Professional Documents
Culture Documents
To cite this article: Ali Can Yilmaz, Cigdem Aci & Kadir Aydin (2016) Traffic accident
reconstruction and an approach for prediction of fault rates using artificial neural networks: A case
study in Turkey, Traffic Injury Prevention, 17:6, 585-589, DOI: 10.1080/15389588.2015.1122760
Traffic accident reconstruction and an approach for prediction of fault rates using
artificial neural networks: A case study in Turkey
Ali Can Yilmaza , Cigdem Acib , and Kadir Aydinc
a
Department of Automotive Engineering, Cukurova University, Adana, Turkey; b Department of Computer Engineering, Mersin University, Mersin,
Turkey; c Department of Mechanical Engineering, Cukurova University, Adana, Turkey
CONTACT Ali Can Yilmaz acyilmaz@cu.edu.tr Department of Automotive Engineering, Cukurova University, P.O. Box , Adana, Turkey.
Associate Editor Alessandro Calvi oversaw the review of this article.
© Taylor & Francis Group, LLC
586 A. C. YILMAZ ET AL.
Overview of methods
and energy-equivalent speed–deformation energy analyses are FITNET is a modified form of MFANN. However, unlike the
of great importance. log-sigmoid transfer function used in MFANN, FITNET uses
default tan-sigmoid transfer function in the hidden layers and a
Scenario linear transfer function in the output layer (MathWorks 2014).
In the third case, an APC (vehicle 1) driving too slow (with speed
under legal limit) in the right lane is hit from the rear-end by Generalized regression neural networks
vehicle 2 at a speed within the legal limits. In this case, vehicle
A GRNN is a kind of radial basis network that is often used for
2 was deemed at fault with a rate of 100% due to a “rear crash”
function approximation. It has a radial basis layer and a special
and vehicle 1 had a rate of 0%, or fault-free. However, no analysis
linear layer. A GRNN does not require an iterative training pro-
was conducted and the low speed of vehicle 1 and crush depths
cedure. It approximates any arbitrary function between input
were not taken into consideration. A snapshot from the crash
and output vectors, drawing the function estimation directly
simulation of the vehicles is depicted in Figure 1c.
from the training data. In addition, it is consistent in that as the
training set size becomes large, the estimation error approaches
Scenario zero, with only mild restrictions on the function (Hannan et al.
In the fourth case, an APC (vehicle 1) egresses in the intersection 2010).
when the signal is on intermittent red and the other APC (vehi- The main function of a GRNN is to estimate a linear or non-
cle 2) enters the intersection on intermittent yellow and they col- linear regression surface on independent variables (input vec-
lide each other as shown in the snapshot in Figure 1d. According tors) U and dependent variables (desired output vectors) X. That
to expert report, vehicle 1 was at fault with a rate of 75% due to is, the network computes the most probable value of an out-
“reckless, incautious driving and not slowing down while enter- put Ox given only training vectors U. Specifically, the network
ing the intersection on intermittent red” and the other driver computes the joint probability density function of U and X. The
with a rate of 25% was deemed minorly at fault due to “careless expected value of X (E[X/U]) is expressed as shown in Eq. (3)
driving and not slowing down while entering the intersection on (Khashei et al. 2012):
intermittent yellow” without conducting a comprehensive anal-
ysis. ∞
X f (U, X )dx
E[X/U ] = −∞
∞ . (3)
−∞ f (U, X )dx
Multilayer feed-forward artificial neural networks and
function fitting neural networks The GRNN model has a 2-layer network. The first layer has
ANNs are designed to solve problems that are difficult using radial basis neurons and calculates weighted inputs with dis-
conventional mathematical methods. ANNs overcome the lim- tance and network inputs. The second layer has pure linear
itation of the conventional approaches by extracting the desired neurons that evaluate the weighted input with norm product
information directly from the data (Haykin 1994). MFANNs and network inputs with network summation (Goyal and Goyal
trained with a back-propagation learning algorithm are the most 2012).
popular neural networks. MFANNs are organized in 3 layers: an
input layer, an output layer, and the layers in between, or hidden Data set generation
layers (Svozil et al. 1997).
In a fully connected artificial neural network, each neuron in In the software step of this study, within scenarios 1 to 4,
a given layer is connected to each neuron in the following layer vehicles with features similar to those in real-world accidents
by an associated numerical weight (wij ). The weight connect- were exposed to different deformations (minimum of 500 data
ing 2 neurons regulates the magnitude of the signal that passes for each scenario) that also correspond to varied speed val-
between them. In addition, each neuron possesses a numerical ues. While one of the speeds was kept constant, the other’s was
bias term, corresponding to an input of −1 whose associated increased by 5 km/h. It was assumed that every 5 km/h incre-
weight has the meaning of a threshold value. ment in speed corresponds to 3 points increment in fault rates
The goal of back-propagation training is to iteratively change of the related vehicle. For every speed value, the simulation was
the weights between the neurons in a direction that mini- repeated and the crush depths were recorded from the simula-
mizes the error E, defined as the squared difference between the tion report. One of the speed values was kept constant and the
desired and the actual outcomes of the output nodes, summed other was changed every 20 trials until 500 data were obtained
over the training patterns (training set data) and the output neu- for every accident scenario. Descriptive statistics for scenarios 1
rons, according to the steepest descent method in Eq. (2): and 2 and 3 and 4 are shown in Tables 1 and 2, respectively.
∂E
wi j (t ) = −η + αwi j (t − 1). (2) Results and discussion
∂wi j
In this article, 3 ANN models (MFANN, FITNET, and GRNN)
The variation of a connection weight at the ith iteration were adopted for the data set. For prediction of fault rates in
depends on the partial derivative of the total error with respect accidents using ANN models, input variables were SDef1 , SDef2 ,
to that weight through a proportionality constant termed the V1 , and V2 ; Fault-rate1 and Fault-rate2 were taken as output
learning rate (η) and on the variation in the same weight dur- variables.
ing the previous iteration by means of a momentum term (α; The MFANN and FITNET models have 3 layers (input, hid-
Marini et al. 2007). den, and output) and the input, hidden, and output layers have 4,
588 A. C. YILMAZ ET AL.
Deformation (m) Deformation (m) Velocity (km/h) Velocity (km/h) Fault rate (%) Fault rate (%)
(vehicle ) (vehicle ) (vehicle ) (vehicle ) (vehicle ) (vehicle )
10, and 2 neurons, respectively. A log-sigmoid activation func- r Because there is no training phase in GRNN, the GRNN
tion is used in the MFANN model and the tan-sigmoid activa- model produced results much faster than the MFANN
tion function is used in the FITNET model in the hidden lay- and FITNET models. However, the GRNN model had the
ers. A pure linear activation function is used in both models in worst prediction results.
the output layer and the Levenberg-Marquardt algorithm (More r The R values for prediction of fault rates were close to 1 for
1978) is utilized for training the networks. The other important all folds and scenarios.
parameters of the MFANN and FITNET models are the number r The MFANN model for scenario 3 yielded the lowest MSE
of epochs (selected as 1,000), learning rate (selected as 0.02), and for prediction of fault rates, where the MSE changed from
momentum (selected as 0.5). 0.000055 to 0.052155.
It should be noted that the most important architectural r The GRNN model for scenario 4 gave the highest MSE
parameter of the GRNN model is the spread constant, which has for prediction of fault rates, where the MSE changed from
a certain effect on prediction efficiency. In this article, the spread 1.999691 to 3.806794.
value was set to 2.0 (default 1.0) empirically. This study focused on exhibiting new aspects and scientific
Each input and output variable of 500 simulations was used approaches for determining fault rates in most frequent PDO
to develop the models. The data were randomly divided into 2 accidents occurring in Turkey by discussing some deficiencies
disjoint subsets using 10-fold cross-validation. The training set in THTA and without regard to the initiative and/or experience
had 450 observations and the testing set had 50. The MSE in of experts. Referring to the direct proportion between speed and
Eq. (3) and R in Eq. (4) were utilized in order to compare the deformation energy, a new systematic approach for determining
prediction ability of the developed models. MATLAB software fault rates was carried out. vCrash software was used to simu-
(MathWorks 2014) was used for the experiments. late the accident scenes by means of 4 scenarios. The obtained
N data set contained 3 parameters (i.e., deformation energy, veloc-
Qexp − Qcal 2 ity, and fault rate) for the accident scenes. MFANN, FITNET,
MSE = (4)
n and GRNN models were developed to predict fault rates in
1
which R and MSE values of the developed models were also cal-
and culated. The MFANN model gave the lowest MSE values for
N the prediction of fault rates for the first 3 scenarios. For only
(Qexp − Qcal )2 the last scenario, the FITNET model gave the lowest MSE and
R = 1 − i=1 _ 2, (5)
N highest R value. Without necessity for expensive reconstruction
i=1 (Q exp − Q exp )
software, with the aid of ANN models (i.e., MFANN, FITNET,
where Qexp is the observed value; Qcal is the predicted value; and GRNN), approximate speed values and fault rates could
_ be determined with regard to deformation (crush depth) at the
Qexp is the mean predicted value; and N is the number of data
accident location. An application for a portable device may be
points.
coded to predict velocities and fault rates of involvements just
Tables 3 and 4 show the R and MSE results of the MFANN,
by entering average crush depths on the collision regions of
FITNET, and GRNN prediction models for scenarios 1 and 2
vehicles to the application. Judicious decisions may be made
and 3 and 4, respectively.
using this scientific approach, especially for court matters and
The following can be concluded from Tables 3 and 4:
r For scenarios 1, 2, and 3, the MFANN model performed forensic investigations. None of the vehicles had black boxes or
a recording camera, which most of the vehicles in Turkey do
better (i.e., higher R and lower MSE) than the FITNET and
not have. An initiative-dependent approach is prevalent, which
GRNN prediction models.
r The results of the FITNET model were best for scenario 4. depicts the importance of this research. On the other hand, this
r The FITNET model showed the second best results for pre- study also yields a probable useful approach for insurance com-
panies that currently consider 100, 50, and 0% fault rates. In
diction for the first 3 scenarios.
Table . Descriptive statistics for scenarios and .
Deformation (m) Deformation (m) Velocity (km/h) Velocity (km/h) Fault rate (%) Fault rate (%)
(vehicle ) (vehicle ) (vehicle ) (vehicle ) (vehicle ) (vehicle )
MSE R
Fold number MFANN FITNET GRNN MFANN FITNET GRNN
MSE R
Fold Number MFANN FITNET GRNN MFANN FITNET GRNN
addition, these algorithms are likely applications to commer- Hannan SA, Manza RR, Ramteke RJ. Generalized regression neural network
cialize. As future work, based on these approaches, various fault and radial basis function for heart disease diagnosis. Int J Comput Appl.
rate studies related to injury/fatal accidents may be conducted 2010;7:7–13.
Haykin S. Neural Networks: A Comprehensive Foundation. New York, NY:
within various prediction methods like Support Vector Regres- MacMillan; 1994.
sion (SVR), fuzzy logic, etc. For pedestrian–vehicle collisions, Islam S, Jones SL, Dye D. Comprehensive analysis of single- and multi-
studies may be carried out on accidents with vehicular skid vehicle large truck at-fault crashes on rural and urban roadways in
marks or definite pedestrian throw distances. In the absence Alabama. Accid Anal Prev. 2014;67:147–158.
of marks or throw distances, some physical dynamic models Khashei M, Hamadani AZ, Bijari B. A novel hybrid classification model of
artificial neural networks and multiple linear regression models. Expert
(i.e., impact dynamics, thin plate theory) may be developed to Syst. 2012;39:2606–2620.
describe relations between pedestrian impact speed and deflec- Marini F, Magri AL, Bucci R. Multilayer feed-forward artificial neural net-
tion of impact region on the vehicle. Similar methods for in- works for class modeling. Chemometr Intell Lab Syst. 2007;88:118–124.
vehicle injury/fatal situations may be compared with real-world MathWorks. MatLab. 2014. Available at: http://www.mathworks.
accidents such as the deformation–speed analyses conducted com/products/matlab/. Accessed May 4, 2015.
More JJ. The Levenberg-Marquardt algorithm: implementation and theory.
in this study. Paper presented at: Biennial Conference; 1978; Dundee, Scotland.
Peng Y, Deng W, Xu P, Yao S. Study on the collision performance of a com-
posite energy-absorbing structure for subway vehicles. Thin-Walled
Funding Structures. 2015;94:663–672.
This study was funded by Cukurova University, Scientific Research Projects Svozil D, Kvasnicka V, Pospichal J. Introduction to multi-layer feed-
Unit, and the Department of Mechanical Engineering. forward neural networks. Chemometr Intell Lab Syst. 1997;39:43–
62.
THTA. No. 2918 Turkish Highway Traffic Act. Official Gazette of the Repub-
References lic of Turkey. 1983;22:687.
Turkish Insurance Information and Monitoring Center. Property dam-
Cangul E. Development of an Expert System for the Quantification of Fault age only motor vehicle accident annual report statistics, Republic of
Rates in Traffic Accidents. [MSc thesis]. Middle East Technical Univer- Turkey Prime Ministry Undersecretariat of Treasury. 2015. Available
sity; 2010. at: http://www.sbm.org.tr/en. Accessed June 2015.
General Directorate of Highways. Traffic accident outlines. 2014. Avail- Turkish Statistical Agency. Road traffic accident statistics, Republic of
able at: http://www.kgm.gov.tr/SiteCollectionDocuments/KGMdocu Turkey Ministry of Internal Affairs. 2015. Available at: http://www.
ments/Trafik/TrafikKazalariOzeti2014.pdf. Accessed June, 2015. tuik.gov.tr/PreTablo.do?alt_id=1051. Accessed June 2015.
Goyal S, Goyal GK. Artificial neuron based models for estimating shelf life Virtual Crash (vCrash). Traffic Accident Reconstruction Tool. 2014. Avail-
of burfi. ARPN J Sci Technol. 2012;2:536–540. able at: http://www.vcrash3.com/ Accessed June 2015.