Transportation Research Part F 32 (2015) 114126

Contents lists available at ScienceDirect

Transportation Research Part F

journal homepage: www.elsevier.com/locate/trf

Models for pedestrian gap acceptance behaviour analysis

at unprotected mid-block crosswalks under mixed trafc
conditions
B Raghuram Kadali a,, P. Vedagiri a,1, Nivedan Rathi b,2
a
Transportation Systems Engineering Group, Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai 400 076, India
b
Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai 400 076, India

a r t i c l e i n f o a b s t r a c t

Article history: Pedestrian safety is an important aspect while crossing the road and it can be explained by
Received 18 November 2014 pedestrian gap acceptance behaviour. The statistical models such as multiple linear regres-
Received in revised form 9 April 2015 sion (MLR) is often used to model linear relationships between dependent variable (viz.,
Accepted 25 May 2015
pedestrian gap acceptance behaviour) and independent variables, due to their ability to
Available online 14 June 2015
quantitatively predict the effect of various factors on the dependent variable. However
such linear models cannot consider the effect of several variables on the output variable,
Keywords:
due to primary assumptions of normality, linear, homoscedasticity and multicollinearity.
Pedestrian
Mid-block
In this regard, the non-linear models based on the articial neural network (ANN), which
Gap are free from assumptions of linear models, can be easily employed for obtaining the effect
Rolling behaviour of several input variables on the pedestrian accepted gap size. However, researchers have
ANN rarely applied ANN modelling technique for predicting the pedestrian gap acceptance
Mixed trafc behaviour, as the pedestrian gap acceptance behaviour depends on several pedestrian, traf-
c and vehicular characteristics. The ANN based models would be quite useful in establish-
ing relationship between these factors on the pedestrian gap acceptance behaviour at
midblock crosswalks under mixed trafc conditions. In this direction, the present study
adopts both MLR as well as ANN with different pedestrian, trafc and vehicular character-
istics to assess the signicant contributing factors for pedestrians gap acceptance beha-
viour at unprotected mid-block crosswalks under mixed trafc conditions. For this
purpose, a video graphic survey was conducted at a six lane divided road at unprotected
mid-block crossing in Mumbai, India. The data such as pedestrian (gender and age), vehic-
ular, trafc and pedestrian behavioural characteristics were extracted to model pedestrian
accepted gaps. The model results show that pedestrian rolling behaviour has a signicant
effect on pedestrian accepted gap size. The model results concluded that ANN has a better
prediction with possibility to consider the effect of more number of variables on the
pedestrian gap acceptance behaviour as compared to the MLR model under mixed trafc
conditions. However, the quantication of signicant contributing variables on pedestrian
accepted gap size is easy by MLR model as compared to the ANN technique. So, both mod-
els have their own signicant role in pedestrian gap acceptance analysis. The developed
models may be useful to enhance the existing mid-block crosswalk facilities or planning

Corresponding author. Tel.: +91 8879 4257 55.

E-mail addresses: raghukadali@civil.iitb.ac.in (B.R. Kadali), vedagiri@civil.iitb.ac.in (P. Vedagiri), rathinivedan@civil.iitb.ac.in (N. Rathi).
1
Tel.: +91 2225 7673 07.
2
Tel.: +91 8286 3488 83.

http://dx.doi.org/10.1016/j.trf.2015.05.006
1369-8478/ 2015 Elsevier Ltd. All rights reserved.
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 115

new facilities by more accurate prediction of the pedestrian gap acceptance behaviour
considering the inuence of various factors under mixed trafc conditions.
2015 Elsevier Ltd. All rights reserved.

1. Introduction

The encouragement of active modes of transportation such as walking and bicycle is more benecial for urban transporta-
tion system to increase the transportation sustainability in developing countries like India. Consequently, walking
(pedestrian trips) are also integrated with other modes of transportation such as public transportation system and other
motorized vehicles (eg. Para transit). The pedestrian walk trips encompass of sidewalks as well as crosswalks during the
course of travel. The pedestrians have freedom to choose their mobility and maneuver on sidewalks when compared to
crosswalks. The mobility of pedestrian mainly depends on availability of adequate gaps in vehicular trafc as well as driver
behaviour due to non-availability of regulations such as sign boards, signals and speed controls at unprotected mid-block
crosswalks. Hence, unprotected mid-block crosswalk locations are chaotic to pedestrians as well as vehicle drivers.
Pedestrians have to cross the road at intersection location or designated crosswalk location (in India designated crosswalk
refers to median openings) in the course of the trip. In India, as per Indian Road congress (IRC) the unprotected mid-block
crosswalk refer to the median opening with or without marking (IRC:103) to access the adjacent land use condition, as
shown in Fig. 1. These mid-block crosswalks rarely have signals as well as sign boards to regulate motorized vehicles.
Further, the driver yield behaviour is not common at such mid-block crossings unless pedestrians are already using cross-
walk area. So, road crossing does not guarantee pedestrians safety, particularly when a pedestrian chooses unprotected
crosswalk locations under mixed trafc conditions. Moreover, the behaviour of pedestrian and vehicular drivers will affect
the pedestrian safety at unprotected mid-block crossings under mixed trafc conditions.
Studies show that nearly 1.2 million persons are killed every year and 50 million people are injured because of road
crashes globally (Peden et al., 2004). It is also found that approximately 65% of the total road crashes relate to pedestrians
(Kareem, 2003). The recent road crashes statistics in the US show an increase of 3 percent in pedestrian fatalities from 2010
to 2011 and almost 73 percent pedestrian crashes occurred in urban areas. Moreover, 70 percent of pedestrian crashes are
related to non-intersection locations (mid-block crossings) (NHTSA, 2011). In India, statistics show that 60 percent victims
are pedestrians and 85 percent pedestrian fatalities occur at mid-block crossings (Mohan, Tsimhoni, Sivak, & Flannagan,
2009). Some studies have shown that pedestrian crossing activities have higher fatality rate as compared to other activities
such as walking and standing on roadside (Kumar & Parida, 2011). All the above collision studies show that pedestrian crash
share is higher at unprotected mid-block crossings when compared to intersection location. In India, it can be noticed that
there are several unprotected mid-block crossings (median opening shown in Fig. 1) in urban areas, and pedestrian may
cross at such mid-block crossings. The developing countries have high population densities in urban areas due to the rapid
growth of economy and it increases the vehicular ow density as well as pedestrian trips substantially. Further, high pedes-
trian density necessitates several unprotected crosswalk trips and hence requires provision of more number of such

Location 1 Location 2 Location 3

Location 4 Location 5 Location 6

Fig. 1. Sample photos of unprotected mid-block crosswalk locations in Mumbai.

116 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

unprotected crosswalk locations, which affects vehicle driver yielding behaviour drastically. Due to lack of sufcient funds in
developing countries like India, most of these crosswalks do not have minimum pedestrian facilities to regulate the pedes-
trian and vehicular ow (Leather, Fabian, Gota, & Mejia, 2011). This further increases the probability of pedestrian fatalities
at un-protected crosswalk locations. Hence, increase in such unprotected crosswalks as well as pedestrian fatalities at such
locations motivated the authors to study the signicant contributing variables on pedestrian road crossing behaviour under
mixed trafc condition.
In India, vehicular trafc density increases tremendously due to rapid growth/inux of population in urban areas and it
makes problems to non-motorized modes in particular to pedestrians. The existing literature shows that there are several
pedestrian related collision statistics reported at mid-block crosswalk locations in India due to non-availability of pedes-
trian facilities. So, the unprotected mid-block crosswalks are critical elements for pedestrian safety in India. Moreover, the
availability of adequate gaps under mixed trafc conditions on the roads in cities of developing countries is rare, and hence
the pedestrians may use different behavioural characteristics (increase in speed, change of path condition, etc.) to accept
the small vehicular gaps at unprotected mid-block crosswalks. There is a possibility of several conict points with this
behaviour of pedestrian further supplemented by the change of vehicle driver behaviour. In addition to this, due to change
in pedestrian behaviour the accepted vehicle gap size drastically decreases and it contributes more in severity to the
pedestrian-vehicle conict at unprotected crosswalks. It is essential to predict the pedestrian accepted gap size more accu-
rately considering different road crossing behaviour to enhance pedestrian safety at unprotected mid-block crosswalk loca-
tions under mixed trafc conditions. Hence, it is necessary to understand the effect of these variables on the pedestrian gap
acceptance. So, in this context the objective of the study is to predict the pedestrian accepted gap size more accurately and
to understand the effect of signicant contributing factors (including pedestrian behavioural characteristics) over the
accepted gap size by using two well-known modelling methods, such as multiple linear regression (MLR) as well as arti-
cial neural networks (ANN). The organization of this research paper is as follows: Section 1 is a brief introduction of
pedestrian crossings and gap acceptance mechanism; Section 2 describes the review of early studies about the pedestrian
gap acceptance models. The data collection and extraction process and model formulation is presented in Section 3.
Section 4 presents the model results. Section 5 describes the discussion of model results. The conclusions are summarized
in Section 6.

2. Review of earlier studies

In general, at unprotected mid-block crossings, pedestrians may wait for more time at the curb and scan the available
vehicular gaps in each lane before crossing the road. Further, vehicular driver may be unable to yield to pedestrians and
pedestrians become impatient, and they try to use forced gaps (by using different pedestrian hand gestures) to cross the
road. Sometimes, pedestrians may not succeed because of the higher vehicle speed as well as non-yielding driver behaviour
(Hakkert, Gitelman, & Ben-Shabat, 2002). In pedestrian gap acceptance mechanism, several components come into the pic-
ture such as pedestrian behaviour (increase in speed, path change condition, etc.) and effect of vehicular characteristics.
Pedestrian road crossing behaviour may be well explained by gap acceptance mechanism. There are two important compo-
nents involved in gap acceptance mechanism; one is the decision-making process of the available gap to cross or not and the
second is a selection of suitable vehicular gap size (Dipietro & King, 1970). Several research studies have been carried out to
explore the pedestrian road crossing behaviour by using the gap acceptance mechanism taking into account several factors
such as pedestrian individual factors (gender and age), driver behaviour, vehicular characteristics and roadway characteris-
tics (Bernhoft & Carstensen, 2008; Oxley, Fildes, Ihsen, Charlton, & Day, 1997). Studies have explored pedestrian road cross-
ing behaviour with a discrete choice model by considering pedestrian age, environmental factors and trafc conditions
(Himanen & Kulmala, 1988). The study found that number of pedestrians, vehicle speed as well as vehicle platoon and pedes-
trian distance from the curb are contributing factors in pedestrian gap acceptance. Researchers have developed binary logit
model and probabilistic models to study the pedestrian accepted gap size as well as choice of accepting the gap with driver
yield behaviour (Sun, Ukkusuri, Benekohal, & Waller, 2003). Studies have addressed the pedestrian decision-making process
at signalized intersection location and results show that pedestrian have shorter gap size while crossing from the median
than the curb side (Das, Manski, & Manuszak, 2005). Researchers have also explored the impact of pedestrians
non-compliant behaviour on road crossing (Cherry, Donlon, Yan, Moore, & Xiong, 2012; Wang, Wu, Zheng, & McDonald,
2010). Further, researchers have investigated the effect of cultural as well as attitude related factors on the pedestrian beha-
viour for evaluation of safety and they concluded the existence of vertical collectivism and minimal risk taking behaviour
(Nordfjrn & S imsekoglu, 2013). Studies have proven that pedestrian behaviour also changes with type of crossing facility
and in this regard the uncontrolled pedestrian crossing locations have higher threat because of increase in non-compliant
behaviour of pedestrians (Sisiopiku & Akin, 2003).
Several researchers have attempted to model pedestrian accepted vehicular gap with numerous factors (pedestrian gen-
der, age, vehicle characteristics, etc.) by using multiple linear regression (MLR) technique, however, these studies included
limited variables in the models (Kadali & Vedagiri, 2012, 2013; Yannis, Papadimitriou, & Theolatos, 2013). The results of
these linear models are quite easy to interpret (viz., the effect of each individual variable on accepted gap size by their sign
or statistics (t-statistic)). However, these linear models have some limitation with their assumptions such as independence,
equal variance, normality of errors and multicollinearity problem. Moreover, linear model results show the low performance
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 117

value (R2) as well as poor correlation between the dependent and independent variables, when compared to non-linear mod-
els (Abyaneh, 2014). As a result of this, the variables could not satisfy the statistical test and several inuencing variables
were ignored by the linear model. The pedestrian accepted gap under mixed trafc conditions is quite different when com-
pared with homogeneous trafc based system. There are numerous contributing factors (trafc and pedestrian behaviour
related characteristics) which inuence the pedestrian accepted gap size, hence omitting such contributing variables
because of statistical insignicance is unreasonable. These limitations of the linear models can be overcome by non-linear
models such as ANN techniques, which can consider the effect of various parameters on pedestrian accepted gap under
mixed trafc scenario.
Some studies have been conducted on neural networks to solve transportation problems in last two decades and these
studies are mainly related to road trafc behaviour (Dougherty, 1995). Researchers have shown that decision making process
is more successfully predicted by ANN technique such as overtaking manoeuvres (Lyons, 1995), decision making for mode
choice (Hensher & Ton, 2000) and pedestrian gap acceptance (Pant & Balakrishnan, 1994). The driver route choice decision is
also important aspect in transportation study as explored by various researchers and these studies were successful in pre-
dictions of driver route choice with user perception survey data (Yang, Kitamura, Jovanis, Vaughn, & Abdel-Aty, 1993).
Studies also investigated the prediction of measurable quantities such as origindestinations (OD) matrix (Lorenzo &
Matteo, 2013) and travel time estimation (Lee, 2009).
Further, the major utilization of ANN modelling in transportation is crash prediction. In this line, several studies were car-
ried such as marine collision prediction with different input data (vehicle type weather condition etc.) (Hashemi, Le Blanc,
Rucks, & Shearr, 1995), crash prediction with different driver characteristics (Kalyoncuoglu & Tigdemir, 2004), urban colli-
sion prediction with vehicular ow, road geometry and vehicle characteristics (Mussone, Ferrari, & Oneta, 1999), freeway
collision prediction (Chang, 2005), and studies were extended to injury severity in trafc collisions (Delen, Sharda, &
Bessonov, 2006).
In transportation, more successful studies with ANN technique was applied for the vehicle detection as well as classi-
cation problems from the collected video graphic data (Bullock, Garrett, & Hendrickson, 1993; Xiong, He, Park, Cooley, & Li,
2009). Researchers have extended ANN prediction techniques in prediction of congested networks (Taylor, 1995), and in
incident detection problems (Dia & Rose, 1997; Ritchie & Cheu, 1993). Researchers have also investigated the effect of noise
near air ports on the residential areas by employing the ANN technique (Collins & Evans, 1994). In this study the prediction
of the important variables was done by using elasticity of neural networks. There are several studies conducted to compare
the neural networks with linear regression models in transportation (Duliba, 1991; Karlaftis & Vlahogianni, 2011) and study
results reported that ANN has strong prediction capability in comparison to the linear regression models.
Moreover, researchers also employed ANN modelling technique in pedestrian gap acceptance behaviour modelling for
decision making (Pant & Balakrishnan, 1994), however no studies have been reported on prediction of accepted gap size
under mixed trafc conditions. ANN models have better prediction capability than the linear models and these models
are free from assumptions like MLR model. Further, non-linear relationship between input and output variables in ANN mod-
els, which simulates the actual relationship between these variables will result better prediction capability. So, the ANN
modelling technique is a better choice instead of MLR models with considerations of wide range of variables, which may
better explain the pedestrian gap acceptance behaviour at unprotected mid-block crossings under mixed trafc conditions.
However, ANN models have difculty in interpretation of the effect of the input variables on the output, which may be easy
with linear models (MLR). Consequently, it is important to identify the important contributing variables as well as effect of
these variables on pedestrian gap acceptance behaviour under mixed trafc conditions and the selection of suitable model
(linear or non-linear) may depend on the purpose of the study such as evaluation of existing facilities (viz., the effect of var-
ious factors on output variable is desired) or facility design (viz., identifying the signicant variables affecting output variable
is desired). In this line, the objective of the present study is to predict the pedestrian accepted gap size by using both models
(ANN and MLR) at unprotected mid-block crosswalk location under mixed trafc conditions.

3. Methodology

In general, unprotected mid-block crosswalks are median openings (usually without signboards or any specic control
over vehicle drivers), for allowing pedestrian crossing at various land-use locations (viz., shopping malls, hospitals, schools,
commercial areas, etc.) which are locations of high pedestrian density. These crosswalks vary from four lanes to six lanes
with trafc barrier. The unprotected mid-block crosswalk selected in this study (refer location 2 in Fig. 1) represents a typical
unprotected mid-block crosswalk in developing countries, and it has well sight distance as well as pavement condition. It is
also connected to the public transport bus stops on either side of vehicular ow which is common component of daily trip of
pedestrians in urban conditions. The selected mid-block crosswalk location was partially protected by white painted mark-
ing and away from the signalized intersection. Also, the selected mid-block crosswalk has uniform roadway width (10.5 m)
on each direction of vehicular ow and signicant sidewalk width on either side of crosswalk. The video graphic survey was
conducted at selected six lanes divided unprotected mid-block crossing in Worli, Mumbai, India (Location 2 in Fig. 1). Three
high resolution cameras were used in order to collect pedestrian individual, behavioural and vehicular characteristics. Two
video cameras were placed along the vehicle ow direction in order to nd out the vehicle speed and type and third camera
was placed perpendicular to ow direction to nd out the pedestrian related characteristics. The data were recorded for a
118 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

period of two hours during peak hour of 4 PM to 6 PM because of high pedestrian crossing density during this time which
leads signicant of pedestrian-vehicle interaction. The recorded video was reviewed in laboratory by using AVS video editor
software for necessary data extraction. The data were extracted by using AVS video editor slow forward option at an accu-
racy of 33 ms (30 frames per second). The total data was extracted by two well-trained persons. The collected discrete data is
given in Table 1 and continuous variables data is summarized in Table 2.
The vehicular speeds were collected at selected location by marking longitudinal known trap length of 20 m. The
vehicular gap size was extracted corresponding to intersecting imaginary line of pedestrian path as well as vehicle
movement. Pedestrian behavioural characteristics may change with each available vehicular gap and due to the behavioural
characteristics at selected location such as rolling behaviour condition where the pedestrian may not wait for long gaps and
they move over successive small vehicular gaps (Kadali & Vedagiri, 2013); pedestrian path as well as speed changes condi-
tion; using a cell phone; group behaviour; frequency of an attempt, etc. The extracted data consists of 384 (accepted) gap
data points.

3.1. Model formulation

The present study considers two different modelling techniques such as linear (MLR) and non-linear (ANN) models to pre-
dict the pedestrian gap acceptance behaviour. The main drawback of the MLR model is multicollinearity problem due to
which several variables are eliminated from model results due to insignicance of test statistics. However, these problems
can be rectied with ANN technique due to its massive structure of input over the output connected architecture. This archi-
tecture can learn and map the relationship between input and the output. This ANN architecture is also free from the prior
assumptions. Moreover, the ANN modelling technique is best option for the predicting pedestrian gap acceptance behaviour
because of non-linear effect of input variables. For evaluation and design of pedestrian crosswalk facilities, both models were
applied to predict pedestrian gap acceptance. Further, both the models are discussed in the following section.

3.1.1. Multiple linear regression model (MLR)

Multiple linear regression (MLR), are the most often used modelling method for developing linear relationship between
independent and dependent variables or for studying the effect of signicant variables on dependent variable. MLR models
are quite simpler and easier as compared to other models (viz., non-linear models based on fuzzy logic, genetic algorithm,
neural networks, etc.) to interpret the results. In general, in the MLR model the deterministic component comprises of inter-
cept, coefcients of each independent variable and residual error terms. The unobserved quantity from the model results can
be represented by the intercept. Further, there are two types of coefcients in MLR models i.e., standardized and unstandard-
ized coefcients. Researchers have shown that standardized coefcients (viz., coefcients of the variables standardized with
respect to their mean and standard deviation values) are unit free coefcients and more appropriate than the unstandardized
and these model results are suitable to compare with ANN model results (Fitrianto & Midi, 2013; Gelman & Hill, 2007; Pao,
2008). Hence, in the present study, MLR model was developed with standardized values (which are free from intercept value)

Table 1
Description and coding of selected input variables.

Serial no. Variables Denition

1 Frequency of attempt
2 Rolling behaviour
3 Speed change condition
4 Age
5 Gender
6 Pedestrian crossing direction
7 Crossing path change condition
8 Type of vehicle
9 Pedestrian platoon size
10 Usage of cell phone
11 Type of gap
The number of attempts made by pedestrian to cross the road
Whether pedestrian rolls over the available small gaps (Yes = 2; No = 1)
Whether a pedestrian speed changes during different stage of crossing (Yes = 2; No = 1)
Age of the pedestrian by visual appearance (Elder = 4; Middle age = 3; Young = 2; Child = 1)
Gender of the pedestrian (Female = 2; Male = 1)
Whether a pedestrian start from curb or median (Median = 2; Curb = 1)
Whether a pedestrian changes crossing path while crossing the road (Yes = 2; No = 1)
Type of vehicle (Heavy = 5; Car = 4; Three wheeler = 3; Two wheeler = 2)
Number of pedestrians in a group (Three or more = 3; Two = 2; Single = 1)
Whether the pedestrian was engaged on his/her cellphone while crossing the road
(Yes = 2; No = 1)
The available vehicular gaps in lane one (near), second lane (Far 1) and third lane (Far 2)
Categorical (Near = 1; Far1 = 2; Far2 = 3)

Table 2
Statistics of continuous input variables.

Serial no. Variable Minimum Maximum Mean Standard deviation

1 Gap size in seconds 2.24 21.96 8.068 3.93
2 Vehicle speed in km/h 16.72 76.69 46.34 12.37
3 Waiting time in seconds 0.04 86.4 17.94 20.89
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 119

and the gap was the response variable and pedestrian individual, behavioural, vehicular and trafc characteristics were
explanatory variables. The explanatory variables were standardized by subtracting the mean value from it and further divid-
ing by the standard deviation, so that the effect of intercept in the model estimation process can be overcome. As a result, the
model can estimate the inuence of each explanatory variable over the response variable (Pao, 2008). The standardized MLR
results can interpret the change in standard deviation of the output variable with unit change of the standard deviation of
corresponding independent variable. In order to quantify the minimum accepted vehicular gap size the following multiple
linear regression (MLR) equation is proposed with selected explanatory variables.
Gap b1 Gen b2 age b3 FA b4 Rbeh b5 PSCC b6 PPCC b7 PCD b8 TV b9 PPS b10 UCP b11 TOG
b12 VS b13 WT 1
where Gap: Accepted vehicular gap size second, b0: intercept value, b1, b2. . . b13 are slopes associated with variables like Gen,
Age . . . WT, Gen: Pedestrian gender, Age: Pedestrian age, FA: Frequency of attempt, Rbeh: Rolling behaviour, PSCC:
Pedestrian speed changes condition, PPCC: Pedestrian path changes condition, PCD: Pedestrian crossing direction, TOV:
Type of vehicle, PPS: Pedestrian platoon size, UCP: Usage cell phone, TOG: Type of gap, VS: Vehicle speed, WT: Pedestrian
waiting time.

3.1.2. Articial neural network (ANN) model

Generally, multi-layer perception with feedforward back propagation (BP) algorithm is a common form of neural network
(Zain, Haron, & Shari, 2010). The back propagation neural network is conveyed with the root mean square error (RMSE) and it
is a general error function that is used in neural networks (Menhaj, 2007). The feed forward BP neural network comprises of
one input layer, one output layer and one or more hidden layers. The non-linear relationship between input and the output
variables can be formulated by hidden layer as a function of weighted summation and the transformation functions. The
layer in the network consists of several neurons and connected with adjacent layer neurons. In order to understand the rela-
tionship of input over the output, the neural network architecture is trained with input and output data. Fig. 2 shows general
architecture of multilayer feed-forward ANN. The input layer consists of thirteen neurons of selected variables (viz., gender,
age, frequency of attempts, rolling behaviour, pedestrian speed change condition, pedestrian path change condition, pedes-
trian crossing direction, type of vehicle, pedestrian platoon size, cell phone usage, type of gap, vehicle speed and waiting
time). The input vector receives the signal from each neuron of the previous layer and each of the input neuron makes col-
lective weight. The net input is calculated at the kth hidden node in the hidden layer as follows:
X
N
net hidden wn;k xn bk 2
i1

where wn,k is the weight between the input neuron nth to kth hidden neuron, xn is the input data from the ith input neuron, N
is the total number of input neurons (thirteen selected variables), and bk is the biases on the kth hidden node. In the hidden
layer, each node uses a transfer function to generate output (accepted gap size in second).
Hk f net hidden 3
The net input from each of the hidden nodes to the output node is expressed as follows:

Input Layer Output Layer

Hidden Layer
Gender X1 w 11

H1
Age
X2
v11

Frequency H2 v21 Accepted

X3
of attempt O1 gap size
v31
H3 vk1

Vehicle
Xi
speed Hk
wnk

Waiting
Xn
time

Fig. 2. The general architecture of multilayer feed-forward articial neural network.

120 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

X
k
net output V k;y Hk bk 4
k1

where k is the total number of hidden nodes, Vk is the weight from the kth hidden node to the output node, Hk is the
weighted sum of the transferred function to the output. In the output layer output generated by transfer function Oy as
follows:
b
Oy f net output Y 5
For different applications, there are different transfer functions such as log-sigmoid transfer function (logsig) linear trans-
fer function (purelin), hyperbolic tangent sigmoid transfer function (tansig), and hard limit transfer function (hardlim) (Zain
et al., 2010). The commonly used with the feedforward BP algorithm are the logsig, tansig and purelin (Demuth, Beale, &
Hagan, 2008). The log-sigmoid transfer function is the most commonly used function because it is self-limiting and has a
simple derivative. Researchers found the same performance with logsig and tansig and suggested the best transfer function
is logsig (Kohli & Dixit, 2005). This study uses logsig as transfer function and it can be written as follows:
1
f 6
1 enet
where the net value is either the net of hidden or output nodes. The next process of BP algorithm is adjustment of the con-
nection weights with respect to the predicted values of Y b and observed values Y. It can be processed with minimizing of
objective function (minimizing error) i.e. sum of square error (SSE).
X
N
SSE b j
Y j  Y 7
j1

where Y b j is the output of the ANN at jth observation.

Another important aspect is which variable is most important variable out of selected thirteen different variables. For this,
sensitivity analysis has been carried by noise to each input method (Gedeon, 1997), partial derivative of output with respect
to the input variable method (Zurada, Malinowski, & Cloete, 1994) and missing variable method (Hunter, Kennedy, Henry, &
Ferguson, 2000) which have been suggested by earlier researchers. The present study follows a sensitivity analysis by miss-
ing variable method, i.e. the variables that are relatively important with corresponding best performance value will be
included and the remaining less important variables may be eliminated (Hunter et al., 2000).

4. Pedestrian gap acceptance models

4.1. Regression model results

The stepwise MLR model was formulated by using the statistical package for the social sciences (SPSS 16). The signicant
variables were conrmed by corresponding t-statistic value and then included as particular variables in the model. A 95%
condence level was chosen with a corresponding critical t-value of 1.96. If the calculated t-values are greater than or equal
to tabulated t-value then corresponding variables were included in the model and other variables were eliminated. Table 3
represents the standardized MLR model results with an accepted gap as dependent variable and remaining selected variables
as independent variables.
In Table 3, the coefcient b represents the change in standard deviation of the output variable (viz., pedestrian accepted
gap size) due to unit change in standard deviation of the input variable (viz., pedestrian platoon size; frequency of attempts;
waiting time; rolling behaviour and vehicle speed). The positive or negative sign of b indicates increase or decrease in output
due to change in input variable. Further, frequency of attempts and rolling behaviour have strong inuence on accepted gap
size with beta values of 0.605 and 0.252 respectively. The negative coefcients indicates decrease in pedestrian accepted
gap size with increase in frequency of attempt and rolling behaviour The R-square of the calibrated MLR model as explained
in Table 3 and expressed in Eq. (8) is obtained as 0.6.
Gap 0:086  age  0:085  PPS  0:605  FA  0:121  WT  0:252  Rbeh 0:098  VS 8
where PPS: pedestrian platoon size; FA: frequency of attempts; WT: waiting time; Rbeh: rolling behaviour; VS: vehicle
speed.

4.2. ANN model results

In this study, the feedforward BP algorithm is chosen to train the network in Matlab 7.12.0 software package. In this ANN
architecture, one input, hidden and output layers were selected. The number of hidden neuron in hidden layer and transfer
functions were optimized based on many trials to predict the best output that satises the acceptable performance value
(RMSE) with a minimum possible time (number of epochs) consumed during the run. Out of all the tested ANN architectures,
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 121

Table 3
MLR model results.

Independent variable Coefcient beta t-value p-value

Pedestrian age 0.086 2.430 0.016*
Pedestrian platoon size 0.085 2.376 0.018*
Frequency of attempts 0.605 15.356 0.000
Pedestrian waiting time 0.121 3.297 0.001
Pedestrian rolling behaviour 0.252 6.397 0.000
Vehicle speed 0.098 2.761 0.006
R2 0.6
Root mean square error (RMSE) 2.579
*
p-value signicance level < 0.05.

Table 4
ANN model results.

Model ANN Architecture (input-hidden-output) RMSE R-square

I 13-9-1 2.115 0.742
II 12-9-1 2.120 0.707
III 11-9-1 2.062 0.722
IV 10-9-1 2.389 0.732
V 09-9-1 1.955 0.847
VI 08-9-1 2.053 0.743
VII 07-9-1 2.098 0.749
VIII 06-9-1 2.278 0.745

ANN architecture with nine hidden neurons in the hidden layer gave the best performance value with selected input nodes
(thirteen) and single output node. The ANN architecture was trained with normalized data. The architecture was trained
with 70% of the data, and 15% data was used for cross validation and 15% was used for testing. At each step, data was selected
randomly from the selected data set. In the network, the learning rule was used as a Levenberg Marquardt algorithm to min-
imize the error. The transfer functions used for the hidden and output layers were log-sigmoid (logsig) and linear (purline),
respectively. If the goal is reached (best performance value i.e., minimum RMSE) the training was stopped.
The following training parameters were used to train the network based on selected Levenberg Marquardt algorithm in
Matlab 7.12: 1000 maximum number of epochs to train; 0 performance goal; 6 Validation checks; 1e-5 Minimum perfor-
mance gradient; 0.001 Initial l; 0.1 l decrease factor; 10 l increase factor; 1e10 Maximum l; 25 Epochs between displays.
In order to nd out best ANN architecture, the ANN architecture was tested with several trails of the xed number of hid-
den neurons (nine) and varied input neurons (varied number of variables). In this sensitivity analysis, the input variables
were selected corresponding to the best performance value (minimum RMSE). Table 4 represents the results of sensitivity
analysis. From Table 4, the ANN model which has lowest RMSE value was selected as best ANN architecture i.e., 9-9-1.
Fig. 3 shows the relative importance of each variable corresponding to the best ANN architecture. From Fig. 3, it can be

Fig. 3. The relative importance of each variable from ANN model best architecture.
122 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

observed that rolling behaviour shows highest importance, while vehicle speed and frequency of an attempt are the next
highest important variables. The pedestrian waiting time, pedestrian platoon size and crossing path condition also shows
signicant contribution on pedestrian minimum accepted gap size.
Moreover, the calibrated models were validated by using 120 data points to check the accuracy of both the MLR as well as
ANN models. For MLR model validation, signicant variables (see Eq. (8)) were used from the calibrated model and validation
check was done with root mean square error (RMSE). Further, the ANN model was also validated with 9-9-1 architecture
with calibrated ANN parameters (Levenberg Marquardt algorithm etc.). The validation result of RMSE is obtained by MLR
model is 3.115 and ANN model is 0.136 respectively. The validation results show that better prediction capability of ANN
model than the MLR model because of signicantly lower RMSE as compared to MLR model.

5. Discussion

5.1. Discussion of MLR and ANN model results

The study evaluates the effect of pedestrian behavioural, trafc and vehicular characteristics on pedestrian accepted gap
size by employing linear (viz., MLR model) and non-linear models (viz., ANN technique). The MLR model results show that
increase in pedestrian age as well as the vehicle speed increases the accepted gap size. The accepted gap size decreases with
increase in waiting time, pedestrian platoon size, number of attempts and rolling behaviour (see in Table 3). The previous
research studies shows that pedestrian age, waiting time, number of pedestrians in the group and vehicle gap signicantly
increases the probability of pedestrian gap acceptance (Sun et al., 2003). However, another research study results shows
pedestrian platoon, near gap increases the probability of gap acceptance and pedestrian age decreases the probability of
gap acceptance (Wang et al., 2010). Researchers stated that the increase in size of vehicle, number of parked vehicles, gender
and pedestrian platoon increases the accepted gap size (Yannis et al., 2013). However, the present study results contradicts
with these results and it has been observed that accepted gap size decreases with increase in pedestrian platoon and the
earlier study did not make any comment on the pedestrian probability of road crossing behaviour. The present study result
suggest the inuence of pedestrian platoon size on accepted gap size is also supported by some of the early research studies
(Das et al., 2005). Moreover, the earlier studies did not consider the pedestrian behavioural characteristics such as rolling
behaviour and frequency of attempt in predicting the accepted gap size, but previous research studies on pedestrian safety
showed that pedestrian behavioural characteristics have signicant impact on pedestrian gap acceptance behaviour while
crossing at unmarked crosswalk locations (Brewer, Fitzpatrick, Whitacre, & Lord, 2006; Zhuang & Wu, 2012).
Further, from the MLR and ANN model results, it can be seen that the ANN model achieves much better performance than
the MLR model as shown in Table 5. The non-linear ANN model is capable of incorporating complex non-linear effects of the
input on the output and provides better t and better performance in comparison to the MLR model for prediction of pedes-
trian behaviour considering the effect of various variables. From Table 5, it can be observed that the eight ANN models per-
form better for different performance criteria such as RMSE, mean absolute percentage error (MAPE), correlation coefcient
and R-square as compared to the MLR model. The correlation coefcient of some ANN models (Model VI and Model VIII, see
in Table 5) is close to the MLR model. However, ANN models achieve better performance than the MLR model in terms of
lower RMSE as well as higher R2 and their values are obtained as 2.053, 2.278 and 0.743, 0.745 respectively. Moreover, from
the ANN results (see Fig. 3) the inuence of input variables such as path change condition, type of gap, pedestrian crossing
direction on output (viz., accepted gap size) can be observed which is eliminated in MLR model due to their insignicance by
t-value. Perhaps, the stepwise MLR model excluded these input variables due to poor correlation or the multicollinearity
problems whereas the ANN model considers the signicant variables corresponding to the minimum MSE. Also, the valida-
tion of the results (lower RMSE of ANN model is 0.136) shows that strong prediction capability of ANN model over the MLR
model (RMSE of MLR model is 3.115).
From MLR model results, the frequency of attempts, pedestrian rolling behaviour, and waiting time show proportionally
higher signicance than the pedestrian individual characteristics. From the ANN model, rolling behaviour and vehicle speed
are found to be of higher importance than the other variables. It is also observed that for both the ANN and MLR models,
pedestrian behavioural characteristics (rolling behaviour and frequency of attempts) are the most important variables for
modelling pedestrian accepted gap size, followed by vehicle speed, waiting time and pedestrian platoon effect. Moreover,

Table 5
Comparison of performance of ANN models and MLR model.

Performance criteria ANN models MLR model

I II III IV V VI VII VIII
RMSE 2.115 2.120 2.062 2.389 1.955 2.053 2.098 2.278 2.579
MAPE 21.47 20.51 19.32 23.59 18.72 18.88 19.90 20.45 26.58
Correlation coefcient 0.852 0.839 0.848 0.828 0.851 0.778 0.812 0.787 0.754
R-square 0.742 0.707 0.722 0.732 0.847 0.743 0.749 0.745 0.6

Note: RMSE: Root mean square error; MAPE: Mean absolute percentage error.
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 123

ANN model results gives more information regarding the contribution of other variables such as pedestrian path change con-
dition, crossing direction and type of gap corresponding to the best ANN architecture (9-9-1). Perhaps, the ANN architecture
13-9-1 represents the importance of each selected input variable over the pedestrian accepted gap size. Also, such important
comparative results cannot be drawn with MLR model due to statistical insignicance of some selected variables. Though,
the ANN model results cannot provide information about the increase or decrease in output variable (viz., pedestrian
accepted gap size) with unit change of signicant input variable (waiting time, vehicle speed, etc.). In this line, the MLR mod-
els are more useful techniques to quantify the effect of signicant input variables on the output variable as from their sta-
tistical signs we can easily predict the effect of the unit change of independent variable (input variable such as vehicle speed,
waiting time, etc.) over the dependent variable (viz., pedestrian accepted gap size). For example from Table 3, it can be seen
that unit change in standard deviation for vehicle speed results in 0.098 change in the standard deviation for accepted vehi-
cle gap size, and positive sign of coefcient b indicates increase in accepted vehicle gap size with increase in vehicle speed.
Early research studies have shown that, pedestrian platoon, parked vehicle and distance (distance between pedestrian
and vehicle) were signicant variables to predict the pedestrian accepted gap size (Yannis et al., 2013). In other studies, it
was found that pedestrian behavioural characteristics (rolling behaviour, speed and path change condition) signicantly
affect the pedestrian accepted gap size as compared to pedestrian individual characteristics (gender and age) (Kadali &
Vedagiri, 2012). However, there are several variables which were not included in the nal model of earlier studies due to
the lack of statistical signicance. Early studies also shows that the increase in pedestrian waiting time increases the viola-
tion at signalized crosswalk locations and it is also observed at unprotected mid-block crosswalk location (Brosseau,
Zangenehpour, Saunier, & Miranda-Moreno, 2013; Hamed, 2001). The increase in pedestrian waiting time leads to increases
the usage of pedestrian rolling behaviour.

Fig. 4. The predicted and observed vehicular gap size with rolling behaviour by ANN and MLR models.
124 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

In this study, the MLR model shows that six variables are signicant out of selected thirteen variables, but with low
performance values (higher RMSE and lower R2). The ANN model with 13-9-1 architecture shows better performance values
than the MLR model. From the sensitivity analysis, it was also found that 9-9-1 is the best ANN architecture and it shows
better performance than the other ANN as well as MLR model. In this best architecture (9-9-1) the rolling behaviour
(pedestrian behavioural characteristic) is observed as a highly important variable when compared to other selected vari-
ables. The study indicates the better prediction of pedestrian accepted gap size by pedestrian behavioural characteristics
with non-linear ANN models when compared to the MLR model. The behavioural characteristics of the pedestrian such as
rolling behaviour, frequency of attempt, etc. (see in Fig. 3) play an important role in pedestrian gap acceptance under mixed
trafc conditions.
The prediction of accepted gap size depends more on the pedestrian behavioural characteristics (rolling behaviour). Fig. 4
compares the pedestrian rolling behaviour for eld observed data, MLR and ANN models. It can be observed that there is a
signicant change in mean accepted gap size with and without pedestrian rolling behaviour. Further, the prediction of
accepted gap size by ANN model results with and without rolling behaviour is given as 5 and 8.1 s respectively (Fig. 4(2))
were more or less similar values as a eld observed values, which are 4.95 and 8.15 respectively (Fig. 4(1). Whereas, the
MLR model predicted values (Fig. 4(3)) are 5.85 and 8.85 s which are higher than the eld observed values corresponding
to with and without pedestrian rolling behaviour condition. Due to non-linearity of the ANN model, ANN model captures
the pedestrian behaviour more accurately to predict the accepted gap size as compared to the linear MLR model.
However, the linear MLR model can easily interpret the inuence of the response variables (selected independent variables)
over the respondent variable (dependent variable).

6. Conclusions

This paper uses linear regression (MLR) and non-linear ANN models with pedestrian behavioural characteristic data to
explain pedestrian accepted vehicular gap size at unprotected mid-block crosswalk location under mixed trafc conditions.
Out of the selected thirteen variables, six variables (age, pedestrian platoon size, frequency of attempts, waiting time, rolling
behaviour and vehicle speed) were included in the MLR model, based on higher signicance level of these variables. In order
to develop the ANN model, the number of hidden neurons in the hidden layer was found with several trails of hidden neu-
rons. Several ANN models were developed with xed nine hidden neurons in the hidden layer and varied input variables in
the input layer to predict the accepted gap size of pedestrian using pedestrian behavioural, vehicle and trafc characteristic
variables as inputs. Further, best ANN architecture was selected from the sensitivity analysis by missing variable method.
The best ANN architecture was used to compare the MLR and ANN model predicting capability. From the sensitivity analysis,
it can be concluded that the ANN model with nine input variables can be utilized for better prediction of accepted gap size.
The developed best ANN architecture based on the performance criteria (minimum RMSE) shows a good prediction of pedes-
trian accepted gap size with pedestrian behavioural characteristics under mixed trafc conditions. From the validation
results, it is also concluded that ANN models predicts the pedestrian accepted gap size better than the MLR model, which
has been veried with eld observed values. Further, it is concluded that the pedestrian rolling behaviour plays a major role
in pedestrian accepted gap size at unprotected mid-block crosswalk under mixed trafc conditions. It is also concluded that
there is signicant contribution of vehicle speed on pedestrian accepted gap size from the model results (viz., MLR and ANN
model results). The study also observed that increase in attempts to cross the road reduces the pedestrian safety as the
pedestrian attempts crossing the road with smaller gap size. Hence, to identify the signicant input variables affecting
the pedestrian accepted gap size the ANN model is the suitable technique. MLR modelling technique is more effective
and easy for the quantication of effect of signicant variables on pedestrian gap acceptance. This study also concludes that
for planning new pedestrian facilities ANN technique is more suitable due to its ability to include inuence of many factors
on output variable, whereas for evaluation of the existing midblock crosswalk facility the MLR models is appropriate tech-
nique, as this technique is more suitable to quantify the effect of various factors on the output variable.
However, there are some limitations in this study. The present study is limited to six lanes divided roadway and needs to
be extended to other roadway characteristics at unprotected mid-block crosswalk locations under mixed trafc condition.
The selected MLR model is limited to linear relationship between dependent and independent variables. Perhaps, some of
the independent variables are correlated with each other and it may lead to loss of signicant variables in model results.
Further, the ANN model studies were performed with only two different transfer functions, the hidden (logsig) as well as
output layers (purline), because of time consumption for calibration as well as validation of results. The effect of change
of these transfer functions on the output needs further studies in order to select the optimum transfer functions. The sen-
sitivity analysis of ANN model was done to nd out the best architecture with a minimum number of variables, however
the quantication of target variable (accepted gap size) with respect to input variables is not possible by ANN model. In spite
of these limitations, the developed ANN model can predict pedestrian accepted gap size more accurately than the linear
models. The developed ANN model results may be also useful to predict accepted gap size more accurately by using pedes-
trian behavioural characteristics because of its non-linear nature. Further, these results may also be helpful in reducing
pedestrian conicts with vehicular movements at unprotected mid-block crossings under mixed trafc conditions. The pre-
dicted accepted gap size can be used for the designing of new pedestrian facilities such as signal design and these are also
useful for the trafc engineers as well as planners in trafc calming techniques to control the pedestrian related safety issues
 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126 125

under mixed trafc conditions. Moreover, these results shows importance of regulating pedestrian behaviour (rolling beha-
viour) at planning level by high barricades at unprotected crosswalk locations. The present study also highlights the fact that
MLR and ANN modelling complement each other well to overcome their individual limitations, and hence study of a
midblock crosswalk by employing both techniques may be quite useful for identifying and quantifying the effect of various
factors for design or evaluation of pedestrian crossing facility.

Acknowledgments

The authors would like to thank the anonymous reviewers and editor for their valuable suggestions that helped to
improve this manuscript.

References

Abyaneh, H. Z. (2014). Evaluation of multivariate linear regression and articial neural networks in prediction of water quality parameters. Journal of
Environmental Health Science & Engineering, 12(40), 18. http://dx.doi.org/10.1186/2052-336X-12-40.
Bernhoft, I. M., & Carstensen, G. (2008). Preferences and behaviour of pedestrians and cyclists by age and gender. Transportation Research Part F, 11(2),
8395. http://dx.doi.org/10.1016/j.trf.2007.08.004.
Brewer, M. A., Fitzpatrick, K., Whitacre, J. A., & Lord, D. (2006). Exploration of pedestrian gap acceptance behavior at selected locations. Transportation
Research Record, 1982, 132140. http://dx.doi.org/10.3141/1982-18.
Brosseau, M., Zangenehpour, S., Saunier, N., & Miranda-Moreno, L. (2013). The impact of waiting time and other factors on dangerous pedestrian crossings
and violations at signalized intersections: A case study in Montreal. Transportation Research Part F, 21, 159172. http://dx.doi.org/10.1016/
j.trf.2013.09.010.
Bullock, D., Garrett, J., & Hendrickson, C. (1993). A neural network for image-based vehicle detection. Transportation Research Part C, 1(3), 235247. http://
dx.doi.org/10.1016/0968-090X(93)90025-B.
Chang, L. Y. (2005). Analysis of freeway accident frequencies: Negative binomial regression versus articial neural network. Safety Science, 43(8), 541557.
http://dx.doi.org/10.1016/j.ssci.2005.04.004.
Cherry, C., Donlon, B., Yan, X., Moore, S. E., & Xiong, J. (2012). Illegal mid-block pedestrian crossings in China: Gap acceptance, conict and crossing path
analysis. International Journal of Injury Control and Safety Promotion, 19(4), 320330. http://dx.doi.org/10.1080/17457300.2011.628751.
Collins, A., & Evans, A. (1994). Aircraft noise and residential property values, an articial neural network approach. Journal of Transport Economics and Policy,
28(2), 175197. http://www.jstor.org/stable/20053034.
Das, S., Manski, C. F., & Manuszak, M. D. (2005). Walk or wait? An empirical analysis of street crossing decisions. Journal of Applied Econometrics, 20(4),
529548. http://dx.doi.org/10.1002/jae.791.
Delen, D., Sharda, R., & Bessonov, M. (2006). Identifying signicant predictors of injury severity in trafc accidents using a series of articial neural networks.
Accident Analysis and Prevention, 38(3), 434444. http://dx.doi.org/10.1016/j.aap.2005.06.024.
Demuth, H., Beale, M., & Hagan, M. (2008). Neural Network ToolboxTM 6 Users Guide. The MathWorks, Inc. 3 Apple Hill Drive Natick, MA 017602098.
Dia, H., & Rose, G. (1997). Development and evaluation of neural network freeway incident detection models using eld data. Transportation Research Part C,
5(5), 313331. http://dx.doi.org/10.1016/S0968-090X(97)00016-8.
Dipietro, C. M., & King, L. E. (1970). Pedestrian Gap-acceptance. Highway Research Record, 308, 8091.
Dougherty, M. (1995). A review of neural networks applied to transport. Transportation Research Part C, 3(4), 247260. http://dx.doi.org/10.1016/0968-
090X(95)00009-8.
Duliba, K. A. (1991). Contrasting neural nets with regression in predicting performance in the transportation industry. In Proceedings of 24th annual Hawaii
international conference on system sciences (Vol. 4, pp. 163170). http://dx.doi.org/10.1109/HICSS.1991.184056.
Fitrianto, A., & Midi, H. (2013). Standardized simple mediation model: A numerical example. World Applied Sciences Journal, 22(8), 11351139. http://
dx.doi.org/10.5829/idosi.wasj.2013.22.08.681.
Gedeon, T. D. (1997). Data mining of inputs: Analyzing magnitude and functional measures. International Journal of Neural Systems, 8(2), 209218. http://
dx.doi.org/10.1142/S0129065797000227.
Gelman, A., & Hill, J. (2007). Data analysis using regression and multilevel/hierarchical models. New York: Cambridge University Press.
Hakkert, A. S., Gitelman, V., & Ben-Shabat, E. (2002). An evaluation of crosswalk warning systems: Effects on pedestrian and vehicle behavior. Transportation
Research Part F, 5(4), 275292. http://dx.doi.org/10.1016/S1369-8478(02)00033-5.
Hamed, M. M. (2001). Analysis of pedestrians behaviour at pedestrain crossings. Safety Science, 38(1), 6382. http://dx.doi.org/10.1016/S0925-
7535(00)00058-8.
Hashemi, R. R., Le Blanc, L. A., Rucks, C. T., & Shearr, A. (1995). A neural network for transportation safety modeling. Expert Systems with Applications, 9(3),
247256. http://dx.doi.org/10.1016/0957-4174(95)00002-Q.
Hensher, D. A., & Ton, T. T. (2000). A comparison of the predictive potential of articial neural networks and nested logit models for commuter mode choice.
Transportation Research Part E, 36(3), 155172. http://dx.doi.org/10.1016/S1366-5545(99)00030-7.
Himanen, V., & Kulmala, R. (1988). An application of logit models in analyzing the behaviour of pedestrians and car drivers on pedestrian crossings. Accident
Analysis and Prevention, 20(3), 187197. http://dx.doi.org/10.1016/0001-4575(88)90003-6.
Hunter, A., Kennedy, L., Henry, J., & Ferguson, I. (2000). Application of neural networks and sensitivity analysis to improved prediction of Trauma survival.
Computer Methods and Programs in Biomedicine, 62(1), 1119. http://dx.doi.org/10.1016/S0169-2607(99)00046-2.
Indian Road Congress (IRC:103) (2012). Guidelines for pedestrian facilities. New Delhi, India.
Kadali, B. R., & Vedagiri, P. (2012). Pedestrians gap acceptance behavior at mid-block location. International Journal of Engineering and Technology, 4(2),
158161. http://www.ijetch.org/papers/339-T0013.pdf.
Kadali, B. R., & Vedagiri, P. (2013). Modelling pedestrian road crossing behaviour under mixed trafc condition. European Transport/Trasporti Europei, 55(3),
117. http://www.istiee.org/te/papers/N55/ET_2013_55_3_Raghuram.pdf.
Kalyoncuoglu, S. F., & Tigdemir, M. (2004). An alternative approach for modelling and simulation of trafc data: Articial neural networks. Simulation
Modelling Practice and Theory, 12(5), 351362. http://dx.doi.org/10.1016/j.simpat.2004.04.002.
Kareem, A. (2003). Review of global menace of road accidents with special reference to Malaysia A social perspective. Malaysian Journal of Medical Science,
10(2), 3139. http://europepmc.org/abstract/MED/23386795.
Karlaftis, M. G., & Vlahogianni, E. I. (2011). Statistical methods versus neural networks in transportation research: Differences, similarities and some
insights. Transportation Research Part C, 19(3), 387399. http://dx.doi.org/10.1016/j.trc.2010.10.004.
Kohli, A., & Dixit, U. S. (2005). A neural-network based methodology for the prediction of surface roughness in a turning process. International Journal of
Advanced Manufacturing Technology, 25(12), 118129. http://dx.doi.org/10.1007/s00170-003-1810-z.
Kumar, P., & Parida, M. (2011). Vulnerable road users in multi-modal transport system for Delhi. Journeys, LTA (pp. 3847). <http://www.lta.gov.sg/
ltaacademy/doc/J11May-p38Pawan_DelhiVulnerableRoadUsers.pdfo>.
126 B Raghuram Kadali et al. / Transportation Research Part F 32 (2015) 114126

Leather, J., Fabian, H., Gota, S., & Mejia, A. (2011). Walkability and pedestrian facilities in Asian Cities. ADB sustainable development working paper series.
Manila: Asian Development Bank. <http://www.adb.org/sites/default/les/publication/28679/adb-wp17-walkability-pedestrian-facilities-asian-cities.
pdf>.
Lee, Y. (2009). Freeway travel time forecast using articial neural networks with cluster method. In 12th International conference on information Fusion, July
69, USA. <http://www.isif.org/fusion/proceedings/fusion09CD/data/papers/0075.pdf>.
Lorenzo, M., & Matteo, M. (2013). OD matrices network estimation from link counts by neural networks. Journal of Transportation Systems Engineering and
Information Technology, 13(4), 8492. http://dx.doi.org/10.1016/S1570-6672(13)60117-8.
Lyons, G. (1995). Calibration and validation of a neural network driver decision model. Trafc Engineering & Control, 36(1), 1015.
Menhaj, M. (2007). The principle of neural networks. Tehran: Amirkabir Press.
Mohan, D., Tsimhoni, O., Sivak, M., & Flannagan, M. J. (2009). Road safety in India: Challenges and opportunities. University of Michigan. Transport Research
Institute. <http://deepblue.lib.umich.edu/handle/2027.42/61504>, Accessed on July 2012.
Mussone, L., Ferrari, A., & Oneta, M. (1999). An analysis of urban collisions using an articial intelligence model. Accident Analysis and Prevention, 31(6),
705718. http://dx.doi.org/10.1016/S0001-4575(99)00031-7.
NHTSA (2011). Trafc safety facts. National center for statistics and analysis. New Jersey Avenue SE, Washington, D.C. <http://www-nrd.nhtsa.dot.gov/Pubs/
811748.pdf>.
Nordfjrn, T., & S imsekoglu, O. (2013). The role of cultural factors and attitudes for pedestrian behaviour in an urban Turkish sample. Transportation
Research Part F, 21, 181193. http://dx.doi.org/10.1016/j.trf.2013.09.015.
Oxley, J., Fildes, B., Ihsen, E., Charlton, J., & Day, R. (1997). Differences in trafc judgements between young and old adult pedestrians. Accident Analysis &
Prevention, 29(6), 839847. http://dx.doi.org/10.1016/S0001-4575(97)00053-5.
Pant, P. D., & Balakrishnan, P. (1994). Neural network for gap acceptance at stop-controlled intersections. ASCE Journal of Transportation Engineering, 120(3),
432446. http://dx.doi.org/10.1061/(ASCE)0733-947X(1994)120:3(432).
Pao, H. T. (2008). A comparison of neural network and multiple regression analysis in modeling capital structure. Expert Systems with Applications, 35(3),
720727. http://dx.doi.org/10.1016/j.eswa.2007.07.018.
Peden, M., Scureld, R., Sleet, D., Mohan, D., Hyder, A. A., & Mathers, C. (2004). World report on road trafc injury prevention. Geneva: World Health
Organization. <http://whqlibdoc.who.int/publications/2004/9241562609.pdf>.
Ritchie, S. G., & Cheu, R. L. (1993). Simulation of freeway incident detection using articial neural networks. Transportation Research Part C, 1(3), 203217.
http://dx.doi.org/10.1016/S0968-090X(13)80001-0.
Sisiopiku, V. P., & Akin, D. (2003). Pedestrian behaviors at and perceptions towards various pedestrian facilities: An examination based on observation and
survey data. Transportation Research Part F, 6(4), 249274. http://dx.doi.org/10.1016/j.trf.2003.06.001.
Sun, D., Ukkusuri, S. V. S. K., Benekohal, R. F., & Waller, S. T. (2003). Modeling of motorist-pedestrian interaction at uncontrolled mid-block crosswalks. In
Proceedings of the 82nd annual meeting (CD-ROM). Transportation Research Board, Washington, D.C.
Taylor, C. (1995). Freeway trafc data prediction using neural networks. In Proceedings of 6th international conference on vehicle navigation and information
systems, in conjunction with the pacic rim transportation technology conference (pp. 225230). http://dx.doi.org/10.1109/VNIS.1995.518843.
Wang, T., Wu, J., Zheng, P., & McDonald, M. (2010). Study of pedestrians gap acceptance behavior when they jaywalk outside crossing facilities. In Annual
conference on intelligent transportation systems, IEEE, Madeira Island, Portugal (pp. 12951300). http://dx.doi.org/10.1109/ITSC.2010.5625157.
Xiong, N., He, J., Park, J. H., Cooley, D., & Li, Y. (2009). A neural network based vehicle classication system for pervasive smart road security. Journal of
Universal Computer Science, 15(5), 11191142. http://dx.doi.org/10.3217/jucs-015-05-1119.
Yang, H., Kitamura, R., Jovanis, P. P., Vaughn, K. M., & Abdel-Aty, M. A. (1993). Exploration of route choice behavior with advanced traveler information using
neural network concepts. Transportation, 20(2), 199223. http://dx.doi.org/10.1007/BF01307059.
Yannis, G., Papadimitriou, E., & Theolatos, A. (2013). Pedestrian gap acceptance for mid-block street crossing. Transportation Planning and Technology, 36(5),
450462. http://dx.doi.org/10.1080/03081060.2013.818274.
Zain, A. M., Haron, H., & Shari, S. (2010). Prediction of surface roughness in the end milling machining using articial neural network. Expert Systems with
Applications, 37(2), 17551768. http://dx.doi.org/10.1016/j.eswa.2009.07.033.
Zhuang, X., & Wu, C. (2012). The safety margin and perceived safety of pedestrians at unmarked roadway. Transportation Research Part F, 15(2), 119131.
http://dx.doi.org/10.1016/j.trf.2011.11.005.
Zurada, J. M., Malinowski, A., & Cloete, I. (1994). Sensitivity analysis for minimization of input data dimension for feedforward neural network. IEEE
International Symposium on Circuits and Systems, 6, 447450. http://dx.doi.org/10.1109/ISCAS.1994.409622.