Hindawi

Journal of Advanced Transportation

Volume 2018, Article ID 3869106, 13 pages
https://doi.org/10.1155/2018/3869106

Research Article
An Improved Deep Learning Model for Traffic Crash Prediction

Chunjiao Dong ,1,2,3 Chunfu Shao,1,2 Juan Li,1 and Zhihua Xiong 1

1
MOE Key Laboratory for Urban Transportation Complex Systems Theory and Technology, Beijing Jiaotong University,
Beijing 100044, China
2
Key Laboratory of Industry of Big Data Application Technology for Comprehensive Transport, Ministry of Transport,
Beijing Jiaotong University, Beijing 100044, China
3
Center for Transportation Research, Tickle College of Engineering, University of Tennessee, 600 Henley Street, Knoxville,
37996, Tennessee, USA

Correspondence should be addressed to Chunjiao Dong; [email protected]

Received 12 June 2018; Accepted 12 November 2018; Published 10 December 2018

Guest Editor: Hamzeh Khazaei

Copyright © 2018 Chunjiao Dong et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Machine-learning technology powers many aspects of modern society. Compared to the conventional machine learning techniques
that were limited in processing natural data in the raw form, deep learning allows computational models to learn representations
of data with multiple levels of abstraction. In this study, an improved deep learning model is proposed to explore the complex
interactions among roadways, traffic, environmental elements, and traffic crashes. The proposed model includes two modules,
an unsupervised feature learning module to identify functional network between the explanatory variables and the feature
representations and a supervised fine tuning module to perform traffic crash prediction. To address the unobserved heterogeneity
issues in the traffic crash prediction, a multivariate negative binomial (MVNB) model is embedding into the supervised fine tuning
module as a regression layer. The proposed model was applied to the dataset that was collected from Knox County in Tennessee
to validate the performances. The results indicate that the feature learning module identifies relational information between the
explanatory variables and the feature representations, which reduces the dimensionality of the input and preserves the original
information. The proposed model that includes the MVNB regression layer in the supervised fine tuning module can better account
for differential distribution patterns in traffic crashes across injury severities and provides superior traffic crash predictions. The
findings suggest that the proposed model is a superior alternative for traffic crash predictions and the average accuracy of the
prediction that was measured by RMSD can be improved by 84.58% and 158.27% compared to the deep learning model without
the regression layer and the SVM model, respectively.

1. Introduction To understand the relationship between the influence

factors and traffic crash outcomes, with the extracted data
Road traffic injuries are a leading cause of preventable death, from police reports and state highway-asset-management
especially, for the young people. In the United States, traffic databases, the analyses of traffic safety estimate and predicate
crashes were the number one cause of death among people the likelihood of a traffic crash. The number of crashes occur-
from 16 to 24 years old for each year from 2012 to 2014 [1]. ring on a defined spatial entity over a specific time period
In 2015, the nation lost 35,092 people in traffic crashes, a 7.2- (for example, the number of crashes per year occurring at a
percent increase from 32,744 in 2014, which is the largest roadway intersection, over a specified roadway segment, or
percentage increase in nearly 50 years [2]. This is an average of in a region) would be considered as the dependent variables
approximately 96 people being killed on the nation’s roadways and some of the many factors affecting the likelihood of
every day of the year, and an average of more than four people a traffic crash are analyzed and examined (see [3–5] for a
per hour. In other words, one person dies on roadways every comprehensive review). Though more and more factors that
15 minutes. are relevant to the traffic crashes have been incorporated and
2 Journal of Advanced Transportation

the proposed models became more and more sophisticated, ANN is a consistent alternative method for analyzing crash
there are still some factors that are not available to the frequency. Abdelwahab and Abdel-Aty [12] employed two
researchers and the models result in bias estimations and well-known ANN paradigms [the multilayer perceptron and
erroneous predictions. In this study, we proposed an innova- radial basis functions (RBF) neural networks] to analyze
tive approach for traffic crash prediction, which incorporates the traffic safety of toll plazas and evaluate the impacts of
a multivariate regression layer into a dynamic deep learning electronic toll collection (ETC) systems on highway safety.
model that contains an unsupervised feature learning module The performance of ANN was compared with calibrated
and a supervised fine tuning module governing the state logit models. Modeling results showed that the RBF neural
dynamics to improve the performances of prediction. network was the best model for analyzing driver injury
severity. Xie, Lord, and Zhang [13] evaluated the application
2. Literature Review of BNN models for predicting traffic crashes by using data
collected on rural frontage roads in Texas. The results showed
The statistical methodologies, such as the Poisson, negative that back-propagation neural network (BPNN) and BNN
binomial (NB), and their variants in univariate and multivari- models perform better than the NB regression model in
ate regression frameworks, have been successfully applied in terms of traffic crash prediction. The results also showed
crash count analyses [4, 6–8], which attempt to deal with the that BNNs could be used to address other issues in high-
data and methodological issues associated with traffic crash way safety, such as the development of crash modification
estimations and predictions, and enhance our understanding factors, and enhance the prediction capabilities for evalu-
on the relationship between the influence factors and traffic ating different highway design alternatives. Kunt, Aghayan,
crash outcomes. However, current research in traffic safety and Noii [14] employed a genetic algorithm (GA), pattern
indicates that the applied statistical modeling fails when search, and ANN models to predict the severity of freeway
dealing with complex and highly nonlinear data [9], which traffic crashes. The results showed that the ANN provided
could suggest that the relationship between the influence the best predictions. Jadaan, Al-Fayyad, and Gammoh [15]
factors and traffic crash outcomes is more complicated than developed a traffic crash prediction model using the ANN
can be captured by a single statistical approach. In addition, simulation with the purpose of identifying its suitability for
most of the statistical methods are based on some strong predicting traffic crashes under Jordanian conditions. The
assumptions, such as specifying a priori and the error dis- results demonstrated that the estimated traffic crashes are
tribution. Moreover, a problematic issue is multicollinearity, close to actual traffic crashes. Akin and Akbas [16] proposed
i.e., the high degree of correlation between two or more an ANN model to predict intersection crashes in Macomb
independent variables. Furthermore, statistical models have County of the State of Michigan. The predictive capabil-
difficulty when dealing with outliers, missing or noisy data ity of the ANN model was determined by classifying the
[10]. crashes into these types: fatal, injury, and property damage
To deal with the limitations of statistical methodologies, only (PDO) crashes. The results were very promising and
the machine learning methods, including Artificial Neural showed that ANN model is capable of providing an accurate
Network (ANN), Support Vector Machine (SVM) models, prediction (90.9%) of the crash types. In summary, though
and deep learning models, have been applied to various ANN and BNN models show better linear/nonlinear approx-
traffic safety problems and used as data analytic methods imation properties than traditional statistical approaches,
because of their ability to work with massive amounts of these models often cannot be generalized to other data sets
multidimensional data. In addition, because of the modeling [3].
flexibility, learning and generalization ability, and good pre- The SVM models have recently been introduced for traffic
dictive ability, the machine learning has been considered as safety analyses [17, 18], which are a new class of models that
generic, accurate, and convenient mathematical models in the are based on statistical learning theory and structural risk
field of traffic safety. minimization [19]. These models are supposed to approxi-
Because the commonly used Poisson or NB regres- mate any multivariate function to any desired degree of accu-
sion models assume the predefined underlying relationship racy with a set of related supervised learning methods. Li et al.
between dependent and independent variables and the viola- [17] evaluated the application of SVM models for predicting
tion of the assumption would lead to erroneous estimation, motor vehicle crashes. The results showed that SVM models
ANN and Bayesian neural network (BNN) models have predict crash data more effectively and accurately than
been employed to analyze the traffic safety problems for traditional NB models. In addition, the findings indicated
many years. Although both ANN and BNN models have that the SVM models provide better (or at least comparable)
similar multilevel network structures, they are different in performance than BPNN models and do not over-fit the data.
predicting the outcome variables. For ANN, the weights To identify the relationship between severe crashes and the
are assumed to fix. However, the weights of BNN follow explanatory variables and enhance model goodness-of-fit,
a probability distribution and the prediction needs to be Yu and Abdel-Aty [20] developed three models to analyze
integrated over all the probability weights. Basically, the ANN crash injury severity, which include a fixed parameter logit
can be characterized by three features: network architecture, model, a SVM model, and a random parameter logit model.
model of a neuron, and learning algorithms. Chang [11] The results showed that the SVM models and the random
compared the performances of NB regression model and parameter models provide superior model fits compared to
ANN in crash frequency analyses. The results showed that the fixed parameter logit model. Findings also demonstrate
Journal of Advanced Transportation 3

that it is important to consider possible nonlinearity and for traffic crash prediction. The proposed model includes
individual heterogeneity when analyzing traffic crashes. Chen two modules: an unsupervised feature learning module and
et al. [21] employed the SVM models to investigate driver a supervised fine tuning module. To discover nonlinear
injury severity patterns in rollover crashes using two-year relationship between the investigated variables and identify
crash data collected in New Mexico. The results showed that the impacts of influence factors on traffic crashes for roadway
the SVM models produce reasonable predictions and the network, a DAE model is proposed in the unsupervised
polynomial kernel outperforms the Gaussian RBF kernel. feature learning module to learn the features of explana-
Dong, Huang, and Zheng [22] proposed a SVM model to tory variables. In addition, a multivariate negative binomial
handle multidimensional spatial data in crash prediction. (MVNB) regression is embedding into the supervised fine
The results showed that the SVM models outperform the tuning module to address the heterogeneity issues. The
nonspatial models in terms of model fitting and predictive proposed model performances are evaluated by comparing
performance. In addition, the SVM models provide better to the deep learning model without the MVNB layer and
goodness-of-fit compared with Bayesian spatial model with
SVM models by using five-year data from Knox County in
conditional autoregressive prior when utilizing the whole
Tennessee.
dataset as the samples. Ren and Zhou [23] proposed a novel
approach that combines particle swarm optimization (PSO)
and SVM for traffic safety prediction. The results showed that 3. The Modeling Framework Formulation
the predictions of PSO-SVM are better than that from BP
neural network. Yu and Abdel-Aty [24] proposed the SVM A novel model is proposed for the traffic crash prediction
models with different kernel functions to evaluate real-time and Figure 1 illustrates the modeling framework. The pro-
crash risk. The results showed that the SVM model with RBF posed model includes two modules. One is the unsuper-
kernel outperformed the SVM model with linear kernel and vised feature learning module and another is the super-
Bayesian logistic regression model. In addition, the findings vised fine tuning module. The obtained encoded feature
showed that smaller sample size could improve the classifi- representations from the unsupervised feature module are
cation accuracy of the SVM models and variable selection used as the input for the supervised fine tuning mod-
procedure is needed prior to the SVM model development. ule.
Overall, it has been found that the SVM models showed
better or comparable results to the outcomes predicted by 3.1. Unsupervised Feature Learning Module. Compared to
ANN/BNN and other statistical models [19]. However, like the commonly used deep learning architectures including
ANN and BNN, the SVM models often cannot be generalized deep belief network [31], stacked autoencoder [32], and
to other data sets and they all tend to behave as black- convolutional neural networks [27], the symmetrical neural
boxes, which cannot provide the interpretable parameters as networks in an unsupervised manner have shown better
statistical models do. performances, which can automatically learn an appropriate
Other than the ANN/BNN and SVM models, other sparse feature representation from the raw data [33]. The
machine learning methods have been introduced in unsupervised feature learning module includes a denoising
traffic safety analyses. Abdel-Aty and Haleem [25] autoencoder (DAE) model to learn the underlying struc-
introduced a recently developed machine learning ture of the dynamic pattern among the characteristics of
technique—multivariate adaptive regression splines (MARS) roadway, traffic, and environment. With the explanatory
to predict vehicle angle crashes using extensive data collected variables, such as roadway geometric design features, traffic
on unsignalized intersections in Florida. The results showed factors, pavement factors, and environmental characteristics
that MARS outperformed the NB models. The proposed
as the input, the designed DAE model can identify the
MARS models showed promising results after screening
nonlinear relationship between the investigated variables in
the covariates using random forest. The findings suggested
an unsupervised and hierarchical manner and the robust
that MARS is an efficient technique for predicting crashes at
unsignalized intersections. feature representations can be obtained. In addition, the
Deep learning is a recently developed branch of machine designed DAE model can encode the explanatory variables
learning method and has been successfully applied in speech into an embedding low-dimensional space. The proposed
recognition, visual object recognition, object detection, and DAE model contains a visible layer, a hidden layer, an output
many other domains such as drug discovery and genomics layer, and a reconstruction layer. Unlike the conventional
[26, 27]. Compared to the conventional machine learning DAE model with K hidden layers [34], the proposed model
techniques that were limited in their ability to process natural uses a reconstruction error optimizing the output layer and
data in their raw form, deep learning constructs compu- the noisy input layer to generate higher level representa-
tational models aiming to extract inherent features in data tions.
from the lowest level to the highest level. Though the deep Assume Φ = {[k𝑖 , u𝑖 ]}𝑛𝑖=1 is a training set that contains
learning methods have shown outstanding performances in n roadway entities, where k𝑖 ∈ R𝐷 is the explanatory
many applications [26], the applications of deep learning variable vector with dimension D and ui is the multi-
in the field of transportation are relatively few and only variate traffic safety outcomes for roadway entity i. Given
focusing on the topic of traffic flow prediction [28–30]. In the training set, the proposed DAE model is trained to
this study, we proposed an improved deep learning model develop a robust feature representation by reconstructing the
4 Journal of Advanced Transportation

(1) Traffic factors

(2) Roadway geometric design features
(3) Pavement factors
(4) Environmental characteristics
(5) Crash data
(6) Other data

Unsupervised feature learning

Hidden layer

Input Visible layer ℎ1 Output layer Output

1 1 x1 x1

ℎ2

..
2 2 . x2 x2

.. .. .. ..
. . ℎj . .
..
D D . xp xp

ℎF

Reconstruction error

Feature representations

Supervised fine tuning

Hidden
Representation
Regression
Input ℎ 1 function Output

x1 y i1 y i1
ℎ2

..
x2 . y i2 y i2

.. .. ..
. ℎj . .

xp .. y im y im
.
ℎF

Traffic crash estimation and prediction

Figure 1: Flowchart of the proposed model for traffic crash prediction.

input vi from its noisy corrupted version ̃V𝑖 , as shown in hidden variables is defined as
Figure 1.
There are D units in the visible layer and F units in the 1
hidden layer and the proposed model can be defined by a 𝑃 (k, h; Θ) = exp [−𝜀 (k, h; Θ)] (1)
𝑍 (Θ)
parameter set Θ = {W, a, b, c}, where W = [𝑊𝑖𝑗 ] ∈ R𝐷×𝐹 is
the interlayer connection weights, a = [𝑎𝑖 ] ∈ R𝐷 is the visible where 𝜀(k, h; Θ) is an energy function defined by symmet-
self-interactions or biases, b = [𝑏𝑗 ] ∈ R𝐹 is the hidden biases, ric interactions between the noisy input variable, hidden
and c = [𝑐𝑘 ] ∈ R𝐾 is the reconstruction error. The joint variables, and a set of interaction parameters Θ; Z(Θ) is a
probability distribution between the noisy input variables and normalized factor.
Journal of Advanced Transportation 5

For a binary variable, a Bernoulli-Bernoulli energy func- The conditional distribution over hidden units can be
tion [34, 35] and the conditional distribution of a single factorized and computed by
𝐹
stochastic hidden variable are given by
𝑃 (h | k; Θ) = ∏𝑃 (ℎ𝑗 | k; Θ) (4)
𝑇 𝑇 𝑇 𝑗=1
𝜀 (k, h; Θ) = k Wh − a k − b h
To estimate the parameters, the method proposed by
1 (2)
𝑃 (ℎ𝑗 = 1 | k; Θ) = Hjelm et al. [35], which maximizes the log-likelihood of
1 + exp [− (𝑏𝑗 + ∑𝐷
𝑖=1 𝑊𝑖𝑗 V𝑖 )] the marginal distribution of the hidden units to find the
gradient of the log-likelihood, is employed in this research.
For a continuous variable, a Gaussian-Bernoulli energy To simplify the estimation process, the free energy in terms
function [31] and the conditional distribution of a single of the probability at a data point vn can be used to replace the
stochastic hidden variable are given by energy function and the gradient has the following form:

𝜀 (k, h; Θ) = −k󸀠𝑇Wh + a󸀠𝑇a󸀠 − b𝑇 h 𝜕 𝜕

∑ log 𝑝 (k) = − ∑ ⟨ 𝜑 (k, h; Θ)⟩
𝜕Θ k∈𝐷 k∈𝐷 𝜕Θ
𝑝(h|k;Θ)
1 (3)
𝑃 (ℎ𝑗 = 1 | k; Θ) = (5)
1 + exp [− (𝑏𝑗 + ∑𝐷 (𝑊𝑖𝑗 V𝑖 /𝜎𝑖 ))] 𝜕
𝑖=1 + ∑ ⟨ 𝜑 (k, h; Θ)⟩
k∈𝐷 𝜕Θ
𝑝(k,h|Θ)
where k󸀠 = (V1 /𝜎1 , V2 /𝜎2 , ⋅ ⋅ ⋅ , V𝑖 /𝜎𝑖 , ⋅ ⋅ ⋅ V𝐷 /𝜎𝐷)𝑇, a󸀠 = ((V1 −
𝑎1 )/𝜎1 , (V2 − 𝑎2 )/𝜎2 , ⋅ ⋅ ⋅ , (V𝑖 − 𝑎𝑖 )/𝜎𝑖 , ⋅ ⋅ ⋅ , (V𝐷 − 𝑎𝐷)/𝜎𝐷), and 𝜎𝑖 where 𝜑(k, h; Θ) = − log ∑ℎ exp[−𝜀(k = k𝑛 , h; Θ)], and the
is the standard deviation of the i-th visible variable vi centered conditional distribution over hidden units should be replaced
on the bias ai . by a loss function.

𝐹
{
{−a𝑇k𝑛 − ∑ log [1 + exp (k𝑛𝑇 𝑊𝑗 + 𝑏𝑗 )] k𝑛 is binary
{
{
{ 𝑗=1
𝜑 (h | k𝑛 ) = { 𝐹 (6)
{ 󵄩
{ 󵄩2
{− 󵄩󵄩󵄩a − k𝑛 󵄩󵄩󵄩 − ∑ log [1 + exp (k𝑛 𝑊𝑗 + 𝑏𝑗 )] k𝑛 is continuous
{ 𝑇

{ 𝑗=1

Considering the hidden layer as the input layer for the particular NB regression model employed in this study has
outcome layer and outcome layer as the hidden layer, the the following form:
joint probability distribution between the hidden variables
and outcome variables, energy function, the conditional dis- 𝑝 (𝑦𝑖𝑗 )
tributions of the outcome variables and units, and parameter
estimation process can be obtained as those for hidden Γ (𝜆 𝑖𝑗 /𝜎 + 𝑦𝑖𝑗 ) 1 𝜆 𝑖𝑗 /𝜎 𝜎 𝑦𝑖𝑗 (7)
layer. The model will stop training when ck satisfies the = ) (( )
reconstruction error requirements or the dimension of the Γ (𝜆 𝑖𝑗 /𝜎) Γ (𝑦𝑖𝑗 + 1) 1 + 𝜎 1+𝜎
feature representation achieves the designed goals.
where Γ(⋅) is the gamma function, yij is the crash number of
3.2. Supervised Fine Tuning Module. The supervised fine crash type j for roadway segment i and E[yij ]=𝜆ij =exp(𝛽𝑗 X𝑖𝑗 +
tuning module is a supervised fine tuning procedure that 𝜀𝑖𝑗 ). exp(𝜀𝑖𝑗) is a multivariate gamma-distributed error term
includes a regression layer on the top of the resulting hidden with mean 1 and variance 𝛼−1 .
representation layers to estimate the likelihood of the crash As described in Shi and Valdez [36] and Anastasopoulos
occurrences, as shown in Figure 1. The obtained encoded et al. [37], with the expected number of crashes 𝜆ij , the MVNB
feature representations from the unsupervised feature mod- model has a joint probability function:
ule are used as the input for the supervised fine tuning
module. To jointly estimate the occurrence likelihood for 𝑝 (𝑦1𝑗 , 𝑦2𝑗 , ⋅ ⋅ ⋅ , 𝑦𝑖𝑗 )
more than one type of crashes simultaneously and address
the potential heterogeneity issues in the interdependent crash min(𝑦1𝑗 ,𝑦2𝑗 ,⋅⋅⋅ ,𝑦𝑖𝑗 ) 𝑁 (8)
data, a multivariate negative binomial (MVNB) model is used = ∑ 𝑝 (𝑘) ∏𝑝 (𝑦𝑖𝑗 − 𝑘)
in the supervised fine tuning module to estimate and predict 𝑘=0 𝑖=1
the traffic crashes across injury severities. Assume yi =(yi1 ,
yi2 ,. . ., yim )󸀠 is a vector of crash occurrence likelihood for The model parameters can be estimated by maximizing the
roadway entities i, which includes m types of crashes. The log-likelihood function:
6 Journal of Advanced Transportation

{ log [Γ (𝜎−1 + 𝑦𝑖𝑗 )] − log [Γ (𝜎−1 )] − log (𝑦𝑖𝑗 !) }

𝐿 = ∑ { −1 −1 } (9)
𝑖,𝑗 −𝜎 ln [1 + 𝜎 exp (𝛽𝑗 X𝑖𝑗 )] − 𝑦𝑖𝑗 log [1 + 𝜎 exp (𝛽𝑗 X𝑖𝑗 )]
{ }

The MVNB regression layer is added on the top of crashes. Because the category of fatal crashes has only a few
the resulting hidden representation layers to perform traffic observations, the categories of fatal crashes and incapacitat-
crash prediction. This yields a deep learning model tailored ing injury crashes have been combined and referred to as
to a task-specific supervised learning. Then we fine tune major injury crashes. The possible-injury crashes and PDO
the module 2 using backpropagation by minimizing the crashes have been combined and referred to as no-injury
following cost function: crashes. The nonincapacitating injury crashes are referred to
as minor injury crashes. A few of pervious literature [38–40]
𝑓 (𝜃)
has used a similar classification for injury outcomes. For those
selected 635 roadway segments, from 2010 to 2014, a total of
1 𝑛 𝑚 exp (ℎ𝜃 (𝑥𝑖 ))
= − ∑∑ 𝑠 (𝑦𝑖𝑗󸀠 = 𝑦𝑖𝑗 ) log [ 𝐹 ] 5365 traffic crashes were reported by the police officers, which
𝑛 𝑖=1𝑗=1 ∑𝑙=1 exp (ℎ𝜃 (𝑥𝑖 )) (10) include 135 (2.51%) major injury crashes, 1312 (24.46%) minor
injury crashes, and 3917 (73.02%) PDO crashes. Individual
𝛼 󵄩󵄩 󵄩2 𝐹 󵄩 󵄩2
[󵄩󵄩󵄩𝑊𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛󵄩󵄩󵄩󵄩𝐹 + ∑ 󵄩󵄩󵄩𝑊𝑙 󵄩󵄩󵄩𝐹 ]
roadway segment experienced from 0 to 23 crashes per
+
2𝑝 𝑙=1
year with a mean of 1.54 and a standard deviation of 1.89.
As expected, a significant amount of zeros is observed.
where s(⋅) is an indicator function, if yij ’=yij , then s(⋅)=1; The dependent variables and their descriptive statistics are
otherwise, s(⋅)=0, 𝛼 is a regularization parameter, and ‖‖𝐹 is shown in Table 1. The descriptive statistics of continuous
the Frobenius norm. The first term refers to the cross entropy independent variables and categorical independent variables
loss for the regression layer, the second term is the weight are shown in Tables 1 and 2, respectively.
decay penalty, and ℎ𝜃 (𝑥𝑖 ) is the output of deep learning for The considered traffic factors include the logarithm of
an input 𝑥𝑖 ’. annual average daily traffic (AADT) per lane, truck traf-
The cost function is minimized with a min-batch gradient fic percentage, and posted speed limits. Roadway segment
descent algorithm [27]. The parameters in module 2 𝜃 = AADT per lane from 2010 to 2014 varies from 851 to 32,359
{W𝑙 , b, W𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛} are estimated by initializing the weights vehicles with a mean of 3,388.44 and a standard deviation
W𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛 of the regression layer to small random values
of 5495.41. Other than the TRIMS dataset maintained by the
and the weights of the F hidden layers are initialized by
DDOT, traffic flow information can be obtained from https://
the encoding weights obtained in the unsupervised feature
www.tdot.tn.gov/APPLICATIONS/traffichistory, which is an
learning module.
AADT map providing traffic volumes based on a 24-hour,
two-directional count at a given location. The website also
4. Data provides the traffic history of any specific count station. The
variable of posted speed limit has a mean of 38.65 and a
The data are obtained from the Tennessee Roadway Informa-
tion Management System (TRIMS) and the Pavement Man- standard deviation of 6.69 with a minimum value of 30 and
agement System (PMS), which are maintained by the Ten- a maximum value of 70. The truck traffic percentage varies
nessee Department of Transportation (TDOT). The dataset from 1 to 33 with a mean of 6.71 and a standard deviation of
includes crash data, traffic factors, geometric design features, 4.98.
pavement factors, and environmental characteristics. The Important measurements of geometric design features
traffic, geometric, pavement, and environmental characteris- considered in this study include segment length, degree
tics are linked to the crash data through the common variable of horizontal curvature, median widths, outsider shoulder
id number. An extensive and comprehensive data screening widths, number of through lanes, lane widths, number of
that includes cleaning, consistency, and accuracy checks is left-turn lanes, median types, and shoulder type. Among
processed and performed to ensure the data are useable, them, the segment length, degree of horizontal curvature,
reliable, and valid for the analyses. After the initial data median widths, and outsider shoulder widths are considered
screening, in total 635 roadway segments in Knox County are as the continuous variables and the others are considered
chosen for the analyses. For the selected roadway segments, as the categorical variables. Other than the traffic factors
each of them has a completed dataset that links to the and geometric design features, the impacts of pavement
crash data. In other words, the dataset contains detailed surface characteristics are considered to better address traffic
information on traffic factors, geometric design features, safety issues for roadway design and maintenance. The
pavement characteristics, and environment factors. considered pavement surface characteristics include inter-
In TRIMS, the crash data have been classified into five national roughness index (IRI) and rut depth (RD). The
categories, fatal crashes, incapacitating injury crashes, nonin- analyzed IRI varies from 25.45 to 182.58 with a mean of
capacitating injury crashes, possible-injury crashes, and PDO 65.85 and a standard deviation of 27.75, which is calculated
Journal of Advanced Transportation 7

Table 1: Summary statistics of analyzed continuous variables.

Variable Mean Std. Dev. Min. Max.

Independent variable
The number of major injury crashes per year per roadway segment 0.04 0.32 0 3
The number of minor injury crashes per year per roadway segment 0.41 0.54 0 9
The number of no injury crashes per year per roadway segment 1.23 1.74 0 22
Traffic factors
The logarithm of AADT per lane 3.53 3.74 2.93 4.51
Truck traffic percentage 6.71 4.98 1 33
Posted speed limits 38.65 6.69 30 70
Geometric design features
Segment length (miles) 0.81 1.03 0.02 12.31
Degree of horizontal curvature 1.51 3.23 0 14.00
Median widths 1.12 2.02 0 12
Outside shoulder widths 3.06 1.88 3.52 8
Pavement factors
International roughness index 65.85 27.75 25.45 182.58
Rut depth (in.) 0.13 0.05 0.06 0.41

Table 2: Summary statistics of analyzed categorical variables.

Variable Category Frequency Percent

Geometric design features
Number of through lanes 3 585 18.43
2 1660 52.28
1 930 29.29
Lane widths (ft) 12 915 28.82
11 1650 51.97
10 610 19.21
Number of left-turn lanes 2 920 28.98
1 650 20.47
0 1605 50.55
Median type 1 for non-traversable median 1185 37.32
0 for traversable median 1990 62.68
Shoulder type 2 for pavement 760 23.94
1 for gravel 1595 50.24
0 for dirt 820 25.83
Environmental factors
Terrain type 1 for mountainous 1206 37.97
0 for rolling 1969 62.03
Land use type 2 for residential 1622 51.10
1 for commercial 775 24.41
0 for rural 778 24.49
Indicator for lighting 1 for lighting exists on the roadway segments 1304 41.06
0 for others 1871 58.94

using a quarter-car vehicle math model and the response is The environmental factors, including terrain types, light-
accumulated to yield a roughness index with units of slope ing condition, and land use type, are considered. Two terrain
(in/mi). Another pavement condition indicator is the RD, types are examined, which include rolling terrace (62.03%)
which is measured at roadway speeds with a laser/inertial and mountainous terrace (37.97%). Lighting condition was
profilograph. The analyzed RD varies from 0.06 to 0.41 with considered as a category variable, which indicated whether
a mean of 0.13 and a standard deviation of 0.05. lighting devises are provided at the roadway segments.
8 Journal of Advanced Transportation

Three types of land use are considered, including commer- pavement feature representation, and environmental feature
cial (24.41%), rural (24.49%), and residential (51.10). These representation.
variables are considered because they might have potential The results show that the signs of the means of four feature
significant effects on traffic safety. representations are negative, which indicate that the feature
representations are associated with crash-prone condition. In
other words, the current traffic, geometric, pavement, and
5. Modelling Results environmental features are the main factors that attribute to
The MATLAB was employed for model development. Four- the risk of crash occurrences. The findings indicate that the
year data, from 2010 to 2013, were used as the training set traffic, geometric, pavement, and environmental factors have
and one-year data, the year of 2014, were used as the testing a direct influence on traffic safety and need to be improved.
set. In order to obtain the model with superior performance, The traffic feature representations have a wide range with
module 1 was developed using 9 Gaussian visible units, 13 a minimum value of -4.420 and a maximum value of 6.650,
binary visible units, and a number of hyperbolic tangent which indicate the traffic factors have a significant impact
hidden units ranging from 32 to 128 in steps of 2. The on traffic safety. The pavement factors have comparative
number of hidden units was setting based on two rules, impacts on traffic safety with a minimum value of -12.076
greater than the input data dimensionality and the powers of and a maximum value of 6.753. The ranges of geometric
two. The parameters for learning rate and weight decay were and environmental feature representations are from -1.386 to
selected to optimize reduction of reconstruction error over 1.280 and from -1.805 to 0.847, respectively, which indicate
training. Module 1 was trained with a sample size of 2540 that the geometric and environmental feature representations
(four-year crash data) to allow for full convergence of the have comparative effects on traffic safety. Compared to those
parameters. The input data were processed by using module 1 crash modeling techniques that use only geometric design
to capture the relationship between traffic factors, geometric features as the input factors, the current research reveals
design features, pavement conditions, and environmental new insights that would benefit the development of updated
characteristics. Module 2 was developed using 4 Gaussian guidelines.
visible units and a number of hyperbolic tangent hidden
units. The initial number of hyperbolic tangent hidden
5.2. Results of the Supervised Fine Tuning Module. For the
units ranging from 8 to 32 was tested and examined. The
supervised fine tuning module, the visible units of the
learning procedure will stop when the number of feature
input layer use the feature representations as the input and
representations is achieved to four or the reconstruction error
the crash counts across injury severities are used as the
is less than 0.01.
training target. The aggregate weights between the input
layer and output layer are shown in Table 5. The results
5.1. Results of Unsupervised Feature Learning Module. For show that the traffic and geometric feature representations
the final model, the reconstruction error is less than 0.01 have positive effects and pavement and environmental feature
and four hidden layers are included. The weights between representations have negative effects on traffic crashes across
the input layer and output layer can be calculated as injury severities. The findings indicate that decreasing the
W32×16 W16×8 W8×4 =W32×4 . The results are shown in Table 3. values of traffic and geometric feature representations will
The negative sign represents a crash-prone condition and a increase the likelihood of crashes and increasing the values
positive sign represents a safe-prone condition. The valued of geometric and environmental feature representations will
number indicates an evaluation score. Because the feature increase the likelihood of crashes. The comparison results
learning module identifies relational information between show that traffic feature representation and geometric feature
input variables and output feature representations, the con- representation have significant impact on PDO crashes. The
nection weights between visible units and hidden units can pavement feature representation and environmental feature
be interpreted as the functional networks [35]. The results representation have significant impact on minor injury
show that each of the output units is significant associated crashes. Among four feature representations, the geometric
with traffic factors, geometric factors, pavement factors, and feature representation, pavement feature representation, and
environmental factors. traffic feature representation have most direct impacts on
Since the proposed feature learning module has sym- major injury, minor injury, and PDO crashes, respectively.
metric connection between visible and hidden layers and Considering the data in the year of 2014 as the input
the units in both layers have the probabilistic characteristics, variables, the developed deep learning model is used to
the proposed feature learning module also can be called as predict the crash counts for the year of 2014. To validate
an auto-encoder. Thus, the values of output units can be the superiority of the proposed models, the predicted results
interpreted as a feature representation and the results are are compared to the observed values. In addition, a deep
shown in Table 4. The number of output units is smaller learning model without the regression layer and a support
than the number of visible units, which indicates the feature vector machine (SVM) model are also developed to predict
representation with the values of output units has reduced the crash counts across injury severities. There are two key issues
dimensionality of the input, but still preserving the original related to the development of SVR models, kernel selection,
information. Hence the four output units are defined as traf- and parameters optimization. To address the nonlinear rela-
fic feature representation, geometric feature representation, tionship between the outcomes and attributes, the commonly
 Table 3: The functional network between the input variables and the feature representations.
Traffic feature Geometric feature Pavement feature Environmental feature
Variable
representation representation representation representation
Traffic factor
The logarithm of AADT per lane -0.529 -0.007 -0.016 -0.034
Truck traffic percentage -0.659 -0.002 -0.024 -0.010
Posted speed limits -0.030 -0.010 0.175 -0.047
Geometric factors
Journal of Advanced Transportation

Segment length (miles) 0.007 -0.580 -0.007 -0.003

Degree of horizontal curvature 0.002 -0.918 -0.006 -0.009
Median widths 0.002 0.826 0.068 0.015
Outside shoulder widths 0.005 0.419 0.016 0.064
Number of through lanes
3 lanes -0.021 -0.891 -0.006 -0.026
2 lanes -0.015 -0.864 -0.022 -0.060
Lane widths
12 ft 0.004 0.776 0.005 0.001
11 ft 0.009 0.592 0.076 0.004
Number of left-turn lanes
2 left-turn lanes -0.003 -0.212 -0.019 -0.017
1 left-turn lane 0.002 -0.783 0.028 0.016
Median type
Non-traversable median 0.013 0.247 0.006 0.007
Shoulder type
Pavement -0.035 -0.144 -0.027 -0.016
Gravel 0.005 -0.743 0.011 0.060
Pavement factor
International roughness index (in./mile) 0.101 0.007 -0.183 -0.001
Rut depth (in.) -0.004 -0.004 -0.341 -0.038
Environmental factor
Terrain type
Mountainous -0.014 -0.011 -0.010 -0.954
Land use type
Residential 0.042 0.015 0.011 0.847
Commercial 0.058 0.001 0.006 -0.915
Indicator for lighting
Lighting exists on roadway 0.018 0.008 0.027 0.781
9
10 Journal of Advanced Transportation

Table 4: Summary statistics of feature representations.

Statistics Mean Std. Dev. Min. Max.

Traffic feature representation -0.697 1.464 -4.420 6.650
Geometric feature representation -1.620 0.521 -1.386 1.280
Pavement feature representation -5.388 2.441 -12.076 6.753
Environmental feature representation -2.167 0.442 -1.805 0.847

Table 5: The functional network between the feature representation and different crash types.

Variable Major injury crashes Minor injury crashes PDO crashes

Traffic feature representation -0.109 -0.421 -1.535
Geometric feature representation 0.129 0.261 0.539
Pavement feature representation -0.033 -0.872 -0.315
Environmental feature representation 0.113 0.298 0.146

used RBF kernel is chosen to develop the SVM model. To proposed model results in the smallest prediction MAE and
precisely reflect the performance on regressing unknown RMSD, no matter for the crash types across injury severities.
data and prevent the overfitting problem, the k-fold cross- We hypothesize that the proposed model better addresses the
validation approach is employed for optimizing the two issue of heterogeneity and allows for excess zero counts in
parameters in RBF kernels—C and 𝜀 [41, 42]. The optimized correlated data.
parameters are (42.30, 37.09) and the optimized SVM model The findings indicate that the predictions from the pro-
yields to a training mean absolute error (MAE) less than posed model have significant improvements over all com-
0.01 and a training root-mean-squared deviation (RMSD) less parison models, in both accuracy and robustness. The best-
than 5%. performing result of the proposed model for major injury
To evaluate the accuracy and reliability of the prediction crashes has a MAE of 0.030, which indicates an 42.105% and
results, the predicted means for each injury severity level by 84.211% improvement from the deep learning model without
the proposed model, the deep learning model without the the regression layer and the SVM model, respectively. The
regression layer, and the SVM model are compared to the proposed models perform worse for the minor injury crashes,
observed means and two evaluation measures are used, which with a 0.080 RMSD for all observed samples. However, it
include MAE and RMSD. The comparison and validation is still better than the evaluation measurements from the
results are shown in Table 6. Results in Table 6 indicate that deep learning model without the regression layer and the
the predicted means from the proposed model (0.110, 0.573, SVM model, which are 0.202 and 0.257, respectively. This
1.335, and 2.019 for major injury, minor injury, no-injury, represents a MAE improvement of 150.980% and 219.608%
and all crashes, respectively) are very close to the observed for minor injury crash prediction. For the PDO crashes,
means (0.096, 0.556, 1.306, and 1.986). We believe that a the MAE improvements of the proposed models over the
very importation feature of the proposed model is that the comparison models range from 471.111% to 564.444%. For
included regression layer can provide a good estimate of the all the observed samples, the MAE improvements range
chance that the roadway segment is in the crash-free state or from 247.368% to 341.053%. Clearly, the improvement is
some crash-prone propensity states. significant for the traffic crash predictions. Therefore, the
For all the observed samples, the proposed model results proposed model seems to be a better alternative for crash
in a MAE of 0.030, 0.080, 0.071, and 0.150 and a RMSD of count predictions.
17.298%, 29.961%, 27.206%, and 43.652% for major injury, The proposed model has better performances in terms
minor injury, PDO, and all crashes, respectively. The deep of small error variances than the comparison models, since
learning mode without the regression layer results in a MAE the regression model is imbedding into the proposed model.
of 0.043, 0.202, 0.405, and 0.520 and a RMSD of 20.620%, The overall performances of the proposed model for all
51.741%, 65.086%, and 82.862% for major injury, minor crashes show an 89.824% RMSD improvement over the deep
injury, PDO, and all crashes, respectively. The SVM model learning mode without regression layer and an 121.378%
results in a MAE of 0.055, 0.257, 0.471, and 0.660 and a RMSD improvement over the SVM model. It is clear that the
RMSD of 26.022%, 61.350%, 72.416%, and 96.636% for major predictions obtained from the proposed models are superior
injury, minor injury, PDO, and all crashes, respectively. The to those obtained from the comparison models. The greatest
results suggest that the proposed model predicts better than difference is demonstrated for the PDO crashes where the
the deep learning model without the regression layer and the proposed model yields a RMSD of 27.206% compared to a
SVM model. In summary, compared to the deep learning 65.086% RMSD value from the deep learning model without
model without the regression layer and the SVM model, the the regression layer and a 72.416% RMSD value from the
Journal of Advanced Transportation 11

Table 6: Results of roadway segment crash count prediction.

Major injury Minor injury No injury Total

Observation mean 0.096 0.556 1.306 1.986
Proposed model
Estimated mean 0.110 0.573 1.335 2.019
MAE 0.030 0.080 0.071 0.150
RMSD (%) 17.298 29.961 27.206 43.652
Deep learning model without the regression layer
Estimated mean 0.101 0.556 1.332 1.989
MAE 0.043 0.202 0.405 0.520
RMSD (%) 20.620 51.741 65.086 82.862
SVM model
Estimated mean 0.101 0.513 1.329 1.943
MAE 0.055 0.257 0.471 0.660
RMSD (%) 26.022 61.350 72.416 96.636

SVM model. The differences in the proposed model and the (4) The findings suggest that the proposed model is a
SVM model for the minor injury crashes are also significant superior alternative for traffic crash predictions. The pro-
(29.961% versus 61.350%). posed model can better account for heterogeneity issues in
traffic crash prediction.
6. Conclusions The proposed models can perform traffic crash prediction
for a given facility. The proposed methodology could be
Because traffic crashes are a big concern of the public, applied to other roadway networks if appropriate attribute
agencies, and policy makers and result in countless fatalities variables are available. Traffic and transportation engineering
and injuries, there is a need to perform a comprehensive agencies can employ the proposed models with relative cases
analysis that aims to understand the relationship between the and develop them to their needs to obtain traffic crash
influence factors and traffic crash outcomes. In this study, we predictions for various time periods. Further investigation
presented an innovative approach for traffic crash prediction. of the proposed models includes the predictions of spatial-
Methodologically, we demonstrated a novel deep learning temporal dynamic pattern in crash data.
technique embedded within a multivariate regression model
can be used to identify relationship between the examined Data Availability
variables and the traffic crashes. Future applications of this
approach have the potential to provide insights into basic The data used to support the findings of this study are
questions regarding roadway spatial and temporal dynamic available from the corresponding author upon request.
function and practical questions regarding countermeasures.
The investigation results provide sufficient evidence for the Conflicts of Interest
following conclusions:
(1) The results show that the feature learning module The authors declare that they have no conflicts of interest.
identifies relational information between input variables and
output feature representations. The findings indicate that the
Acknowledgments
feature representations have reduced the dimensionality of
the input, but still preserving the original information. Special thanks are due to TDOT for providing the TRIMS
(2) The proposed model that includes a MVNB regression data. This research is supported by funding provided by
layer in the supervised fine tuning module can better account the Southeastern Transportation Center—a Regional UTC
for different patterns in crash data across injury severities funded by the USDOT—Research and Innovative Technol-
and provide superior traffic crash predictions. In addition, ogy Administration. Additional funding was provided by the
the proposed model can perform multivariate estimations National Natural Science Foundation of China (Grant Nos.
simultaneously with a superior fit. 51678044, 51338008, 71621001, 71210001, and 71501011).
(3) The proposed model has superior performances in
terms of prediction power compared to the deep learning
model without a regression layer and the SVM model. The References
overall performances of the proposed model for all crashes [1] NHTSA, Motor Vehicle Traffic Crashes as a Leading Cause of
show an 89.824% RMSD improvement over the deep learning Death in the United States, 2012-2014, Department of Trans-
model without a regression layer and an 121.378% RMSD portation, National Highway Traffic Safety Administration,
improvement over the SVM model. Traffic Safety Facts, DOT HS 812 297, 2016.
12 Journal of Advanced Transportation

[2] NHTSA, 2015 Motor Vehicle Crashes: Overview, U.S. Depart- [18] Y. Zhang and Y. Xie, “Forecasting of short-term freeway vol-
ment of Transportation, National Highway Traffic Safety ume with v-support vector machines,” Transportation Research
Administration, Traffic Safety Facts, DOT HS 812 318, 2016. Record, vol. 2024, no. 1, pp. 92–99, 2007.
[3] D. Lord and F. Mannering, “The statistical analysis of crash- [19] V. Kecman, “Support Vector Machines – An Introduction,” in
frequency data: a review and assessment of methodological Support Vector Machines: Theory and Applications, vol. 177 of
alternatives,” Transportation Research Part A: Policy and Prac- Studies in Fuzziness and Soft Computing, pp. 1–47, Springer
tice, vol. 44, no. 5, pp. 291–305, 2010. Berlin Heidelberg, Berlin, Heidelberg, 2005.
[4] F. L. Mannering and C. R. Bhat, “Analytic methods in accident [20] R. Yu and M. Abdel-Aty, “Analyzing crash injury severity for
research: methodological frontier and future directions,” Ana- a mountainous freeway incorporating real-time traffic and
lytic Methods in Accident Research, vol. 1, pp. 1–22, 2014. weather data,” Safety Science, vol. 63, pp. 50–56, 2014.
[5] F. L. Mannering, V. Shankar, and C. R. Bhat, “Unobserved [21] C. Chen, G. Zhang, Z. Qian, R. A. Tarefder, and Z. Tian,
heterogeneity and the statistical analysis of highway accident “Investigating driver injury severity patterns in rollover crashes
data,” Analytic Methods in Accident Research, vol. 11, pp. 1–16, using support vector machine models,” Accident Analysis &
2016. Prevention, vol. 90, pp. 128–139, 2016.
[6] J. C. Milton, V. N. Shankar, and F. L. Mannering, “Highway [22] N. Dong, H. Huang, and L. Zheng, “Support vector machine in
accident severities and the mixed logit model: an exploratory crash prediction at the level of traffic analysis zones: assessing
empirical analysis,” Accident Analysis & Prevention, vol. 40, no. the spatial proximity effects,” Accident Analysis & Prevention,
1, pp. 260–266, 2008. vol. 82, pp. 192–198, 2015.
[7] C. Dong, D. B. Clarke, X. Yan, A. Khattak, and B. Huang, “Mul- [23] G. Ren and Z. Zhou, “Traffic safety forecasting method by
tivariate random-parameters zero-inflated negative binomial particle swarm optimization and support vector machine,”
regression model: an application to estimate crash frequencies Expert Systems with Applications, vol. 38, no. 8, pp. 10420–10424,
at intersections,” Accident Analysis & Prevention, vol. 70, pp. 2011.
320–329, 2014a. [24] R. Yu and M. Abdel-Aty, “Utilizing support vector machine in
[8] C. Dong, D. B. Clarke, S. S. Nambisan, and B. Huang, “Analyzing real-time crash risk evaluation,” Accident Analysis & Prevention,
injury crashes using random-parameter bivariate regression vol. 51, pp. 252–259, 2013.
models,” Transportmetrica A: Transport Science, vol. 12, no. 9, [25] M. Abdel-Aty and K. Haleem, “Analyzing angle crashes at
pp. 794–810, 2016. unsignalized intersections using machine learning techniques,”
[9] M. G. Karlaftis and E. I. Vlahogianni, “Statistical methods Accident Analysis & Prevention, vol. 43, no. 1, pp. 461–470, 2011.
versus neural networks in transportation research: differences, [26] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature,
similarities and some insights,” Transportation Research Part C: vol. 521, no. 7553, pp. 436–444, 2015.
Emerging Technologies, vol. 19, no. 3, pp. 387–399, 2011. [27] J. Schmidhuber, “Deep learning in neural networks: an
[10] J. C. Principe, N. R. Euliano, and W. C. Lefebvre, Neural and overview,” Neural Networks, vol. 61, pp. 85–117, 2015.
Adaptive Systems: Fundamentals through Simulations, vol. 672, [28] W. Huang, G. Song, H. Hong, and K. Xie, “Deep architecture
Wiley, New York, NY, USA, 2000. for traffic flow prediction: deep belief networks with multitask
[11] L.-Y. Chang, “Analysis of freeway accident frequencies: Negative learning,” IEEE Transactions on Intelligent Transportation Sys-
binomial regression versus artificial neural network,” Safety tems, vol. 15, no. 5, pp. 2191–2201, 2014.
Science, vol. 43, no. 8, pp. 541–557, 2005. [29] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic Flow
[12] H. T. Abdelwahab and M. A. Abdel-Aty, “Artificial neural Prediction with Big Data: A Deep Learning Approach,” IEEE
networks and logit models for traffic safety analysis of toll Transactions on Intelligent Transportation Systems, vol. 16, no. 2,
plazas,” Transportation Research Record, no. 1784, pp. 115–125, pp. 865–873, 2015.
2002. [30] N. G. Polson and V. O. Sokolov, “Deep learning for short-
[13] Y. Xie, D. Lord, and Y. Zhang, “Predicting motor vehicle term traffic flow prediction,” Transportation Research Part C:
collisions using Bayesian neural network models: an empirical Emerging Technologies, vol. 79, pp. 1–17, 2017.
analysis,” Accident Analysis & Prevention, vol. 39, no. 5, pp. 922– [31] G. E. Hinton, S. Osindero, and Y. Teh, “A fast learning algorithm
933, 2007. for deep belief nets,” Neural Computation, vol. 18, no. 7, pp. 1527–
[14] M. M. Kunt, I. Aghayan, and N. Noii, “Prediction for traffic 1554, 2006.
accident severity: Comparing the artificial neural network, [32] P. Swietojanski, A. Ghoshal, and S. Renals, “Convolutional
genetic algorithm, combined genetic algorithm and pattern neural networks for distant speech recognition,” IEEE Signal
search methods,” Transport, vol. 26, no. 4, pp. 353–366, 2011. Processing Letters, vol. 21, no. 9, pp. 1120–1124, 2014.
[15] K. S. Jadaan, M. Al-Fayyad, and H. F. Gammoh, “Prediction [33] M. M. A. Rahhal, Y. Bazi, H. Alhichri, N. Alajlan, F. Melgani, and
of Road Traffic Accidents in Jordan using Artificial Neural R. R. Yager, “Deep learning approach for active classification of
Network (ANN),” Journal of Traffic and Logistics Engineering, electrocardiogram signals,” Information Sciences, vol. 345, pp.
vol. 2, no. 2, pp. 92–94, 2014. 340–354, 2016.
[16] D. Akin and B. Akbaş, “A neural network (NN) model to predict [34] H.-I. Suk, C.-Y. Wee, S.-W. Lee, and D. Shen, “State-space
intersection crashes based upon driver, vehicle and roadway model with deep learning for functional dynamics estimation
surface characteristics,” Scientific Research and Essays, vol. 5, no. in resting-state fMRI,” NeuroImage, vol. 129, pp. 292–307, 2016.
19, pp. 2837–2847, 2010. [35] R. D. Hjelm, V. D. Calhoun, R. Salakhutdinov, E. A. Allen, T.
[17] X. Li, D. Lord, Y. Zhang, and Y. Xie, “Predicting motor vehicle Adali, and S. M. Plis, “Restricted Boltzmann machines for neu-
crashes using Support Vector Machine models,” Accident Anal- roimaging: An application in identifying intrinsic networks,”
ysis & Prevention, vol. 40, no. 4, pp. 1611–1618, 2008. NeuroImage, vol. 96, pp. 245–260, 2014.
Journal of Advanced Transportation 13

[36] P. Shi and E. A. Valdez, “Multivariate negative binomial models

for insurance claim counts,” Insurance: Mathematics & Eco-
nomics, vol. 55, pp. 18–29, 2014.
[37] P. C. Anastasopoulos, V. N. Shankar, J. E. Haddock, and F. L.
Mannering, “A multivariate tobit analysis of highway accident-
injury-severity rates,” Accident Analysis & Prevention, vol. 45,
pp. 110–119, 2012.
[38] L. Chang and J. Chien, “Analysis of driver injury severity in
truck-involved accidents using a non-parametric classification
tree model,” Safety Science, vol. 51, no. 1, pp. 17–22, 2013.
[39] J. Pahukula, S. Hernandez, and A. Unnikrishnan, “A time of
day analysis of crashes involving large trucks in urban areas,”
Accident Analysis & Prevention, vol. 75, pp. 155–163, 2015.
[40] Q. Wu, F. Chen, and G. H. Zhang, “Mixed logit model-
based driver injury severity investigations in single- and multi-
vehicle crashes on rural two-lane highways,” Accident Analysis
& Prevention, vol. 72, pp. 105–115, 2014.
[41] N. Cristianini and J. Shawe-Taylor, An Introduction to Support
Vector Machines, Cambridge University Press, Cambridge,UK,
2000.
[42] C. Campbell, “Kernel methods: A survey of current techniques,”
Neurocomputing, vol. 48, pp. 63–84, 2002.
 International Journal of

Rotating Advances in
Machinery Multimedia

The Scientific
Engineering
Journal of
Journal of

Hindawi
World Journal
Hindawi Publishing Corporation Hindawi
Sensors
Hindawi Hindawi
www.hindawi.com Volume 2018 http://www.hindawi.com
www.hindawi.com Volume 2018
2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Journal of

Control Science
and Engineering

Advances in
Civil Engineering
Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

Submit your manuscripts at

www.hindawi.com

Journal of
Journal of Electrical and Computer
Robotics
Hindawi
Engineering
Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018

VLSI Design
Advances in
OptoElectronics
International Journal of

International Journal of
Modelling &
Simulation
Aerospace
Hindawi Volume 2018
Navigation and
Observation
Hindawi
www.hindawi.com Volume 2018
in Engineering
Hindawi
www.hindawi.com Volume 2018
Engineering
Hindawi
www.hindawi.com Volume 2018
Hindawi
www.hindawi.com www.hindawi.com Volume 2018

International Journal of
International Journal of Antennas and Active and Passive Advances in
Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration
Hindawi Hindawi Hindawi Hindawi Hindawi
www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018