You are on page 1of 9

Analysis of Road Traffic Accident Using Data

Mining
Nidhi Chandel
Asst. Professor
Computer Science Department,
Career College, Bhopal(M.P.)

Abstract – Engineers and researchers in the automobile industry have tried to design and
build safer automobiles, but traffic accidents are unavoidable. Patterns involved in
dangerous crashes could be detected if we develop a prediction model that automatically
classifies the type of injury severity of various traffic accidents. These behavioral and
roadway patterns are useful in the development of traffic safety control policy. We
believe that to obtain the greatest possible accident reduction effects with limited
budgetary resources, it is important that measures be based on scientific and objective
surveys of the causes of accidents and severity of injuries. This paper presents some
models to predict the severity of injury that occurred during traffic accidents using three
machine-learning approaches. We considered neural networks trained using hybrid
learning approaches, decision trees and a concurrent hybrid model involving decision
trees and neural networks. Experiment results reveal that among the machine learning
paradigms considered the hybrid decision tree-neural network approach outperformed
the individual approaches.

KEYWORDS
Traffic accident data mining, Support Vector Machines and Hybrid Decision Tree-ANN (DTANN),.

I. INTRODUCTION AND RELATED RESEARCH


The costs of fatalities and injuries due to traffic accidents have a great impact on
society. In recent years, researchers have paid increasing attention at determining
the factors that significantly affect driver injury severity in traffic accidents.Applying
data mining techniques to model traffic accident data records can help to understand the characteristics of
drivers’ behaviour, roadway condition and weather condition that were causally connected with different
injury severity. This can help decision makers to formulate better traffic safety control policies.
Related important works can be summarized as follows.
Abdelwahab et al. [3] studied the 1997 accident data for the Central Florida area
focusing on two-vehicle
accidents that occurred at signalized intersections. The injury severity was divided
into three classes: no
injury, possible injury and disabling injury. The performance of Neural Network (NN)
trained by Levenberg-
Marquardt algorithm and Fuzzy ARTMAP were compared, and found that NN (65.6%
and 60.4%
classification accuracy for the training and testing phases) performed better than
Fuzzy ARTMAP (56.1%).
Shankar, et al. [22] applied a nested logic formulation for estimating accident
severity likelihood
conditioned on the occurrence of an accident. The study found that there is a greater
probability of evident
injury or disabling injury/fatality relative to no evident injury if at least one driver did
not use a restraint
system at the time of the accident.
Kim et al. [11] developed a log-linear model to clarify the role of driver
characteristics and behaviors in the causal sequence leading to more severe injuries.
They found that driver behaviors of alcohol or drug use and lack of seat belt use
greatly increase the odds of more severe crashes and injuries.
Bedard et al. [4] applied a multivariate logistic regression to determine the
independent contribution of driver, crash, and vehicle characteristics to drivers’
fatality risk. It was found that increasing seatbelt use,
reducing speed, and reducing the number and severity of driver-side impacts might
prevent fatalities.
Yang, et al. used NN approach to detect safer driving patterns that have less chances
of causing death and injury when a car crash occurs [26]. They performed the
Cramer’s V Coefficient test to identify significant
variables that cause injury, therefore, reduced the dimensions of the data for the
analysis.
The 1997 Alabama interstate alcohol-related data was used and was found that by
controlling a single variable (such as driving speed, or light conditions) fatalities and
injuries could be reduced by up to 40%.
Ossiander et al. [19] use Poisson regression to analyze the association between the
fatal crash rate (fatal crashes per vehicle mile traveled) and the speed limit increase
[3] and found that the speed limit increase was associated with a higher fatal crash
rate and more deaths on freeways in Washington State. Some researchers studied
the relationship between drivers’ age, gender, vehicle mass, impact speed or driving
speed measure with fatalities.
Mussone et al. [17] used NN to analyze vehicle accident that occurred at
intersections in Milan, Italy. Results showed that the highest accident index for
running over of
II. ROAD TRAFFIC ACCIDENT AND INJURY ANALYSIS
A road traffic accident is defined as any vehicle accident occurring on a public
highway. It includes collisions between vehicles and animals, vehicles and
pedestrians, or vehicles and fixed obstacles. Single vehicle accidents, which involve a
single vehicle, that means without other road user, are also included (Safecarguide,
2004). At all levels, whether at national or international level, road traffic accidents
continue to be a growing problem. In connection with this, according to a World
Health organization/World Bank Report, deaths from non-communicable diseases are
expected to grow from 28.1 million a year in 1990 to 49.7 million by 2020, which is
an increase in absolute number of 77%. Traffic accidents are the main cause of this
rise. Road traffic injures are expected to take higher place in the rank order of
disease burden in the near future. . The rise in automobileownership together with
the poor condition of the roads has resulted in the high level of traffic safety and
congestion problems. In Ethiopia, above 1,800 people died while above 7,000 were
crippled or injured in 2003. Moreover the death rate is 136 per 10,000 vehicles and
Ethiopia is loosing over 400 million birr yearly as a result of road traffic accidents.
The share of Addis Ababa city in the total number of accidents was 60 percent in
1989 with annual average traffic accident growth 31.4 percent [Addis Ababa
Transport Authority (AATA), 2004

III. ADAPTIVE REGRESSION TREES


Data mining by itself is a process of finding, understanding and interpreting
interesting and useful knowledge. Usually, data mining tasks can be categorized into
either prediction or description. Directed data mining, which is a top-down approach,
is used when the ultimate result is expected. It is often also called predictive
modeling. Undirected data mining, basically used to identify patterns in a given data
and let the user to determine the interestingness of it. It is normally a bottom-up
approach of describing the data.
The Classification and Adaptive Regression Trees (CART) methodology is technically
known as binary recursive partitioning. The process is binary because parent nodes
are always split into exactly two child nodes, and recursive because the process can
be repeated by treating each child node as a parent. The key elements of a CART
analysis are a set of rules for splitting each node in a tree:
deciding when a tree is complete
assigning each terminal node to a class
outcome (or predicted value for regression) CART is the most advanced decision-tree
technology for data analysis, pre-processing and predictive modeling. CART is a
robust data-analysis tool that automatically searches for important patterns and
relationships and quickly uncovers hidden structure even in highly complex data.
CART's binary decision trees are more sparing with data and detect more structure
before further splitting is impossible or stopped. Splitting is impossible if only one
case remains in a particular node, or if all the cases in that node are exact copies of
each other (on predictor variables). CART also allows splitting to be stopped for
several other reasons, including that a node has too few cases 0. Once a terminal
node is found we must decide how to classify all cases falling within it. One simple
criterion is the plurality rule: the group with the greatest representation determines
the class assignment. CART goes a step further: because each node has the potential
for being a terminal node, a class assignment is made for every node whether it is
terminal or not. The rules of class assignment can be modified from simple plurality
to account for the costs of making a mistake in classification and to adjust for over –
or under – sampling from certain classes.

IV. Data Collection


Collecting, analyzing and understanding the content and structure of the data
available is one of the most important tasks that need close attention. With respect
to this specific research the sole source of data about an accident is the daily
accident report form to be filled and reported by the traffic police officers. It consists
of full details about a given accident. Through successive update, the office keeps
this data in excel flat file format. This helps a great deal in understanding the data
and viewing it from different angle in exploring its potential. This study used data
from the Addis Ababa Traffic Office (AATO). The initial data source for the study
contained traffic accident records from September, 1995 up to March 1997 E.C a
total number of 25,560 cases. The first attempt of the office to electronically record
the accident data in September, 1995 was limited to some attributes like labels of
date and time, accident id , driver’s name, driver’s age, driver’s gender, driver’s
license status, relation of the driver and vehicle, driver’s experience, possession of
the vehicle, vehicle type, accident area, accident road name, road segment
separation, road direction, road surface type, roadway surface condition, vehicle
maneuver, accident type, total vehicles involved, total number of victims, accident
victims category, victims profession, victims health condition, pedestrian maneuver,
vehicle plate number, cost estimate of the damage and injury severity. Through time
specifically starting from September 1995, attempt has been made to incorporate
some more attributes like, driver’s educational level, vehicle defect, vehicle age,
weather condition, light condition, cause for accident. The accident record’s database
AATO is still under revision to incorporate as many attribute as possible so that to
create a more comprehensive database. Because of the unavailability of important
attributes in the accident data of the previous years, the data set selected for this
specific research covers the time from September 1995 to March 2005, which is a
total of 5,207 records.

V. ACCIDENT DATA SET


A. Data Understanding
A good understanding of the data at hand leads to a better success in achieving the
data mining goal. The success criterion for this data mining research is the discovery
of accident severity classification rules that would find out and differentiate accidents
which are serious to those which are potentially not serious in different levels.
Provided that reasonable accident severity classification rules are discovered, the
office could device a means to reduce the number of fatal and serious injuries and be
able to recognize the level of severity when an accident has occurred. In short, this
can help decision makers to formulate better traffic safety control policies. As
indicated above the organization keeps detailed information about an accident
through various attributes. Types of information that the office records about an
accident includes driver and vehicle characteristics, road and weather conditions,
date and time of the accident, type, injury severity and possible causes of such
accidents. The specific attributes by which a given accident can be described are
date and time, accident id , driver’s name, vehicle type, driver’s age, driver’s gender,
driver’s educational level, driver’s license status, relation of the driver and vehicle,
driver’s experience, possession of the vehicle, vehicle defect, vehicle age, accident
area, accident road name, road segment separation, road direction, road surface
type, roadway surface condition, light condition, weather condition, vehicle
maneuver, accident type, total vehicles involved, total number of victims, accident
victims category, victims profession, victims health condition, pedestrian maneuver,
vehicle plate number, cost estimate of the damage and cause for accident. In
addition to the input variables mentioned above the output variable for this research
that is injury severity is also another attribute of a given accident. The target
attribute, injury severity, has four classes: fatal injury, property damage (no injury),
serious injury and slight injury.
B.Defining the Data Mining Function
Each individual accident record in the data set is an input/output pair with each
record having an associated output. The output variable, the injury severity, is
categorical and as described above, has four classes. A supervised learning algorithm
employed maps an input vector to the desired output class. Accurate results of such
data analysis could provide crucial information for the road accident prevention
policy.
This study used data from the National Automotive Sampling System (NASS) General Estimates System
(GES) [21]. The GES datasets are intended to be a nationally representative probability sample from the
annual estimated 6.4 million accident reports in the United States. The dataset for the study contains traffic
accident records from 1995 to 2000, a total number of 417,670 cases. According to the variable definitions
for the GES dataset, this dataset has drivers’ only records and doesn’t include passengers’ information. It
includes labels of year, month, region, primary sampling unit, the number describing the police jurisdiction,
case number, person number, vehicle number, vehicle make and model; inputs of drivers’ age, gender,
alcohol usage, restraint system, eject, vehicle body type, vehicle age, vehicle role, initial point of impact,
manner of collision, rollover, roadway surface condition, light condition, travel speed, speed limit and the
output injury severity. The injury severity has five classes: No Injury, Possible Injury, Non-incapacitating
Injury, Incapacitating Injury, and Fatal Injury. In the original dataset, 70.18% of the cases have output of no
injury, 16.07% of the cases have output of possible injury, 9.48% of the cases have output of non
incapacitating injury, 4.02% of the cases have output of incapacitating injury, and 0.25% of the cases have
fatal injury. Our task was to develop machine learning based intelligent models that could classify the
severity of injuries (5 categories) more accurately. This can in turn lead to greater understanding of the
relationship between the factors of driver, vehicle, roadway, and environment and driver injury severity.
Accurate results of such data analysis can provide crucial information for the road accident prevention
policy.
B. Data Preparation
The input and output variables are considered for the model building. There are no conflicts between the
attributes since each variable represents its own characteristic. The variables are already categorized and
are represented by numbers. The manner in which the collision occurred has 7 categories: not collision,
rear-end, head-on, rear-to-rear, angle, sideswipe same direction, and sideswipe opposite direction. For these
7 categories the distribution of the fatal injury is as follows: 0.56% for not collision, 0.08% for rear-end
collision, 1.54% for head-on collision, 0.00% for rear-to-rear collision, 0.20% for angle collision, 0.08%
for sideswipe same direction collision, 0.49% for sideswipe opposite direction collision. Since head-on
collision has the highest percent of fatal injury; therefore, the dataset was narrowed down to head-on
collision only. Head-on collision has a total of 10,386 records. There are 160 records of head-on collision
with fatal injury; all of these 160 records have the initial point of impact categorized as front. The initial
point of impact has 9 categories: no damage/non-collision, front, right side, left side, back, front right
corner, front left corner, back right corner, back left corner. The head-on collision with front impact has
10,251 records; this is 98.70% of the 10,386 head-on collision records. We have therefore decided to focus
on front impact only and removed the remaining 135 records. Travel speed and speed limit will not be used
in the model because there are too many records with unknown value; for 67.68% of records the travel
speed during accident and local speed limit were unknown. This means that the remaining input variables
are: drivers’ age, gender, alcohol usage, restraint system, eject, vehicle body type, vehicle role, vehicle age,
rollover, road surface condition, light condition. Vehicle age with values 37, 41, 46 and 56 each has only
one record and these were the only records representing such old cars. These four records
were therefore deleted from the dataset since they were clear outliers. Thus, finally, the dataset for
modeling had 10,247 records. There were 5,171 (50.46%) records with no injury, 2138 (20.86%) records
with possible injury, 1721 (16.80%) records with non-incapacitating injury, 1057 (10.32%) records with
incapacitating injury, and 160 (1.56%) records with fatal injury. We have separated each output class and
used one-against-all approach. This approach selects one output class to be the positive class, and all the
other classes are combined to be the negative class. We set the output value of the positive class to 1, and
the (combined) negative class(es) to 0. We divided the datasets randomly into 60%, 20%, and 20% for
training, crossvalidation, and testing respectively.

VI. Data mining techniques


Descriptive techniques were useful to classify the large amount of analyzed accidents, and between the two
techniques considered, K-means clustering appeared to be more effective. Clusters from the K-means
algorithm appear to be more clear-cut defined and classification is less trivial than the one obtained with the
Kohonen networks, where the occurrence of the accident during day or during night seemed to be the most
relevant factor to identify similarities among the collisions. The limit of descriptive techniques is that for
safety analysis they do not provide indications different from the mere description of the main
characteristics of the accidents, and for this reason predictive methods have to be considered more suitable.
Neural networks presented a similar problem of interpretation of the results, especially in terms of
understanding the relative importance of the inputs and of the intermediate neurons that constituted the two
hidden levels connecting inputs and outputs. The definition of the outputs was a general problem in the
prediction phase, as neural networks and decision trees require the a priori definition of the output variables
and consequently highly depend on this variable choice.

From a general perspective, once a problem is assigned and an output variable is accordingly defined,
decision trees appear the most promising techniques to investigate the problem and provide rules able to
explain the phenomenon. The extraction of the rules is extremely simpler than with neural networks, and
the different algorithms help treating any different type of variables (nominal, ordinal, categorical etc.), just
as in this research the CHAID algorithm was the most suitable to the available data.

PERFORMANCE ANALYSIS
A. Decision Trees
The decision tree model consists of a hierarchy of univariate binary decisions . Each internal node in the
tree specifies a binary test on a single variable, branch represents an outcome of the test, each leaf node
represent class labels or class distribution. The decision tree algorithm operates by choosing the best
variable for splitting the data into two groups at the root node, partitioning the data into two disjoint
branches in such a way that the class labels in each branch are as homogeneous as possible, and splitting is
recursively applied to each branch, and so forth. Once a maximal tree is generated, it examines smaller
trees obtained by pruning away branches of the maximal tree. Once the maximal tree is grown and a set of
sub-trees is derived from it, the decision tree algorithm determines the best tree by testing for
misclassification error rates.
We trained each class with Gini goodness of fit measure, the prior class probabilities was set to equal, the
stopping option for pruning was misclassification error, the minimum n per node was set to 5, fraction of
objects was 0.05, the maximum number of nodes was 1000, the maximum number of level in tree was 32,
the number of surrogates was 5, we used 10 fold cross-validation, and generated the comprehensive results.
The cross-validation testing ensures that the patterns found will hold up when applied to new data. The
performances for no injury, possible injury, non-incapacitating injury, incapacitating injury and fatal injury
were 67.54%, 64.39%, 60.37%, 71.38%, and 89.46% respectively. Empirical results including
classification matrix are illustrated in Table 1.
TABLE 1. DECISION TREE PERFORMANCE
Injury Class DT Accuracy (%)

No Injury: 67.54

Possible Injury : 64.40

Non-incapacitating Injury: 60.37

Incapacitating Injury: 71.38

Fatal Injury: 89.46


B.Support Vector Machines (SVM)
SVMs are kernel-based learning algorithms in which only a fraction of the training examples are used in
the solution (these are called the support vectors), and where the objective of learning is to maximize a
margin around the decision surface. The flexibility of kernel functions allows the SVM to search a wide
variety of hypothesis spaces. The basic idea of applying SVMs to pattern classification can be stated briefly
as: first map the input vectors into one feature space (possibly with a higher dimension), either linearly or
nonlinearly, which is relevant with the selection of the kernel function; then within the feature space, seek
an optimized linear division, i.e. construct a hyperplane which separates two classes. We experimented
with polynomial kernel and radial basis function kernel. For some reason, polynomial kernel was not that
successful and hence we only focused on radial basis function (RBF) kernels. Table 2 illustrates the SVM
performance for the different parameter settings and the accuracy of each experiment for each class.

TABLE 2: PERFORMANCE OF SVM USING RADIAL BASIS FUNCTION KERNEL

g=0.0001 g=0.001 g=1.2 g=1.5 g=2 g=0.00001 g=0.0001 g=0.001


c=42.8758 c=4.6594 c=0.5 c=2 c=10 c=100 c=100 c=100

No injury

Class 0 59.76 59.80 57.95 57.65 53.6 54.12 57.34 59.76 60.46
2
Class 1 60.14 60.14 60.82 55.63 55.7 55.53 62.88 60.14 60.14
3
Possible injury

Class 0 100.00 100.00 100.0 99.88 95.3 95.58 100.00 100.00 100.00
0 3
Class 1 0.00 0.00 0.00 0.00 3.67 3.42 0.00 0.00 0.00
Non-incapacitating

Class 0 100.00 100.00 100.0 100.0 97.4 97.49 100.00 100.00 100.00
0 0 3
Class 1 0.00 0.00 0.00 0.00 3.21 2.92 0.00 0.00 0.00
Incapacitating

Class 0 100.00 100.00 100.0 99.89 98.0 98.11 100.00 100.00 100.00
0 6
Class 1 0.00 0.00 0.00 2.83 0.00 0.00 0.00 0.00 0.00
Fatal Injury

Class 0 100.00 100.00 100.0 100.0 99.9 99.95 100.00 100.00 100.00
0 0 5
Class 1 0.00 0.00 3.33 0.00 0.00 0.00 0.00 0.00 0.00

C.Neural Networks
Hyperbolic activation function was used in the hidden layer and logistic activation function in the output
layer. The models were trained with BP (100 epochs, learning rate 0.01) and SCGA (500 epochs) to
minimize the Mean Squared Error (MSE). For each output class, we experimented with different number of
hidden neurons, and selected the model with highest classification accuracy for the class. From the results
of experiments, for no injury class the best model had 65 hidden neurons, and achieved training and testing
performance of 63.86% and 60.45% respectively. For possible injury class, the best model had 65 hidden
neurons achieving it’s training and testing performance of 59.34% and 57.58% respectively. For
nonincapacitating injury class, the best model had 75 hidden neurons achieving training and testing
performance of 58.71% and 56.8% respectively. For incapacitating injury class, the best model had 60
hidden neurons achieving training and testing performance of 63.40% and 63.36% respectively. Finally, for
fatal injury class, the best model had 45 hidden neurons achieving training and testing performance of
No Injury
Neurons Accuracy % Fatal Injury
# Train Test 78.61% and 78.17% respectively. These results are the
Neurons Accuracy % summary of multiple experiments as illustrated in Table 3.
60 63.5 59.67
7 # Train Test
TABLE 3. NEURAL NETWORK PERFORMANCE
65 63.8 60.45
6 45 77.2 75.17
70 63.9 60.25 6
Non-incapacitating Incapacitating
3 57 74.7 70.65
Neurons Accuracy % Neurons Accuracy %
75 64.3 57.43 8
# Train Test # Train Test
8 65 69.8 69.73
80 63.6 58.89 1
60 57.8 55.25 60 63.4 63.36
4 75 60.1 59.62
8 65 62.2 61.32
Possible Injury 9
65 57.6 54.66 3
Neurons Accuracy % 80 74.3 71.77
9 75 61.0 61.52
# 3
Train Test 70 58.7 56.80 6
1 84 63.2 58.41
60 63.5 59.67 D.Hybrid Decision Tree- 75 57.7 54.13 3
7 ANN (DTANN) 8 90 59.3 59.08
65 63.8 60.45 We used a concurrent 80 57.8 55.59 2
6 hybrid model where traffic 3
75 63.9 60.25 accidents data are fed to
3 the decision tree to generate the node information. Terminal nodes are numbered left to
80 64.3 57.43 right starting with 1. All the data set records are assigned to one of the terminal nodes,
8 which represent the particular class or subset. The training data together with the node
95 63.6 58.89 information were supplied for training ANN. For the hybrid decision tree–ANN, we used
4 the same hybrid learning algorithms and parameters setting as we used for ANN (except
for the number of hidden neurons). Experiments were performed with different number of
hidden neurons and models were selected with the highest classification accuracy for the output class.
From the experiment results, for no injury class the best model had 70 hidden neurons, with training and
testing performance of 83.02% and 65.12% respectively. For possible injury class, the best model had 98
hidden neurons with training and testing performance of 74.93% and 63.10% respectively. For non-
incapacitating injury class, the best model had 109 hidden neurons with training and testing performance of
71.88% and 62.24% respectively. For incapacitating injury class, the best model had 102 hidden neurons,
with training and testing performance of 77.95% and 72.63% respectively. For fatal injury class, the best
model had 76 hidden neurons with training and testing performance of 91.53% and 90.00% respectively.
These are the best models out of multiple experiments. Empirical results are presented in Table 4 .For all
the output classes, the hybrid DTANN outperformed the ANN. For non-incapacitating injury,
incapacitating injury, and fatal injury classes, the hybrid DTANN outperformed both ANN and DT.
TABLE 4. TEST PERFORMANCE OF DTANN

Injury type %
Accuracy

DTANN

No Injury 65.12

Possible Injury 63.10

Non-Incapacitating 62.24
Injury
Incapacitating Injury 72.63

Fatal Injury 90.00

V. CONCLUSIONS
The conclusion of this research is that the importance of the data is extremely crucial when data mining and
accident analysis are involved. Further, more independence should be given to the analyst, in the sense that
data should be as general as possible in the description of the information, in order for the analyst to work
on the definition of the classification. Of course, more accuracy in the collection of the information
regarding the accidents would be welcome, as not only missing data are frequent, but also errors in the data
gathering.

In this paper, we studied the GES automobile accident data from 1995 to 2000 and investigated the
performance of neural network, decision tree, support vector machines and a hybrid decision tree - neural
network for predicting drivers’ injury severity in head-on front impact point collisions. The classification
accuracy on the test results reveals that, for non-incapacitating injury, incapacitating injury, and fatal injury
classes, the hybrid approach performed better than neural network, decision trees and support vector
machines. For no injury and possible injury classes, the hybrid approach performed better than neural
network. The no injury and possible injury classes could be best modeled by decision trees. Previous
researches had focus mainly on no injury and injury (including fatality) classes. In this paper, we extended
the research to possible injury, non-incapacitating injury, incapacitating injury, and fatal injury. Our
experiments showed that the model for fatal and non-fatal injury performed better than other classes. The
ability of predicting fatal and non-fatal injury is very important since drivers’ fatality has the highest cost to
society economically and socially. One very important factor of causing different injury level is the actual
speed that the vehicle was going when the accident happened. Our dataset doesn’t provide enough
information on the actual speed since speed for 67.68% of the data records’ was unknown. If the speed was
available, it might help to improve the models performance. From an intelligent systems point of view it is
interesting to note about the failure of SVMs to model the complexity of the different injury classes.

VI. REFERENCES

[1]Abdalla, I.M., Robert, R., Derek, B. and McGuicagan, 1997, D.R.D. “An investigation into the
relationships between area social characteristics and road accident casualties”. Accidents Analysis and
Preventions, 5, pp. 583–593.
[2]Abdel-Aty, M., and Abdelwahab, H. 2003, “Analysis and Prediction of Traffic Fatalities Resulting From
Angle Collisions Including the Effect of Vehicles’ Configuration and Compatibility”. Accident Analysis
and Prevention.
[3]Abdelwahab, H. T. and Abdel-Aty, M. A. 2001, “Development of Artificial Neural Network Models to
Predict Driver Injury Severity in Traffic Accidents at Signalized Intersections”. Transportation Research
Record 1746, Paper No. 01-2234.
[4]Bedard, M., Guyatt, G. H., Stones, M. J., and Hireds, J. P. 2002, “The Independent Contribution of
Driver, Crash, and Vehicle Characteristics to Driver Fatalities”. Accident analysis and Prevention, 34, pp.
717-727.
[5]Breiman, L., Friedman, J., Olshen, R., and Stone, C. J. 1984, “Classification and Regression Trees”,
Chapman and Hall, New York.
[6]Buzeman, D. G., Viano, D. C., and Lovsund, P. 1998, “Car Occupant Safety in Frontal Crashes: A
Parameter Study of Vehicle Mass, Impact Speed, and Inherent Vehicle Protection”. Accident Analysis and
Prevention, 30 (6), pp. 713-722.
[7]Chong M., Abraham A., Paprzycki M. 2004, “Traffic Accident Analysis Using Decision Trees and
Neural Networks”, In Proceedings of International Conference on Applied Computing, (IADIS), Nuno
Guimarães and Pedro Isaías (Eds.), pp. 39-42, Portugal.
[8]Dia, H., and Rose, G. 1997, “Development and Evaluation of Neural Network Freeway Incident
Detection Models Using Field Data”. Transportation Research C, 5(5), pp. 313-331.
[9]Evanco, W. M.1999, “The Potential Impact of Rural Mayday Systems on Vehicular Crash Fatalities”.
Accident Analysis and Prevention, 31, pp. 455-462.
[10]Hand, D., Mannila, H., and Smyth, P. 2001, ‘Principles of Data Mining”. The MIT Press.
[11]Kim, K., Nitz, L., Richardson, J., and Li, L. 1995, “Personal and Behavioral Predictors of Automobile
Crash and Injury Severity”. Accident Analysis and Prevention, 27(4), pp. 469-481.
[12]Kweon, Y. J., and Kockelman, D. M. 2003, “Overall Injury Risk to Different Drivers: Combining
Exposure, Frequency, and Severity Models”. Accident Analysis and Prevention, 35, pp. 441-450.
[13]Martin, P. G., Crandall, J. R., and Pilkey, W. D., 2002, “Injury Trends of Passenger Car Drivers In the
USA”. Accident Analysis and Prevention, 32, pp. 541-557.
[14]Mayhew, D. R., Ferguson, S. A., Desmond, K. J., and Simpson, G. M. 2003, “Trends In Fatal Crashes
Involving Female Drivers”, 1975-1998. Accident Analysis and Prevention, 35, pp. 407-415.
[15] Mekitew Mola. 2000, “Traffic Information System: The case of Addis Ababa Traffic Police
Department”. Master’s Thesis. Addis Ababa University. Addis Ababa.
[16] Miaou, S.P. and Harry, L. 1993, “Modeling vehicle accidents and highway geometric design
relationships”. Accidents Analysis and Prevention, 25 (6), pp. 689–709.
[17] Mussone, L., Ferrari, A., and Oneta, M. 1999, “An analysis of urban collisions using an artificial
intelligence model”. Accident Analysis and Prevention, 31, pp. 705-718.
[18] Ossenbruggen, P.J., Pendharkar, J. and Ivan, J. 2001, “Roadway safety in rural and small urbanized
areas”. Accidents Analysis and Prevention, 33 (4), pp. 485–498.
[19] Ossiander, E. M., and Cummings, P. 2002, “Freeway speed limits and Traffic Fatalities in Washington
State”. Accident Analysis and Prevention, 34, pp. 13-18.
[20] Peltzman, S. 1975, “The effects of automobile safety regulation”. Journal of Political Economy 83, pp.
677–725.
[21] Roh J.W., Bessler D.A. and Gilbert R.F. 1998, “Traffic fatalities, Peltzman’s model, and directed
graphs”. Accident Analysis and Prevention, 31 (1-2), pp. 55-61.
[22] Shankar, V., Mannering, F., and Barfield, W. 1996,“Statistical Analysis of Accident Severity on Rural
Freeways”. Accident Analysis and Prevention, 28(3), pp.391-401.
[23] Sohn, S. Y., and Lee, S. H. 2003, “Data Fusion, Ensemble and Clustering to Improve the
Classification Accuracy for the Severity of Road Traffic Accidents in Korea”. Safety Science, 4(1), pp. 1-
14. Steinberg, D. and Colla, P. L. 1995, “CART: Tree-Structured Nonparametric Data Analysis”, San
Diego, CA: Salford Systems.
[24] Tavris, D. R., Kuhn, E. M, and Layde, P. M. 2001, “Age and Gender Patterns in Motor Vehicle Crash
injuries: Importance of Type of Crash and Occupant Role”. Accident Analysis and Prevention, 33, pp. 167-
172.
[25] Taddele, D. and Larson, C.P . 1991, “The occurrence and driver characteristics associated with motor
vehicle injuries in Addis Ababa, Ethiopia”. Journal of Tropical Medicine and Hygiene, 94, 395-400.
[26] Yang, W.T., Chen, H. C., and Brown, D. B. 1999, “Detecting Safer Driving Patterns By A Neural
Network Approach”. In Proceedings of Smart Engineering System Design Neural Network, Evolutionary
Programming, Complex Systems and Data Mining, pp 839-844.