See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/309591840

Detection of Motorway Disorders by Processing and Classification of

Smartphone Signals Using Artificial Neural Networks

Article in International Journal of Natural Sciences Research · November 2016

DOI: 10.18488/journal.63/2016.4.3/63.3.56.67

CITATIONS READS

3 435

3 authors:

Yusra Jaff Ali Kattan

Tishk International University Code of Science
6 PUBLICATIONS 16 CITATIONS 33 PUBLICATIONS 270 CITATIONS

SEE PROFILE SEE PROFILE

Taner Cevik
Istanbul Rumeli University
37 PUBLICATIONS 345 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ali Kattan on 01 November 2016.

The user has requested enhancement of the downloaded file.

 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

International Journal of Natural Sciences Research

ISSN(e): 2311-4746/ISSN(p): 2311-7435

URL: www.pakinsight.com

DETECTION OF MOTORWAY DISORDERS BY PROCESSING AND

CLASSIFICATION OF SMARTPHONE SIGNALS USING ARTIFICIAL NEURAL
NETWORKS

Yusra Mohammed M. Salih1† --- Ali Kattan2 --- Taner Çevik3

1
Sulaymaniyah, Iraq
2
Erbil, Iraq
3
Istanbul, Turkey

ABSTRACT
Potholes, debris, sunken manhole covers and others are common street safety hazards which drivers experience daily
as they bump into them unexpectedly while driving. Pavement roughness is usually evaluated based on the
International Roughness Index (IRI), which is considered the most prevalent metric. In this work, IRI values are
collected by using a smartphone, with built-in vibration sensor, placed on the car’s dashboard while driving around
the city. The classification process of IRI values is primarily performed using an Artificial Neural Network (ANN) for
the detection of diverse predefined street safety hazards. The designed ANN is a backpropagation pattern classifier,
that must be trained to yield either a detected “disorder” area of the road or “normal” area based on the IRI data
collected. The process of preparing the training and testing datasets involves a number of pre-processing operations.
The IRI values are pre-processed in order to extract the most effective features. Then the network is trained with the
normalized feature set by using supervised learning method. The performance of the designed network is compared to
a similar works in the literature. Results show that the designed network can successfully classify the street
conditions by using IRI values with a success rate that outperforms the classification rates obtained by other works.
© 2016 Pak Publishing Group. All Rights Reserved.

Keywords: Road safety, Smartphone, International roughness index, Artificial neural network, Backpropagation, Supervised training.

Received: 15 August 2016/ Revised: 31 August 2016/ Accepted: 16 September 2016/ Published: 21 October 2016

1. INTRODUCTION
Road roughness, also referred to as smoothness, is an important pavement characteristic. It affects not only
driving quality but also vehicle performance, fuel consumption and maintenance costs. Road maintenance is a vital
and challenging service that city municipals must handle carefully. Potholes, debris and sunken manhole covers are
some of the common road disorders that poses safety risks and also sources of complaints by drivers [1]. Detecting
the road condition has gained significant importance in the last few years. There are various reasons for attracting the
focus of the researchers to this field. Firstly, it will ensure safety and comfort and secondly it will save considerable
costs related to maintenance as problems can be addressed as soon as possible. Smooth and safe roads will lead to less
vehicle damage and government investment [2].
Various proposed techniques require dedicated hardware installed in vehicles or at several road joints [3-5].
These methods typically focus on roads of developed countries that owe relatively simple road and traffic flow

56
† Corresponding author
DOI: 10.18488/journal.63/2016.4.3/63.3.56.67
ISSN(e): 2311-4746/ISSN(p): 2311-7435
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

patterns. Furthermore, these methods are not cheap since they require specialized hardware, time and effort.
Attaching sensors to a large number of vehicles and various road joints is impractical due to the great financial cost
and human effort involved.
In the past few years, the widespread of low-cost smartphones and globally prevalent Internet access through
these devices have inspired developers to write promising applications [6]. Large memory capacity, powerful
processing capabilities, high-speed internet connectivity through 3G/4G or Wi-Fi, various built-in sensors such as an
accelerometer, magnetometer, gyroscope, microphone, camera, GPS to sense and capture data from the environment
are the prominent specifications that make smartphones an integral part of our daily lives [7, 8]. The simplicity of
using smartphones and the availability of on-the-move Internet access have led application developers to develop
applications which can access real-time information offered by Cloud Servers.
The smartphone’s built-in sensors utilization is proposed in several works to detect road hazards and conditions
[9-11]. The established method using smartphones eliminates the need for installing special sensors in vehicles.
Smartphones became cheap commodities as newer, and cheaper models with greater capabilities are being released
[12]. These features make smartphones better candidates to use for detecting and analyzing road hazards. The concept
also allow for scalability as the number of mobile users increases and users themselves play a potential role in
collecting and sharing data about the road conditions.
However, some sort of software intelligence is needed to identify whether the detected road condition is really a
road hazard or normal driving maneuver performed by the driver. The smartphone application should comprise
machine learning techniques to be able to detect such disorders. By utilizing Cloud processing, that is associated with
geospatial information, the application could access such a service to identify the road conditions to give the best
performance results [1].
This study considers the problem of identifying and analyzing the road conditions as an essential step that would
help city municipals to gather proper data, based on international standards, and take the suitable action to maintain
roads. The main aim of this work is collect suitable data in order to design and train an ANN that will distinguish
such road hazards. In the data gathering stage an Android smartphone application is used to collect the IRI patterns
while driving on actual roads. The application uses the smartphone’s accelerometer to measure and produce the IRI
patterns. Then the data is analyzed, filtered and normalized in order to efficiently train an ANN using multi-layer
perceptron with back-propagation algorithm to improve the identification and proper classification of road surface
conditions. The resultant outcome can be used as part of a cloud-based service to facilitate the road maintenance
process by associating such process with geospatial services (such as GPS) offered by the phone to indicate the
location of such road hazard.
The rest of the paper is organized as follows: Section 2 is the related work whereby it covers some of the
research studies that have been previously performed for detecting and analyzing road bump conditions using special
instruments or vehicles. The proposed method is presented in section 3. The implementation and results are given in
section 4. Discussion is given in section 5. Lastly, the conclusion and future work are in section 6.

2. RELATED WORK
Driving became an essential routine in our daily lives and has a significant effect safety and financial costs.
Inherently, roads are the primary element of the driving system. Continuous rolling under the wheels, natural factors
such as snow, rain and others lead to developing various impairments and anomalies on the roads which significantly
affect the quality of driving.
Some measurement such as roughness is required to determine the road condition. Roughness is measured
usually using some form of index values such as International Roughness Index (IRI), Present Serviceability Rating
(PSR), Power Spectral Density (PSD), Riding Quality Index (RQI), etc. [13]. Several techniques are used to gather
data and calculate the indices mentioned above.

57
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

An excessive amount of research activities has been proposed in this field using special instruments or vehicles
[13, 14]. This includes proprietary embedded sensors in vehicles [3, 5] or by using smartphone applications that make
use of the smartphone’s deployed sensors for detecting road conditions [15, 16]. Machine learning plays a crucial part
in analyzing such roughness index data using either special propose algorithms [4] or other machine learning
techniques to identify road safety hazards [1, 17].

3. THE PROPOSED METHOD

A similar system was originally proposed in Kattan and Aboalmaaly [1]. The architecture comprises of four main
steps. Data Gathering, Data Transmission via Cloud Service, Dataset Preparation, ANN Design. The proposed system
is different from the one proposed in Kattan and Aboalmaaly [1] in that it uses IRI values as a standard rather than
processing resultant signals directly (waves analogous to the sensor output values) and as shown in Fig. 1.

Fig-1. General overview of the proposed architecture [1]

3.1. Data Gathering

The smartphone uses an application that records IRI values while driving. This process involves driving on roads
with and without anomalies. Such data is to be filtered to and the anomalies are to be isolated as “segments”. Patterns
are created from such data in order to prepare for the dataset. The accepted segments of IRI samples are transferred to
a spreadsheet file for further processing. The IRI points for normal and disorder areas are selected to be 21 IRI values
in total and become the attributes of road samples. The attribute sample for the disordered area is created by selecting
the highest IRI value with ten points after and before it which will be 21 in total. For creating attribute sample for the
normal area, the IRI value greater than zero and less than 6 is selected with ten points after and before it which again
will be 21 points in total. Measurements for disorder and normal portions of the road are depicted in Fig.2 (a-b)
respectively.

Fig-2. IRI signals for normal and disorder area.

(The line chart of the IRI (m/Km) value generated using Excel)

58
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

3.2. Dataset Preparation

The collected IRI values for normal and disorder area would have different ranges. All data should be normalized
(as a real number) to be in the range [0, 1] to prepare for a suitable training and testing data for the proposed ANN
design. The normalization is done by dividing each IRI value by 100 to scale the values to the range [0-1]. Since there
is no vertical movement of the vehicle, the accelerometer reads the IRI value to be less than zero. To avoid this
situation, the negative IRI values in each selected normal and disorder samples are neglected and assigned to the
value of zero.
After the pre-processing stage of the IRI values, the next step involves the process of creating the datasets. The
collected IRI values are divided into two groups “Disorder Area” and “Normal Area”. The “Disorder Area”
represents any street which has anomalies and it is assigned the value of 0. The “Normal Area” represent smooth
areas of roads and it is assigned the value of 1.

3.3. ANN Design

A Feed-Forward Artificial Neural Network (FFANN) with back-propagation gradient-decent based search
method has been designed for the classification of IRI pattern signals. The designed FFANN consists of three layers:
input layer, log- sigmoid hidden layer and log-sigmoid output layer.
Each input is weighted with an appropriate weight. The sum of the weighted inputs and the bias forms the input
to the sigmoid transfer function which is calculated using Eq. (1) [18].

( ) (1)

Back-propagation learning algorithm is applied to the neural network for supervised training so that it can
minimize the differences between the simulated and target output(s). The Gradient Descent (GD) algorithm is used
for the backpropagation training momentum term, and variable learning rate is applied to speed up the BP. The
training of the neural network is made as generalized to avoid overfitting [19].
According to Eq. (2), the termination of the network occurs when the total mean squared error (MSE) becomes
less than or equal to 0.001 [19] or when it reaches to a different number of epochs which is specified according to the
network design. For each iteration and session, the weight and bias values associated with different layers are saved
automatically. If the simulation results from ANN are not satisfactory, the network has been tuned and trained again
with the last saved weight and bias values. This has been done to improve the network performance and reduce time
consumption for training.

∑( ) ∑( ) (2)

The classification of IRI values is considered through different designs of the ANN network. For the purpose of
finding an optimum number of hidden layers and corresponding neuron size for best classification performance, a
different number of hidden layers with the same number of neurons has been selected for both type of the design and
their classification accuracy are also reported accordingly.
The confusion matrix [20] is used to estimate the performance of the proposed ANN design. Accuracy,
sensitivity, and specificity are evaluated to evaluate the performance.
True Positive Rate (TPR): Roads correctly identified as Disorder
True Negative rate (TNR): Roads correctly identified as Normal
False Negative rate (FNR): Road incorrectly identified as Normal
False Positive rate (FPR): smooth roads are mistakenly classified as a disorder.
The confusion matrix on the road analysis is given in Table 1:

59
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

Table-1. Confusion matrix for road analysis (indicating the values for
normal and disorder)
Actual value Expected value
Disorder Normal
Disorder TP FN
Normal FP TN

The dataset is to be divided into training and testing dataset and applied to the proposed design to test the
accuracy. Fig. 3 shows the training process involved for the selected set of training patterns.

Fig-3. The ANN training process.

The proposed ANN considered two designs, single hidden layer design and two hidden layers design as shown in
Fig. 4 and Fig. 5 respectively. The single hidden layer design ANN comprises 21 neurons in the input layer, each
representing one of the sampled IRI points, 10 neurons in the hidden layer and two neurons in the output layer which
represent the classified Disorder or Normal areas.

Fig-4. The structure of the designed single hidden layer ANN (generated by Weka using
the proposed model )

The two hidden layer design is the same as the single layer in terms of input and output but instead has two
hidden layer each with 10 hidden neurons.

60
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

Fig-5. The structure of the designed ANN with two hidden layers (Generated by using Weka
for the proposed model)

4. IMPLEMENTATION & DISCUSSION

The data is gathered by first designating some streets with anomalies in the cities of Sulaymaniyah and Erbil, at
the north of Iraq. A total of 58 different types of streets were tested and some samples are shown in Fig.6. The driving
speed is considered to be the speed limits set for that street, which ranged between 40 to 80 km/h. Data collection
considered highway streets, main city streets, and neighborhood streets.
In order to create a proper dataset, various types of roads are considered for data collection such as crosswalks
using extra thick point brick, or strips of powers and metal expansion joints in bridges and overpasses. Different types
of speed bumps were also considered such as raised dots and extended dots or lines which are placed on some streets
of the city by municipals to forcibly limit the driving speed in those area. Such speed bumps are to be considered as
“normal” and they do not represent “anomalies”. Anomalies would consider driving over things such as potholes,
broken manhole covers, missing chunks of pavements severely sunk in protruding manhole covers, and other
significant road surface anomalies such as moderate cracking, damaged pavements and rough unpaved roads, which
occur after putting the roads into use after construction.

Fig-6. Diverse street samples (showing normal speed bumps, manholes, potholes, etc. )
Photos were taken by the others in the cities of Sulaimaniyah and Erbil (North of Iraq).

The RoadBump Pro © Android application was used to collect data and records the roughness (IRI values) of the
road using the accelerometer sensor in the smartphone [21]. The application also utilizes GPS to capture the location
area of the road. The produced IRI values generated by the Smartphone match the real IRI results which previously
produced by very high-cost inertial profilers.
RoadBump Pro © was installed on a Samsung galaxy A5 Duos © smartphone and the smartphone is placed on
the car's dashboard using a non-slipping mat and data are recorded by driving over actual road comprising normal and
disorder areas with different speed considerations. A car smartphone charger was continuously plugged in to ensure
not running out of battery [1].

61
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

The recorded data were equally divided into two set; Normal and Disorder data. A dataset of collective IRI
patterns is created by mixing these. Each pattern would have 21 attributes of IRI sampling points with a total of 195
IRI sample patterns were prepared as described earlier in the previous section. Table 2 is a summary of the obtained
dataset.

Table-2. Selected data to be used for the MLP NN design

Segment types Number of segment samples
Disorder sample 98
Normal sample 97
Total 195

The dataset is divided into two separate datasets using percentage split. Each dataset comprises of equal
percentages of “Disorder” and “Normal” samples. The Diagram in Fig.7 explains the working procedure on the
datasets.

Fig-7. Working procedure on training and testing sample datasets. It shows the diagram for training & testing sample datasets on the
designed neural network and testing the accuracy for both training sample and testing sample.

A map of the sampled roads with the IRI moving average graph (m/Km) is shown in Fig.8. IRI moving average
graph uses a variable-sized moving average. The number of accelerometer readings being averaged into each point on
the graph varies with the number of accelerometer records between the markers [22].

Fig-8. Map with IRI graph. (collected by authors)

Screenshot of the application used for collecting data using
IRI moving average graph which uses a variable sized
average.

Long paths in each drive are subsequently filtered to save smaller sections between the Start and End markers as
individual recordings. These recordings include both “Normal” and “Disorder” areas of the roads targeted as shown

62
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

in Fig.9. The segments which do not satisfy the data collection considerations are neglected and others are saved to be
tested in the ANN network design.

Fig-9. Targeted segments on the smartphone application. (Collected by authors) Screenshot of

the mobile application used for collecting the data)

Weka © tool bench is used to create the ANNs proposed earlier and to train and use the ANN for the
classification of collected dataset. Weka © is a powerful open source software, yet easy collection of machine
learning algorithms for data mining tasks supplied under the GNU General Public License. The algorithms, such as
SVM and multilayer perceptron, can either be applied directly to a dataset or called from our Java code for
classification on big datasets [23]. Training would consider using 70% of the dataset as training set and the rest 30%
as testing set.

4.1. Performance Evaluation using Single-Hidden Layer

The results show that 30 instances of the test split dataset are correctly classified with percentage 96.774 %, and
one instance of the test split dataset is incorrectly classified with a mean squared error (MSE) value of 0.2094.
The ROC area under the curve of percentage test split for normal and disorder area is shown in Fig. 10.

Fig-10. ROC Area under the curve of percentage split test for single hidden layer ANN design. The threshold
value for 30 instances of the split test set for training dataset visualized using Weka, x-axis: False positive rate
and y-axis: True positive rate

To test the accuracy of the ANN design, a validation test sample is run based on the saved design and the
performance evaluation is calculated. The results clarify that 87 instances out of 90 instances of user supplied test
dataset are correctly classified with an accuracy of %96.667, and three instances of the supplied test dataset are
incorrectly classified with a mean squared error (MSE) value of 0.1969. The ROC area under the threshold curve of
disorder and normal class is shown in Fig. 11.

63
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

Fig-11. ROC Area under the curve for user supplied test set for single hidden layer ANN design. The
threshold value for 90 instances of user supplied test set visualized using Weka to see the accuracy of the
ANN design. x-axis: False positive rate and y-axis: True positive rate

4.2. Performance Evaluation Using Two Hidden Layers

The network parameters are set for two hidden layer design and the evaluation performance on the test set is
calculated. The results show that 29 instances of the test split dataset are correctly classified with an accuracy of
%93.5484, and two instances of the test split dataset are incorrectly classified with a mean squared error (MSE) value
of 0.3485. The ROC area that visualizes the threshold curve of each Disorder and Normal class is shown in Fig. 12.

Fig-12. ROC Area under the curve of percentage split test for two hidden layer ANN design. The threshold value
for 30 instances of the split test set for the training dataset Visualized using Weka, x-axis: False positive rate and y-
axis: True positive rate.

To test the accuracy of the saved ANN design, a validation test sample is run based on the saved design and the
performance evaluation is calculated. The result shows that 84 instances out of 90 instances of the user supplied test
dataset are correctly classified with an accuracy of %93.333 %, and three instances of the supplied test dataset are
incorrectly classified with mean squared error (MSE) value of 0.229. The ROC area that visualizes the threshold
curve of disorder and normal class is shown in Fig. 13.

Fig-13. ROC Area under the curve for user supplied test set for two hidden layer ANN design. The threshold
value of 90 instances for the user supplied test set visualized using Weka x-axis: False positive rate and y-
axis: True positive rate.

4.3. Overall performance

The summary of the classification performance using the back-propagation neural network is given in Table-3.
The single-hidden layer design of the network with ten neurons achieves the best classification rate and requires less

64
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

processing time. The single-hidden-layer design achieves a classification performance of 96.774% for the test split
dataset and 96.6667% for the user defined test dataset with MSE value of 0.1969 and true positive rate of 0.967.

Table-3. The experimental and performance comparison of the network design (collected by authors: shows the properties of each design with
result accuracy for both training sample and user supplied test set)
Network Design Training Time Hidden Neurons Time elapsed Split Test Result User supplied test result
1 Hidden Layer 15 10 0.1 96.774 96.6667
2 hidden layer 35 10, 10 0.1 93.5484 93.333

The proposed method is compared with others similar works as given in Table 4. The results expressively show
the superiority of the proposed design. The proposed ANN performs its classification of IRI values to output the
condition of the road as normal or disorder. The proposed single-hidden layer backpropagation ANN architecture
outperformed with an accuracy of 97%.

5. CONCLUSION
In this work, an ANN with supervised learning architecture is designed to detect and identify road disorders. For
the successful recognition of such hazards, the ANN training dataset is created by driving through designated roads
with known anomalies. The vibrations arise from driving over such disorder, and normal road areas are recorded
based on the International Roughness Index (IRI). The recording was carried out using an Android smartphone
application that utilizes the device’s built-in sensors, namely the accelerometers. The collected IRI-based data is
applied through rigorous pre-processing operations to filter the collected data. The patterns are then extracted and
normalized for the most effective features for the supervised training of the ANN.
The trained ANN successfully classifies different road conditions with a high success rate. The performance of
the designed network is also compared with similar research work, whereby it shows the superiority of the results
obtained. The approach used in this work and proposed method are considered to be cost-effective in comparison to
other traditional methods that involve using proprietary equipment and vehicles to collect such road data.
Future work should consider the proper design of the cloud-service system to automatically train the artificial
neural network (based on IRI values). Furthermore, some other evolutionary training algorithms can be integrated
into the supervised learning part to train the network [24, 25] which would overcome some known shortages of the
backpropagation training method. Moreover, the smartphone application could incorporate built-in pre-processing
capabilities and the ability to connect to a cloud service that would help the municipals to take proper action to repair
roads suffering from hazards as soon as possible.

Table-4. Comparison with related research work

Researcher Data collection Methods Classification rate
Kattan and Aboalmaaly [1] Smartphone application ANN 61%
(No road standard measurements)
Eriksson [4] Attached sensor on the dashboard Private machine learning 90%
of the car (No road standard algorithm (P2)
measurements)
Bhoraskar, et al. [17] Smartphone sensors (No road K-mean clustering and 90%
standard measurements) Support Vector Machine
(SVM).
Proposed Method RoadBump pro smartphone ANN-backpropagation 97%
application (Road standard algorithm
measurements IRI)

65
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

Funding: This study received no specific financial support.

Competing Interests: The authors declare that they have no competing interests.
Contributors/Acknowledgement: All authors contributed equally to the conception and design of the study. Our
sincere thanks go to Grimmer Software especially Mr. David Grimmer for letting us to use their, RoadBumpPro ©,
smartphone application freely.

REFERENCES
[1] A. Kattan and M. Aboalmaaly, "A smartphone-cloud application as an aid for street safety inventory," in Proceedigns of
the 11th International Conference on Electronics, Computer & Computation (ICECCO / IEEE), Abuja, Nigeria, 2014,
pp. 6-7.
[2] G. Chugh, D. Bansal, and S. Sofat, "Road condition detection using smartphone sensors : A survey," International
Journal of Electronic and Electrical Engineering, vol. 7, pp. 595–601, 2014.
[3] B. Hull, "Cartel: A distributed mobile sensor computing system," in Proceedings of the 4th International Conference on
Embedded Networked Sensor Systems (SenSys '06), Boulder, Colorado, USA, 2006, pp. 125-138.
[4] J. Eriksson, "The pothole patrol: Using a mobile sensor network for road surface monitoring," in Proceedings of the 6th
International Conference on Mobile Systems, Applications and Services (MobiSys 2008), Breckenridge, Colorado, USA,
2008, pp. 29–39.
[5] K. Chen, "Road condition monitoring using on-board three-axis accelerometer and GPS sensor," in 6th International
ICST Conference on Communications and Networking in China (CHINACOM), Harbin, China, 2011.
[6] M. K. Hasan, N. Ahmed, and A. H. M. S. Islam, "Android mobile application: Remote monitoring of blood pressure," in
15th International Conference on Computer and Information Technology (ICCIT), University of Chittagong,
Bangladesh, 2012.
[7] K. Rachuri, "Smartphones based social sensing: Adaptive sampling, sensing and computation offloading," Ph.D Thesis,
University of Cambridge, UK, 2012.
[8] N. Kalra and D. Bansal, "Analyzing driver behavior using smartphone sensors : A survey," International Journal of
Electronic and Electrical Engineering, vol. 7, pp. 697–702, 2014.
[9] F. Orhan and E. P. Erhan, "Road hazard detection and sharing with multimodal sensor analysis on smartphones," in 7th
International Conference on Next Generation Mobile Apps, Services and Technologies (NGMAST), Prague, Czech
Repulbic, 2013, pp. 56-61.
[10] M. K. Hasan, "Road structure analysis using GPS information," in Proceedings of the International Conference on
Electrical Information and Communication Technology (EICT), Khulna, Bangladesh, 2013.
[11] B. J. D. Sutter, "Street bump app detects potholes, tells city officials, white paper, official website of the city of Boston."
Retrieved from http://www.cityofboston.gov/news/Default.aspx?id=5490 2012.
[12] J. Whipple, W. Arensman, and M. S. Boler, "A public safety application of GPS-enabled smartphones and the android
operating system," in Proceedings of the 2009 IEEE International Conference on Systems, Man and Cybernetics
(SMC'09), San Antonio, USA, 2009, pp. 2059-2061.
[13] Pavement Interactive, "Roughness, pavementinteractive.org. Available:
http://www.pavementinteractive.org/article/roughness/," 2007.
[14] Minnesota Department of Transportation, "Rating overview." Retrieved from
http://www.dot.state.mn.us/materials/pvmtmgmtdocs/Rating_Overview_State.pdf, 2003.
[15] A. Mednis, "Real time pothole detection using android smartphones with accelerometers," in Proceedings of the 7th
IEEE International Conference on Distributed Computing in Sensor Systems and Workshops (DCOSS), Barcelona,
Spain, 2011, pp. 1-6.
[16] P. Mohan, R. Ramjee, and V. N. Padmanabhan, "Nericell: Rich monitoring of road and traffic conditions using mobile
smartphones," in Proceedings of the 6th ACM Conference on Embedded Network Sensor Systems (SenSys '08), Raleigh,
NC, USA, 2008, pp. 323-336.

66
© 2016 Pak Publishing Group. All Rights Reserved.
 International Journal of Natural Sciences Research, 2016, 4(3): 56-67

[17] R. Bhoraskar, N. Vankadhara, B. Raman, and P. Kulkarni, "Wolverine: Traffic and road condition estimation using
smartphone sensors," in Proceedings of the 4th International Conference on Communication Systems and Networks
(COMSNETS 2012), Bangalore, India, 2012, pp. 1-6
[18] M. Cilimkovic, "Neural networks and back propagation algorithm," presented at the White Paper, Institute of
Technology Blanchardstown, Dublin, Ireland, 2015.
[19] Y. M. Salih, "Detection of motorway disorders by processing of smart phone signals using artifıcial neural networks,"
Fatih University, M.Sc. Thesis, Department of Comptuer Engineering, 2016.
[20] I. O. Dataschool, "Simple guide to confusion matrix terminology." Retrieved from http://www.dataschool.io/simple-
guide-to-confusion-matrix-terminology/," 2014.
[21] Grimmersoftware, "Road bump pro." Retrieved from http://www.grimmersoftware.com/roadbump.html, 2015.
[22] Grimmersoftware, "Road bump user’s guide." Retrieved from
http://nebula.wsimg.com/5d5f829c0bc5d7b768a5c2769c424535?AccessKeyId=77FD3AE629701C2E9990&dispositio=
0&alloworigin=1, 2015.
[23] M. A. Hall, I. H. Witten, and E. Frank, Data mining: Practical machine learning tools and techniques, 3rd ed.: Morgan
Kaufmann Publishers, 2011.
[24] A. Kattan, R. Abdullah, and R. A. Salam, "Harmony search based supervised training of artificial neural networks," in
Proceedings of the 1st International Conference on Intelligent Systems, Modelling and Simulation (ISMS2010), 2010,
pp. 105-110.
[25] A. Kattan and R. Abdullah, "Training feed-forward artificial neural networks for pattern-classification using the
harmony search algorithm," in Proceedings of the 2nd International Conference on Digital Enterprise and Information
Systems (DEIS2013), 2013, pp. 84-97.

Views and opinions expressed in this article are the views and opinions of the author(s), International Journal of Natural Sciences Research
shall not be responsible or answerable for any loss, damage or liability etc. caused in relation to/arising out of the use of the content.

67
© 2016 Pak Publishing Group. All Rights Reserved.

View publication stats