Road traffic accidents are the most common accidents that annually Endangers lives of many people in the world. Our country Iran is one of the countries with highest incidence and mortality due to accidents that has been introduced. So it’s requires identification of underlay in dimensions in this field. Due to the increasing amount of car accidents in order to increase volume of information related to car accidents and needs to explore and reveal hidden dependencies and very long time among this information. So using traditional methods to discover these complex relations don't response between involved factors and we need to use new techniques. Considering that main aim of this paper is to find best relationship between volumes of information in shortest time. So, in this paper, we classify accidents in West Azerbaijan province in Iran by accident type (damage, injury, death) and we describe it by using Particle Swarm Optimization (PSO) algorithm
EVALUATION OF PARTICLE SWARM OPTIMIZATION ALGORITHM IN PREDICTION OF THE CAR ACCIDENTS ON THE ROADS: A CASE STUDY
1. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
EVALUATION OF PARTICLE SWARM
OPTIMIZATION ALGORITHM IN PREDICTION
OF THE CAR ACCIDENTS ON THE ROADS: A
CASE STUDY
Farhad Soleimanian Gharehchopogh1, Zahra Asheghi Dizaji2, and Zahra Aghighi3
1
Department of Computer Engineering, Urmia Branch, Islamic Azad University, Iran
Bonab.farhad@gmail.com
2,3
Department of Computer Engineering, Science and Research Branch, IslamicAzad
University, West Azerbaijan, Iran.
2
Zahra_ashegi@yahoo.com, 3zaghighi@yahoo.com
ABSTRACT
Road traffic accidents are the most common accidents that annually Endangers lives of many people in the
world. Our country Iran is one of the countries with highest incidence and mortality due to accidents that
has been introduced. So it’s requires identification of underlay in dimensions in this field. Due to the
increasing amount of car accidents in order to increase volume of information related to car accidents and
needs to explore and reveal hidden dependencies and very long time among this information. So using
traditional methods to discover these complex relations don't response between involved factors and we
need to use new techniques. Considering that main aim of this paper is to find best relationship between
volumes of information in shortest time. So, in this paper, we classify accidents in West Azerbaijan
province in Iran by accident type (damage, injury, death) and we describe it by using Particle Swarm
Optimization (PSO) algorithm.
KEYWORDS
Particle Swarm Optimization, prediction of traffic accidents, Data Mining, classify by type of Accident,
machine learning
1. INTRODUCTION
Accidents are important part of daily life disasters take the lives of many people. Today, the issue
of economic and social costs of accidents and traffic fatalities because of that is the base problem
where challenges the transportation Specialists. This is more important for developing countries,
since, number of road accidents in developing countries is increasing and direct (such as medical
expenses caused by accidents and accident disability care) and indirect costs (such as mental
health problems and depression in family members, loss of the active labor force permanently or
temporarily) is more in comparison to developed countries. Injuries and damages caused by road
accidents is important and significant matter, which unfortunately is usually ignored. Public
health of society is a challenge that demands coordinated and integrated efforts and acts for
effective and continual prevention.
Accident includes errors and defects in function in one of these four factors:human, road,
environment around road and vehicle, While each of these four factors carry out their
DOI:10.5121/ijcsa.2013.3401
1
2. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
responsibilities correctly and be impeccable, Probability of accident happens nears to zero, But
when at least one of these four factors find a failure, any bad event are not unexpected. Among
themain elements ofan accident, human is the intelligent one, so that, in addition to its role as an
agent, controlling two other factors role in incidence of accidents is also possible by him [1-2].
Therefore, any accident happens, all judgments is facing to human factor, even if the other two
pillars of the vehicle and road cause an accident to happen. This subject is very important that
repeatedly through the mass media and road and safety officials of country to have 70% share of
human error in incidence of accidents is referred in country [3]. In Figure (1) the chain of
significant factors that caused incident or accident is shown.
Human
Vehicle
Feature, behavior-skills
and habits, Disability,
mental status,…
Segmentstechnical fault
Safety, Convenience
Accident
Environment
Weather condition,light
radiation status,…
Road
Geometric properties،
Road surface condition،
...Traffic
Figure(1): a chain of affecting factors of the incidence of accident [1]
Considering thecomplexity of accidents and probability ofits occurrence and also the multiple
involved factors and heterogeneous in this field, we see a different data products which each of
them made from various sources. In order toexploitthelarge numberof data and produced data we
need to connect this data and their hidden available patterns. Conventional methods do not have
feature analysis and detailed, significant and important information extraction in accident analysis
and probability of its occurrence. In this context it is necessary to use the information theory and
data analysis, one of new methods in large and various data analysis is using data mining
techniques.[4-5] One Data bases that contains complete information about the subject of complex
factors possibly associated with accidents, are 114forms. 114form is a type of form that is
completed by the police during a road accident in which a number of important factors and
information about the crash occurred is mentioned in the form .This information's include: This
information includes: type of accident, type of approach, the effective vehicle accidents, human
factors affecting the accidents, the total result of accidents, way affecting faults, In this paper, we
have used a method for prediction of PSO in road accidents based on the data. The rest of the
paper is organized as follows: section2 reviews the work done in this field, the proposed method
is presented in section 3, the proposed method is evaluated in section 4, and conclusions is
presented in section 5.
2
3. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
2. PREVIOUS WORKS
Accident data analysis using data mining techniques have been considered by many researchers
[5-6-7-8-9], used techniques in most of this studies is performed by using predictive analysis
means that the set of inputs is a particular output And predictive analysis to explore the
development of a set of observations or variables based on the similarity or lack of similarity
between these sets. Brijsest applied Search association rules to identify and distinguish high-risk
and low-risk areas, traffic patterns. In this paper, a technique were used for classifying a set of
association rules based on data accident and features of the site that had been divided into black
and white points. Analysis shows that association rules facilitate the identification of events that
often occur together that this will lead to a better understanding of road accidents. The results also
show that using the connect algorithm in addition to analysis possibility accident patterns in highrisk areas, give the possibility of finding features that lies between low-risk and high-risk areas
which are Discriminatory. The most important results show that although human behavioral
characteristics and role plays an important role in traffic accidents, but accidents in the high risk
areas are not affectless. The main difference in accident patterns between high-risk and low-risk
areas can be expressed in terms of infrastructure or location. It should be noted that this is a
(Belgium) case study [10].
Marukatat used Adaptive Regression Trees to build decision support systems that handle Road
accident analysis [11]. In these studies, classifying, selecting and filtering data, using rules
structure analysis has been done, so rule-based structure analysis of various parameters is done
many times. Based on the results we identified candidate rules. Researchers [12] were studied
cause of occurred accidents using association rules algorithm based on PSO algorithm. In this
paper, the concept of the association entropy is used to compare the characteristics of accidents
and also T-test model and Delphi method [12] to test the health of enhanced algorithm. In this
paper, an algorithm on a database of more than twenty thousand cases, each with 56 features was
tested. The end result showed that the improved algorithm is accurate and stable.
Halim Ferit Bayata et al. were employed Artificial Neural Networks (ANN) to model traffic
accidents. In these studies, the best model according to Akaike statistical criteria, select and
statistical analysis were tested in error condition. It should be noted that this is a case study (in
Turkey) [13]. According to accomplished studies by researchers [14] intensity of driving injuries
in accidents examined by using ANN s and decision trees. In these studies, according to the
results of the model a decision tree method was introduced better than ANN method, and
according to analysis three important factors were identified as follows: non-use of safety belts,
road lighting conditions and use of alcohol.
3. PROPOSED METHOD:
3-1. PSO Optimization Algorithm
PSO algorithm, the first time introduced in 1995 by Eberhart and Kennedy, is inspired by mass
movement of birds that are looking for food. In the first step of the algorithm, a random value is
generated for each particle, in this problem a Set of random values generated for the involved
factors in this type of accident then the fitness value is calculated based on the values which in
this problem fitness value is calculated by comparing random values with available facts (training
data).Now if the value of the fitness function is against zero, values for each particle is updated
from two optimum value, the optimal values are obtained based on the specified values for
involved factors in types of accidents .The first optimum value is the best position that the particle
is successful in reaching to it until now (The best values for the involved parameter sin type of
3
4. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
accident in particle itself|), this situation is called best and maintained. The Second optimal value
is used by the algorithm, is the best position ever achieved by a particle population (Best values
for the involved parameters in the accidents types among all particles). This situation is displayed
by gbest. After finding the optimal values, speed and position of each particle is updated by using
(1) and (2) equation.Then based on updated values since fitness function value is not zero or a
number of steps are not finish the cycle continues.
Right side of equation(1) is composed of three: the first parts is the current speed of particle and
second and third are the particle speed change and its rotation toward individual experience and
the best experience of group. If we do not consider the first part of this equation is determined
according to the current position and the best experience and the best experience of group. Thus
the best particles gathered and remain fixed in its situation and the others moving to toward that
particle. In fact, the collective motion of particles without the first part of equation(1), will be the
process which because of that search space gradually comes small, and local search forms around
best particle. In contrast, if we only consider the first part of equation (1) particles go their routine
way until reach to the walls and they do national search.
Equation(1) vi,j(t+1)=wvi,j(t)+c1r1,j(t)(pbesti,j(t)-xi,j(t))+c2r2(t)(gbesti(t)-xi,j(t))
Part 1
Part 2
Part 3
Equation (2) xi(t+1)=xi(t)+vi(t+1)
Part 1 Part 2
4
W: represents the inertia weight, the standard value is 1for local searches considers low value and
for public searches considers high value.
r1, r2: usually is between [0, 1].
C1, c2: are learning coefficients to control the pbest and gbest. The sum of these two numbers
should not be more than 4. Decreases1and increase c2 is a cause to moving toward optimum
particle, and as a result search space becomes smaller. Equation (2) is for getting a place for
particle that it will be transferred there. The right side of the equation consists of two parts: the
first part is the current location of the particle and second part is Particle momentum.
In this paper, we by using PSO algorithm that is a sample of evolutionary algorithms, we tried to
classify the accidents by type of it. Process of this prediction by assuming this the data collected
and additional data are omitted is shown in Figure 2.
4
5. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Read pre-processed
data
Start
Initialization
The data set used for
testing and training
Fitness Calculate
Send training data for
PSO algorithm to
learn
Experiences updates
N
o
Calculate
new values
N
o
You have
reached the
desired criteria?
Evaluation
of training
data
Y
es
Yes
Number of correct
answers
Finish
Assessment results
and determine the
percent error
Figure( 3)- PSO algorithm
Figure( 2)- The proposed method
As shown in Figure2 in first step processing data read by the model then based on need number of
data is considered randomly as training data. The rest of data used as test data to evaluate the
proposed method. After determining the training data based on this information for each and
every particle a series of laws defined and then based on this Initialization the fitted function of
each particle is calculated and chosen the best bit.PSO algorithm finishes by finishing number of
irritations or obtaining the desired criteria. Finally by law using samples of test data examined
and error percentage is determined. The main steps of the model are shown in Table (1).
5
6. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Table (1): The simulated proposedmethod
1-Ds=read data from dataset
2-Train_ds=get many random record for training
3-Exam_ds=get many random record for examine
4-Use pso algorithm for train with train_ds
1. Set the swarm size. Initialize the velocity and the position of each particle randomly.
2. For each j, evaluate the fitness value of xj and update the individual best position pbest if
better fitness is found.
3. Find the new best position of the whole swarm. Update the swarm best position gbest if
the fitness of the new best position is better than that of the previous swarm.
4. If the stopping criterion is satisfied, then stop.
Rules=Get best rules that outputposition algorithm
5. For each particle, update the of pso and the velocity according (1) and (2). Go to step
2.
5-Result= Evaluated according to the rules
6-Review and evaluate result
4. EVALUATION AND ASSESMENT
Examines data of this paper is related to accidents taking place in West Azerbaijan province axis
in a period of the first 6 months of 1390. The received data is from the Transportation and
Terminals organization West Azerbaijanas 114com forms. After an initial preprocessing of the
data some information extracted from this information. The results of these studies indicate that
among the four main causes of accidents Human factors, environmental factors, and the way the
vehicle), the influence of human factors, environmental factors, and the way is far greater than the
vehicle. In Figures (4-5-6-7-8-9-10-11) displayed result of the accidents data analysis.
Figure (4)-climate and its relationship with accidents
6
7. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Most accidents occurred on Clear days. But according to the frequency of accidents occurring and
a period of time, actually the most of fatal crashes occurred on rainy and snowy days.
Figure(5) –Member of the site and its relationship with accidents
Considering that production of travel around agricultural and residential land are much so rates of
fatal crashes around these sites are higher too.
Figure(6) –type of lineation and its relationship with accidents
Due to the lack of Non-compliance of precedence laws by drivers on the road, especially in
"lineation places" the amount of fatal crashes at locations with drawn and cross lineation is more.
7
8. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Figure (7) - type of the shoulder and its relationship with accident
Considering that the majority of vehicles on the parts of road which have soil shoulder,
passengers get in or get out so The rate of fatal crashes is more in this places
Figure(8) –lighting Status and its relationship with accidents
The figure (8) revealed that most accidents happened during the day.
8
9. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Figure (9) - treatment and its relationship with accidents
Most accidents happen at axis as a form of a dealing with another vehicle which figure (9) is also
confirmed this.
Figure(10) -Human factors to blame for the accident and its relationship with accidents
The human factor is one of the three major accident occurrence, As you can see drivers hurry that
has the highest proportion of deaths from traffic accidents
9
10. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
Figure (11) –Cause and its relationship with accidents
According to the results, the main reason for most accidents is, "non conformity of Priority” and
"inattention to up ".
After performing data preprocessing, PSO algorithm used for extract behavior and information
dependency. Using this algorithm, based on the available data, we can have accurate prediction
based on the type of happened accident. The results of the implementation of PSO algorithm are
shown in Table (2).
Table (2)–Resultsontraining data
Precision
Data that classified correctly
Data that classified incorrectly
Kappa percentage
Training data
Testing data
Whole data
Numbers
130
15
579
145
724
Percentage
89%
11%
80%
80%
20%
5. CONCLUSION AND FUTURE WORK
ANNs are suitable tools to adapt, learning and classification information. Collective intelligence
and neural networks can be regarded as a response to the challenge. In this paper, we investigated
using PSO algorithm to classify the accident based on their kind, considering the accident
occurrence in terms of significant parameters. According to performed simulations on the input
data (724 counts of accidents taking place in West Azerbaijan Province in Iran) using the
MATLAB tool for prediction results on kind of accidents were obtained with 89% Precision.
Based on these results, PSO algorithm can be used in software and future statistical program for
various data mining.
10
11. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
6. ACKNOWLEDGEMENTS
We take this opportunity to thank staff members of West AzerbaijanTransportation and Terminals
Organization for the valuable datasets presented by them in their respective fields. We grateful
for their cooperation during the period of our research .
7. REFERENCES
[1]
Oh C., Oh J. S. Ritchie S. G. and Chang M, “Real-time estimation of freeway accident likelihood”,
Annual Meeting CD-Rom 80th Annual Meeting of The Transportation Research Board, pp.1–16, January
7-11, 2001.
[2]
Juan de O˜n, Griselda López, Joaquín Abellán,” Extracting decision rules from police accident
reports through decision trees”, science direct,pp. 1151– 1160,2013
[3]
Ilna site, http://ilna.ir/news/news.cfm?id=44716,( News code 44716), {Last Available:
2013/02/05}
[4]
Hui-Huang Hsu,”Advanced Data Mining technologies in Bioinformatics”, Idea Group Publishing,
pp.1-324, United States of America, 2006
[5]
Bastian Bohn, JochenGarcke, Rodrigo Iza-Teran, Alexander Paprotny, Benjamin Peherstorfer, Ulf
Schepsmeier, Clemens-August Thole,” Analysis of Car Crash Simulation Data with Nonlinear Machine
Learning Methods”, Procedia Computer Science, ,pp. 000– 000,2013
[6]
Abdelwahab, H., Abdel-Aty, M.,“Development of artificial neural network models to
predict driver injury severity in traffic accidents at signalized intersections”, Transportation Research
Record 1746, pp. 6– 13 , 2001
[7]
Chang, Li-Y., Chen, Wen-C., “Data mining of tree-based models to analyze freeway accident
frequency”, Journal of Safety Research, Volume 36 (4), pp. 365– 375, 2005.
[8]
Chang, Li-Y., “Analysis of freeway accident frequencies: negative binomial regression
versus artificial neural network”, Safety Science, Volume 43 (8), pp. 541 –557, 2005.
[9]
MarkusDeublein, Matthias Schubert, Bryan T. Adey, Jochen Kohler, Michael H. Faber,”
Prediction of road accidents: A Bayesian hierarchical approach”, science direct, pp. 274– 291, 2013.
[10]
K. Geurts, G. Wets, T.Brijs, K. Vanhoof, “profiling high frequency accident locations using
association rules”, Transportation Research Record 1840, pp:123-130.,2003
[11]
R.Marukatat, “Structure-based rule selection frame work for association rule mining of traffic
accident data”, in Computational Intelligence and Security, vol.4456, pp.231–239, 2007.
[12]
Jianfeng Xi, ZhenhaiGao, ShifengNiu, Tongqiang Ding, and GuobaoNing,” A Hybrid Algorithm
of Traffic Accident Data Mining on Cause Analysis”, Mathematical Problems in Engineering, Volume
2013, pp.1-8, 2012
[13]
HalimFeritBayata, FatihHattatoglu and NeslihanKarsli ” Modeling of monthly traffic accidents
with the artificial neural network method”,” International Journal of the Physical Sciences”, Vol. 6(2), pp.
244-254, 18 January, 2011
[14]
S.Krishnaveni, M.Hemalatha,”A Perspective Analysis of Traffic Accident using Data Mining
Techniques”,” International Journal of Computer Applications (0975 – 8887)”, Volume 23, pp.40-48, June
2011.
11
12. International Journal on Computational Sciences & Applications (IJCSA) Vol.3, No.4, August 2013
AUTHORS
Farhad Soleimanian Gharehchopogh is currently Ph.d candidate in department
of computer engineering at Hacettepe University, Ankara, Turkey. And he works
an honors lecture in computer engineering department, science and research and
Urmia branches, Islamic Azad University, West Azerbaijan, Iran. For more
information please visit www.soleimanian.net
Zahra Asheghi Dizaji is a M.Sc. student in Computer Engineering Department,
Science and Research Branch, Islamic Azad University, West Azerbaijan, Iran. Her
interested research area are Data Mining and Meta-heuristic Algorithms. Email:
Zahra_ashegi@yahoo.com
Zahra Aghighii is a M.Sc. student in Computer Engineering Department, Science
and Research Branch, Islamic Azad University, West Azerbaijan, Iran. Her
interested research area are Networks, Software Developing and Meta-heuristic
Algorithms. Email: zaghighi@yahoo.com
12