Professional Documents
Culture Documents
To cite this article: Sangram Nirmale, Anshuman Sharma & Abdul Rawoof Pinjari (2023): Multi-
vehicle anticipation-based driver behavior models: a synthesis of existing research and future
research directions, Transportation Letters, DOI: 10.1080/19427867.2023.2231212
Article views: 60
Introduction
and multiple sources of stimuli (or information). The terms ‘multi-
Driver behavior models are used to describe driver decisions in anticipative’, ‘multi-vehicle anticipation’, and ‘spatial anticipation’
various traffic scenarios ranging from free-flow to congested traffic, refer to drivers’ ability to consider stimuli from several vehicles
city traffic to highway traffic. A few examples of well-known driver ahead in their maneuvering decisions. The term ‘temporal antici
behavior models are Newell’s model (Newell 1961), Gipps’ model pation’ refers to drivers’ ability to anticipate the traffic situation for
(Gipps 1981), Optimal velocity model (Bando et al. 1995), and the next few seconds and react accordingly. Whereas the term
Intelligent driver model (Treiber, Hennecke, and Helbing 2000). ‘multiple sources of stimuli’ indicates drivers’ ability to consider
At large, these models are founded on the stimulus-response frame stimuli from lead vehicles as well as from other sources (mentioned
work that was first introduced in the 1950s at the General Motors above) in their maneuvering decisions. This review focuses on
research laboratories (Chandler, Herman, and Montroll 1958; models incorporating the multi-vehicle anticipation (MVA) beha
Gazis, Herman, and Rothery 1961). According to this framework, vior of drivers.
a driver’s response is proportional to the stimulus from a vehicle The first question that arises is why do drivers consider stimuli
ahead. In general, an event or a quantity that evokes a response from several vehicles or anticipate their behavior? Drivers antici
from the driver is the stimulus. Researchers have considered various pate the surrounding vehicles’ behavior to adjust their maneuvering
stimuli such as speed difference between the subject vehicle and the decisions and perform them safely. In homogeneous traffic streams,
immediate leader, spacing between the subject vehicle and the while moving in the longitudinal direction, drivers adjust their
immediate leader, and other sources such as gradient, traffic signals, acceleration/deceleration and its extent by observing the dynamics
lane closures, lane markings, etc. A majority of available driver of vehicles directly ahead. For instance, by noticing the slowing
behavior models assume that drivers respond to the stimulus down of vehicles ahead (second leader or third leader), drivers
from the immediate lead vehicle only (Brackstone and McDonald proactively regulate their acceleration to avoid sudden jerks.
1999; Saifuzzaman and Zheng 2014). However, it has been demon When it comes to lane-changing, consideration of stimuli from
strated theoretically and empirically that drivers anticipate down multiple surrounding vehicles becomes more important because
stream traffic conditions, consider stimuli from surrounding a lane change is a complex and riskier maneuver and involves
vehicles (front and side) and respond accordingly (Hoogendoorn interactions with a greater number of vehicles than car-following.
and Ossen 2006; Nirmale, Pinjari, and Sharma 2021). Therefore, Further, it is well known that the characteristics of heterogeneous
models that consider stimuli from more than one vehicle to disordered (HD) traffic streams are substantially different than
describe driver behavior offer a higher degree of realism. These those of homogeneous traffic streams in terms of traffic composi
driver behavior models are the focus of this review. tion, lane discipline, and overall driver behavior (more on this in
Before moving further, we define some terminologies related to Section 3). One can witness longitudinal, lateral, and a combination
the anticipation by drivers that are commonly and interchangeably of longitudinal and lateral (also known as two-dimensional) move
used in the literature. Existing literature refers to the following ments in HD traffic streams. For a two-dimensional movement to
terms in the context of anticipation: multi-anticipative, multi- be safe, drivers must consider the dynamics of vehicles that are not
vehicle anticipation, spatial anticipation, temporal anticipation, only directly ahead but also laterally placed. The anticipation of
CONTACT Sangram Nirmale nirmalek@iisc.ac.in Department of Civil Engineering, Indian Institute of Science (IISc), Bengaluru 560012, India
© 2023 Informa UK Limited, trading as Taylor & Francis Group
2 S. K. NIRMALE ET AL.
vehicular movements becomes far more important in HD traffic Note that the studies that examine models with multi-vehicle
streams because vehicles can cut-in at any time (typical HD stream anticipation behavior also assisted us in understanding the concept
characteristic) and may cause a safety-critical situation. Hence, it is of multi-vehicle anticipation behavior.
no exaggeration to say that multi-vehicle anticipation is an integral
aspect of drivers’ maneuvering decisions. It makes drivers proactive
and thereby better and safer decision-makers (Sharma et al. 2017). Classification of identified studies
Studies on MVA enumerate various advantages of incorporating A reference management software package is utilized to manage the
MVA in driver behavior models, such as superior numerical and large number of references collected for this study and to perform
behavioral soundness, plausible parameter estimates, and model sorting and categorizing tasks. In addition, the data extraction
outputs, and an improved model realism (Bexelius 1968; procedure included a two-step screening procedure. First, the
Hoogendoorn and Ossen 2006; Lenz, Wagner, and Sollacher 1999; downloaded papers are categorized based on their Titles,
Treiber, Kesting, and Helbing 2006). Despite the importance of Abstracts, and Keywords; then, they are organized based on the
considering MVA, models that incorporate MVA are underex specific topics addressed in each paper. We observe that the major
plored. Although there are some excellent review studies on driver ity of MVA-based models are extensions of well-known single-
behavior models (Asaithambi, Kanagaraj, and Toledo 2016; Azam, leader driver behavior models such as the Gazis-Herman-Rothery
Bhaskar, and Haque 2020; Brackstone and McDonald 1999; (GHR) car-following model, optimal velocity model (OVM), full
Chakroborty, Maurya, and Vikram 2019; Das and Maurya 2018; velocity difference (FVD) model, and intelligent driver model
Mahapatra, Maurya, and Chakroborty 2018; Moridpour, Sarvi, and (IDM). Hence, we review the MVA-based driver behavior models
Rose 2010; Munigety and Mathew 2016; Saifuzzaman and Zheng under the classifications of extensions of these well-known single-
2014; Toledo 2007), none of them comprehensively reviews MVA- leader models. The functional forms of these single-leader driver
based driver behavior models. An in-depth exploration of MVA- behavior models are provided in Table 1.
based models will reveal prevalent theories behind such models,
approaches to incorporate MVA behavior, shortcomings of the
available approaches, and a way forward to better mimic the Evaluation of the identified studies
MVA behavior of drivers. In this step, the identified studies are critically reviewed and sum
This paper, hence, attempts to fill this gap by providing a critical marized in Tables 2 and 3. Specifically, the evaluation focuses on the
review of MVA-based driver behavior models developed for both strength and limitations of each model, components of the model
homogeneous and heterogeneous traffic streams. The remainder of ing framework, input and output variables, and data utilized for
this paper is organized as follows. Section 2 presents a methodology calibration. This systematic review led to a firm grasp of state of the
adopted for collecting the relevant literature. Sections 3 and 4 literature and insights into the multi-vehicle anticipation-based
review models that incorporate MVA by drivers in homogeneous driving behavior modeling requirements for both homogeneous
traffic conditions and HD traffic conditions, respectively. Section 5 and HD traffic conditions.
provides critical findings from the current review paper. Section 6
synthesizes the gaps with the current approaches to incorporate
MVA behavior and identifies directions for future research. Finally, Models that incorporate multi-vehicle anticipation by
Section 7 concludes the paper. drivers in homogeneous traffic streams
We observe that the majority of MVA-based models are extensions
Review framework of well-known single-leader driver behavior models, such as the
Gazis-Herman-Rothery (GHR) car-following model, optimal velo
This paper adopts a systematic review approach as done by Das and city model (OVM), full velocity difference (FVD) model, and intel
Maurya (2018), Azam, Bhaskar, and Haque (2020), and Arun et al. ligent driver model (IDM). Hence, we review the MVA-based
(2021), which includes the following steps: 1) identification of driver behavior models under the classifications of extensions of
studies using different inclusion and exclusion criteria, 2) classifica these well-known single-leader models. The functional forms of
tion of identified studies, and 3) evaluation of the identified studies. these single-leader driver behavior models are provided in Table 1.
The following subsections discuss these three steps.
Note – Refer to the Appendix A for details on the notation used in the above table.
following model with a single lead vehicle. An advantage of this decisions (acceleration or deceleration). To fill these gaps,
model is its simplicity. However, the model presented in Equation Hoogendoorn, Ossen, and Schreuder (2006) proposed
(1) suffers from the same limitations as the original GHR model. a generalized version of Helly’s (1959) model by including
These are: (i) inter-driver heterogeneity is not captured since the a distance-dependent factor in it as provided below:
model assumes the identical reaction time values for all drivers;
and (ii) it assumes that the drivers’ maneuvering actions are only
N1
X N2
X � �
ðjÞ ðjÞ ðdesÞ
ai ðt þ Ti Þ ¼ αj ΔVi ðtÞ þ βj ΔXi ðtÞ ΔXi ðt þ Ti Þ
dependent on relative speed. Further, the earlier studies did not
j¼1 j¼1
provide empirical evidence for the presence of MVA behavior
(Saifuzzaman and Zheng 2014). (4)
To empirically investigate MVA-based models, Hoogendoorn and ðjÞ
where, ΔXi ðtÞ represents the distance between the lead vehicle j and
Ossen (2006), Hoogendoorn, Ossen, and Schreuder (2006) and ðdesÞ
Zhang (2014) analyzed vehicle trajectory data using regression ana subject vehicle i at time t; ΔXi ðt þ Ti Þ is desired distance between
lysis. Hoogendoorn and Ossen (2006) utilized the multivariate linear the subject vehicle i and its jth leader at time ðt þ Ti Þ. In Equation (4),
regression analysis to estimate parameters of Bexelius’s multi- the additive expression brings in relative speed and spacing as different
anticipatory car-following model (Equation (1)). They empirically sources of stimuli. Hoogendoorn, Ossen, and Schreuder (2006)
demonstrated that drivers not only consider the vehicle directly employed the maximum likelihood estimation approach to calibrate
ahead but also the second leader. Furthermore, they concluded that the sensitivities (aka, coefficients) of the stimuli from different vehicles,
the degree of driver’s reaction to the second leader is dependent on and used a likelihood ratio test to evaluate the benefits of accounting for
both the types of the following and followed vehicles. Drivers follow MVA. Their empirical results suggested that incorporating the MVA
ing a truck, for example, had a weaker reaction to the second leader effect improved the extent to which the models could explain observed
on average than those behind a car, perhaps because drivers cannot driver behavior.
easily sight past a truck. Similarly, truck drivers have a significant Following Hoogendoorn, Ossen, and Schreuder (2006), Zhang
reaction to the second leader, maybe because truck drivers’ elevated (2014) provided more insights on the linear-type MVA driver beha
vantage point offers a greater sight distance. vior model. Specifically, Zhang (2014) proposed a modified general
Next, Hoogendoorn, Ossen, and Schreuder (2006) proposed the ized linear MVA car-following model (Equation (5)) and addressed
following two modifications to the Bexilus’ MVA model. two issues – multicollinearity between explanatory variables and the
Bexelius Type 2 model: serial correlation of time series data.
n o N1
ð1Þ ð2Þ ðNÞ
ai ðt þ Ti Þ ¼ α min ΔVi ðtÞ; ΔVi ðtÞ; . . . ; ΔVi ðtÞ (2) X ðjÞ
ai ðt þ Ti Þ ¼ α0 þ αj ΔVi ðtÞ
j¼1
Bexelius Type 3 model: N2
X � �
ðjÞ ðdesÞ
n o þ βj ΔXi ðtÞ ΔXi ðt þ Ti Þ þ γVi ðtÞ þ εit
ð1Þ ð2Þ ðNÞ
ai ðt þ Ti Þ ¼ min α1 ΔVi ðtÞ; α2 ΔVi ðtÞ; . . . ; αN ΔVi ðtÞ (3) j¼1
(5)
These models assume that the driver may only respond to the
first, second, or third leader based on the speed difference This study empirically demonstrated that the traffic congestion
between their vehicle and the lead vehicles. However, these level (extremely congested and less congested) affects driver’s reac
models do not consider relative spacing as a stimulus and tions to different stimuli with respect to different lead vehicles. For
ignore behavioral differences between different maneuvering example, the study’s estimation results depicted that drivers in
4
Table 2. Summary on models that incorporate multi-vehicle anticipation by drivers in homogeneous traffic streams.
Model category Author(s) Proposed model Data utilized and region Strengths Weakness
Extensions of GHR Car- Bexelius (1968) and Gazis, N
P – ● Simple model is proposed ● The model is too simplis
ðjÞ
ai ðt þ Ti Þ ¼ αj ΔVi ðtÞ ● Special case of this model
Following Model Herman, and Potts (1959) j¼1
tic to describe actual traf
gives the GHR model fic phenomena
● Stability conditions are accurately.
derived analytically ● Inter-driver heterogeneity
is not captured since the
model assumes identical
S. K. NIRMALE ET AL.
(Continued)
5
6
Table 2. (Continued).
Model category Author(s) Proposed model Data utilized and region Strengths Weakness
!
Li and Liu (2006) N
P � – ● The study confirms that ● Model was not estimated
ð1Þ ðjÞ 1 j
ai ðt þ Ti Þ ¼ VOV ΔXi ðtÞ; αj ΔVi ðtÞ ; where αj ¼ 5 incorporating the relative using any data
j¼1
speed as a stimulus can
stabilize the traffic flow
just as in the forward-
S. K. NIRMALE ET AL.
extremely congested traffic conditions would react to relative Wilson et al. (2004) proposed multiple look-ahead models that
speeds with respect to the first, second, and even the third leader. consider information from multiple leaders as provided below.
Model A:
( )
X
N � �
ðjÞ
Extensions of models based on optimal velocity model ai ðtÞ ¼ α0 αj VOV ΔXi ðtÞ Vi ðtÞ (11)
j¼1
Extensions of optimal velocity model
The optimal velocity model (OVM) was also widely utilized to Model B:
mimic the MVA behavior of drivers. For example, Lenz, Wagner, ( ! )
X
N ðjÞ
and Sollacher (1999) extended the OVM to include multiple vehicle ΔXi ðtÞ
ai ðtÞ ¼ α0 αj VOV Vi ðtÞ (12)
responses. Following Bexelius (1968), Lenz, Wagner, and Sollacher j
j¼1
(1999) assumed that drivers react to the dynamics of their leading
P
vehicle and an arbitrary number of vehicles ahead with a sensitivity In Model A, αj > 0 and Nj¼1 αj ¼ 1 so that uniform flow solutions
αj . The mathematical formulation is provided below: are the same as in the standard OVM, thus, if the space gap between
" ! # the subject vehicle i and the lead vehicle j is small, the subject
XN ðjÞ
ΔXi ðtÞ vehicle i will have a lower optimal velocity than the standard
ai ðtÞ ¼ αj VOV Vi ðtÞ (6)
j¼1
j OVM. Hence, the subject vehicle i will tend to brake earlier when
approaching congested traffic. Model B is taken from Lenz,
Here, N ¼ 1 leads to the original optimal velocity model. They Wagner, and Sollacher (1999) where the optimal velocity of the
borrowed the following functional form from Bando et al. (1998)1 subject driver i is given by a weighted sum of the original optimal
for VOV ð:Þ. velocity function, evaluated at a distance to each lead vehicle ahead.
� � � �
ð1Þ ð1Þ The study exhibited that the proposed model did not predict
VOV ΔXi ðtÞ ¼ tanh Δ Xi ðtÞ C
unrealistic values of acceleration. Note that we observe two types
tanhðCÞ where C is aconstant (7) of representations when MVA behavior is considered in the OVM.
In Equations (11) and (12), the sensitivity coefficient is directly
The study demonstrated increased stability with the MVA-based
multiplied to the optimal velocity function, whereas in Equations
model. Same as in the case of the GHR model, the MVA is incor
(6) and (8), the sensitivity coefficient is multiplied by the difference
porated by summing up the weighted stimuli from vehicles ahead.
between optimal velocity function and velocity of the subject vehi
However, a functional form is assumed to compute the sensitivity
cle. Note that formal analysis is missing in the existing literature
coefficient αj .
regarding which formulation is better. We believe that the latter is
Notably, a few studies have extended Lenz’s model by consider
more behaviorally sound since it considers driver’s sensitivity to the
ing reaction time and desired distance (Chen et al. 2016;
stimuli; i.e. the difference between the optimal velocity and the
Hoogendoorn, Ossen, and Schreuder 2006; Hu, Ma, and Chen
velocity of the subject vehicle rather than the optimal velocity only.
2014). For instance, Hoogendoorn, Ossen, and Schreuder (2006)
Furthermore, Chen, Shi, and Hu (2012) incorporated driver’s
integrated the reaction time as per Equation (8).
" ! # reaction time in Wilson et al. (2004)’s model and proposed
XN ðjÞ
ΔXi ðtÞ a multiple look-ahead model with driver reaction delay as below:
ai ðt þ Ti Þ ¼ αj VOV Vi ðtÞ (8) ( )
j¼1
j X
N
ðjÞ
ai ðt þ Ti Þ ¼ α0 αj VOV ðΔXi ðtÞÞ Vi ðtÞ (13)
Later, Ge et al. (2004) proposed an extended car-following model by j¼1
incorporating headways of arbitrary numbers of lead vehicles in
optimal velocity function itself. The functional form of the model is Following Ge et al. (2004), the following functional form was
given below: employed by Chen, Shi, and Hu (2012) for αj .
! X N �N 1
X N j for j�N
N
ai ðt þ Ti Þ ¼ VOV
ðjÞ
αj ΔXi ðtÞ (9) αj ¼ 1; αj ¼ 1 (14)
j¼1 N j 1 for j ¼ N:
j¼1
ðjÞ
where αj is a weighted function of ΔXi ðtÞ with the following The study utilized the following optimal velocity function from the
properties: Bando et al. (1995) study.
� � h � � i
ðJÞ ðjÞ
(1) αj decreases monotonically as j increases, i.e. α1 > α2 indicat VOV ΔXi ðtÞ ¼ 0:5Vmax tanh ΔXi ðtÞ ΔXc þ tanhðΔXc Þ
ing that the influence of the vehicle ahead of the subject (15)
vehicle reduces gradually as the distance between the subject
vehicle and lead vehicle increases. where, Vmax is the maximum velocity and ΔXc is the safety distance.
(2) αj takes the following functional form: Hasebe, Nakayama, and Sugiyama (2003) also extended the
�6 OVM to incorporate MVA behavior and called this model
X N
for j�N a forward-looking optimal velocity model (FLOVM). Their formu
αj ¼ 1; α1 ¼ 1 "j ¼ 1; αj ¼ 7j1 (10)
7 j 1 for j ¼ N: lation is given below:
j¼1
h � � i
ð1Þ ð2Þ ðJÞ
The authors performed a simulation-assisted stability analysis and ai ðtÞ ¼ α0 VOV ΔXi ðtÞ; VðΔXi ðtÞ; . . . ; VðΔXi ðtÞ Vi ðtÞ
found that MVA further stabilizes the traffic flow compared to the
(16)
single leader model. Additionally, the simulation results from the
study confirmed that only the information of three cars ahead of the The authors employed the optimal velocity function of Bando et al.
subject vehicle is enough for cooperative driving. However, empiri (1995) and demonstrated that the FLOVM provides greater stability
cal evidence is missing from this study. than the original OVM.
TRANSPORTATION LETTERS 9
where, VOV ð:Þ is the optimal velocity function and Gð:Þ is assumed where, λ0 > 0, and λ1 ; λ2 2 ½0; 1� are different sensitivity coeffi
ðjÞ
as a monotonically increasing function and their formulations are cients; Δai ðtÞ represents acceleration difference between the
given below: lead vehicle j and subject vehicle i at the time t; αj , βj , and γj
� � are different weighting coefficients. This study used Equation
ðjÞ ðjþ1Þ ðjþN 1Þ ðjÞ
VOV ðΔXi ðtÞ; ΔXi ðtÞ; . . . ; ΔXi ðtÞÞ ¼ α1 V � ΔXi ðtÞ (15) for VOV ð:Þ. Moreover, they assumed αj as a decreasing
� � � � PN
ðjþ1Þ ðjþN 1Þ function with j, i.e. αj > αjþ1 and
þα2 V � ΔXi ðtÞ þ . . . þ αN V � ΔXi ðtÞ j¼1 αj ¼ 1. The functional
10 S. K. NIRMALE ET AL.
� � qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
form for αj is the same as that from Peng and Sun (2010). ð2�Þ ð2�Þ
Vsafe V2 ; ΔXi ¼ bTi þ b2 Ti2 þ V22 þ 2bΔXi (34)
Moreover, Li et al. (2011) illustrated that their model had
a greater stable region than that of the FVD model. Then, Vpred is used to calculate the safe speed for the subject vehicle
i as below:
Vi;safe ¼ bTi
Extensions of Intelligent Driver Model (IDM) rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi�ffiffi
� �
þ b2 Ti2 þ Vpre 2 þ 2b ΔX ð1Þ þ V
pred Ti min Vpred Ti ; ΔXconstant
To incorporate MVA, Treiber, Kesting, and Helbing (2006) con i
sidered cumulative stimuli between the subject vehicle i and lead (35)
vehicle j for N lead vehicles as given below:
where, a and b represent acceleration and deceleration, respectively.
0 12
2 The speed of the first and second lead vehicles are represented as V1
!β 3 ð1Þ
Vi ðtÞ XN B ΔXðj;desÞ ðtÞ C and V2 , respectively. ΔXi is the space gap between the first lead
41 5 B i C
ai ðtÞ ¼ amax
i amax
i BN C (29) ð2�Þ
ðdesÞ
Vi ðtÞ @ P ðjÞ A vehicle and subject vehicle i whereas ΔXi is the space gap between
j¼1 ΔX ðtÞ i the second lead vehicle and the first lead vehicle. The parameter ε
j¼1
represents fluctuations in units of acceleration (a).
P ðjÞ
where, Nj¼1 ΔXi ðtÞ represents net gaps between the subject vehi
ðj;desÞ
cle i and lead vehicles j and ΔXi ðtÞ is represented below: Extensions of other models
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðjÞ
ðj;desÞ Vi ðtÞ ðdesÞ Vi ðtÞΔVi ðtÞ Extensions of piece-wise linear car-following model
ΔXi ðtÞ ¼ C0 þ C1 ðdesÞ
þ V i ðtÞT i q ffi ffi ffi ffi ffi ffi ffi ffi ffi ffiffiffiffiffiffiffiffi A piece-wise linear car-following model was extended to incorpo
Vi ðtÞ 2 amax bi
conf
i rate MVA in driver behavior (Farhi, Haj-Salem, and Lebacque
(30) 2012). It was a first-order discrete-time model, where speeds were
modeled as a function of spacings. In this study, the minimum form
Chen et al. (2010) showed that this formulation leads to an issue at was used rather than an additive form, which is commonly incor
equilibrium flow; i.e. the space gap at equilibrium flow is dependent porated in the context of MVA-based driver behavior models.
on the number of lead vehicles leading to different desired head Moreover, the spacing to the jth lead vehicle was divided by j to
ways at the same uniform space gap at equilibrium. To overcome make sensitivities uniform across different lead vehicles. The math
this drawback, Chen et al. (2010) proposed the following IDM- ematical formulation is presented below:
based model:
2 Xi ðt þ 1Þ ¼ Xi ðtÞ
!β 3 (
ðjÞ
! )
max 4 Vi ðtÞ 5 j 1 ΔXi ðtÞ
ai ðt þ Ti Þ ¼ ai 1 ðdesÞ
þ min ð1 þ λÞ min min αuw þ βuw
Vi ðtÞ 1�j�N u2U w2W j
!2
XN
ΔXi
ðjamÞ equ
ðtÞ þ Vi ðtÞTi (36)
max
ai αj (31)
j¼1
ΔXequ where, t þ 1 represents the next discrete time step, αuw and βuw , for
ðu; wÞ 2 U � W are parameters to be estimated, and U and W are
where, αj represents weight coefficients which decrease mono two finite sets of indices. With this formulation, this study con
tonically as j increases to reflect that the influence of lead cluded that the minimum form used for considering more than one
vehicles decrease with an increase in the space gaps. They also lead vehicle and with a discounting factor ðλÞ used to favor the
concluded that the stable region increases after incorporat closest leader over distant ones. The proposed model was tested
ing MVA. only for single driver trajectory data. The authors reported para
meter identification issues when a large number of vehicles (the
limit was not reported) were considered. Also, their modeling
Extension of collision-avoidance models framework did not account for heterogeneity in driving behavior.
Table 3. (Continued).
Model Data utilized and
category Author(s) Proposed model region Strengths Weakness
Nirmale R
P
and Uijt ¼ βi0 þ βjr xir� þ �ij ; xir� : perceived traffic environment variable • Vehicular • The study addresses the issue of • Extent of acceleration is not modeled
r¼1 trajectory data parameter identification in MVA • The study only considers rectangular-
Pinjari collected at based driver behavior model shaped influence zone
(2023) Chennai city, • The empirical application of the • The model only considers longitudinal
India model provided interesting insights movements
into drivers’ perception errors. For
example, the perception errors are
greater for relative longitudinal
speeds than the longitudinal space
gaps and greater for vehicles
obliquely placed than those directly
ahead.
TRANSPORTATION LETTERS 13
models, e) Hazard-based models, f) Fuzzy-logic-based models, and modeling driving behavior in HD traffic streams. Moreover, we
g) game theory-based models. Gipps (1986) was the first to model divide this section into two subsections − 1) studies those model
a driver’s lane change behavior. His model is based on the notion of only longitudinal movements and 2) studies those models’ two-
collision avoidance, treats lane changing as a deterministic process, dimensional movements.
and ignores inconsistency in driver behavior over time. Building on
Gipp’s model’s shortcomings, Yang and Koutsopoulos (1996)
developed and implemented a probabilistic lane-changing model Multi-vehicle anticipation-based driver behaviour model for
similar to Gipps’ model in the microscopic traffic simulator describing longitudinal movements
MITSIM. They defined lane changes as mandatory or discretionary
and modeled the lane changing process as four consecutive steps: Extensions of Full Velocity Difference (FVD) model
deciding on a lane change, selecting a target lane, examining an Jin et al. (2010) proposed a non-lane-based FVD model by taking
acceptable distance, and performing the lane shift. Another weak into consideration the effect of the lateral gap between the subject
ness of Gipps (1986) is the assumption that lane change occurs only vehicle and the lead vehicle. It was hypothesized that a driver is
when a large enough gap exists in the target lane. However, in heavy affected by the vehicle directly ahead as well as adjacent to it.
or congested traffic, this assumption would be unrealistic. To over h � � i
ð1Þ ð2Þ
come this limitation, Hidas (2002, 2005) proposed an improved ai ðtÞ ¼ α U ΔXi ðtÞ; ΔXi ðtÞ Vi ðtÞ
modeling framework to capture the vehicular interaction induced h � �i
ð1Þ ð2Þ
by lane change, which was explicitly classified into three categories þ β G ΔVi ðtÞ; ΔVi ðtÞ (37)
based on observations from video-recording vehicular trajectory:
free, cooperative, and forced lane changes, to overcome this The following functional forms for U ð:Þ and Gð:Þ are employed in
limitation. their study:
Next, Ahmed et al. (1996) implemented utility theory to � � � �
model the decision process of lane change. The proposed ð1Þ ð2Þ
U ΔXi ðtÞ; ΔXi ðtÞ ¼ VOV ð1
ð1Þ ð2Þ
pi ÞΔXi ðtÞ þ pi ΔXi ðtÞ
model structure has four latent levels of decision hierarchy,
similar to steps given by Yang and Koutsopoulos (1996). (38)
Later, Toldeo (2003) provided an integrated modeling frame
work where the car-following and lane-changing behavior mod
� �
els were joined together in a single model. The integrated ð1Þ ð2Þ ð1Þ ð2Þ
G ΔVi ðtÞ; ΔVi ðtÞ ¼ ð1 pi ÞΔVi ðtÞ þ pi ΔVi ðtÞ (39)
model captured the trade-offs between the utility of being in
the correct lane and the speed advantage offered by a faster
lane. Furthermore, Kesting, Treiber, and Helbing (2007) pro
posed a novel logic for modeling lane changing decisions based ð1Þ
LSi
on lane change’s anticipated advantages and disadvantages. For pi ¼ (40)
LSmax
example, the driver attempts to minimize overall braking
induced by lane changes. Recently, game theory has received
where, VOV ð:Þ is the optimal velocity function provided in Equation
plenty of attention to model lane change behavior. Talebpour, ð1Þ
Mahmassani, and Hamdar (2015) and Ali et al. (2019) are a few (15); pi captures the effect of lateral separation distance; LSi is the
notable attempts in this regard. lateral separation distance between the center line passing through
the subject vehicle i and the center line passing through the first
lead vehicle (i.e. j ¼ 1). LSmax is the maximum lateral separation
Models that incorporate multi-vehicle anticipation by distance beyond which the lead vehicle has no influence on the
drivers in heterogeneous disordered traffic streams subject vehicle and set to 3.6 m, i.e. typical lane width. Further,
Most driver behavior models assume that vehicles follow lane when LSi ¼ 0 (i.e. pi ¼ 0), the proposed model can be simplified as
discipline, follow the center line of the lane, and consider stimuli the FVD model. When LSi ¼ LSmax (i.e. pi ¼ 1), the first lead
solely from vehicles in front since they are developed for homo vehicle is on another lane, and the subject vehicle follows
geneous traffic conditions. However, in many developing coun the second lead vehicle. This study demonstrated that incorporating
tries (such as India, Bangladesh, China, etc.), lanes may not be the effect of the lateral gap in the car-following model stabilizes
well delineated, or lane discipline may not be effectively main traffic flow, suppresses traffic jams, and increases capacity.
tained, allowing vehicles to occupy any lateral position on the The above model assumed that only one side of the subject
road and encouraging two-dimensional movement for the vehi vehicle had a laterally separated vehicle. However, the subject
cles to navigate ahead. In addition, vehicles from the side can cut- vehicle may be influenced by vehicles traveling laterally to either
in to the front of a subject vehicle anytime. Furthermore, HD side of it. Considering this, Li et al. (2015) developed a two-sided
traffic streams comprise a wide variety of vehicle classes (such as lateral gap FVD model for non-lane discipline traffic as given
passenger cars, motorbikes, buses, trucks, three-wheeled auto- below:
rickshaws, and non-motorized vehicles) with considerably differ h � � i
ð1Þ ð2Þ ð3Þ
ent physical and operational characteristics. Most of these classes ai ðtÞ ¼ α U ΔXi ðtÞ; ΔXi ðtÞ; ΔXi ðtÞ Vi ðtÞ
have substantial representation in the traffic streams. In contrast, h � �i
ð1Þ ð2Þ ð3Þ
homogeneous traffic streams are dominated mostly by passenger þ β G ΔVi ðtÞ; ΔVi ðtÞ; ΔVi ðtÞ (41)
cars having similar physical and operational characteristics. Such
distinctive characteristics of HD traffic streams cause differences It was assumed that the first ðj ¼ 1Þ, second ðj ¼ 2Þ, and third lead
in the driver behavior between homogeneous and HD traffic ðj ¼ 3Þ vehicles are traveling on the right front side, left front side,
streams (Nirmale, Pinjari, and Sharma 2022). The same can be and immediate front of the subject vehicle, respectively.
reflected in MVA behavior. Therefore, in this section, we mainly Accordingly, stimuli were calculated for each of the lead vehicles.
focus on the studies that have considered the MVA effect while The following functional forms for Uð:Þ and Gð:Þ were utilized:
14 S. K. NIRMALE ET AL.
� �
ð1Þ ð2Þ ð3Þ
U ΔXi ðtÞ; ΔXi ðtÞ; ΔXi ðtÞ Subsequently, Nirmale, Pinjari, and Sharma (2022) proposed
8 h i
< ð1Þ ð3Þ ð1Þ
V ð1 2pi ÞΔXi ðtÞ þ 2pi ΔXi ðtÞ "LSi 2 ð0; 0:5LSmax Þ
a panel data version of Nirmale, Pinjari, and Sharma (2021)’s
¼ h i multi-vehicle anticipation-based discrete-continuous choice mod
: V ð2pi 1ÞΔX ð2Þ ðtÞ þ 2ð1 pi ÞΔX ð3Þ ðtÞ "LSð1Þ 2 ð0:5LSmax ; LSmax Þ
i i i eling framework. Specifically, the modeling framework was
(42) enhanced by incorporating the influence of vehicle- and driver-
specific unobserved factors on driver behavior. In addition,
� � Nirmale, Pinjari, and Sharma (2022) study addressed the serial
ð1Þ ð2Þ ð3Þ
G ΔVi ðtÞ; ΔVi ðtÞ; ΔVi ðtÞ
( correlation issue by employing a simple empirical strategy. Using
ð1Þ ð3Þ ð1Þ
ð1 2pi ÞΔXi ðtÞ þ 2pi ΔXi ðtÞ"LSi 2 ð0; 0:5LSmax Þ this proposed model, driving behavior in HD and homogeneous
¼ ð2Þ ð3Þ ð1Þ (43)
ð2pi 1ÞΔXi ðtÞ þ 2ð1 pi ÞΔXi ðtÞ"LSi 2 ð0:5LSmax ; LSmax Þ traffic streams were compared by estimating empirical models
using trajectory data from Chennai, India (HD traffic) and recon
LS1i structed NGSIM data (homogeneous traffic).
pi ¼ (44) In another study, Nirmale and Pinjari (2023) incorporated dri
LSmax
ver errors in perceiving stimuli from vehicles directly ahead, obli
Similar to Jin et al. (2010), this study also employed the optimal quely placed, and on either side in their driver behavior model. The
velocity function as per Equation (15). Stability analysis of the empirical application of their model provided interesting insights
proposed model revealed that the model is more efficient in dis on drivers’ perception errors. For example, the perception errors
sipating perturbation than previous FVD-based models. are greater for relative longitudinal speeds than the longitudinal
space gaps, and greater for vehicles obliquely placed than those
Utility theory-based models directly ahead.
In HD traffic streams, the subject vehicle is surrounded by many
vehicles, resulting in multiple stimuli sources from multiple vehi
cles. Therefore, drivers might react to stimuli from a governing
Driver behaviour model for describing two-dimensional
leader. However, only the driver’s final actions (such as applied
movement
acceleration) are observed from trajectory data, and the governing
leader is latent to the analyst. This prompts the need to develop Models that can describe simultaneous lateral and longitudinal
a modeling framework to consider the latent leader while analyzing movements offer more realism in describing HD traffic. Yet, such
driver behavior. To do so, a latent leader approach was proposed by studies are rare (Mahapatra, Maurya, and Chakroborty (2018) and
Choudhury and Islam (2016) to model acceleration decisions. Chakroborty, Maurya, and Vikram (2019) provide a review of such
Particularly, a random utility-based modeling framework was pro efforts). In the ensuing paragraphs, a brief discussion of these
posed with two components: latent leader component and accel models is provided, along with some recent studies that are not
eration component. The former was modeled as a random utility- included in previous reviews.
based discrete choice model, and the probability of the front lead Chakroborty, Agrawal, and Vasishtha (2004) developed
vehicle l (which can be front left, front direct, front right) being a comprehensive microscopic model for two-way traffic using
a governing leader of the subject vehicle i was expressed as: a potential field approach. The proposed model had two major
� � components, namely, (1) Steering Response Model (SRM, for pre
j
exp βj Zi ðtÞ dicting the steering angles adoption with time) and (2) Acceleration
Pðli ðtÞÞ ¼ P � 0 0 � j; j0 2 J ¼ FL; FD; FR (45) Response Model (ARM, to predict the rate of acceleration and
j
exp βj Zi ðtÞ deceleration over time). Later, a Comprehensive Unidirectional
J
j
Traffic Simulation Model (CUTSiM) was proposed by Maurya
where, Zi ðtÞrepresents
the explanatory variables associated with the (2007) that considered the Indian traffic conditions explicitly. The
j
lead vehicle j and β is an estimated parameter vector associated proposed model included a lateral control model component that
with respect to the lead vehicle j. Whereas the acceleration compo describes the driver’s decision to choose a suitable steering angle
nent was modeled using the GM model. Strictly speaking, the latent based on the hypothesized best path along with its longitudinal
leader model is not an MVA-based model, as it assumes that single control model. However, this model was not extensively calibrated
(unknown) leader. and validated. Furthermore, Kanagaraj and Treiber (2018) pro
More recently, Nirmale, Pinjari, and Sharma (2021) proposed posed a two-dimensional time-continuous model for the mixed
an MVA-based discrete-continuous choice modeling framework traffic flow of motorized and non-motorized vehicles based on the
for describing driver behavior in HD traffic conditions. To incor force-field model. The trajectory data from Chennai, India, was
porate MVA, they introduced the concept of an influence zone utilized to calibrate this model. Recently, Delpiano et al. (2019)
around a vehicle (subject vehicle) where vehicles within the developed a two-dimensional microscopic car-following model
influence zone can potentially influence the subject vehicle’s using the social force approach. They argued that the distance
driving behavior. Moreover, driving decisions were separated maintained by the driver to avoid collisions in all directions is
into discrete and continuous components – discrete component a critical factor and thus, proposed the multi-directional collision
for the decision to accelerate, decelerate, or maintain a constant avoidance behavior model where two-dimensional repulsive force
speed and the continuous component for the extent of accelera between vehicles was modeled. Specifically, three forces were con
tion or deceleration. The estimation results from this study sidered that act on each vehicle: the acceleration force (willingness
underscored the importance of considering the MVA effect for to accelerate), lane force (tendency to be in the center in a specific
describing driving behavior in HD traffic conditions. Specifically, lane), and the repulsive force for collision avoidance. Simulation
their study concluded that drivers in HD traffic conditions not experiments were also performed to reproduce two-dimensional
only consider vehicles that are ahead of their vehicle but also collision avoidance behavior. They concluded that the proposed
consider those vehicles that are on either side. Also, they found model is a sound starting point for building autonomous vehicles
that the extent of influence of stimuli with respect to lead vehicles traffic flow models and can improve autonomous driving
is different on discrete decisions and continuous decisions. algorithms.
TRANSPORTATION LETTERS 15
Next, Mathew, Munigety, and Bajpai (2013) proposed a strip- etc.) and, thus, gives low to no weightage to stimuli from those
based approach for the simulation of HD traffic conditions. They vehicles. It has also been demonstrated that drivers not only react
developed a simulator named Simulation of Mixed Traffic Mobility to the vehicles directly ahead but also those that are placed
using a traditional lane-based simulator SUMO. In the proposed obliquely ahead or on either side of the subject vehicle, specifi
approach, the road was divided into thin strips allowing continuous cally in HD traffic streams. Furthermore, incorporating percep
lateral movement rather than conventional discrete lane changing. tion errors in MVA-based driver behavior models improved the
They observed that the reduction of strip width increases the behavioral soundness of these models. Nirmale and Pinjari (2023)
throughput indicating improved utilization of road space. found that the stimuli perception errors are greater for vehicles
Furthermore, Lee, Polak, and Bell (2009) proposed an agent- that are obliquely placed than vehicles that are directly ahead.
based model to simulate motorcycle behavior in HD traffic condi This may be because drivers pay greater attention to vehicles
tions. They developed three models to mimic motorcycle move directly ahead than those obliquely placed. In addition, all studies
ment patterns, namely, the longitudinal headway model, the have concluded that MVA-based models perform better in stabi
oblique and lateral headway model, and the path choice model. lizing the traffic and suppressing traffic oscillations than single-
The longitudinal headway model described the motorcyclist driving lead vehicle-based models. Treiber, Kesting, and Helbing (2006)
behavior that they maintain a shorter headway when aligning to the revealed that the model with MVA has a higher threshold for
edge of the lead vehicle. The oblique and lateral headway model traffic stability over the single leader model, i.e. traffic remains
described the headway distribution of motorcycles when they are stable for higher values of reaction times for the model with
following the lead vehicles obliquely. MVA, which is consistent with real-world traffic observations.
The virtual lane-based movement of motorcycles was modeled as Most of the studies were restricted to three lead vehicles ahead of
a multinomial logit. Next, Shiomi et al. (2012) proposed a utility the subject vehicle and that too those that were directly ahead of the
theory-based approach to describe the two-dimensional movement subject vehicle. The number of lead vehicles to be considered in the
of the two-wheelers. This approach captured the characteristics of MVA framework may depend on the type of lead vehicle as con
driver perception of the surrounding traffic situation albeit, it failed jectured by Hoogendoorn and Ossen (2006). In addition, surround
to capture the heterogeneity across the drivers. Following a similar ing vehicles that are not only straight ahead of the driver but also
utility theory-based approach, Sarkar et al. (2020) proposed a modeling obliquely placed influence drivers’ decisions (Nirmale, Pinjari, and
framework that models driver’s two-dimensional movement behavior. Sharma 2021). Specifically, in the HD traffic streams, it becomes
This framework proposed two components: area selection and vehicle more important to consider vehicles around the subject vehicle
movement. In the area selection component, the possible movements because of its typical characteristics (see Section 4). Surprisingly, the
of subject vehicles for next time steps were considered, and two- influence of vehicle type is largely ignored by all studies reviewed here
dimensional space ahead of the subject vehicle was divided into the except two (Hoogendoorn and Ossen 2006; Nirmale, Pinjari, and
number of realistic cones and treated as choice alternatives for the Sharma 2021). In fact, the number of vehicles considered by the
subject vehicle. A multinomial logit model was used to model the driver is always kept constant irrespective of the type of vehicle ahead.
driver’s discrete decisions. For the vehicle movement component,
a modified intelligent driver model was proposed, which simulates
the subject vehicle’s next position along the selected direction. Both
Modeling requirements, challenges and future research
components of the proposed framework were calibrated separately
directions
using a real trajectory data set collected at Chennai, India.
Recently, Amrutsamanvar (2020) analyzed the lateral movement In this section, we discuss common issues in the present MVA-based
decisions of two-wheeler drivers in HD traffic conditions using driver behavior modeling framework and future research needs.
different machine-learning (ML) models. The study framed the
lateral movement decision of two-wheeler drivers as a multiclass
classification problem, where the driver must select one alternative How many vehicles ahead to consider in MVA-based driver
from a set of three discrete options: turn left, turn right, or move behavior models?
straight. The study used vehicle trajectory data from Chennai,
India, to train the ML models. The study revealed that the lateral As discussed earlier, the influence of only a fixed number (say 3
distance to a ready-to-overtake position is more important than the or 5) of vehicles ahead is considered in these models. Hence, future
longitudinal gap between the lead and subject vehicles in influen endeavors can focus on investigating the impact of considering the
cing the lateral movement decisions of two-wheeler drivers. different number of vehicles ahead in both homogeneous and HD
traffic streams on the numerical and behavioral soundness of the
model. Similarly, the influence of side vehicle’s impact on driving
Critical findings behavior can be examined leading to modifications in the existing
This paper provides a critical review of notable attempts to MVA-based models.
incorporate MVA in driving behavior models. Primarily, MVA- In the recent study by Nirmale, Pinjari, and Sharma (2021),
based driver behavior models follow a similar pattern of extend a concept of an influence zone was proposed, and it was assumed
ing the single-leader car-following model by adding weighted that vehicles within the influence zone (irrespective of the number
stimuli from vehicles ahead. Interestingly, the influence of only and position with respect to the subject vehicle) might impact the
a fixed number (say 3 or 5) of vehicles ahead is considered in microscopic driving behavior of the driver. One can employ this
these models. Moreover, vehicles far ahead have a low influence approach to incorporate MVA. One of the limitations of the
on drivers’ maneuvering decisions, and hence, weights on the Nirmale, Pinjari, and Sharma (2021) study is that they used fixed
stimuli decrease as the distance from the subject vehicle increases. size and shape influence zone for all drivers. However, the shape
Furthermore, the type of vehicle ahead (a car or a truck) also and size might be different for drivers and different vehicle types
influences the MVA behavior of drivers and, thereby, their man of the subject vehicle. Hence, future research endeavors can
euvering decisions. For example, a car driver following a truck is explore different shapes and sizes for different drivers and vehicle
blind to vehicles ahead of the truck (second leader, third leader, types.
16 S. K. NIRMALE ET AL.
Additional critical factors to be considered in MVA-based Another important point is that the above enumerated human
driver behavior models factors remain mainly unobserved because the data collection pro
cess for building, calibrating, and validating driver behavior models
Influence of vehicle type
majorly focuses on gathering the trajectory data from videos or GPS
The traffic stream consists of different categories of vehicles, includ
devices. Another reason could be that some of these factors are
ing cars, motorbikes, buses, heavy vehicles, etc. The relative per
unknown to the analyst. Little effort has been made to acknowledge
centages of these vehicle types change depending on the type of
drivers’ unobserved factors and their implications on the modeling
stream (homogeneous or HD), type of facility (urban streets or
framework. A recent study by Nirmale, Pinjari, and Sharma (2022)
highways), geographical locations (hilly terrain vs plain terrain),
incorporated drivers’ unobserved factors latent to analysts to better
etc. MVA also depends on the type of vehicle. Let’s first consider the
mimic driver behavior in HD traffic streams. In future endeavors,
type of subject vehicle. The distance up to which vehicles ahead will
considering the effect of human factors explicitly or as unobserved
influence the driver will be larger for bus/truck drivers than car
factors is important to improve driver behavior models.
drivers because the eye level of bus/truck drivers is at a greater
Incorporating the aforementioned factors will make MVA-based
height; thus, can see for a larger distance as compared to car drivers
driver behavior models more realistic.
leading to differences in their MVA behavior. Second, the type of
vehicles in the visual field also influences the MVA behavior.
Imagine a situation that a car is following a bus/truck. Vehicles
ahead of the bus/truck will not be visible to the car driver; thus, Data needs for calibration and validation of MVA based driver
stimulus from those vehicles shall not be considered in the car behavior models
driver’s decision-making. Surprisingly, this factor is largely ignored
by all studies reviewed here except one (Hoogendoorn and Ossen We did not find studies focussing on collecting data specifically to
2006). In fact, the number of vehicles considered by the driver is capture multi-vehicle anticipation behavior, probably because most
always kept constant irrespective of the type of vehicle ahead. For of the time MVA-based models are simple extensions of single-lead
realistic modeling of MVA, the type of vehicle both at the subject vehicle-based models. As revealed in this study, MVA is not
vehicle level and at the traffic stream ahead shall be considered. a straightforward process. Before incorporating it, one needs to
answer how many vehicles a driver considers? What are the driver’s
focus areas? How to quantify a driver’s useful visual field? How does
Drivers’ useful visual field the useful visual field vary for a particular driver? How does MVA
As per Mackworth (1965), ‘the useful visual field can be defined change based on the type of subject vehicle (bus or car or motor
as the area around the fixation point (the point in space on bike) and on vehicle classes in front? To answer all these questions,
which the eyes are focused) inside which information can be it is important to collect detailed driver behavior data. For instance,
perceived.’ If the quantity of information to be processed researchers shall focus on collecting data up to a trap length of 100
increases, the useful visual field decreases (Rogé et al. 2004). m to 500 m so that kinematics of a few vehicles ahead can be
For example, the useful visual field decreases if the number of captured. Notably, driving simulators offer various advantages,
vehicles within the visual field increases. Moreover, it specifically the collection of detailed driver-level data. Various
decreases with the increase in speed. This indicates that the scenarios with a varying number of vehicles ahead, different vehicle
useful visual field will have a strong bearing on the MVA classes, varied placement of vehicles in the driving scene, etc., can
behavior of drivers because it directly influences the stimulus be created and tested. Next, eye trackers can be utilized to identify
perception by drivers. Therefore, future endeavors on building driver focus points within the visual field. Combining driving
MVA-based driver behavior models shall consider this impor simulators and eye-tracking data can assist in a better understand
tant factor. ing and modeling of the MVA behavior.
Journal of Modern Physics C 21 (5): 647–668. https://doi.org/10.1142/ Lee, T.-C., J. W. Polak, and M. G. Bell. 2009. “New Approach to Modeling Mixed
S0129183110015397. Traffic Containing Motorcycles in Urban Areas.” Transportation Research
Choudhury, C. F., and M. M. Islam. 2016. “Modelling Acceleration Decisions in Record 2140 (1): 195–205. https://doi.org/10.3141/2140-22.
Traffic Streams with Weak Lane Discipline: A Latent Leader Approach.” Lenz, H., C. Wagner, and R. Sollacher. 1999. “Multi-Anticipative Car-Following
Transportation Research Part C: Emerging Technologies 67:214–226. https:// Model.” The European Physical Journal B-Condensed Matter and Complex
doi.org/10.1016/j.trc.2016.02.010. Systems 7 (2): 331–335. https://doi.org/10.1007/s100510050618 .
Das, S., and A. K. Maurya. 2018. “Modelling of Motorised Two-Wheelers: Li, Z. P., and Y. C. Liu. 2006. “Analysis of Stability and Density Waves of Traffic
A Review of the Literature.” Transport Reviews 38 (2): 209–231. https://doi. Flow Model in an ITS Environment.” The European Physical Journal
org/10.1080/01441647.2017.1317049. B-Condensed Matter and Complex Systems 53 (3): 367–374. https://doi.org/
Delpiano, R., J. Herrera, J. Laval, and J. Coeymans-Avaria. 2019. “A 10.1140/epjb/e2006-00382-7.
Two-Dimensional Car-Following Model for Two-Dimensional Traffic Flow Li, Y., D. Sun, W. Liu, M. Zhang, M. Zhao, X. Liao, and L. Tang. 2011. “Modeling
Problems.” Transportation Research Part C: Emerging Technologies and Simulation for Microscopic Traffic Flow Based on Multiple Headway,
114:504–516. https://doi.org/10.1016/j.trc.2020.02.025. Velocity and Acceleration Difference.” Nonlinear Dynamics 66 (1–2): 15–28.
Eissfeldt, N., and P. Wagner. 2003. “Effects of Anticipatory Driving in a Traffic Flow https://doi.org/10.1007/s11071-010-9907-z.
Model.” The European Physical Journal B-Condensed Matter and Complex Li, Y., L. Zhang, S. Peeta, H. Pan, T. Zheng, Y. Li, and X. He. 2015. “Non-Lane-
Systems 33 (1): 121–129. https://doi.org/10.1140/epjb/e2003-00149-8 . Discipline-Based Car-Following Model Considering the Effects of Two-Sided
Farhi, N., H. Haj-Salem, and J. P. Lebacque. 2012. “Multianticipative Lateral Gaps.” Nonlinear Dynamics 80 (1): 227–238. https://doi.org/10.1007/
Piecewise-Linear Car-Following Model.” Transportation Research Record: s11071-014-1863-6.
Journal of the Transportation Research Board 2315 (1): 100–109. https://doi. Lu, B., S. Ni, and S. S. Washburn. 2015. “A Support Vector Regression Approach
org/10.3141/2315-11. for Investigating Multianticipative Driving Behavior.” Mathematical
Gazis, D. C., R. Herman, and R. B. Potts. 1959. “Car-Following Theory of Problems in Engineering 2015:1–10. https://doi.org/10.1155/2015/584954.
Steady-State Traffic Flow.” Operations Research 7 (4): 499–505. https://doi. Mackworth, N. H. 1965. “Visual Noise Causes Tunnel Vision.” Psychonomic
org/10.1287/opre.7.4.499. Science 3 (1): 67–68. https://doi.org/10.3758/BF03343023.
Gazis, D. C., R. Herman, and R. W. Rothery. 1961. “Nonlinear Follow-The- Mahapatra, G., A. K. Maurya, and P. Chakroborty. 2018. “Parametric Study of
Leader Models of Traffic Flow.” Operations Research 9 (4): 545–567. https:// Microscopic Two-Dimensional Traffic Flow Models: A Literature Review.”
doi.org/10.1287/opre.9.4.545. Canadian Journal of Civil Engineering 45 (11): 909–921. https://doi.org/10.
Ge, H. X., S. Q. Dai, L. Y. Dong, and Y. Xue. 2004. “Stabilisation Effect of Traffic 1139/cjce-2017-0686.
Flow in an Extended Car-Following Model Based on an Intelligent Mathew, T. V., C. R. Munigety, and A. Bajpai. 2013. “Strip-Based Approach for
Transportation System Application.” Physical Review E 70 (6): 066134. the Simulation of Mixed Traffic Conditions.” Journal of Computing in Civil
https://doi.org/10.1103/PhysRevE.70.066134. Engineering 29 (5). https://doi.org/10.1061/(ASCE)CP.1943-5487.0000378.
Gipps, P. G. 1981. “A Behavioural Car-Following Model for Computer Maurya, A. K. 2007. Development of a comprehensive microscopic model for
Simulation.” Transportation Research Part B: Methodological 15 (2): simulation of large uninterrupted traffic streams without lane discipline. PhD
105–111. https://doi.org/10.1016/0191-2615(81)90037-0. thesis, Kanpur: Indian Institute of Technology.
Gipps, P. G. 1986. “A Model for the Structure of Lane-Changing Decisions.” Moridpour, S., M. Sarvi, and G. Rose. 2010. “Lane Changing Models: A Critical
Transportation Research Part B: Methodological 20 (5): 403–414. https://doi. Review.” Transportation Letters 2 (3): 157–173. https://doi.org/10.3328/TL.
org/10.1016/0191-2615(86)90012-3. 2010.02.03.157-173.
Hasebe, K., A. Nakayama, and Y. Sugiyama. 2003. “Dynamical Model of Munigety, C. R., and T. V. Mathew. 2016. “Towards Behavioral Modeling of
a Cooperative Driving System for Freeway Traffic.” Physical Review E Drivers in Mixed Traffic Conditions.” Transportation in Developing
68 (2): 026102. https://doi.org/10.1103/PhysRevE.68.026102. Economies 2 (1): 6. https://doi.org/10.1007/s40890-016-0012-y.
Helly, W. 1959. “Simulation of Bottlenecks in Single-Lane Traffic Flow.” Newell, G. F. 1961. “Nonlinear Effects in the Dynamics of Car Following.”
Proceedings of the Symposium on Traffic Flow Theory, Research Operations Research 9 (2): 209–229. https://doi.org/10.1287/opre.9.2.209.
Laboratories, General Motors 207–238. Nirmale, S. K. 2022. Multi-Vehicle Anticipation-based Models for Describing
Hidas, P. 2002. “Modelling Lane Changing and Merging in Microscopic Traffic Driver Behaviour in Heterogeneous and Disorderly Traffic Conditions. PhD
Simulation.” Transportation Research Part C: Emerging Technologies 10 (5– thesis, Indian Institute of Science.
6): 351–371. https://doi.org/10.1016/S0968-090X(02)00026-8. Nirmale, S. K., and A. R. Pinjari. 2023. “Discrete Choice Models with
Hidas, P. 2005. “Modelling Vehicle Interactions in Microscopic Simulation of Multiplicative Stochasticity in Choice Environment Variables: Application
Merging and Weaving.” Transportation Research Part C: Emerging for Accommodating Perception Errors in Driver Behavior Models.”
Technologies 13 (1): 37–62. https://doi.org/10.1016/j.trc.2004.12.003. Transportation Research Part B: Methodological 170:169–193. https://doi.
Hoogendoorn, S., and S. Ossen. 2006. “Empirical Analysis of Two-Leader org/10.1016/j.trb.2023.02.014.
Car-Following Behavior.” European Journal of Transport and Infrastructure Nirmale, S. K., A. R. Pinjari, and A. Sharma. 2021. “A Discrete-Continuous
Research 6 (3): 229–246. Multi-Vehicle Anticipation Model of Driving Behaviour in Heterogeneous
Hoogendoorn, S., S. Ossen, and M. Schreuder. 2006. “Empirics of Disordered Traffic Conditions.” Transportation Research Part C: Emerging
Multianticipative Car-Following Behavior.” Transportation Research Record Technologies 128:103144. https://doi.org/10.1016/j.trc.2021.103144.
1965 (1): 112–120. https://doi.org/10.1177/0361198106196500112. Nirmale, S. K., A. R. Pinjari, and A. Sharma. 2022. “A Panel Data-Based
Hu, Y., T. Ma, and J. Chen. 2014. “An Extended Multi-Anticipative Delay Model Discrete-Continuous Modelling Framework to Analyse Longitudinal Driver
of Traffic Flow.” Communications in Nonlinear Science and Numerical Behavior in Homogeneous and Heterogeneous Disordered Traffic
Simulation 19 (9): 3128–3135. https://doi.org/10.1016/j.cnsns.2014.02.006. Conditions.” Transportation Letters 1–14. https://doi.org/10.1080/19427867.
Jiang, R., Q. Wu, and Z. Zhu. 2001. “Full Velocity Difference Model for a 2022.2132058.
Car-Following Theory.” Physical Review E 64 (1): 017101. https://doi.org/ Peng, G., and D. Sun. 2010. “A Dynamical Model of Car-Following with the
10.1103/PhysRevE.64.017101. Consideration of the Multiple Information of Preceding Cars.” Physics
Jin, S., D. Wang, P. Tao, and P. Li. 2010. “Non-Lane-Based Full Velocity Difference Letters A 374 (15–16): 1694–1698. https://doi.org/10.1016/j.physleta.2010.
Car Following Model.” Physica A: Statistical Mechanics and Its Applications 02.020 .
389 (21): 4654–4662. https://doi.org/10.1016/j.physa.2010.06.014. Punzo, V., M. Montanino, and B. Ciuffo. 2014. “Do We Really Need to Calibrate All
Jin, Y., M. Xu, and Z. Gao. 2011. “KdV and Kink-Antikink Solitons in an the Parameters? Variance-Based Sensitivity Analysis to Simplify Microscopic
Extended Car-Following Model.” Journal of Computational and Nonlinear Traffic Flow Models.” IEEE Transactions on Intelligent Transportation Systems
Dynamics 6 (1): 011018. https://doi.org/10.1115/1.4002336. 16 (1): 184–193. https://doi.org/10.1109/TITS.2014.2331453.
Kanagaraj, V., and M. Treiber. 2018. “Self-Driven Particle Model for Mixed Rogé, J., T. Pébayle, E. Lambilliotte, F. Spitzenstetter, D. Giselbrecht, and
Traffic and Other Disordered Flows.” Physica A: Statistical Mechanics and Its A. Muzet. 2004. “Influence of Age, Speed and Duration of Monotonous
Applications 509:1–11. https://doi.org/10.1016/j.physa.2018.05.086. Driving Task in Traffic on the Driver’s Useful Visual Field.” Vision
Kesting, A., M. Treiber, and D. Helbing. 2007. “General Lane-Changing Model Research 44 (23): 2737–2744. https://doi.org/10.1016/j.visres.2004.05.026.
MOBIL for Car-Following Models.” Transportation Research Record: Journal Saifuzzaman, M., and Z. Zheng. 2014. “Incorporating Human-Factors in
of the Transportation Research Board 1999 (1): 86–94. https://doi.org/10. Car-Following Models: A Review of Recent Developments and Research
3141/1999-10. Needs.” Transportation Research Part C: Emerging Technologies 48:379–403.
Krauss, S. 1998. Microscopic modeling of traffic flow: Investigation of collision https://doi.org/10.1016/j.trc.2014.09.008.
free vehicle dynamics. PhD thesis. Sarkar, N. C., A. Bhaskar, Z. Zheng, and M. P. Miska. 2020. “Microscopic
Krauss, S., P. Wagner, and C. Gawron. 1996. “Continuous Limit of the Modelling of Area-Based Heterogeneous Traffic Flow: Area Selection and
Nagel-Schreckenberg Model.” Physical Review E 54 (4): 3707. https://doi. Vehicle Movement.” Transportation Research Part C: Emerging Technologies
org/10.1103/PhysRevE.54.3707. 111:373–396. https://doi.org/10.1016/j.trc.2019.12.013.
TRANSPORTATION LETTERS 19
Sharma, A., Y. Ali, M. Saifuzzaman, Z. Zheng, and M. M. Haque. 2017. “Human Walker, J. 2001. Extended discrete choice models: integrated framework, flexible error
Factors in Modelling Mixed Traffic of Traditional, Connected, and structures, and latent variables. PhD thesis, Massachusetts Institute of Technology.
Automated Vehicles.” International Conference on Applied Human Factors Wang, T., Z. Y. Gao, and X. M. Zhao. 2006. “Multiple Velocity Difference Model
and Ergonomics 262–273. and Its Stability Analysis.” Acta Physica Sinica 55 (2): 634. https://doi.org/10.
Shiomi, Y., T. Hanamori, N. Uno, and H. Shimamoto. 2012. “Modeling Traffic 7498/aps.55.634.
Flow Dominated by Motorcycles Based on Discrete Choice Approach.” WHO. (2018). Global Status Report on Road Safety 2018: Summary.
Proceedings of 1st LATSIS Conference. Wilson, R., P. Berg, S. Hooper, and G. Lunt. 2004. “Many-Neighbour Interaction
Talebpour, A., H. S. Mahmassani, and S. H. Hamdar. 2015. “Modeling and Non-Locality in Traffic Models.” The European Physical Journal
Lane-Changing Behavior in a Connected Environment: A Game Theory B-Condensed Matter and Complex Systems 39 (3): 397–408. https://doi.org/
Approach.” Transportation Research Part C: Emerging Technologies 10.1140/epjb/e2004-00205-y.
59:216–232. https://doi.org/10.1016/j.trc.2015.07.007. Yang, Q., and H. N. Koutsopoulos. 1996. “A Microscopic Traffic Simulator for
Toledo, T. 2003. Integrated driving behavior modeling. PhD thesis, Evaluation of Dynamic Traffic Management Systems.” Transportation
Massachusetts Institute of Technology. Research Part C: Emerging Technologies 4 (3): 113–129. https://doi.org/10.
Toledo, T. 2007. “Driving Behaviour: Models and Challenges.” Transport 1016/S0968-090X(96)00006-X.
Reviews 27 (1): 65–84. https://doi.org/10.1080/01441640600823940 . Yu, L., Z. Shi, and B. Zhou. 2008. “Kink–Antikink Density Wave of an Extended
Treiber, M., A. Hennecke, and D. Helbing. 2000. “Congested Traffic States in Car-Following Model in a Cooperative Driving System.” Communications in
Empirical Observations and Microscopic Simulations.” Physical Review E Nonlinear Science and Numerical Simulation 13 (10): 2167–2176. https://doi.
62 (2): 1805–1824. https://doi.org/10.1103/PhysRevE.62.1805. org/10.1016/j.cnsns.2007.07.008.
Treiber, M., and A. Kesting. 2013. Traffic Flow Dynamics. Traffic Flow Zhang, X. 2014. “Empirical Analysis of a Generalised Linear Multianticipative
Dynamics: Data, Models and Simulation. Springer-Verlag Berlin Car-Following Model in Congested Traffic Conditions.” Journal of
Heidelberg. https://doi.org/10.1007/978-3-642-32460-4. Transportation Engineering 140 (6): 04014018. https://doi.org/10.1061/
Treiber, M., A. Kesting, and D. Helbing. 2006. “Delays, Inaccuracies and (ASCE)TE.1943-5436.0000667.
Anticipation in Microscopic Traffic Models.” Physica A: Statistical Zheng, Z. 2014. “Recent Developments and Research Needs in Modeling Lane
Mechanics and Its Applications 360 (1): 71–88. https://doi.org/10.1016/j. Changing.” Transportation Research Part B: Methodological 60:16–32.
physa.2005.05.001. https://doi.org/10.1016/j.trb.2013.11.009.
20 S. K. NIRMALE ET AL.
Appendix A. Notations
ΔXi
ðj;desÞ
ðtÞ Desired following distance (m) for subject vehicle i at time t with respect to lead vehicle j
ðjamÞ
ΔXi ðtÞ Minimum spacing at the standstill situation
ðminÞ
ΔXi Minimum space headway for subject vehicle i
ðsafeÞ
ΔXi ðtÞ Safe distance
ΔXequ Space gap at equilibrium
Li Length of vehicle i
LSi
ðjÞ Lateral separation between subject vehicle i and lead vehicle j
LSmax Maximum lateral separation
ðjÞ
Δai ðtÞ Acceleration difference between lead vehicle j and subject vehicle i at time t
ðcomf Þ
di Comfortable or desired deceleration
Ti Reaction time or delay time of subject vehicle i
Ti
ðdesÞ Desired time headway of vehicle i
TGap Time gap
Δt Update time
ðjÞ
ðdesÞ
Vi ðΔXi ðtÞÞ
ðjÞ Desired speed of subject vehicle i for a given space headwayΔXi ðtÞat time t
Uij Utility that the driver of the ith vehicle perceives from choosing a maneuvering decision j
Zij ; Kij Traffic environment variables (such as space gap, relative speed) with respect to the driver of the ith vehicle and maneuvering
decision j
Senð:Þ Sensitivity function
Stimð:Þ Stimulus function