Development of a Prototype Smart City System for Refuse Disposal Management

Full text

Mathematics and Computer Science
2019; 4(1): 6-23
http://www.sciencepublishinggroup.com/j/mcs
doi: 10.11648/j.mcs.20190401.12
ISSN: 2575-6036 (Print); ISSN: 2575-6028 (Online)
Development of a Prototype Smart City System for Refuse
Disposal Management
Joke O. Adeyemo*, Oludayo O. Olugbara, Emmanuel Adetiba
ICT and Society Research Group, Durban University of Technology, Durban, South Africa
Email address:
*Corresponding author
To cite this article:
Joke O. Adeyemo, Oludayo O. Olugbara, Emmanuel Adetiba. Development of a Prototype Smart City System for Refuse Disposal
Management. Mathematics and Computer Science. Vol. 4, No. 1, 2019, pp. 6-23. doi: 10.11648/j.mcs.20190401.12
Received: August 20, 2018; Accepted: October 6, 2018; Published: May 15, 2019
Abstract: The future of modern cities largely depends on how well they can tackle problems that confront them by
embracing the next era of digital revolution. A vital element of such revolution is the creation of smart cities. Smart city is an
evolving paradigm that involves the deployment of information communication technology wares into public or private
infrastructure to provide intelligent data gathering and analysis. To align concretely with the smart city revolution in the area of
environmental cleanliness, this paper involves the development of a smart city system for refuse disposal management. The
architecture of the proposed system is an adaptation of the Jalali reference smart city architecture. It features four essential
layers, which are signal sensing and processing, network, intelligent user application and Internet of Things (IoT) web
application layers. A proof of concept prototype was implemented based on the designed architecture of the proposed system.
The signal sensing and processing layer was implemented to produce a smart refuse bin that contains the Arduino
microcontroller board, Wi-Fi/GSM transceiver, proximity sensor, gas sensor, temperature sensor and other relevant electronic
components. The network layer provides interconnectivity among the layers via the internet. The intelligent user application
layer was realized with non-browser client application, statistical feature extraction method and pattern classifiers. Whereas the
IoT web application layer was realised with ThingSpeak, which is an online web application for IoT based projects. The
sensors in the smart refuse bin generate multivariate dataset that corresponds to the status of refuse in the bin. Training and
testing features were extracted from the dataset using first order statistical feature extraction method. Afterward, multilayer
perceptron artificial neural network and support vector machine were trained and compared experimentally. The multilayer
perceptron artificial neural network model gave the overall best accuracy of 98.0% and the least mean square error of 0.0036.
Keywords: Disposal, Embedded, Feature Extraction, IoT, Pattern Classifier, Smart City
1. Introduction
The International Telecommunications Union (ITU)
describes a smart city as an innovative city that uses
Information Communication Technology (ICT) and other
means to improve the quality of life of the citizens [1]. There
have been various initiatives and studies about smart cities by
multinational companies and research institutes with the aim
of making cities smarter through effective utilization of
available resources [2]. Smart city projects primarily rely on
data gathering from the city infrastructure. Such data include
weather information, electrical energy consumption pattern
and refuse disposal infrastructure usage to mention just a few
[3, 9].
Most communities across the globe are experiencing
serious budgetary challenges on refuse disposal [4, 5].
Despite the provision of refuse bin within most
municipalities, the fundamental responsibility to maintain a
culture of environmental sanitation is left in the hands of
individuals. There are also challenges with respect to
collection services. Consequently, refuse bins often spill over
into accumulation. An environmental cleanliness solution is
therefore mandatory for any community seeking progress in
the well-being of its populace [5].
Given that refuse accumulation usually exists subsequent
to recurrent service delay or untimely refuse collection, smart
city based solutions can assist in the detection of majority of

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 7
refuse accumulation problems [6-8]. Collection services can
therefore be more effective through an automated alert
procedure that readily indicates refuse bin status and prompt
response for refuse collection based on such alert can help to
promote long term communal sanitation.
The requirements of a smart refuse disposal system include
system simplicity, cost effectiveness, easy management,
flexibility and portability [9]. Smart refuse disposal systems
should also be able to make automatic decision with respect
to the status of the refuse bin with little or no input from
humans [10-11]. Hence, the aim of this study is to develop
smart city technology based system architecture and
prototype for refuse disposal management. Based on this
overarching aim, the objectives of the study are: a) To
provide timely and accurate information on the status of
refuse bins to the relevant refuse disposal management
agents. b) To enhance the ease with which full refuse bins are
located by refuse collection agents.
Related Works
Refuse management is obviously a concern in modern
cities because of service cost and storage problem that are
associated with disposing refuse in landfills. Published works
from researchers have indicated different ways that
technology have been applied in an attempt to achieve smart
refuse disposal management.
The Wireless Sensor Networks (WSN), Global System for
Mobile Communications (GSM) and Radio Frequency
Identification (RFID) technologies have been employed by
authors to develop smart refuse disposal management
systems. For instance, real time activity monitoring system
for waste bins was developed by Chowdhury and Chowdhury
[13]. The systems is based on multi-tier platform of wireless
sensors and RFID technology. The authors utilised
knowledge driven database to filter sample datasets and an
extensible graphical user interface for user friendliness.
Bashir et al. [14] designed a vehicular smart transport using
Infra-Red sensor, radio frequency encoder along with
decoder base station nodes for smart bin monitoring. The
system ensures that refuse collection service agents are
accountable through data entry for bin status and location.
Longhi et al. [15] proposed a method to determine smart bin
position via a wireless sensor unique received signal strength
information. According to the authors, the work was aimed at
improving the accuracy in bin activity monitoring by means
of a public open sensor innovation. Glouche et al. [16]
proposed a new approach by using RFID on smart bin. The
refuse items were configured as self-describing objects as a
quick option for easy recycling.
Moreover, pattern recognition algorithms have also been
recently employed to realise refuse disposal management.
Abbasi et al. [17] developed a method for forecasting
Municipal Solid Waste (MSW) generation. The researchers
combined the Support Vector Machine (SVM) classifier and
Partial Least Square (PLS) for feature selection to determine
the quantity of MSW that are generated on a weekly basis in
Tehran, Iran. Comparing SVM and PLS-SVM models, the
authors reported that PLS-SVM gave superior performance
to the SVM model with respect to predictive ability and
computational time. Livani et al. [18] developed a prediction
method in which the weighted K-means clustering algorithm
was utilized for clustering refuse dataset of 63,000 records.
Afterwards, linear regression model was used to build the
prediction models from each of the clusters. They submitted
that the combination of the two methods gave better
performance than using each of the methods individually.
Additionally, the authors produced a hybrid of the prediction
model to develop a decision support system for refuse
collection and recycling services.
Essentially, Nuortio et al. [19] and Zanella et al. [20] are of
the opinion that through the use of smart bin that can detect
load level and allow collector truck route optimization, cost
reduction and improved quality of the refuse management
service can be achieved.
2. Materials and Methods
The various smart city architectures in the literature
contains Radio Frequency (RF) components and sensors,
liaison networks as well as service-oriented information
systems as the first, second and third layers respectively [20-
23]. A vital gap in those earlier architectures is the lack of
specific intelligent data processing in the service–oriented
information system layer. The topologies of most of the
architectures are also cumbersome, which may impose some
difficulties for practitioners that desire to implement them.
These gaps provide the motivation to develop a system
architecture in this study, which is an adaptation of the Jalali
et al. (2015) smart city reference architecture for refuse
disposal management.
The proposed architecture (Figure 1) is more compact than
the reference architecture Jalali et al. [25] and it incorporates
pattern recognition methods for bin status decision making.
The compactness provides ease of implementation for the
architecture while the pattern recognition methods provide
better decision by processing multivariate dataset extracted
from the internal state of the refuse bin. The four layers in the
proposed architecture are the Signal Sensing and Processing
Layer (SSPL), Intelligent User Application Layer (IUAL),
IoT Web Application Layer (IWAL) and the Network Layer
(NL). The NL provides wide area wireless communication
channel for the other three layers using the Internet
technology. However, each of the other three layers (SSPL,
IUAL and IWAL) offers a template for the design of the
functional hardware and software units of the prototype in
this paper.
2.1. Signal Sensing and Processing Layer of the Proposed
Architecture
As shown in Figure 1, the Signal Sensing and Processing
Layer (SSPL) is made up of the Refuse Bin (RB), DC Power
Supply Unit (PSU), electronic sensors, Wireless Fidelity (Wi-
Fi) or GSM (Global System for Mobile) / GPRS (General
Packet Radio Service) transceiver and the microcontroller
development board that runs an embedded software. The RB

8 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
is calibrated into five different levels as shown in Figure 2.
Level1, level2, level3, level4 and level5 represent empty,
quarter-full, half-full, three-quarter-full and full statuses
respectively. The PSU contains rectification circuits to
convert Alternating Current (AC) from the main supply to
Direct Current (5V DC), which powers the SSPL. The
electronic sensors capture raw information from the RB and
transduce them into electronic signals. Such signals include:
The distance from the lid to the level of refuse in the RB
using proximity sensor.
The gaseous emission from the decomposing refuse using
gas sensor.
The temperature of the refuse in the RB using temperature
sensor.
Figure 1. The proposed smart refuse disposal system architecture.
Figure 2. The Refuse Bin.
The microcontroller in the SSPL was realised with an
Arduino microcontroller board. The Arduino board contains
an ATmega328 microprocessor, memory interface
controllers, timers, interrupt controller and General Purpose
Input/Output (GPIO) pins. Six electronic sensors were
utilised in the SSPL. These include two Sharp
GP2Y0A02YK0F proximity sensors, two MQ 135 gas sensors
and two LM35 temperature sensors (GP2Y0A02YK0F
Datasheet, MQ135 Datasheet, and LM35 Datasheet). The six
sensors have three pins - Vcc, GND and Vo, which are
connected using male and female jumper wires to the 5V,
GND and Analog pins 0, 1, 2, 3, 4 and 5 respectively on the
microcontroller board. The duplication of each of the sensors
is to incorporate redundancy into the SSPL circuitry, eliminate
data sparsity and provide holistic coverage of the entire RB.
The proximity sensor covers a distance of up to 150 cm.
This functional range makes the sensor suitable for the refuse
bin in Figure 2, which is only 50cm in length. The sensor’s
input voltage (4.5V to 5.5 V) is obtained by connecting it to
any of the power pins on the microcontroller, which also
generates an output voltage of 5V. The sensor produces an
analog output voltage that corresponds to the distance between
the lid of the RB and the top of the refuse in the bin. The MQ
135 gas sensor operates at 5V± 0.1 DC and its resistance varies
at various concentrations of gases. Gaseous emissions that can
be measured by the sensor include benzene, methane (CH4),
hexane, carbon monoxide (CO) and air. The sensor is suitable
for this work because the most significant greenhouse gas that

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 9
is produced from refuse is methane [26]. The LM35
temperature sensor operates from 4V to 30V. This implies that
the sensor can be powered conveniently from any of the
microcontroller’s power pins. It is calibrated directly in Celsius
with linear +10-mV/°C scale factor and rated for full −55°C to
150°C. The characteristic curve in the datasheet [27] shows
that the functioning of the temperature sensors exhibit a
positive exponential behavior over time. While the refuse get
decomposed in the RB, microorganisms consume the organic
matter and generate both heat and other gases. Once the
temperature sensors are powered, they read the internal heat
within the RB. Because the RB was located in the laboratory
for this work (being a prototyping phase), the heat will largely
depend on the quantity of refuse in the bin.
The GSM/GPRS or Wi-Fi transceiver in Figure 1 provides
optional internet access for the SSPL to transmit data stream
to the IoT web application (which will be discussed later in
this section). The transmitted data stream contains
information concerning the status of the refuse in the RB.
Nevertheless, the Wi-Fi interface card on the development
PC was used to emulate a Wi-Fi transceiver for connecting
the microcontroller board to the internet wirelessly. The Wi-
Fi option became paramount because using a GSM/GPRS
transceiver is too cost intensive for a work of this nature,
since the mobile operators require the GSM/GPRS
transceiver to always have data subscription in order to gain
connection to the internet. However, the GSM/GPRS
transceiver option will be very handy for real-time
deployment, which is beyond the scope of the current study.
The SSPL circuitry is incorporated into the RB shown in
Figure 2 to produce the Smart Refuse Bin (SRB). The
pictorial view of the SRB is presented in Section 3.
The ATmega328 microprocessor translates the electrical
signals from the sensors into digital data streams using the
embedded program in the memory of the microcontroller.
The algorithm of the embedded program is shown in
Algorithm 1. The algorithm was implemented using the
Arduino hardware support package in MATLAB R2015a and
loaded into the microcontroller for full functionality.
Figure 3. Embedded program of the microcontroller unit.
2.2. IoT Web Application Layer of the Proposed
Architecture
The IoT Web Application Layer (IWAL) shown in
Figure 1 is primarily made up of IoT based web
application services and protocols. Internet of Things
(IoT) is a technological scenario in which connected
systems acquire, process, store or communicate data
gathered by electronic sensors from machines, objects,
environments, infrastructure and other physical entities.
The data is thus processed into beneficial information that
can be applied to “command and control” things to
improve human lives on the planet [29-30]. Some
examples of IoT web applications are ThingsSpeak,
Carriots, SmartObject, Skynet Sensorthings, Nimbits,
SensorCloud, Exosite, iDi, EVRYTHNG, Paraimpu,
Manybots and Pachube [31-32]. ThingSpeak is selected
for this work out of these numerous IoT web applications
because of its notable advantages and strengths.
ThingSpeak is an open source IoT web application which
uses Phusion Passenger Enterprise web application server. It
provides Application Programming Interfaces (APIs) for
storing and retrieving data from sensors and devices over the
internet. According to Maureira et al. [32], the servers being
used by some of the other IoT applications are not clearly
documented. Another clear advantage of ThingSpeak that
motivates its selection for this work is that its APIs provide
support for programming languages such as Ruby, Python,
MATLAB and Node.js.
Furthermore, ThingSpeak also provides free hosting for
data channels. Each channel in ThingSpeak support data
entries of 8 data fields, latitude, longitude, elevation,
description and status. Incoming data from sensors into the
channels are communicated via Hyper Text Transport
Protocol (HTTP) POSTs through plaintext, JASON or XML
formats. The channels are set to private by default and
machines or users need read or write key, which is generated
by ThingSpeak to access the data [31-32].
Given the above mentioned strengths of ThingSpeak over
its rivals, the writing/reading of sensors’ data to/from
ThingSpeak were implemented in this study with functions in
the ThingSpeak Support Toolbox for MATLAB R2015a. The
results obtained are presented in Section 3.
2.3. Intelligent User Application Layer of the Proposed
Architecture
The Intelligent User Application Layer (IUAL) in Figure 1
comprises of the pattern recognition and non-browser client
application with web capability (hereafter referred to as client
application for convenience). Nevertheless, the intricacy of
pattern recognition depends on two primary factors. The first
is the dimension and discriminatory power of the extracted
features from the dataset. The second is the choice and
configuration of pattern classifiers [33]. We present the
features extraction method, pattern classifiers and the design
of the client application in this section.

10 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
2.3.1. Dataset and First Order Statistical Feature
Extraction
To obtain the training and testing data sets for this work,
the authors collected wastes from refuse containers that are
located within the Ritson Campus at the Durban University
of Technology, South Africa. The refuse containers contain
decomposing materials, papers and different varieties of solid
wastes that were dumped by students, staff as well as visitors
within the University. The embedded program within the
microcontroller of the SRB extract data in 200 iterations
from each of the 6 sensors to produce a 200x6 dataset for
each level of the SRB. The procedure for generating the data
sets using the embedded program has been presented in
Algorithm 1. Data were extracted in 200 iterations to cater
for any possible electrical variation within the sensors.
Irrefutably, pattern classification using original sensor
measurements is often inefficient and may even hinder
proper interpretation [34]. Feature extraction therefore
reduces the dimensionality of raw dataset by keeping the
most discriminatory information. In addition, the
performance of the feature extraction stage sturdily impacts
the design and performance of pattern classifiers. If the best
features are selected from the raw data, the task of the
subsequent pattern classifiers becomes trivial. Conversely, if
the features with little discriminatory ability are chosen, a
more complex pattern classifier may be required [35].
Principal Component Analysis (PCA) is one of the most
commonly used algorithms for features extraction and
dimensionality reduction in fields like image processing,
pattern recognition, bioinformatics, telecommunications and
etcetera. It is an optimal linear transformation algorithm for
keeping the subspace of a dataset that has the largest
variance. However, the computational requirement for PCA
is high [34] [36-37]. First Order Statistical Features (FOSF)
is another strategy being employed for discriminatory
features extraction with lower computational requirements
(Indonesia 2011). The computational advantage of the FOSF
over PCA motivate its selection as the feature extraction
method for this research work.
The FOSF utilised include the mean, standard deviation,
kurtosis and skewness values computed from each of the 200
values generated for each of the 6 sensors per SRB level [34,
37-38]. The computed mean, standard deviation, skewness
and kurtosis for the six sensors were concatenated to obtain
FOSF vectors in 24- dimensional space for each level of the
smart refuse bin. The procedure for computing the FOSF was
implemented in MATLAB R2015a programming
environment.
2.3.2. Pattern Classification
The feature vectors obtained with the FOSF extraction
method discussed in the preceding sub section is transmitted
to train the selected pattern classifiers. A trained pattern
classifier will thus be able to categorise an incoming raw data
stream from the SSPL, which have been processed with the
FOSF into one of empty, quarter full, half full, three quarter
full and full bin status.
Two state-of-the-art pattern classifiers investigated for the
classification of the features in this work are the Multilayer
Perceptron Artificial Neural Network (MLP-ANN) and
Support Vector Machine (SVM). They are extensively
engaged in the literature to solve pattern classification
problems [42]. Nevertheless, each of the classifiers have
inherent merits and demerits. MLP neural networks have the
capability to detect complex non-linear associations between
variables, they are very fast to use for classification problems
and generally achieve good performance. However, because
MLP is based on the traditional empirical risk minimization
principle, it often suffers from overfitting and multiple local
minimal. Conversely, SVM has the advantage of good
generalization performance because it is based on the
structural risk minimization principle. It has a modest
geometric interpretation, its computational involvedness do
not depend on the dimensionality of the input space and its
solution is global and unique. SVM is however a shallow
model and its performance result rely heavily on the selected
kernel function. The joint strength of both MLP and SVM is
that they yield good performance in high-dimensional
classification problems. Generally, the best classifier for a
given problem depends on the problem domain, the number
of classes to be classified and the number of example(s)
available per class. Furthermore, experimentation becomes
handy to compare the performances of the selected classifiers
for this work [39-41].
2.3.3. Configuration of the Pattern Classifiers
The MLP-ANN and SVM were tuned in this study through
established parameters in the literature as well as via
experimentations to determine the configuration that gives
the most optimal performance [33, 42].
An MLP-ANN consists of an input, one or more hidden
and an output layers of neurons with each layer fully
connected to the next one. The size and number of classes in
the dataset determines the number of neurons in the input and
output layers respectively [33]. Each feature vector in this
work is a 24- dimensional vector and there are 5 different
levels for the SRB, hence, the MLP-ANN contains 24
neurons in the input layer and 5 neurons in the output layer.
Popescu et al. [43] recommended that more hidden layers
with several neurons often lead to fewer local minima. Thus,
the MLP-ANN in this study were configured with two hidden
layers. In codicil, 20, 40, 60, 80 as well as 100 neurons per
hidden layer were investigated to experimentally determine
the appropriate number of neurons for each of the hidden
layers. Except for the input neurons, which contain linear
activation function, each neuron in neural network has
nonlinear activation function [33]. For the MLP-ANN in this
work, the neurons in the hidden and output layers were
configured with the hyperbolic tangent function. The
nonlinearity and differentiability properties of the function
have been deemed as essential qualities for optimal
performance of neural networks [43]. The MLP-ANN in this
study utilizes a supervised learning technique called scale
conjugate gradient backpropagation based on its outstanding

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 11
performance record [33] [44-45]. The FOSF were partitioned
based on a data partitioning ratio of 70% training, 15%
validation and 15% testing. The MLP-ANN contains
additional configurations as follows:
Training epochs = 10,000,
Learning rate = 0.1,
Maximum training time = 180sec,
Minimum performance gradient = 1e-6,
Validation checks = 10,000
SVM possesses the ability to capture both linear and
nonlinear patterns in feature space by employing a mapping
function which transforms the feature space into a higher
domain that exhibits multiple dimensionalities [46]. The
mapping function is usually implemented by using
specialised kernels. Minh et al. (2006) opined that SVM
kernels must satisfy Mercer condition for it to be effective.
Examples of kernels that satisfy the Mercer condition are
linear, polynomial, radial basis function and perceptron
kernels [48]. Since the performance of the SVM is
determined by the choice of kernel selected for the
experiment, the four listed kernels were tested using 10 fold
cross validation and their performances were compared. The
multi-class SVM techniques that was employed for this study
is the One-Against-All (1AA) because it is acclaimed to be
simple and highly efficient [48]. The same data partitioning
strategy for MLP-ANN was used for training the SVM.
Implementation, training as well as all the experimental
configurations of the MLP-ANN and SVM pattern classifiers
in this work were carried out in MATLAB R2015a. The
computer system for the implementation and experiments
contains Intel Core i5-2540MCPU @2.60GHz speed with
4.00GB RAM and 64-bit Windows 8 operating system.
2.3.4. Client Application Design
In this study, the client application, which is a graphical
user interface (non-browser-based) web client, was designed
with the Universal Modelling Language (UML). UML2
defines 13 basic diagram types that are divided into two
generic groups, namely behavioral and structural modeling
diagrams. Behavioral diagrams are used to capture the
interactions and state instantiations within a model over its
execution time. On the other hand, structural diagrams are
used to emphasize the things that must be present in the
software application being modeled. These include classes,
objects and interfaces. Structural diagrams are also used to
represent the relationships and dependencies between the
various elements of the modeled software [49]. According to
Bell [50], the most useful standard UML diagrams are use
case, class, sequence, statechart, activity, component and
deployment diagrams. However, four of these diagrams
sufficiently captures the different views of the IUAL client
application as well as the prototype of the proposed
architecture (Figure 1). Hence, in this study, use case diagram
was used to model the behavioral view while class diagram
was used to model the structural view of the client
application. Furthermore, activity diagram was adopted to
model the operational step-by-step workflow of the hardware
and software components of the full prototype. Deployment
diagram was also adopted to model how the entire prototype
can be deployed, even though, the deployment of the
prototype is beyond the scope of the current study.
Nevertheless, due to space constraint, only the activity
diagram is described in this paper.
Activity diagram illustrates the flow of signal across the
different aspects of a system. It usually consist of shapes
such as round-cornered rectangle, diamond, circle, can and
interconnection lines, which symbolise action state, decision,
start or stop, database and association respectively [49].
Figure 3 shows the activity diagram of the proposed
architecture’s prototype. The diagram consists of 10 action
states, 3 decisions, 3 forks, 1 database, 1 start and 1 stop
symbols. The operation of the entire activity diagram starts
with the loading of the client application in the Load Client
App action state. A decision on the appropriate user group is
taken, which branches out to Janitor and Others. The Janitor
user group branches out to 2 action states namely,
View/Update Task, and Check Bin. The second decision
symbol splits into the Administrator and Supervisor user
groups. The Administrator user group links to action states,
which are Create Users and Edit/Delete Users. The
Supervisor also links to 2 action states, namely Assign Task
and Edit/Delete Tasks. Similar to Janitor, both the
Administrator and Supervisor also link to the Check Bin
action state. The Smart Refuse Bin section has only 1 Signal
Sensing and Transmission action state which sends refuse
status and location data into the ThingSpeak Services action
state at the IoT Web Application section. The ThingSpeak
Services action state transmits the data it receives to the Bin
Status and Location View action state for the Administrator,
Supervisor and Janitor to view the status and location of the
SRB. The last decision symbol is for the different user group
to determine if more tasks should be performed or all
activities should end. If more tasks are desired by any of the
user group, a link is returned back to the first decision
symbol and there is a link to end the activities if otherwise.
2.4. Performance Evaluation
In order to compare pattern classifiers, it is important to
define and specify the evaluation metrics. The performance
evaluation metrics of accuracy and Mean Square Error
(MSE) were adopted in this work for the MLP-ANN and
SVM because they are commonly used in the literature. The
accuracy is the degree of closeness of the measurements of a
quantity to its true value. Whereas MSE is the mean of the
square of the difference between the expected output and the
actual output [42, 51].
Evaluation of a software application involves the
investigation of whether the developed application satisfies
specified design requirements. The client application in this
work was evaluated through laboratory experiments using
unit testing method. Koomen and Pol [52] defined unit
testing as “a test executed by the developer in a laboratory
environment that should demonstrate that the program meets
the requirement set in the design specification”. The authors

12 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
are the developers in the context of this work. Unit testing is
mostly engaged to test the functionality and quality of
software components or a collection of components because
it allows easy identification of bugs within the code [54].
Unit testing was realized in this work using the xUnit-style
MATLAB unit testing framework in MATLAB R2015a.
Although the researcher is aware of other unit testing
frameworks such as MS Test, JUnit and NUnit, the xUnit-
style framework becomes handy because all the codes in this
work were implemented using MATLAB R2015a.
Figure 4. Activity diagram of the system prototype.
3. Results
3.1. Hardware Unit
The testing results of the electronic sensors and the
arduino microcontroller development board while the
prototype is being developed in the laboratory are presented
in this sub-section. The Arduino Uno development board,
which contains the ATmega 328 microcontroller was
connected through USB 2.0 cable to the development PC to
evaluate its functionality. The board gets powered up by the
5V voltage that is supplied to it from the PC via the USB 2.0
cable. The On LED situated on the board emits a steady
green light to indicate the device is in good and proper
working conditions as shown in Figure 4.
Once the embedded program (Algorithm 1) starts running
on the microcontroller, it commences data capturing from the
SRB using the electronic sensors. Consequently, the Arduino
board transmission (TX) and reception (RX) LED emits a
steady amber coloured light as shown in Figure 4. Data
transmission from the SRB to the MATLAB R2015a
workspace during the development phase and to the IoT web

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 13
application at the operational phase are deemed successful
once the amber coloured LED are on. It is noteworthy that
the pins on the Arduino Uno board connect directly to the
corresponding pin on the ATmega 328 microcontroller.
Figure 5. Arduino development board at ready state and TX/RX states.
The functionality of the proximity, gas and temperature
sensors were evaluated by connecting them to the appropriate
pins on the Arduino board and capturing the sensors’
readings at different levels of the SRB. The plots of the mean
voltages obtained from each of the sensors against the
distances marked on the SRB are shown in Figure 5. The
graph of the proximity sensor is highly similar to the
characteristics curve of the sensor in the datasheet, in which
the voltage output was plotted against distance
(GP2Y0A02YK0F Datasheet). The graph of the gas sensor
indicates an initial spike at level 1 but reduces drastically and
almost remains constant thereafter. The spike is obviously
due to high gaseous concentration at that level. This
behaviour may also be obtained in real life scenario if the
initial refuse that was dumped in the bin contains
decomposing waste while those that were dumped afterwards
until the bin is full have minimal decomposing refuse. The
temperature sensor graph also peaked at level 1 but reduces
steadily to lower values at level 2 and beyond. With this
graph, it can be deduced that the initial refuse in the bin
generated more heat since more gaseous emission was also
indicated by the gas sensor. As the concentration of the gas
reduced, the heat also reduced up to level 2 and afterwards.
The test results that were obtained from the three sensors
apparently illustrate that they are all suitable for acquiring the
appropriate signals, which represent the status of the refuse in
the SRB. The SRB picture, showing the refuse bin,
microcontroller board, sensor array and connection wires, as it
was been constructed in the laboratory is shown in Figure 6.
Figure 6. Mean voltage against the bin levels.
3.2. Software Units
This sub-section presents the configuration, experimental
evaluation, implementation and unit testing of the prototype’s
software units.
3.2.1. ThingSpeak Configuration
In order for the ThingSpeak web application to be ready
for use, a user account was first of all created with username
and password. Thereafter, a channel of six fields was created
and configured to accept data from each of the six sensors in
1 1.5 2 2. 5 3 3. 5 4 4.5 5
0
0.5
1
1.5
2
2.5
Bin lev el
Mean voltage of the sensors
Proxim ity se nsor
Gas sensor
Tem perature sensor

14 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
the SRB. Figure 7 shows the home page of a fully configured
and customized ThingSpeak web application for this work.
The page contains a generic menu bar at the top, which
contains the commands that were used for the configuration.
After the menu bar, there are labels such as Page title,
ChannelID, Author and Access. It also contains link tabs in
the middle with labels; Private View, Public View, Channel
Settings, API Keys, Data Import/Export. After this, there are
clickable buttons such as Add Visualization, Data Export,
MATLAB Analysis and MATLAB Visualization. The bottom
of the page contains the Channel Stats. There are 200 entries
indicated under the Channel Stats for each of the six fields.
This clearly illustrates that the data from the SRB were
successfully streamed to the ThingSpeak web application.
The Access option was initially set to public in order to test
the functionality of the application. However, it was
eventually set to private in order to secure the data and
settings for this research on the ThingSpeak web application
domain on the internet.
Figure 7. SRB construction in the laboratory.
Figure 8. The customised home page on ThingSpeak web application.
The streamed data from the SRB to each of the data fields
are graphically captured as shown in Figure 8. Each graph
shows the real-time streaming of data from each of the
sensors with values on the Y axis and the date/time stamps
on the X axis. The data from each sensor were fully available
in the ThingSpeak web application within 90 seconds. This
shows that the SRB as well as the configured ThingSpeak
web application are functioning optimally. The ThingSpeak
web application provides ubiquitous computing for this work.
This implies that the data generated from the SRB is
available to users at anytime and from anywhere.

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 15
Figure 9. ThingSpeak fields data visualization.

16 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
Figure 10. Dataset histogram of the first instance of each SRB level.
3.2.2. Acquired Dataset from the Smart Refuse Bin
The SRB is configured to generate 200 values for each of
the six sensors so as to produce a 200x6 dataset. However, to
acquire sufficient data that caters for possible variations in
the electronic sensors while taking the measurement, the
researcher deemed it necessary to run several trials for the
data acquisition. Overall, there were 15 different data
acquisition trials for each of the 5 levels in the SRB (i.e.
empty, quarter full, half full, three quarter full and full). This
implies that for each bin level, there were 15 different
instances of the 200x6 dataset. The entire dataset cannot be
presented here because of the huge size. However, to grasp a
graphical view of a portion of the data, the histogram plots of
the first training instance for each level is presented in Figure
9. The histogram plots, which represent the distribution of
the dataset, clearly show that the acquired data are different
for each of the levels. For instance, the distribution of the
data with high frequencies are concentrated at the middle for
level 1 and it contains unique values from 0.1 up to 2.3 with
0.5 having the highest frequency of about 320. Even though
the distribution of the data with the highest frequencies are
majorly skewed to the left for level 2, level 3, level 4 and
level 5, the range of the data distributions and the frequencies
within the range are uniquely different. However, the raw
dataset could not be used to train the pattern classifiers
directly so as to avoid structural complexity, high
computational time, overfitting as well as the curse of
dimensionality [34, 42].
3.2.3. First Order Statistical Features
The first order statistics which include mean, standard
deviation, skewness and kurtosis were employed to compute
discriminatory features from the extracted dataset in this
work. These values were computed for each level of the SRB
and for all the 15 instances. Due to space constraint, the
computed values for the first instances of all the levels are
shown in Tables 1. The 4 statistical values were concatenated
for all the 6 sensors, which culminated in a feature vector of
24 elements per level. Meanwhile, 10 instances of the feature
vectors for each level were earmarked to train the pattern
classifiers. The remaining 5 instances were reserved to test
the pattern classifiers after they were successfully trained.
This strategy helped to examine the generalization ability of
the selected pattern classifiers.
Table 1. First order statistical features for the first training instance of each level.
S/N Level 1 Level 2 Level 3 Level 4 Level 5
1 0.79922 2.29736 2.10643 1.80908 1.18054
2 0.25591 0.30659 0.39469 0.37803 1.47689
3 -0.6346 3.24151 2.28962 2.38496 0.7239
4 -0.7738 9.34752 3.602 4.03413 -1.237
5 1.4192 1.97227 1.80042 1.67344 1.32198
6 0.23541 1.64759 1.73307 1.75043 0.25404
7 2.6402 0.54842 0.64183 0.71787 2.2597
8 5.21212 -1.4193 -1.3964 -1.3321 3.34301
9 1.61653 0.82096 0.79427 0.81667 0.7859
10 0.06831 0.04342 0.04046 0.03733 0.04257
11 1.47168 1.46928 1.55846 1.49372 1.33149
12 2.27797 1.67282 2.44121 2.27213 2.24741
13 1.16522 0.60343 0.63106 0.679 0.53038
14 0.07955 0.037 0.04017 0.03955 0.04184
15 1.16358 1.61973 1.63679 1.63694 1.49151
16 1.49495 3.39342 3.31708 3.13864 2.59734
17 0.53325 0.40236 0.40093 0.36862 0.39347
18 0.07663 0.04902 0.04762 0.04663 0.05071
19 2.43467 1.36455 1.89009 2.02748 1.65768

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 17
S/N Level 1 Level 2 Level 3 Level 4 Level 5
20 4.85811 1.74582 4.58913 4.74427 2.66225
21 0.53723 0.1368 0.13356 0.11932 0.12775
22 0.07701 0.05264 0.05377 0.05118 0.05825
23 2.30185 1.59718 1.57779 1.33749 1.38151
24 4.29572 3.28227 4.18363 2.07824 1.66323
3.2.4. Pattern Classifiers Training
The basic configurations of the classifiers as well as the
motivation for the choice of the evaluation metrics have been
presented in Section 2. As shown in Table 2, the one-per-
class coding method was used to encode the MLP-ANN
target output while decimal values were used for the SVM
target output [48, 53].
Table 2. Target output of the MLP-ANN and SVM classifiers.
Level Bin Status MLP-ANN Target Output SVM Target Output
1. Empty 1 0 0 0 0 1
2. Quarter Full 0 1 0 0 0 2
3. Half Full 0 0 1 0 0 3
4. Three Quarter Full 0 0 0 1 0 4
5. Full 0 0 0 0 1 5
The first order statistical features earlier computed were
used to train the MLP-ANN and SVM classifiers and the
outcome of the trainings were evaluated using the accuracy
and Mean Square Error (MSE) metrics. The performance
results of the trained MLP-ANN for 20, 40, 60, 80 and 100
neurons in the hidden layers are reported in Table 3, while
the results of SVM for polynomial, linear, RBF and
perceptron kernels are shown in Table 4.
Table 3 shows that the MLP-ANN with 60 neurons in the
hidden layers produced the best result among other MLP-
ANN configurations with an accuracy of 98% and a MSE of
0.0036. Table 4 also shows that the SVM classifier with
polynomial kernel functions gave the highest accuracy
among the other SVM kernel functions, with an accuracy of
88.89% and MSE of 0.1558. The MLP-ANN with 60 neurons
in the hidden layers is therefore nominated as the pattern
classifier for this work based on its superior performance
over the others.
Table 3. Performance result of the MLP- ANN pattern classifier.
ANN Type Number of Hidden
layer neurons Accuracy (%) MSE
ANN1 20 94 0.0322
ANN2 40 96 0.0171
ANN3 60 98 0.0036
ANN4 80 98 0.0168
ANN5 100 98 0.0191
Table 4. Performance result of the SVM pattern classifier.
Kernel Accuracy MSE
Polynomial 88.89 0.1558
Linear 77.78 0.1519
RBF 22.22 0.1920
Perceptron 44.44 0.1557
The confusion matrix of the nominated MLP-ANN pattern
classifier is shown in Figure 10 to illustrate its performance
for each class in the training dataset. The 10 training
instances in each of the 1st, 2nd, 4th and 5th classes were
correctly classified. However, 9 of the training instances in
the 3rd class were correctly classified while 1 instance was
wrongly classified as belonging to the 2nd class. The sum of
the percentages of correctly classified classes across the
diagonal culminated in the 98% accuracy earlier reported in
Table 3.
Figure 11. Confusion matrix of the best MLP-ANN in Table 3.
3.2.5. Testing of the Best Pattern Classifier
The five testing instances were used to test the nominated
MLP-ANN classifier so as to investigate how it generalises.
These testing instances were not part of the dataset used for
training the classifier. The same results were obtained for all
the testing instances as shown in Tables 5 to 9. Generally, the
expected and actual outputs are the same for levels 1, 2, 3
and 5. These implies that these levels were correctly
classified. However, for all the 5 testing instances, the
expected and actual outputs for level 4 are not the same.

18 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
Notably, level 4 was wrongly classified as level 3 in all cases.
Assuming the decision for refuse collection is taken when the
classification output produces level 5 (full status), it can
therefore be stated that the wrong classification of level 4
(three quarter full) as level 3 (half full) does not have any
major negative implication. For example, wrongly classifying
level 4 (three quarter full) as level 5 (full status) would have
resulted in more adverse effects like wasted effort, time and
resources by the refuse collection agents. Essentially, since
only one level was wrongly classified for all the testing
instances, it can be inferred that the nominated MLP-ANN
generalizes well on unseen dataset. The performance of the
nominated pattern classifier for this work is therefore
adequate and acceptable.
Table 5. Testing result of the best MLP-ANN using the first testing instance.
Level Bin status Expected output Actual output Remark
Level 1 Empty 1 0 0 0 0 1 0 0 0 0 Correct
Level 2 Quarter Full 0 1 0 0 0 0 1 0 0 0 Correct
Level 3 Half Full 0 0 1 0 0 0 0 1 0 0 Correct
Level 4 Three Quarter Full 0 0 0 1 0 0 0 1 0 0 Incorrect
Level 5 Full 0 0 0 0 1 0 0 0 0 1 Correct
Table 6. Testing result of the best MLP-ANN using the second testing instance.
Level Bin status Expected output Actual output Remark
Level 1 Empty 1 0 0 0 0 1 0 0 0 0 Correct
Level 2 Quarter Full 0 1 0 0 0 0 1 0 0 0 Correct
Level 3 Half Full 0 0 1 0 0 0 0 1 0 0 Correct
Level 4 Three Quarter Full 0 0 0 1 0 0 0 1 0 0 Incorrect
Level 5 Full 0 0 0 0 1 0 0 0 0 1 Correct
Table 7. Testing result of the best MLP-ANN using the third testing instance.
Level Bin status Expected output Actual output Remark
Level 1 Empty 1 0 0 0 0 1 0 0 0 0 Correct
Level 2 Quarter Full 0 1 0 0 0 0 1 0 0 0 Correct
Level 3 Half Full 0 0 1 0 0 0 0 1 0 0 Correct
Level 4 Three Quarter Full 0 0 0 1 0 0 0 1 0 0 Incorrect
Level 5 Full 0 0 0 0 1 0 0 0 0 1 Correct
Table 8. Testing result of the best MLP-ANN using the fourth testing instance.
Level Bin status Expected output Actual output Remark
Level 1 Empty 1 0 0 0 0 1 0 0 0 0 Correct
Level 2 Quarter Full 0 1 0 0 0 0 1 0 0 0 Correct
Level 3 Half Full 0 0 1 0 0 0 0 1 0 0 Correct
Level 4 Three Quarter Full 0 0 0 1 0 0 0 1 0 0 Incorrect
Level 5 Full 0 0 0 0 1 0 0 0 0 1 Correct
Table 9. Testing result of the best MLP-ANN using the fifth testing instance.
Level Bin status Expected output Actual output Remark
Level 1 Empty 1 0 0 0 0 1 0 0 0 0 Correct
Level 2 Quarter Full 0 1 0 0 0 0 1 0 0 0 Correct
Level 3 Half Full 0 0 1 0 0 0 0 1 0 0 Correct
Level 4 Three Quarter Full 0 0 0 1 0 0 0 1 0 0 Incorrect
Level 5 Full 0 0 0 0 1 0 0 0 0 1 Correct
3.2.6. Client Application Implementation and Unit Testing
The non-browser client application was implemented in a
prototype form using the GUIDE in MATLAB R2015a. The
implementation is based on the UML design described in
Section 2. Some of the key pages in the implemented
application are shown in Figure 11. These include Welcome
page, Main window, Bin status page, Bin location page,
Create user page and Update/Delete user page. Appropriate
codes were written to ensure that all the interfaces perform
the functions in the requirement specifications. The manual
testing of the functionality of the different commands in all
the interfaces to ensure the integrity of the codes was
successful.
For instance, to test the functionality of the Bin status
page, one of the authors specified the ID of the SRB in the
combo box and clicked on the Check Status button. This
produces the bin status information and the date/time as
shown in the figure. Notably, the returned status actually
corresponded with the actual status of the SRB. This shows
that the page as well as the pattern classifier are functioning
properly. Furthermore, specifying the Bin ID on the Bin
location page using the combo box and clicking on the Check
Location button on the page produces the coordinates of the
bin location as well as its Google earthTM image. These
information can be seen clearly on the figure. Similar to the

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 19
bin status information, the location information is also very
essential. This is because it provides accurate direction for
the Janitor to locate bins that are ready for collection.
Figure 12. Client application implementation interfaces.
The output of the unit test carried out for the components
of the client application is shown in Figure 12. All the
modules in the client application passed the test successfully
as shown and the average execution time for any of the tested
module is 225.52ms. This clearly illustrates that the
implemented client application is devoid of coding errors and
the running time is optimal. The evaluation of the
implemented client application in this work was limited to
unit testing because the deployment is beyond the scope of
the present work.

20 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
Figure 13. The client application’s unit testing result.
In summary, Figure 13 provides an overview of the
different parts of the prototype and illustrates the
communication of signals among them to produce a
functional system. The SRB sends data on its current status
to the ThingSpeak web application periodically via the
internet. This provides ubiquitous data availability. A user
sends relevant queries to the ThingSpeak web application to
retrieve the SRB data. Applicable functions (methods) on the
client application process the retrieved data and display the
bin status as well as the location information for the user to
view and take necessary actions.
Figure 14. Signal communications among the prototype software and hardware units.

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 21
4. Discussion
The research objectives of this study have been succinctly
achieved. The objectives are hereby restated along with how
they have been realised.
a) To provide timely and accurate information on the status
of refuse bins to the relevant refuse disposal agents. This
objective was realised through the hardware and software
components of the prototype that was developed based on the
proposed architecture. SRB sends regular updates on the bin
status to the ThingSpeak web application and through the
client application, any of the user groups can send a query to
ThingSpeak for the latest data. The data is processed in the
pattern recognition module (FOSF and MLP-ANN) to return
the status of the bin via a user interface on the client
application. The time to achieve this procedure based on the
evaluation is about 90.05s and the accuracy of the status
detection is 98%. Hence, the objective to provide timely and
accurate information on the status of refuse bins has been
achieved.
b) To enhance the ease with which full refuse bins are
located by refuse collection agents.
This objective was also realised through the prototype. The
user can query the location of any refuse bin via the client
application once the status has been obtained. The
ThingSpeak web application returns the coordinates of the
SRB location as well as the corresponding google earth
image. This helps refuse collection officials to easily find the
route for disposing the refuses in the bin.
ilar to the development of the proposed architecture for
refuse disposal management in the current study, previous
authors have also developed multi-tier architectures of five
layers in response to the general demand for refuse disposal
in a smarter way [12]. Furthermore, refuse management
process which implements an on-site handling and transfer
optimization has been previously studied using sensor nodes
and browser based remote monitoring solutions. Even though
the authors did not implement decision support system, they
projected it as a future feature that could boost the
performance of the developed solution [6]. In codicil, pattern
recognition techniques such as Partial Least Square with
SVM [16] and linear regression with weighted k-means
clustering [17] have been recently employed for forecasting
the quantity of municipal refuse generation. Based on the
results obtained in the current work, it has become more
apparent that pattern recognition techniques like multi-layer
perceptron neural networks are viable for accurate detection
of refuse bin status. This fulfils the expectation of a boost in
the performance of refuse disposal systems through decision
support and pattern recognition techniques.
Energy efficiency cannot be over emphasised when it
comes to the development of initiatives for smart
environment. In this light, a previous work was presented on
the development of an energy efficient sensing algorithm for
measuring the parameters of bin status [10]. The current
work inherently caters for energy efficiency through the
choice of the hardware components that was used for
implementing the prototype. The microcontroller board as
well as all the electronic components require only 5V DC to
function accurately and optimally. The location vis-à-vis the
bin status detection in this work also provide energy saving
medium for operators of refuse disposal management.
The accuracy achieved through the use of first order
statistical features and MLP-ANN in this work strongly
corroborates the position of other authors, who reported
improved performance when feature selection algorithm is
implemented alongside pattern classifiers [16-17]. Adetiba
and Olugbara [42] have also previously reported that MLP-
ANN is often used as a pattern classifier because of its
accuracy and effectiveness.
5. Conclusion
Thus far, we have been able to realise the research
objectives in this study. Electronic principles, IoT
technological paradigm and pattern recognition techniques
provided a rich set of theoretical toolkits for the realisation
of these objectives. The results obtained in this paper have
proven the efficacy of the prototype and its readiness for
upgrade into a full system. Such a system can be deployed
within communities in any country. The system also has
huge commercial prospects for government authorities and
business organisations. Normally, citizens pay for waste
disposal services to the municipal authorities through
taxation or levies. The value addition that will derive from
the system will encourage citizens to make payment on a
promptly basis. Electronics and plastic companies may also
leverage on the prototype to manufacture smart refuse bins.
The client application with the associated IoT web
application may be deployed using the Software-as-a-
Service (SaaS) cloud computing model. Apart from the
heathier environment that will be derived through the
deployment of the system, more jobs can be created for the
citizens at large.
Acknowledgements
E. Adetiba is on a postdoctoral fellowship at the ICT and
Society (ICTAS) Research Group, Durban University of
Technology, South Africa. He is on postdoctoral research
leave from the Department of Electrical & Information
Engineering, College of Engineering, Covenant University,
Ota, Ogun State, Nigeria.
References
[1] I. T., Union. 2014. International Telecommunication Union
2014. Measuring the Information Society Report 2014, ISBN
978-92-61-15291-8, pp. 1-270
https://www.itu.int/en/ITUD/Statistics/Documents/publication
s/mis2014/MIS2014_without_Annex_4.pdf (accessed
17.07.2016)

22 Joke O. Adeyemo et al.: Development of a Prototype Smart City System for Refuse Disposal Management
[2] T. Nam, T. A. Pardo, Conceptualizing smart city with
dimensions of technology, people, and institutions, in:
Proceedings of the 12th Annual International Digital
Government Research Conference: Digital Government
Innovation in Challenging Times, ACM, 2011, pp. 282-291.
[3] M. Roscia, M. Longo, and G. C, Lazaroiu, Smart city by
multi-agent systems, in: Proceedings of International
Conference on Renewable Energy Research and Applications
(ICRERA), IEEE, 2013, pp. 371-376.
[4] A. Ohri, P. Singh, Development of decision support system for
municipal solid waste management in India: a review.
International Journal of Environmental Sciences, 2010, 1 (4):
440.
[5] E. Muzenda, F. Ntuli, T. J. Pilusa, Waste management,
strategies and situation in South Africa: an overview. World
Academy of Science, Engineering and Technology, 2012, pp.
1-5.
[6] S. Longhi, D. Marzioni, E. Alidori, G. Di Buo, M. Prist, M.
Grisostomi, M. Pirro, Solid waste management architecture
using wireless sensor network technology, in: 5th International
Conference on New Technologies, Mobility and Security
(NTMS), IEEE, 2012, pp. 1-5.
[7] A. Patsakis, R. Venanzio, P. Bellavista, A. Solanas, M.
Bouroche, Personalized medical services using smart cities'
infrastructures, in: Medical Measurements and Applications
(MeMeA), IEEE International Symposium on, IEEE, 2014,
pp. 1-5.
[8] A. Domingo, B. Bellalta, M. Palacin, M. Oliver, E. Almirall,
Public open sensor data: Revolutionizing smart cities,
Technology and Society Magazine, IEEE, 2013, 32 (4): pp.50-
56.
[9] W. A. Arbab, F. Ijaz, T. M. Yoon, C. Lee, A USN based
Automatic waste collection system, in: Advanced
Communication Technology (ICACT), 2012 14th
International Conference on, IEEE, 2012, pp. 936-941.
[10] M. A. Al Mamun, M. Hannan, A. Hussain, H. Basri, Wireless
sensor network prototype for solid waste bin monitoring with
energy efficient sensing algorithm, in: 16th International
Conference on Computational Science and Engineering
(CSE), IEEE, 2013, pp. 382-387.
[11] M. Jiménez, A. Jiménez, P. Lozada, S. Jiménez, C. Jimenez,
Using a Wireless Sensors Network in the Sustainable
Management of African Palm Oil Solid Waste, in: Tenth
International Conference on Information Technology: New
Generations (ITNG), IEEE, 2013, pp. 133-137.
[12] A. Chowdhury, M. U. Chowdhury, RFID-based real-time
smart waste management system, in: Telecommunication
Networks and Applications Conference, ATNAC Australasian,
IEEE, 2007, pp. 175-180.
[13] A. Bashir, S. A. Banday, A. R. Khan, M. Shafi, Concept,
Design and Implementation of Automatic Waste Management
System, International Journal on Recent and Innovation
Trends in Computing and Communication, 2013, 1(7): pp.
604–609.
[14] S. Longhi, D. Marzioni, E. Alidori, G. D. Buò, M. Prist, M.
Grisostomi, M. Pirro, Solid waste management architecture
using wireless sensor network technology. in: Proceedings of
2012 5th International Conference on New Technologies,
Mobility and Security (NTMS), IEEE 2012, pp.1-5.
[15] Y. Glouche, A. Sinha, P. Couderc, A Smart Waste
Management with Self-Describing Complex Objects,
Publisher, City, 2015. International Journal on Advances in
Intelligent Systems, 2013, 8 (1-2): pp1-16.
[16] M. Abbasi, M. Abduli, B. Omidvar, A. Baghvand, Forecasting
municipal solid waste generation by hybrid support vector
machine and partial least square model. International Journal
of Environmental Research, 2012, 7 (1): pp.27-38.
[17] E. Livani, R. Nguyen, J. Denzinger, G. Ruhe, S. Banack, A
hybrid machine learning method and its application in
municipal waste prediction, in: Industrial Conference on Data
Mining, Springer, 2013, pp. 166-180.
[18] T. Nuortio, J. Kytöjoki, H. Niska, O. Bräysy, Improved route
planning and scheduling of waste collection and transport,
Expert systems with applications 2006, 30 (2): pp.223-232.
[19] A. Zanella, N. Bui, A. Castellani, L. Vangelista, M. Zorzi,
Internet of things for smart cities, Internet of Things Journal,
IEEE, 2014, 1 (1): 22-32.
[20] L. Filipponi, A. Vitaletti, G. Landi, V. Memeo, G. Laura, P.
Pucci, Smart city: An event driven architecture for monitoring
public spaces with heterogeneous sensors, in: Sensor
Technologies and Applications (SENSORCOMM), 2010
Fourth International Conference on, IEEE, 2010, pp. 281-286.
[21] Z. Khan, S. L. Kiani, A cloud-based architecture for citizen
services in smart cities, in: Proceedings of the 2012
IEEE/ACM fifth international conference on utility and cloud
computing, IEEE Computer Society, 2012, pp. 315-320.
[22] P. Bellini, M. Benigni, R. Billero, P. Nesi, N. Rauch, Km4City
ontology building vs data harvesting and cleaning for smart-
city services, Journal of Visual Languages & Computing,
201425 (6): pp.827-839.
[23] Arduino Uno Datasheet specification 2012.
<https://www.arduino.cc/en/Main/ArduinoBoardUno>
(accessed 17.07.2016).
[24] ATMega 328 Processor Datasheet.
<http://www.atmel.com/devices/atmega328.aspx> (accessed
17.07.2016).
[25] R. Jalali, K. El-Khatib, C. McGregor, Smart city architecture
for community level services through the internet of things, in:
Intelligence in Next Generation Networks (ICIN), 2015 18th
International Conference on, IEEE, 2015, pp. 108-113.
[26] K. Spokas, J. Bogner, J. Chanton, M. Morcet, C. Aran, C.
Graff, Y. Moreau-Le Golvan, I. Hebe, Methane mass balance
at three landfill sites: What is the efficiency of capture by gas
collection systems?, Waste Management, 2006, 26 (5):
pp.516-525.
[27] LM 35 Precision centigrade temperature sensor datasheet.
Texas Instrument.
<http://www.ti.com/lit/ds/symlink/lm35.pdf> (accessed
17.07.2016).
[28] J. S. Atkinson, O. Adetoye, M. Rio, J. E. Mitchell, G. Matich,
Your WiFi is leaking: Inferring user behaviour, encryption
irrelevant, in: Wireless Communications and Networking
Conference (WCNC), IEEE, 2013, pp. 1097-1102.
[29] K. Karimi, G. Atkinson, What the Internet of Things (IoT)
needs to become a reality, White Paper, FreeScale and ARM,
2012: pp.1-16.

American Journal of Mechanical and Industrial Engineering 2019; 4(1): 6-23 23
[30] J. O. Adeyemo, E. Adetiba, O. Olugbara. Smart City
Technology based Architecture for Refuse Disposal
Management. In: IST-Africa 2016 Conference Proceedings,
2016, pp. 1-8.
[31] A. Doukas, Building Internet of Things with the ARDUINO,
CreateSpace Independent, in Proceedling of Sixth
International Conference on Innovative Mobile and Internet
Services (IMIS) in Ubiquitous Computing, IEEE, 2012, pp.
922-926.
[32] Maureira MA, Oldenhof D, Teernstra L. ThingSpeak–an API
and Web Service for the Internet of Things, 2014.
<https://staas.home.xs4all.nl/t/swtr/documents/wt2014_things
peak.pdf> (accessed 17.07.2016)
[33] A. Abayomi, O. O. Olugbara, E. Adetiba, D. Heukelman,
Training Pattern Classifiers with Physiological Cepstral
Features to Recognise Human Emotion, in: Advances in
Nature and Biologically Inspired Computing, Springer, 2016,
pp. 271-280.
[34] Y. Indonesia, Principal component analysis combined with
first order statistical method for breast thermal images
classification, IJCST 2011, 2(2).
[35] H. Cevikalp, Feature extraction techniques in high-
dimensional spaces: in: Linear and nonlinear approaches,
2005, Vanderbilt University.
[36] IT. Jolliffe, Rotation of principal components: some comment,
Journal of Climatology 1986, 7(5), pp.507-510.
[37] S. Kalkan, F. Wörgötter, N. Krüger, First-order and second-
order statistical analysis of 3d and 2d image structure,
Network: Computation in Neural Systems, 2007, 18(2),
pp.129-160.
[38] N. Aggarwal, R. Agrawal, First and second order statistics
features for classification of magnetic resonance brain images,
2012, pp1-8.
[39] R. Plamondon, S. N. Srihari, Online and off-line handwriting
recognition: a comprehensive survey, IEEE Transactions.
Pattern Analysis and Machine Intelligence, 2000 22, pp. 63–
84.
[40] N. Cristianini, J. Shawe-Taylor, An introduction to support
vector machines and other kernel-based learning methods,
Cambridge university press, 2000.
[41] Lippmann, R. P., 1987. An introduction to computing with
neural nets. IEEE ASSP Magazine,
<http://viveksingh.in/softcomp/Lippmann%20neural%20com
puting-intro.pdf > (accessed 17.07.2016 ).
[42] E. Adetiba, O. O. Olugbara, Lung cancer prediction using
neural network ensemble with histogram of oriented gradient
genomic features, in: Advances in Nature and Biologically
Inspired Computing, Springer, 2016, pp. 271-280.
[43] M.-C. Popescu, V. E. Balas, L. Perescu-Popescu, N.
Mastorakis, Multilayer perceptron and neural networks,
WSEAS Transactions on Circuits and Systems, 2009, 8 (7):
579-588.
[44] D. E. Rumelhart, G. E. Hinton, R. J. Williams, Learning
representations by back-propagating errors, 1986, Nature, 323,
pp. 533–536
[45] G. Cybenko, Approximation by superpositions of a sigmoidal
function, Mathematics of control, signals and systems, 1992, 2
(4): 303-314.
[46] E. E. Jordaan, Development of robust inferential sensors:
industrial applications of support vector machines for
regression, in, Technische Universiteit Eindhoven, 2002.
[47] H. Q. Minh, P. Niyogi, Y. Yao, Mercer’s theorem, feature
maps, and smoothing, in: International Conference on
Computational Learning Theory, Springer, 2006, pp. 154-168.
[48] S. Oyewole, O. Olugbara, E. Adetiba, T. Nepal, Classification
of Product Images in Different Color Models with Customized
Kernel for Support Vector Machine, Third International
Conference on Artificial Intelligence, Modelling and
Simulation, 2015, pp. 153 -157.
[49] B. Padmanabhan, 2012. Unified Modeling Language (UML)
Overview. Available:
http://people.eecs.ku.edu/~saiedian/Teaching/Fa13/810/Readi
ngs/UML-diagrams.pdf (Accessed 4 July 2016)
[50] D. Bell, 2003. UML basics: An introduction to the Unified
Modeling Language. The Rational Edge.
[51] Z. Lu, D. Szafron, R. Greiner, P. Lu, D. S. Wishart, B. Poulin,
J. Anvik, C. Macdonell, R. Eisner, Predicting subcellular
localization of proteins using machine-learned classifiers,
Bioinformatics, 2004, 20(4): 547–556. PMID: 1499045.
[52] T. Koomen, M. Pol, Test process improvement: a practical
step-by-step guide to structured testing, Addison-Wesley
Longman Publishing Co., Inc., 1999.
[53] T. G. Dietterich, G. Bakiri, Solving multiclass learning
problems via error-correcting output codes, Journal of
artificial intelligence research, 1995, 2, pp.263-286.
[54] J. A. Whittaker, What is software testing? And why is it so
hard? IEEE software 2000, 17 (1): pp.70-79.