You are on page 1of 6

Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)

IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

Real-Time Garbage, Potholes and Manholes


Monitoring System using Deep Learning
Techniques
Shahzia Sayyad Shaily Parmar Mrinalini Jadhav Karan Khadayate
Department of Computer Department of Computer Department of Computer Department of Computer
Engineering Engineering Engineering Engineering
Shah & Anchor Kutchhi Shah & Anchor Kutchhi Shah & Anchor Kutchhi Shah & Anchor Kutchhi
Engineering College Engineering College Engineering College Engineering College
Mumbai, Maharashtra Mumbai, Maharashtra Mumbai, Maharashtra Mumbai, Maharashtra
shahzia.sayyad@sakec.ac.in shaily.parmar@sakec.ac.in mrinalini.jadhav@sakec.ac.in karan.khadayate@sakec.ac.in

Abstract— A major challenge in urban cities is waste methods. An article published in ‘The Times of India'
management, as the pace of urbanization is growing rapidly dated 02/08/2018 states that "1,600 metric tons of garbage
sustainable urban development strategies are therefore is generated every day in the city". An article in
required. One of the main concerns with our environment 'www.dailymail.co.uk' dated 28/03/2017 states "30 deaths
has been improper garbage containment, unrepaired take place daily due to potholes". An article in
potholes, open manholes, and stagnant water which are
harmful to the well-being of the residents. S ince the idea of 'https://www.mid-day.com' dated 29/09/2016 states "A
smart cities is very trendy these days and without a smart 21-year-old person died after his bike hit a pothole". In
waste management system, smart cities can't be complete. the traditional system, manual monitoring every area
Traditional manual monitoring is a cumbersome process proves to be a cumbersome procedure that utilizes a lot of
and utilizes more effort, time, and cost which can easily be effort, time, and cost. Sometimes the daily activities of
assisted with present technologies. Monitoring such tasks localities become difficult due to odor, infections , and
using IoT is one such solution, but incorporating IoT deals many other factors caused by such scenarios. At such
with the use of sensitive electronic devices which are difficult times contacting or complaining to the respective
to maintain and hence add on some additional costs. Hence authorities seems to be the only approach. But these
this paper describes the development of a system using deep
learning techniques which can cut down huge costs in facilities aren’t available easily and instantly which may
effectively monitoring the surroundings and lend a hel ping worsen situations. Thus, there is a need to design an
hand for a better, safer present and future. automated system which will assist the concerned
authorities to efficiently tackle the problem of garbage,
Keywords - IoT, Faster-RCNN, potholes, garbage, manholes open manholes, and potholes. In this paper, a real-time
application is proposed to fill the requirements using
I. Introduction advanced techniques of Artificial Intelligence and image
processing to automate the monitoring process.
The concept of a city integrates the use of
modern technologies to optimize various city operations II. Literature Survey
and services. Street cleanliness is one of the primary tasks
of urban civilization. Cleanliness, management of Varioue research works that have implemented
potholes and manholes represents the spiritual outlook automated garbage, manholes and potholes monitoring
and humanistic atmosphere of a city. However, sites of systems and discussed this below. This will help one gain
widespread debris, open manholes, unrepaired potholes , a comprehensive idea of all the different types of systems
and stagnant water are common nowadays in many proposed or developed and can help one develop an
urbanizing areas. The Zurburg 2002 report states that in efficient system based on previously proposed or
2010 the amount of waste generated was 20,000 tons per developed systems.
day, and it was estimated that the amount will be no less
than around 47,000 tons per day by 2025. With the A. Anitha [1] developed an Internet of Things
existing methods of collecting and disposal, it is nearly (IoT) based system that consists of a sensor over the trash
impossible to manage such an amount of waste in the container which will distinguish the absolute degree of
future as around 30% of waste end up on the roads and trash inside it. At the point when the trash will arrive at
public places due to ineffective disposing and collecting the greatest level, a notice will be sent to the corporation's

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 826

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)
IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

office, at that point the workers can take further activities and battery. The maximum threshold levels are set for
to exhaust the container. The workers can check the status each sensor. The sensors monitor the changing conditions
of the containers on their cell phones. Ultrasonic sen sors when drainage gets blocked or sewage water floods or
are being utilized in the system to check the degree of manhole covers are open. When a specific sensor arrives
trash in the dustbins. at the particular threshold level, the respective value of
Mustafa M.R et al.[2] also proposed an IoT the sensor will be sent to the microcontroller. The
based system using ultrasonic sensors that measures the microcontroller sends the signal and the precise area of
trash level, an Advanced RISC Machine (ARM) the drainage through IoT to the managing station and a
microcontroller that controls system activity while programmed mail is sent by Raspberry Pi. This alarms the
everything will be associated with Thing Speak. This individual in-charge to take the necessary actions in
system enables the waste administration to screen, regards to the issue happening inside the drainage.
dependent on the degree of the trash inside the dustbin. [7] developed a system that uses sensors (for
The system shows the status of various four sorts of trash; example, accelerometers present in smartphones) which is
local waste, paper, glass, and plastic on a Liquid Crystal capable of automatically detecting potholes in real-time
Display (LCD) screen. and also monitor the road traffic conditions. Various
Liu Ying et al. [3] developed a system that Artificial Intelligence (AI) procedures like Support
automatically detects and identifies garbage directly using Vector Machines (SVM) based on K-means clustering are
a front-end embedded monitoring module which uses the applied to the information acquired from such sensors to
improved You Only Look Once Version 2 (YOLOv2) estimate street/traffic conditions.
network model to do garbage detection and recognition, Seung-Ki Ryu et al.[8] proposed a pothole
and manages thousands of monitoring front-ends through detection method based on two-dimensional (2D) images.
narrow-band Internet of Things and background server. This system is applied to ITS (Intelligent Transportation
Means of target box dimension clustering and System) service and road management system. Survey
classification network pre-training are used to improve vehicles in Korea were used to collect 2D road images.
the YOLOv2 model performance. The network becomes The system gave promising results and the information
lightweight after replacing the feature extraction network extracted using the proposed method was used in
by YOLOv2. determining the maintenance of the road management
Hiroya Maeda et al. [4] launched a system with system and in taking immediate action for their repair and
a large-scale road damage dataset. This dataset is maintenance. It also provides alert information of
composed of 9,053 road damage images captured with a potholes to drivers as one of ITS services.
Smartphone installed on a car, with 15,435 instances of Arul Ananth T S et al. [9] proposed a system that
road surface damage included in these road images. In uses various types of sensors like waft, level, and gas
each image, they annotated the bounding box representing sensors by interfacing with ARM7. This system is
the location and type of damage. They also used a state- evolved to monitor the water level, water waft, and gases.
of-the-art object detection method using convolution If the drainage gets blocked and water overflows this
neural networks to train the damage detection model with blockage could be recognized by the sensor and will
the dataset, and compared the accuracy and runtime speed display the message at the 16x2 LCD, then this
on both, using a web server and a Smartphone. Finally, information is dispatched via Global System for Mobile
the paper demonstrates the type of damage that can be Communication (GSM) to the nearby municipality
classified into eight types with high accuracy by applying provider for further corrective movement.
the proposed object detection method. On surveying the papers one would notice the
[5] proposed the installation of high-resolution evolution from using sensors installed in vehicles to using
cameras on vehicles to collect the street images. Mobile smartphones, using IoT based systems, detection based
edge servers are used to store and extract street image on threshold value to using artificial intelligence and
information temporarily. This street data is transmitted to machine learning approaches, post-processing to real-time
the cloud data center for analysis through city networks. detection of garbage, manhole or potholes. However, each
At the same time, Faster Region-Convolutional Neural system is designed for detecting either garbage, manhole,
Network (Faster R-CNN) is used to identify the street or pothole; they are unable to detect all the three problems
garbage categories and count the number of garbage. in a single system. To enable a single system to detect
Finally, the results are incorporated into the street garbage, manholes, and potholes have developed a system
cleanliness calculation framework to ultimately visualize using a combination of Image Processing and Deep
the street cleanliness levels, which provides convenience Learning techniques and have introduced an improved
for city managers to arrange clean-up personnel technique to reach the desired goal.
effectively. An online survey is conducted through Google
Chandraprabha R. et al. [6] developed a system Forms in which it gains more information about the
that uses Wireless Sensor Networks (WSN), consisting of hygiene problems faced by residents in our localities. In a
sensor nodes made out of controller, memory, transceiver, day, about 165 responses were submitted. 80% of

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 827

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)
IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

responses were in agreement about the presence of public


bins nearby. Despite the availability, 75% of responses Labeling Images
agreed that there is a presence of overflowing bins and The downloaded images were then labeled using the
widespread garbage in some areas. 51% of responses LabelImg Tool[11]. This tool generated an Extensible
were with the fact of potholes being kept unrepaired for Markup Language (XML) file for each image stating the
as long as 3 months. It is also found that 49.7% of Xmin, Ymin, Xmax and Ymax coordinates along with the
manholes were left open. Moreover, 74% of respondents Class Label for each class occurrence in that image.
voted for the development of a mobile application which
could help them to tackle such problems efficiently.

III. Proposed System

Fig. 2. Label Image

Training the image classifier


A pre-trained Faster-RCNN model was downloaded from
Common Objects in Context(COCO) Trained Models[12]
repository. This model is trained to detect 90 classes on
the MS COCO dataset. The Tensorflow Application
Program Interface API[13] was used for seamless training
and testing of the Image Classifier with Tensorflow
library in Python language. TensorBoard was used to
precisely monitor the training process. Tensorboard
provides a graphical representation of the following losses
Fig. 1. Flowchart against the step number:

The overall idea is to develop an alert system. ● Region Proposal Network (RPN) Loss
An application is created in which users can upload RPN Loss [10], LRP N calculated by Eq. (1) is to monitor
photos of untidy locations, manholes, and potholes. The the training of the RPN. It is the sum of two components,
images will be uploaded along with the location and the classification loss, LR1 shown in Eq. (2) which
timestamp. An Image Classifier based on the Faster R- measures the inaccuracy of RPN in deciding whether an
CNN model is built for validation of the images for the anchor consists of an object or not and the localization
occurrence of debris, potholes, and manholes and if loss, LR2 as shown in Eq. (3) which measures the
found, those details are then stored in a database, which inaccuracy of RPN in deciding the anchor coordinates.
will then be used to alert the concerned authorities. If left
unchecked, a counter will be kept for each image-location = + ∗ (1)
pair which will indicate the delay of necessary actions.
= ∑ ( , ∗)
Prolonged delay will increase the severity (and hence
priority) of that location. (2)
= ∑ ∗
∗ ( , ∗
) (3)
IV. Methodology

A. Training the Image Classifier Here, i is the anchor index in mini-batch, Ncls stands for
the batch size, Nreg is the number of anchor boxes, Pi is the
Preparing Dataset probability that anchor i has an object, Pi * stands for the
ground truth of anchor i containing an object (i.e. value is
The dataset was created by downloading 70 images from
1 if ith anchor contains an object, 0 otherwise), T i
Google Images for each class i.e. Garbage, Potholes , and represents the vector of parameterized coordinates (i.e.
Manholes. These images were then divided as 70% for Xmin , Ymin , Width and Height of the anchor), Ti * is the
training and 30% for testing.

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 828

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)
IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

ground truth box associated with that positive anchor, TABLE 4. BOX-CLASSIFIER LOCALISATION LOSS
is the balancing factor, Lcls is the log loss of 2 classes (LC2 )
(object and non-object) and Lreg is a smooth L1 loss Step Loss
function. 23210 0.070
While training the classifier, the RPN losses were 35000 0.040
observed as follows: 45000 0.010
64030 0.005
TABLE 1. RPN CLASSIFICATION LOSS (LR1 )
Step Loss The total loss (sum of LR and LCN ) was around 2.0 when
25220 0.008 the training initiated. The training process was terminated
35410 0.010 when the total loss dropped around 0.02. It took
45180 0.007 approximately 65,000 iterations to reach this milestone.
62490 0.001
B. Back -end implementation
TABLE 2. RPN LOCALISATION LOSS (LR2 )
Step Loss Web Server
26940 0.047 Flask Web Framework is used to develop a HyperText
35000 0.018 Transfer Protocol (HTTP) server. This server is used to
45000 0.003 provide the following services:
62580 0.002 ● User management and authorization
● Accept images from the users along with the
● Classifier Network Loss Geolocation and Timestamp to store them in a
The classifier network loss [14], LCN is used to measure database.
● Feed the images to the Image Classifier for
the inaccuracy of the classifier network. It is calculated
classification and store the classification details in
similarly as the RPN loss, i.e. it has 2 components,
classification loss LC1 shown in Eq. 4 and regression loss the database.
● Assist the users to track their complaints.
LC2 . The main difference is that in RPN, there are only 2
● Provide a summary of the complaints in a graphical
classes (object and non-object), but in the classifier
network there are multiple classes (equal to the class form to the concerned authorities.
labels for which the model is being trained). The
classification component is cross -entropy loss and is Database System
calculated as: Relational Database Management System (RDBMS) was
used to store the user and complaint details on the server.

[ ][ ] This database helps the concerned authorities to monitor
= (∑ )
[ ][ ] the cleanliness of the locality by storing the Geo-Location
(4) and time stamp of the complaints.
C. Front-end implementation
Where, Nsamples is the number of samples, i stands for the The front end is an Android App which consists of the
sample number, j stands for the index of class labels, following modules:
x[i][j] stands for the probability that the sample i belongs
to class j and Ci is the class index for which sample i Login & Register Activity
shows maximum belongingness. Using this module, users can sign up and log in to use the
The regression loss LC2 is calculated in a similar way to app for posting complaints. A minimum 8 character
LR2 except, now the regression coefficients are class - password is used for authentication of the users.
specific, i.e. for each object class, the coefficient of
regressions is calculated. Home Activity
In this module, the user can capture images using the
TABLE 3. BOX-CLASSIFIER CLASSIFICATION camera and upload them on the Web Server. The
LOSS (LC1 ) Geographic location of the user and Time Stamp of the
Step Loss capture action is noted and sent along with the image data
23410 0.14 and user's ID in the complaint body. The image data is
35000 0.100 encoded into Base-64 format for a seamless transfer over
46640 0.090 the Internet.
61990 0.020

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 829

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)
IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

image classifier for classification and uploads the


classification results back to the server. Hardware
isolation of this module from the Web Server helps in
reducing the computational load on both modules and
thus, enhances performance and user experience of the
system.
V. Result

Fig. 3. Home Page

History Activity
Through this activity, users can view a list of all
complaints they have registered before.
Fig. 5. Garbage Detected
Details Activity
This activity helps the users to keep track of the selected
complaint. It shows the status of the complaints to the
user along with the classification results of the image
submitted with that complaint in graphical form.

Fig. 6. Pothole Detected

Fig. 4. Details Page


Admin Panel
This section is used by officials of the concerned
authorities to monitor the complaints.

D. Image classification Fig. 7. Pothole, Manhole, and Garbage Detected

A script is written in python which regularly queries the As shown in figures 5, 6 and 7, the garbage, manhole and
Web Server after fixed intervals for complaints that pothole objects identified are marked in rectangular boxes
haven't been filtered yet. For each of such complaints, the and the value on the rectangle represents an accuracy of
script downloads the corresponding image feeds it to the

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 830

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.
Proceedings of the Fourth International Conference on Trends in Electronics and Informatics (ICOEI 2020)
IEEE Xplore Part Number: CFP20J32-ART; ISBN: 978-1-7281-5518-0

determining whether the object is a garbage category,


[14].Object Detection and Classification using R-CNNs [Online].
manhole category or pothole category.
Available: http://www.telesens.co/2018/03/11/object-detection-and-clas
sification-using-r-cnns/
VI. Conclusion

The challenges embedded in the garbage,


potholes, and manhole management as detailed in the
introduction leading to the development of the system is
significant. To address these challenges, using the
mentioned methodologies to implement a real-time
scalable and reliable system and lend a helping hand in
solving global problems. Hence, the system shows how
advancements in technologies can be used to solve a
variety of such issues easily.

ACKNOWLEDGM ENT

The research was supported by Shah & Anchor


Kutchhi Engineering College, Mumbai. We express our
gratitude to the Management and Principal, Dr. Bhavesh
Patel for providing resources that we required in this
research. We also thank our HoD Prof. Uday Bhave for
his immense support.
REFERENCES

[1]. Anitha, A. (2017, November). Garbage monitoring system using


IoT . In 14th ICSET-2017, IOP Conf. Series: Materials Science and
Engineering (Vol. 263, p. 042027).
[2]. Mustafa, M. R., & Azir, K. K. (2017). Smart bin: internet-of-things
garbage monitoring system. In MATEC Web of Conferences (Vol. 140,
p. 01030). EDP Sciences.
[3]. Liu, Y., Ge, Z., Lv, G., & Wang, S. (2018, August). Research on
Automatic Garbage Detection System Based on Deep Learning and
Narrowband Internet of T hings. In Journal of Physics: Conference
Series (Vol. 1069, No. 1, p. 012032). IOP Publishing.
[4]. Maeda, H., Sekimoto, Y., Seto, T ., Kashiyama, T ., & Omata, H.
(2018). Road damage detection using deep neural networks with images
captured through a smartphone. arXiv preprint arXiv:1801.09454.
[5]. Zhang, P., Zhao, Q., Gao, J., Li, W., & Lu, J. (2019). Urban Street
Cleanliness Assessment Using Mobile Edge Computing and Deep
Learning. IEEE Access, 7, 63550-63563.
[6] Chandraprabha R, Ashwini C.V, Dharani M, Harshitha G, Kruti
Mohan. "Smart Real T ime Manhole Monitoring System" ,Volume:06,
Issue:07,July 2019,e-ISSN:2395-0056 ,p-ISSN:2395-0072.
[7]. Akinwande, V., Adewole, K., Bello, O., & Akintola, A. (2015).
Automatic and real-time pothole detection and traffic monitoring system
using smartphone technology. In International Conference on Computer
Science Research and Innovations (CoSRI 2015).
[8]. Ryu, S. K., Kim, T ., & Kim, Y. R. (2015). Image-based pothole
detection system for IT S service and road management
system. Mathematical Problems in Engineering, 2015.
[9]. Arul Ananth T S, Ramya Laxmi G, Renuka K, Karthik K. "Smart
Sensors and Arm Based Drainage Monitoring System", Digital Object
Identifier:10.35940/ijitee.K1183.09811S19, September 2019, Volume-8,
ISSN:2278-3075.
[10]. Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun “ Faster R-
CNN: T owards Real-T ime Object Detection with Region Proposal
Networks” (Submitted on 4 Jun 2015 (v1), last revised 6 Jan 2016 (v3))
arXiv:1506.01497
[11]. LabelImg T ool [Online]. Available: https://github.com/tzutalin/lab
elImg
[12]. COCO T rained Models [Online]. Available: https://github.com/ten
sorflow/models/blob/master/research/object_detection/g3doc/detection_
model_zoo.md
[13]. T ensorFlow Models API [Online]. Available: https://github.com/
tensorflow/models/tree/master/research/object_detection/legacy

978-1-7281-5518-0/20/$31.00 ©2020 IEEE 831

Authorized licensed use limited to: Western Sydney University. Downloaded on August 16,2020 at 12:38:40 UTC from IEEE Xplore. Restrictions apply.

You might also like