ordinal association between two

measurements and it does not rely on any assump-

tions about the distributions. Rank standardized Re-

gression usually evaluates which of the independent

variables has a more signiﬁcant effect on the depen-

dent variable in a multiple regression analysis based

on the assumption that the parameters could linearly

inﬂuence the cost function.

Figure 2: Sensitivity Analysis Results.

The statistical results are shown in Figure 2 where

the ﬁrst ﬁgure shows the inﬂuence to the speed match-

ing and the second ﬁgure shows the inﬂuence to the

time headway matching. From Figure 2, it could

be seen that the safe time headway parameter is the

most inﬂuential parameter of the IDM to mimic hu-

man drivers’ time headway and speed trajectories for

all three statistical tests. For the other four parame-

ters, it has different inﬂuential ranking to the speed

and time headway. The minimum distance s

ranks

as the second inﬂuential parameter for speed trajec-

tory, while it is the least inﬂuential parameter for time

headway. On the contrary, the desired speed v

is the

second inﬂuential parameter for time headway, while

it is the second least inﬂuential parameter for speed.

It is noted that apart from safe time headway, the

other four parameters’ correlation value is less than

0.3, which means that compared to the safe time head-

Does the Intelligent Driver Model Adequately Represent Human Drivers?

117

way which plays the most signiﬁcant role in the IDM

to mimic human drivers’ car-following behaviors, the

inﬂuence of other four parameters is limited.

4.2 Time Headway of the IDM and

Human Drivers

Figure 3: The calibrated distribution of the variance and

mean of time headway for human drivers.

As the safe time headway is identiﬁed as the most in-

ﬂuential parameter in the IDM model, we further in-

vestigated the time headway trajectory of real-world

human drivers and compared it with that of the IDM

generated. Human drivers with an average time head-

way within 2 seconds (533 pairs) are extracted from

NGSIM data for evaluations and comparisons. To ob-

serve the time headway of human drivers, the mean

and variance of time headway proﬁles of 533 hu-

man drivers are calculated and shown in Figure 3.

In addition, the change of time headway also needs

to be investigated for analyzing the stochastic human

driver behaviors. The time headway and time head-

way change △ T (t) = T (t+ △ t) −T (t) for each time

step (0.1 seconds) are also shown in Figure 4.

From Figure 3, it could be seen that human drivers

have different characteristics. The safe time head-

way of different human drivers have a wide range of

mean and variance, while the time headway changes

of most human drivers are within [-0.1, 0.1] seconds.

From these observations, the human drivers tend to

have more ﬂuctuating time headway than that the

IDM generates. An example of a comparison be-

tween real-world human drivers and the calibrated

IDM is shown in Figure 1. From Figure 1, it could

be seen that IDM generated a more stable time head-

way than human drivers. This is mainly attributed

to the constant time headway setting. The IDM was

able to reach the desired time headway and maintain it

steadily to follow the preceding vehicle, while human

Figure 4: Histograms of drivers’ time headway and their

time headway change at each time step (533 pairs).

drivers were not. Therefore, constant time headway is

one reason that limits the IDM’s capability to mimic

human driver behaviors.

4.3 Calibration Performance

Comparison

Based on the conclusion from the previous section,

instead of using the constant safe time headway T in

the IDM, a dynamic time headway, T (t), is proposed

for IDM to mimic human driver behaviors (IDM-T

To utilize dynamic time headway and evaluate the im-

pacts to the IDM, the data of human drivers were ana-

lyzed. Based on the time headway at each step shown

in Figure 4, the dynamic time headway is assumed

to follow a normal distribution, and the variance and

mean are generated for each human driver’s character-

istics. To exclude unrealistic time headway changes

in the simulation, the time headway change is con-

strained within [-0.1, 0.1] seconds based on Figure 4.

The rest of the parameters of the IDM with dynamic

time headway shown in Figure 4 are the same as the

parameters of the standard IDM which are calibrated

using the optimization method mentioned in Section

3.2.

The performances of the standard IDM, IDM with

dynamic time headway, ACC, and human drivers are

shown in Table 3. From Table 3, it could be seen that

the standard IDM performed similarly to ACC in ac-

celeration but had better fuel consumption than ACC.

When the IDM is compared with human drivers, it

performs signiﬁcantly better in acceleration and fuel

consumption. This means that the IDM cannot ade-

quately represent human drivers in terms of comfort

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

118

Table 3: Performance Comparison of IDM, IDM-T

, ACC, human drivers.

Model

Speed Mean

[m/s]

Speed. Std

Gap Mean

[m]

Gap. Std

Accl. Mean

[m/s

]

Accl. Std

Fuel

[ml]

Fuel Std

Human

Drivers

12.62 2.72 21.10 4.71 0.78 1.12 60.45 10.43

IDM-T

12.63 2.53 20.17 4.12 0.90 1.11 61.42 10.92

IDM 12.59 2.54 20.92 4.79 0.38 0.47 38.84 9.12

ACC 12.6 2.66 18.30 3.73 0.43 0.52 41.47 9.73

Table 4: Calibrated parameters of IDM (533 pairs).

Parameter Name (Unit) Mean Std

Safe time headway (T,s) 1.14 0.34

Maximum acceleration (a

max

,m/s

) 2.19 1.25

Maximum deceleration (b

max

,m/s

) 4.25 1.20

Desired speed (v

,m/s) 24.48 2.25

Minimum distance (s

,m) 2.18 0.96

and fuel efﬁciency. If the IDM were used in a mi-

croscopic simulation to represent human drivers, it is

unlikely to accurately estimate the improvement from

ACC or other control systems of automated vehicles.

However, with dynamic time headway, the IDM-T

performed more similarly to that of human drivers

compared with IDM, which could be seen from Table

3. Based on the t-test, the fuel efﬁciency and acceler-

ation of the IDM with dynamic time headway (IDM-

) is not signiﬁcantly different from human drivers.

The IDM with constant time headway and ACC sys-

tems are signiﬁcantly different from human drivers in

all metrics.

4.4 Validation Performance

Comparison

This subsection performed the validation of the cali-

brated IDM-T

using the NGSIM human drivers’ tra-

jectory data that were not used during the calibra-

tion. This is necessary to ensure the calibrated IDM

with dynamic headway is applicable to general human

driven vehicles. A total of 100 pairs of human drivers’

trajectories were used. To validate the proposed dy-

namic time headway in the IDM-T

, the calibrated

time headway and other parameters are implemted us-

ing different trajectories in the NGSIM data set and

compared with those of human driven vehicles. The

parameter settings shown in Figure 4 are analyzed to

create a probability distribution for the variance of the

safe time headway by ﬁtting an appropriate distribu-

tion. The distribution is identiﬁed as shown in Fig-

ure 3. The variance of the time headway follows a

Gamma distribution with a = 1.32,b = 0.057. The

mean value of the time headway follows a normal dis-

tribution with a mean of 1.34 and a standard deviation

of 0.34 and the value generated by this distribution is

constrained within [0.8, 2] s.

Based on the calibrated distribution of time head-

way and the calibrated parameters in Table 4, the

IDM, IDM-T

, and ACC were simulated and com-

pared with human drivers shown in Table 5. The sim-

ulation process is the same with the previous section.

From Table 5, it could be seen that the IDM-T

was

not able to perform as well as the calibrated results,

while IDM-T

still performed more similar to human

drivers compared with the standard IDM.

5 CONCLUSIONS AND FUTURE

WORK

In this study, we investigated the limitations and po-

tential of IDM to represent human drivers for micro-

scopic trafﬁc simulation. To evaluate the reliability of

IDM to represent human drivers, the calibrated IDM

is compared with real-world human drivers and the

ACC system. The parameters of the IDM were cal-

ibrated in terms of speed and time headway match-

ing to human drivers based on real-world NGSIM

data set. The simulation results showed that the

IDM matches well with the speed and gap of human

drivers, while it performs signiﬁcantly better than hu-

man drivers and ACC systems in comfort and fuel ef-

ﬁciency, which makes it inadequate to represent hu-

man drivers. Therefore, in order to improve IDM

application for human drivers, the safe time head-

way, which is the most inﬂuential parameter in the

IDM, was proposed to be dynamic instead of using

a constant static value. To evaluate the impacts of

the dynamic time headway to IDM application on hu-

man drivers, 633 pairs of car-following behaviors of

human drivers from the NGSIM data set were com-

pared with IDM, IDM with dynamic time headway,

and ACC control mode with calibration and valida-

tion. With dynamic time headway, IDM can be more

similar to human drivers than the standard IDM in

fuel consumption based on the paired t-test and have a

signiﬁcant improvement in acceleration similarity. It

is expected that the proposed dynamic safe headway-

Does the Intelligent Driver Model Adequately Represent Human Drivers?

119

Table 5: Validation Performance of IDM, IDM-T

, ACC and human drivers.

Model

Speed Mean

[m/s]

Speed. Std

Gap Mean

[m]

Gap. Std

Accl. Mean

[m/s

]

Accl. Std

Fuel

[ml]

Fuel Std

Human

Drivers

12.52 1.85 20.25 3.32 0.92 1.24 44.54 14.97

IDM-T

12.52 1.64 19.41 2.66 0.84 1.03 39.11 14.84

IDM 12.46 1.63 20.32 3.28 0.36 0.42 24.55 9.97

ACC 12.48 1.73 14.82 2.39 0.43 0.52 26.67 11.58

based IDM improves the evaluation of mixed trafﬁc

interacting with human-driven vehicles and connected

automated vehicles (Chen and Park, 2020).

In this study, safe time headway is the only param-

eter we adjusted to represent human drivers. While

the other parameters in the IDM were less inﬂuen-

tial, they should be considered to be adjusted for more

accurate human driver modeling. The dynamic time

headway is based on the normal distribution and the

change is constrained based on real-world data. How-

ever, the value and the direction of time headway

change have not been well-investigated, which will be

investigated in future research. Besides, vehicle stop

and vehicle catch-up behaviors are not considered in

this study. Given these behaviors are also essential to

human driver behavior modeling in microscopic sim-

ulation, these behaviors should be investigated for hu-

man driver modeling.

ACKNOWLEDGEMENTS

This research is supported by the National Science

Foundation under Grant CMMI-2009342.

REFERENCES

Administration, U. D. O. T. F. H. (2017). Next Generation

Simulation (NGSIM) Vehicle Trajectories and Sup-

porting Data. Type: dataset.

Alhariqi, A., Gu, Z., and Saberi, M. (2022). Calibration

of the intelligent driver model (idm) with adaptive pa-

rameters for mixed autonomy trafﬁc using experimen-

tal trajectory data. Transportmetrica B: Transport Dy-

namics, 10(1):421–440.

Bando, M., Hasebe, K., Nakayama, A., Shibata, A., and

Sugiyama, Y. (1995). Dynamical model of trafﬁc

congestion and numerical simulation. Phys. Rev. E,

51:1035–1042.

Bhattacharyya, R., Senanayake, R., Brown, K., and

Kochenderfer, M. (2020). Online parameter estima-

tion for human driver behavior prediction.

Chen, C., Li, L., Hu, J., and Geng, C. (2010). Calibration of

mitsim and idm car-following model based on ngsim

trajectory datasets. In Proceedings of 2010 IEEE In-

ternational Conference on Vehicular Electronics and

Safety, pages 48–53.

Chen, Z. and Park, B. B. (2020). Cooperative adaptive

cruise control with unconnected vehicle in the loop.

IEEE Transactions on Intelligent Transportation Sys-

tems, pages 1–11.

Ciuffo, B., Punzo, V., and Montanino, M. (2014). Global

sensitivity analysis techniques to simplify the calibra-

tion of trafﬁc simulation models. methodology and ap-

plication to the idm car-following model. IET Intelli-

gent Transport Systems, 8(5):479–489.

Coleman, T. F. and Li, Y. (1996). An interior trust region ap-

proach for nonlinear minimization subject to bounds.

SIAM Journal on Optimization, 6(2):418–445.

Ding, H., Pan, H., Bai, H., Zheng, X., Chen, J., and

Zhang, W. (2022). Driving strategy of connected

and autonomous vehicles based on multiple preced-

ing vehicles state estimation in mixed vehicular trafﬁc.

Physica A: Statistical Mechanics and its Applications,

596:127154.

Eggert, J., Damerow, F., and Klingelschmitt, S. (2015). The

foresighted driver model. In 2015 IEEE Intelligent

Vehicles Symposium (IV), pages 322–329.

Gazis, D. C., Herman, R., and Potts, R. B. (1959). Car-

Following Theory of Steady-State Trafﬁc Flow. Op-

erations Research, 7(4):499–505. Publisher: IN-

FORMS.

Greenland, S., Maclure, M., Schlesselman, J. J., Poole, C.,

and Morgenstern, H. (1991). Standardized Regression

Coefﬁcients: A Further Critique and Review of Some

Alternatives. Epidemiology, 2(5).

Guo, L. and Jia, Y. (2021). Anticipative and predic-

tive control of automated vehicles in communication-

constrained connected mixed trafﬁc. IEEE Transac-

tions on Intelligent Transportation Systems.

Hegde, B., O’Keefe, M., Muldoon, S., Gonder, J., and

Chang, C.-F. (2021). Real-World Driving Features

for Identifying Intelligent Driver Model Parameters.

pages 2021–01–0436.

Jiang, H., Hu, J., An, S., Wang, M., and Park, B. B. (2017).

Eco approaching at an isolated signalized intersection

under partially connected and automated vehicles en-

vironment. Transportation Research Part C: Emerg-

ing Technologies, 79:290–307.

Jiang, H., Hu, J., Park, B., Wang, M., and Zhou, W. (2019).

An extensive investigation of an eco-approach con-

troller under a partially connected and automated ve-

hicle environment. Sustainability (Switzerland), 11.

VEHITS 2023 - 9th International Conference on Vehicle Technology and Intelligent Transport Systems

120

Kendall, M. and Gibbons, J. D. (1990). Rank Correlation

Methods. A Charles Grifﬁn Title, 5 edition.

Kesting, A. and Treiber, M. (2008). Calibrating car-

following models by using trajectory data. Trans-

portation Research Record: Journal of the Trans-

portation Research Board, 2088(1):148–156.

Kesting, A., Treiber, M., and Helbing, D. (2010). Enhanced

intelligent driver model to access the impact of driving

strategies on trafﬁc capacity. Philosophical Transac-

tions of the Royal Society A: Mathematical, Physical

and Engineering Sciences, 368(1928):4585–4605.

Marsden, G., McDonald, M., and Brackstone, M. (2001).

Towards an understanding of adaptive cruise control.

Transportation Research Part C: Emerging Technolo-

gies, 9(1):33–51.

Mu, Z., Chen, Z., Ryu, S., Avedisov, S. S., Guo, R.,

and Park, B. B. (2022). Cooperative platooning with

mixed trafﬁc on urban arterial roads. In 2022 IEEE In-

telligent Vehicles Symposium (IV), pages 1578–1583.

Punzo, V., Borzacchiello, M. T., and Ciuffo, B. (2011).

On the assessment of vehicle trajectory data accu-

racy and application to the Next Generation SIMula-

tion (NGSIM) program data. Transportation Research

Part C: Emerging Technologies, 19(6):1243–1262.

Rahman, M. S. and Abdel-Aty, M. (2018). Longitudi-

nal safety evaluation of connected vehicles’ platoon-

ing on expressways. Accident Analysis & Prevention,

117:381–391.

Rakha, H. A., Ahn, K., Moran, K., Saerens, B., and den

Bulck, E. V. (2011). Virginia tech comprehensive

power-based fuel consumption model: Model devel-

opment and testing. Transportation Research Part D:

Transport and Environment, 16(7):492–503.

Ro, J. W., Roop, P. S., Malik, A., and Ranjitkar, P. (2018).

A Formal Approach for Modeling and Simulation of

Human Car-Following Behavior. IEEE Transactions

on Intelligent Transportation Systems, 19(2):639–648.

Conference Name: IEEE Transactions on Intelligent

Transportation Systems.

Rutherford, A. (2005). Research design and statistical anal-

ysis. jerome l. myers,and arnold d. well,(2nd edn)

mahwah, nj: Lea 2003 , £43.10 . isbn: 08058-4037-

0. British Journal of Mathematical & Statistical Psy-

chology - BRIT J MATH STATIST PSYCHOL, 58:187–

189.

Sharma, A., Zheng, Z., Bhaskar, A., and Haque, M. M.

(2019). Modelling car-following behaviour of con-

nected vehicles with a focus on driver compli-

ance. Transportation research part B: methodologi-

cal, 126:256–279.

Sharma, A., Zheng, Z., Kim, J., Bhaskar, A., and Mazharul

Haque, M. (2021). Assessing trafﬁc disturbance, ef-

ﬁciency, and safety of the mixed trafﬁc ﬂow of con-

nected vehicles and traditional vehicles by consider-

ing human factors. Transportation Research Part C:

Emerging Technologies, 124:102934.

Stern, R. Modeling stochastic human-driver car following

behavior in oscillatory trafﬁc conditions. page 27.

Sun, C., Guanetti, J., Borrelli, F., and Moura, S. J. (2020).

Optimal eco-driving control of connected and au-

tonomous vehicles through signalized intersections.

IEEE Internet of Things Journal, 7(5):3759–3773.

Treiber, M., Hennecke, A., and Helbing, D. (2000). Con-

gested Trafﬁc States in Empirical Observations and

Microscopic Simulations. arXiv:cond-mat/0002177.

Vahidi, A. and Eskandarian, A. (2003). Research advances

in intelligent collision avoidance and adaptive cruise

control. IEEE Transactions on Intelligent Transporta-

tion Systems, 4(3):143–153.

Wang, Z., Wu, G., and Barth, M. J. (2019). Coopera-

tive eco-driving at signalized intersections in a par-

tially connected and automated vehicle environment.

IEEE Transactions on Intelligent Transportation Sys-

tems, 21(5):2029–2038.

Yi, Z.-w., Lu, W.-q., Xu, L.-h., Qu, X., and Ran, B.

(2020). Intelligent back-looking distance driver

model and stability analysis for connected and auto-

mated vehicles. Journal of Central South University,

27(11):3499–3512.

Zhang, J., Wu, K., Cheng, M., Yang, M., Cheng, Y., and

Li, S. (2020). Safety evaluation for connected and au-

tonomous vehicles’ exclusive lanes considering pene-

trate ratios and impact of trucks using surrogate safety

measures. Journal of advanced transportation, 2020.

Does the Intelligent Driver Model Adequately Represent Human Drivers?

121