Traffic Injury Prevention

ISSN: 1538-9588 (Print) 1538-957X (Online) Journal homepage: https://www.tandfonline.com/loi/gcpi20

Traffic accident reconstruction and an approach

for prediction of fault rates using artificial neural
networks: A case study in Turkey

Ali Can Yilmaz, Cigdem Aci & Kadir Aydin

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

To link to this article: https://doi.org/10.1080/15389588.2015.1122760

Accepted author version posted online: 13

Jan 2016.
Published online: 01 Jul 2016.

Submit your article to this journal

Article views: 320

View related articles

View Crossmark data

Citing articles: 4 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=gcpi20
TRAFFIC INJURY PREVENTION
, VOL. , NO. , –
http://dx.doi.org/./..

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

ABSTRACT ARTICLE HISTORY

Objective: Currently, in Turkey, fault rates in traffic accidents are determined according to the initiative of Received  August 
accident experts (no speed analyses of vehicles just considering accident type) and there are no specific Accepted  November 
quantitative instructions on fault rates related to procession of accidents which just represents the type of KEYWORDS
collision (side impact, head to head, rear end, etc.) in No. 2918 Turkish Highway Traffic Act (THTA 1983). The traffic accident
aim of this study is to introduce a scientific and systematic approach for determination of fault rates in most reconstruction; neural
frequent property damage–only (PDO) traffic accidents in Turkey. network; fault rate;
Methods: In this study, data (police reports, skid marks, deformation, crush depth, etc.) collected from the prediction
most frequent and controversial accident types (4 sample vehicle–vehicle scenarios) that consist of PDO
were inserted into a reconstruction software called vCrash. Sample real-world scenarios were simulated on
the software to generate different vehicle deformations that also correspond to energy-equivalent speed
data just before the crash. These values were used to train a multilayer feedforward artificial neural network
(MFANN), function fitting neural network (FITNET, a specialized version of MFANN), and generalized regres-
sion neural network (GRNN) models within 10-fold cross-validation to predict fault rates without using soft-
ware. The performance of the artificial neural network (ANN) prediction models was evaluated using mean
square error (MSE) and multiple correlation coefficient (R).
Results: It was shown that the MFANN model performed better for predicting fault rates (i.e., lower MSE and
higher R) than FITNET and GRNN models for accident scenarios 1, 2, and 3, whereas FITNET performed the
best for scenario 4. The FITNET model showed the second best results for prediction for the first 3 scenarios.
Because there is no training phase in GRNN, the GRNN model produced results much faster than MFANN
and FITNET models. However, the GRNN model had the worst prediction results. The R values for prediction
of fault rates were close to 1 for all folds and scenarios.
Conclusions: This study focuses on exhibiting new aspects and scientific approaches for determining fault
rates of involvement in most frequent PDO accidents occurring in Turkey by discussing some deficiencies in
THTA and without regard to initiative and/or experience of experts. This study yields judicious decisions to
be made especially on forensic investigations and events involving insurance companies. Referring to this
approach, injury/fatal and/or pedestrian-related accidents may be analyzed as future work by developing
new scientific models.

Introduction accident experts (no speed analyses of vehicles, just procession

of accident) and there are no specific quantitative instructions
Traffic accidents play a substantial role in fatalities in Turkey
on fault rates related to procession of accidents in No. 2918 Turk-
according to annual reports. According to the Turkish Statisti-
ish Highway Traffic Act (THTA 1983).
cal Agency (TSA) reports, these accidents are the second cause
There are several studies on determining the fault status of
of deaths at ages 5–29 and the third for those age 30–44. The
drivers. Islam et al. (2014) made a comprehensive analysis of
General Directorate of Highway reports indicate 3,524 fatalities
single- and multivehicle large truck at-fault crashes on rural and
and 285,059 injuries involving 1,199,010 accidents in 2014 (TSA
urban roadways in Alabama and they deduced that there existed
and General Directorate of Highway 2015).
important factors that significantly impact the injury severity
Traffic accident reconstruction has been crucial in recent
resulting from single-vehicle and multivehicle large truck at-
years to scientifically anticipate necessary information as accu-
fault accidents in urban and rural locations based on the esti-
rately as possible (speeds, fault rates, procession of accident, etc.)
mated values of average direct pseudo-elasticity. Cangul (2010)
based on physics laws. Partial or initiative-dependent analyses
developed an expert system for the quantification of fault rates
of experts on accident scenes bring to light challenging issues
in traffic accidents and assessed the critical factors affecting fault
regarding appropriate decisions, especially regarding what is a
rates with a quantitative questionnaire.
matter for the courts or forensic investigations. Currently, in
The motivation of this study is to exhibit an approach
Turkey, fault rates are determined according to the initiative of
referring to 3 different artificial neural network (ANN)-based

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.

systems for prediction of fault rates of the 4 most frequent acci-

dent types with regard to severity and dissipation of deformation
energy corresponding to the observed vehicles’ residual crushes
and speeds, without requiring simulations of probable accidents.
In this study, necessary data were obtained for the 4 most fre-
quent and controversial property damage–only (PDO) accident
scenarios from police reports, skid marks, vehicle deformation,
and crush depths. These data were inserted in a validated recon-
struction software to conduct simulations in 3D, called vCrash
(2014). ANN models were used to predict fault rates based on
vehicle deformation and speed data in order to develop a sys-
tematic and scientific approach to traffic accidents. The defor-
mations and velocities were regarded as input variables in the
prediction models and the fault rates were regarded as output
variables. The results of the prediction models were evaluated
by calculating their mean square error (MSE) and R values.
The results were compared with those obtained from multilayer
feedforward artificial neural network (MFANN), function fit-
ting neural network (FITNET), and generalized regression neu-
ral network (GRNN). It was shown that FITNET performed bet-
ter than MFANN and GRNN for predicting fault rates.

Overview of methods

Traffic accident reconstruction using vCrash software

Referring to TSA (2015) and the Turkish Insurance Informa-
tion and Monitoring Center’s (2015) accident reports for the
last 10 years, 4 most frequent PDO accidents were observed and
data relevant to these were used in the software. In the first step,
scenarios were examined in order to reconstruct the accident
as closely as possible. Related impulse–momentum and energy
transfer analysis of the software, type of accident, vehicle loca-
tions and situations after the crash, and damage were examined
and interpreted. Each vehicle brand in the software’s library has
material and strength features similar to its real-world version
(vCrash). Expert reports for similar accident scenes and at-fault
and not-at-fault (minor fault) definitions in No. 2918 THTA
were used.
Scenario 
In the first accident case, crash simulation of 2 average passen-
ger cars (APCs) on an equal-arm intersection (no traffic lights,
“stop”, “yield”, warning signs, and/or officer) was carried out. A
sample snapshot and its instant data related to accident proces-
sion is shown in Figure 1a.
Referring to THTA, at an intersection such as that in the
example, passing priority belongs to the vehicle coming from the
right (vehicle 1). In such cases, the at-fault vehicle is the one on
the left side (vehicle 2) and vehicle 1 is noted as minorly at fault
due to careless egressing into the intersection. The fault rates
are generally allocated to 75–25% or 80–20% by the initiative
of experts. However, there several questions arise in this case:
• What were the speeds of the vehicles just before the acci-
dent?
• Are there any skid marks on the road surface in order to
compute the collision velocities of the vehicles?
• If the right side vehicle enters the intersection above the
legal speed limit, can he or she be deemed not at fault?
• Is there any systematic method to determine the fault rates?
Figure . Sample snapshots from the first contact moments of vehicles from soft-
ware simulations (specifications that are not legible within the figures are shown
under each case’s figure).
At the time of a collision, a large deceleration is inevitable
that is caused by the initial peak force. The energy absorption
is a measure of the material’s capability to absorb impact energy.
The energy absorption (ED ) of a structure measures the capacity
to absorb impact energy, which can be computed from Eq. (1)
(Peng et al. 2015):
 δ
ED = F (δ)dδ, (1)
0
where F(δ) is instantaneous crashing force (N) and δ is
displacement (m).
Scenario 
The second accident case is a snapshot from the crash simulation
of a truck (vehicle 1) in the right lane that egresses in front of an
APC (vehicle 2) approaching in the left lane. A sample snapshot
and its instant data related to accident procession are shown in
Figure 1b. According to the expert report, the truck driver was
at fault with a rate of 75% for “not obeying lane changing rules”
and the other driver with a rate of 25% was minorly at fault due
to “careless driving” as stated in THTA without conducting a
comprehensive analysis. Similar questions a in the first case arise

and energy-equivalent speed–deformation energy analyses are
of great importance.
Scenario 
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
vehicle 2 at a speed within the legal limits. In this case, vehicle
2 was deemed at fault with a rate of 100% due to a “rear crash”
and vehicle 1 had a rate of 0%, or fault-free. However, no analysis
was conducted and the low speed of vehicle 1 and crush depths
were not taken into consideration. A snapshot from the crash
simulation of the vehicles is depicted in Figure 1c.
Scenario 
In the fourth case, an APC (vehicle 1) egresses in the intersection
when the signal is on intermittent red and the other APC (vehi-
cle 2) enters the intersection on intermittent yellow and they col-
lide each other as shown in the snapshot in Figure 1d. According
to expert report, vehicle 1 was at fault with a rate of 75% due to
“reckless, incautious driving and not slowing down while enter-
ing the intersection on intermittent red” and the other driver
with a rate of 25% was deemed minorly at fault due to “careless
driving and not slowing down while entering the intersection on
intermittent yellow” without conducting a comprehensive anal-
ysis.
Multilayer feed-forward artificial neural networks and
function fitting neural networks
ANNs are designed to solve problems that are difficult using
conventional mathematical methods. ANNs overcome the lim-
itation of the conventional approaches by extracting the desired
information directly from the data (Haykin 1994). MFANNs
trained with a back-propagation learning algorithm are the most
popular neural networks. MFANNs are organized in 3 layers: an
input layer, an output layer, and the layers in between, or hidden
layers (Svozil et al. 1997).
In a fully connected artificial neural network, each neuron in
a given layer is connected to each neuron in the following layer
by an associated numerical weight (wij ). The weight connect-
ing 2 neurons regulates the magnitude of the signal that passes
between them. In addition, each neuron possesses a numerical
bias term, corresponding to an input of −1 whose associated
weight has the meaning of a threshold value.
The goal of back-propagation training is to iteratively change
the weights between the neurons in a direction that mini-
mizes the error E, defined as the squared difference between the
desired and the actual outcomes of the output nodes, summed
over the training patterns (training set data) and the output neu-
rons, according to the steepest descent method in Eq. (2):
∂E
wi j (t ) = −η + αwi j (t − 1). (2)
∂wi j
The variation of a connection weight at the ith iteration
depends on the partial derivative of the total error with respect
to that weight through a proportionality constant termed the
learning rate (η) and on the variation in the same weight dur-
ing the previous iteration by means of a momentum term (α;
Marini et al. 2007).
TRAFFIC INJURY PREVENTION 587
FITNET is a modified form of MFANN. However, unlike the
log-sigmoid transfer function used in MFANN, FITNET uses
default tan-sigmoid transfer function in the hidden layers and a
linear transfer function in the output layer (MathWorks 2014).
Generalized regression neural networks
A GRNN is a kind of radial basis network that is often used for
function approximation. It has a radial basis layer and a special
linear layer. A GRNN does not require an iterative training pro-
cedure. It approximates any arbitrary function between input
and output vectors, drawing the function estimation directly
from the training data. In addition, it is consistent in that as the
training set size becomes large, the estimation error approaches
zero, with only mild restrictions on the function (Hannan et al.
2010).
The main function of a GRNN is to estimate a linear or non-
linear regression surface on independent variables (input vec-
tors) U and dependent variables (desired output vectors) X. That
is, the network computes the most probable value of an out-
put Ox given only training vectors U. Specifically, the network
computes the joint probability density function of U and X. The
expected value of X (E[X/U]) is expressed as shown in Eq. (3)
(Khashei et al. 2012):
∞
X f (U, X )dx
E[X/U ] = −∞
∞ . (3)
−∞ f (U, X )dx
The GRNN model has a 2-layer network. The first layer has
radial basis neurons and calculates weighted inputs with dis-
tance and network inputs. The second layer has pure linear
neurons that evaluate the weighted input with norm product
and network inputs with network summation (Goyal and Goyal
2012).
Data set generation
In the software step of this study, within scenarios 1 to 4,
vehicles with features similar to those in real-world accidents
were exposed to different deformations (minimum of 500 data
for each scenario) that also correspond to varied speed val-
ues. While one of the speeds was kept constant, the other’s was
increased by 5 km/h. It was assumed that every 5 km/h incre-
ment in speed corresponds to 3 points increment in fault rates
of the related vehicle. For every speed value, the simulation was
repeated and the crush depths were recorded from the simula-
tion report. One of the speed values was kept constant and the
other was changed every 20 trials until 500 data were obtained
for every accident scenario. Descriptive statistics for scenarios 1
and 2 and 3 and 4 are shown in Tables 1 and 2, respectively.
Results and discussion
In this article, 3 ANN models (MFANN, FITNET, and GRNN)
were adopted for the data set. For prediction of fault rates in
accidents using ANN models, input variables were SDef1 , SDef2 ,
V1 , and V2 ; Fault-rate1 and Fault-rate2 were taken as output
variables.
The MFANN and FITNET models have 3 layers (input, hid-
den, and output) and the input, hidden, and output layers have 4,
588 A. C. YILMAZ ET AL.

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 )

Minimum ./. ./. / / / /

Maximum ./. ./. / / / /
Mean .. ./. ./. ./. ./. /
SD ./. ./. ./. ./. ./. ./.

10, and 2 neurons, respectively. A log-sigmoid activation func-
tion is used in the MFANN model and the tan-sigmoid activa-
tion function is used in the FITNET model in the hidden lay-
ers. A pure linear activation function is used in both models in
the output layer and the Levenberg-Marquardt algorithm (More
1978) is utilized for training the networks. The other important
parameters of the MFANN and FITNET models are the number
of epochs (selected as 1,000), learning rate (selected as 0.02), and
momentum (selected as 0.5).
It should be noted that the most important architectural
parameter of the GRNN model is the spread constant, which has
a certain effect on prediction efficiency. In this article, the spread
value was set to 2.0 (default 1.0) empirically.
Each input and output variable of 500 simulations was used
to develop the models. The data were randomly divided into 2
disjoint subsets using 10-fold cross-validation. The training set
had 450 observations and the testing set had 50. The MSE in
Eq. (3) and R in Eq. (4) were utilized in order to compare the
prediction ability of the developed models. MATLAB software
(MathWorks 2014) was used for the experiments.
 N   data set contained 3 parameters (i.e., deformation energy, veloc-
 Qexp − Qcal 2
MSE = (4)
n and GRNN models were developed to predict fault rates in
1
and
r Because there is no training phase in GRNN, the GRNN
model produced results much faster than the MFANN
and FITNET models. However, the GRNN model had the
worst prediction results.
r The R values for prediction of fault rates were close to 1 for
all folds and scenarios.
r The MFANN model for scenario 3 yielded the lowest MSE
for prediction of fault rates, where the MSE changed from
0.000055 to 0.052155.
r The GRNN model for scenario 4 gave the highest MSE
for prediction of fault rates, where the MSE changed from
1.999691 to 3.806794.
This study focused on exhibiting new aspects and scientific
approaches for determining fault rates in most frequent PDO
accidents occurring in Turkey by discussing some deficiencies
in THTA and without regard to the initiative and/or experience
of experts. Referring to the direct proportion between speed and
deformation energy, a new systematic approach for determining
fault rates was carried out. vCrash software was used to simu-
late the accident scenes by means of 4 scenarios. The obtained
ity, and fault rate) for the accident scenes. MFANN, FITNET,
which R and MSE values of the developed models were also cal-
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 )

Minimum ./. ./. / / / /

Maximum ./. ./. / / / /
Mean ./. ./. ./. ./. ./. ./
SD ./. ./. ./. ./. ./. ./.
 TRAFFIC INJURY PREVENTION 589

Table . Results of scenarios  and  by means of -fold cross-validation (rounded values).

MSE R
Fold number MFANN FITNET GRNN MFANN FITNET GRNN

 ./. ./. ./. ./. ./. ./.

 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
Average ./. ./. ./. ./. ./. ./.

Table . Results of scenarios  and  by means of -fold cross-validation (rounded values).

MSE R
Fold Number MFANN FITNET GRNN MFANN FITNET GRNN

 ./. ./. ./. ./. ./. ./.

 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
 ./. ./. ./. ./. ./. ./.
Average ./. ./. ./. ./. ./. ./.

addition, these algorithms are likely applications to commer-
cialize. As future work, based on these approaches, various fault
rate studies related to injury/fatal accidents may be conducted
within various prediction methods like Support Vector Regres-
sion (SVR), fuzzy logic, etc. For pedestrian–vehicle collisions,
studies may be carried out on accidents with vehicular skid
marks or definite pedestrian throw distances. In the absence
of marks or throw distances, some physical dynamic models
(i.e., impact dynamics, thin plate theory) may be developed to
describe relations between pedestrian impact speed and deflec-
tion of impact region on the vehicle. Similar methods for in-
vehicle injury/fatal situations may be compared with real-world
accidents such as the deformation–speed analyses conducted
in this study.
Funding
This study was funded by Cukurova University, Scientific Research Projects
Unit, and the Department of Mechanical Engineering.
References
Cangul E. Development of an Expert System for the Quantification of Fault
Rates in Traffic Accidents. [MSc thesis]. Middle East Technical Univer-
sity; 2010.
General Directorate of Highways. Traffic accident outlines. 2014. Avail-
able at: http://www.kgm.gov.tr/SiteCollectionDocuments/KGMdocu
ments/Trafik/TrafikKazalariOzeti2014.pdf. Accessed June, 2015.
Goyal S, Goyal GK. Artificial neuron based models for estimating shelf life
of burfi. ARPN J Sci Technol. 2012;2:536–540.
Hannan SA, Manza RR, Ramteke RJ. Generalized regression neural network
and radial basis function for heart disease diagnosis. Int J Comput Appl.
2010;7:7–13.
Haykin S. Neural Networks: A Comprehensive Foundation. New York, NY:
MacMillan; 1994.
Islam S, Jones SL, Dye D. Comprehensive analysis of single- and multi-
vehicle large truck at-fault crashes on rural and urban roadways in
Alabama. Accid Anal Prev. 2014;67:147–158.
Khashei M, Hamadani AZ, Bijari B. A novel hybrid classification model of
artificial neural networks and multiple linear regression models. Expert
Syst. 2012;39:2606–2620.
Marini F, Magri AL, Bucci R. Multilayer feed-forward artificial neural net-
works for class modeling. Chemometr Intell Lab Syst. 2007;88:118–124.
MathWorks. MatLab. 2014. Available at: http://www.mathworks.
com/products/matlab/. Accessed May 4, 2015.
More JJ. The Levenberg-Marquardt algorithm: implementation and theory.
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
Structures. 2015;94:663–672.
Svozil D, Kvasnicka V, Pospichal J. Introduction to multi-layer feed-
forward neural networks. Chemometr Intell Lab Syst. 1997;39:43–
62.
THTA. No. 2918 Turkish Highway Traffic Act. Official Gazette of the Repub-
lic of Turkey. 1983;22:687.
Turkish Insurance Information and Monitoring Center. Property dam-
age only motor vehicle accident annual report statistics, Republic of
Turkey Prime Ministry Undersecretariat of Treasury. 2015. Available
at: http://www.sbm.org.tr/en. Accessed June 2015.
Turkish Statistical Agency. Road traffic accident statistics, Republic of
Turkey Ministry of Internal Affairs. 2015. Available at: http://www.
tuik.gov.tr/PreTablo.do?alt_id=1051. Accessed June 2015.
Virtual Crash (vCrash). Traffic Accident Reconstruction Tool. 2014. Avail-
able at: http://www.vcrash3.com/ Accessed June 2015.