Pothole Detection using Accelerometer and

Computer Vision with Automated Complaint

Redressal
Aditi Kandoi, Harsh Agarwal, Abhishek Revadekar, Bhalchandra N Chaudhari
Department of Computer Engineering,
Sardar Patel Institute of Technology,
Mumbai, 400058
kandoiaditi0@gmail.com, harshagarwal2703@gmail.com, abhishek.revadekar@spit.ac.in, bnc@spit.ac.in

Abstract—Road transport is the most widely used means of

transportation around the world. With this high use of road
transport, the safety of travellers’ becomes the prime concern for
any governing authority. While some safety concerns arise from
driver errors and environmental factors, most cases are a result
of poor maintenance of these roads. Potholes, specifically, are one
of the leading causes of road accidents throughout the world and
need to be taken care of immediately, by the authorities. This
paper presents a solution that makes use of civilians’ mobile
sensors, along with image-based alternatives to detect potholes
in real-time, using Machine Learning. The concerned authorities
are then notified about the same through a web-based portal,
to take the necessary action. The solution also incorporates
pivoting existing complaints, location tagging and prioritization.
Additionally, the solution provides a forecast of the likelihood of
issues regarding potholes, constantly updating time series data
of the locations.
Index Terms—Road Safety, Pothole Detection, Sensors, Com-
puter Vision, Machine Learning, Complaint Redressal, Predictive Fig. 1: Analysis of Death by Accidents
Maintenance

I. I NTRODUCTION authorities and the citizens. Furthermore, a robust tracking

Road transport is the most common mode of transportation. system is also equally necessary to track the progress of repair
Especially countries like India, which are highly populous, and allowing citizens to be informed of such poor conditioned
the number of vehicles are constantly increasing , causing roads, thereby reducing the fatality caused by such situations.
several accidents on the streets. One of the main reasons Fortunately, using some of the latest technologies, we can
being, uneven, road surface conditions. Moreover, people face avoid the loss of human lives due to road accidents. This
difficulties on the roads be it walking or driving, especially paper aims to make the streets safe for everyone. We address
during the monsoons when the roads are completely dilapi- this problem by bridging the gap of communication between
dated. According to India’s Ministry of Statistics and Program the government and the civilians. We aim at crowd-sourcing
Implementation, majority of the accidental deaths in India take the data so that every person is actively involved in the
place because of mishaps on roads as shown in Figure 1. The development and they have a platform to raise concerns.
research in [1] mentions that India has a total of two million We propose a solution that uses a real-time pothole detection
kilometre roads, out of which, one million kilometre roads system for automated detection of potholes using accelerom-
are poorly constructed. As per the data set obtained by the eter and gyroscope sensors inbuilt in smartphones as well as
Ministry of Road Transport and Highways, approximately fifty through computer vision. The proposed system collects and
thousand accidents take place because of the potholes every analyses this data, which in turn can be used by the civic
year. A large percentage (4̃0%) of this often leads to death. authorities responsible for maintaining the roads. The system
The main problem in addressing the road maintenance keeps in check if the problems are being worked on to make
concerns is the lack of communication between the authorities it effective. The uniqueness of the proposed solution is that it
and the civilians facing the issues, and poor tracking of predicts the localities susceptible to the problem using machine
locations where repairs are required. Involvement of civilians learning besides reporting the existing conditions of the road
further enhances societal living and cooperation between the network.
 II. L ITERATURE S URVEY work which involved reproducing road surfaces as disparity
maps, by extracting undamaged road surfaces as references
Artis Mednis et al. [2] and Fatjon Seraj et al. [3] used and evaluating their difference with coarser potholes. It also
accelerometer sensors in smartphones to measure the accel- used stereo vision for 3- dimensional perception. Conventional
eration values to detect potholes. They have described the use detection algorithms such as R-CNN and its variants make use
of different algorithms like Z-THRESH, X-DIFF, STDEV(Z), of two-stage detectors as a part of detection. Recent works
G-ZERO, SVM and compared those algorithms to accurately on YOLO and SSD show promising results that one-stage
detect potholes on the road. By using different algorithms in detectors might also yield an accuracy within the range of
optimal scenarios both were able to achieve accuracy up to the two-stage detectors. However, these algorithms face the
90%. Ghadge et al. [4] used K-means clustering algorithm for consequences of class imbalance. The work by Tsung-Yi Lin
training and then was tested using Random forest algorithm. It et.al. [12] introduced the Focal Loss function, implemented in
was very feasible to implement because no external hardware their proposed RetinaNet detector for the dense sampling of
components were required, and the accuracy was quite high input data. Our proposed methodology for Pothole detection
for detecting potholes. The paper [5] proposed a method that makes use of this approach, owing to the promising accuracy
uses Euler angle to normalize the data from the sensors and the over state-of-the-art algorithms.
combined Z-THRESH and G-ZERO methods to achieve better Ayush Vora et. al [13] used a smartphone to report the
precision scores. The Real-time system accurately detected the road quality to maintenance authorities by facilitating the
pothole without false positives hence increasing the precision citizens’ involvement for real-time data collection. Afify et al.
of the algorithm and then using GPS of the phone to mark the [14] proposed the implementation of an electronic Customer
location of the detected anomaly. Thitaree et al. [6] proposed Complaint Management System (e-CCMS) which dealt with
an algorithm that used unsupervised learning techniques with all the complexities of a generic grievance system. The model
Neural Networks to develop a Real-Time system for the helped them manage the various issues and solved the time-
drivers. The training data showed that the system could iden- critical problems on time. The authors of the paper [15]
tify the pothole within 7 seconds but during real-life training, proposed the use of a system linking the users and the
they achieved an accuracy of 81% within 13 seconds which government. The app kept track of all the complaints and
is independent of the speed and road traffic. The authors [7] increased the transparency between the two by enabling the
used a different machine learning approach to identify potholes users to know more information about their complaints and
using data from accelerometer and gyroscope sensors built-in how the government was going to take action on the same.
smartphones. They tested the results on different classification A smartphone-based solution is proposed in [16] to solve
algorithms like Naive Bayes, C4.5, SVM and were able to the problem of complaint redressal. Citizens could send a
achieve the best accuracy of 98.6% using the C4.5 decision photo of the issue with the GPS location of it. The authorities
tree. could reject the complaint but only with a valid reason. The
The work proposed by S. Nienaber et al. [8] did not dashboard maintained a record of active, reported and resolved
rely on machine learning to detect potholes. Instead, they cases to help the users understand the steps taken by authorities
used the Canny edge detection algorithm accompanied by to resolve their issue.
dilation of the resultant edges to abort unwanted edges, and After reviewing the existing work, we found out that
provided a precision of 81.8% and recall of 74.4% and used there are various methods to detect potholes but there is a
a dash-mounted GoPro camera for road surveillance. This lack of effective communication between the citizens and
approach, however, has to restrict the portion of the image to the government officials. The aim of the proposed solution
be observed as it depends on camera placement and distance is to develop an integrated system that detects, reports and
from the car. Pothole detection approaches include various addresses the grievances effectively. In the existing solutions,
methods such as Vibration-based, 2D-Vision- Based, 3D Scene it was observed that even if the complaints were registered,
Reconstruction and Learning methods. Amita Dhiman and some of them might be overlooked by the officials and they do
Reinhard Klette [9] presented the use of stereo-vision cameras not have a clear insight about the criticality of it. The proposed
that considerably enhanced the accuracy of the results. The system resolves it by setting priority to the complaints and
four methods proposed in the paper included two based on filtering them for the officials. Also, reports are generated for
stereo vision, while the other two being implemented with all the complaints which are displayed on the website and can
the help of Transfer Learning. Emir Buza et al [10] proposed be sent to higher authorities. The dashboard on the website
a method that used normalized spectral clustering to extract provides the forecast of the roads which might need repair
non-linear shapes and regions on the road. Using segmented and hence dealing with pothole problems before they occur.
road data utilised the unsupervised learning approach for shape
III. M ETHODOLOGY
extraction. With a well-distributed set of seeds obtained from
the clusters, the algorithm located the farthest points, vertically A. Architecture Model Description
and horizontally, in the shape with the same pixel value as The core of this research involves detecting potholes
the seed point, which provided a nearly accurate shape of through various methods on roads and its redressal. The
the pothole on the road. Rui Fan et al. [11] proposed their aim is to implement a real-time pothole detection system
that facilitates this. Therefore, a mobile application is used from the moment the driver started the application until
for interaction between the users and civic authorities about the stop button was pressed. Once the driver stops the data
the queries put forward by the users and a web portal is acquisition, the data set is stored and used for training. For
developed for redressal of the queries. Smartphones sensors consistent results sampling rate of the accelerometer sensor
like accelerometer and GPS are used for the collection of the in the system is set to 50Hz. To maintain uniformity in
data. The system proposed uses two methods for detecting all cases, the cars were made to drive over the potholes
potholes; by using accelerometer sensors, and by using images for data collection at different speeds ranging from 20-40
uploaded by the users. Detection algorithms are applied to this km/hr simulating real-life road scenarios in urban cities. Four
data and coordinates of the potholes are obtained using Google driving journeys of about 1 hour each were used for data
Maps API. The complaints are displayed on the feed in the collection. Twenty-five thousand samples of data having 7
application for the users. Users can upvote any complaint to features were obtained in the experimental study.
increase its priority. Several features are added in the system
which allows the users to check the status of the complaints Preprocessing: To remove the irrelevant data, we filter
raised by them. out some of the factors. Different positions of smartphones
cause changes in the magnitude of gravity. So, a low-pass
WEB APPLICATION filter is applied to isolate the force of gravity and filter out
Dashboard
Viewing and updating Status
high-frequency signals by using a filter constant. Hence,
Notifying Users
the filtered accelerometer and gyroscope sensors give the
resultant sample data.
DJANGO REST FRAMEWORK
(Hosted on pythonanywhere.com)

GOOGLE MAPS
SETTING
PRIORITY
Labelling: The features of the data are timestamp, 3-axes
MySQL Database coordinates of accelerometer and gyroscope sensors. By taking
ISSUE
into consideration the z-axis of the accelerometer, the thresh-
old value was found which was used to identify potholes on
ANDROID APPLICATION the journeys recorded.
DETECTION
Real-Time Detection
Geo-Location with Upvoting
Verifying Status of Issues
ALGORITHMS
To detect potholes in real-time, consecutive samples are
examined to find road anomalies by comparing with the
ACCELEROMETER
threshold and the resultant down-sampled points are stored
IMAGE DATA
GYROSCOPE
DATA separately. When a pothole is detected, the z-axis of the
accelerometer sensor has a pattern by which there is a sudden
Fig. 2: Architecture increase followed by a decrease in its value. Therefore, a
vertical and a horizontal limit were set to create a window and
As illustrated in Figure 2, the complaints are sent to the were obtained by using the consecutive down sampling points.
Django server and stored in the MySQL database. An algo- Since the z-axis of the accelerometer sensor was used for
rithm is developed to set a priority to each complaint which is detection, the horizontal limit was used to find the “h-up” and
displayed on the portal using Google Maps API, for the civic “h-dn” values which created a window as shown in Figure 3.
authorities to verify and facilitate the redressal of them. A Here, the blue lines denote the sample data that was collected
status is added for each complaint which can be edited only for training and the red marker denotes the downsampled data
by the civic authorities based on the actions taken for each that was obtained after comparing with the threshold value.
complaint, which is displayed on mobile. Once the pothole is The X-axis denotes the time series data converted to integer
resolved, the complaint is removed from the database, which for convenience and the Y-axis denotes the acceleration values
is thus reflected in the application. in the z-direction for each sample. A window is also created
whose upper limit is “h-up” and the lower limit is “h-dn”.
B. Mobile Application Once these values are detected, the downsampled data
The solution deals with the detection of potholes using is compared with the window and a label is assigned for
various methods which are done using Android OS based each sample which gives the detail of the sample whether
smartphones. The popularity of smartphones makes the it is a pothole or not. The label ‘1’ is assigned if it is a
system very user-friendly. It enables strong participation, and pothole, otherwise ‘0’, For a continuous value of ‘1’ in the
interaction of users to solve the problem. Several features sample data, it denotes one pothole. In Figure 4, the green
have been added to make the system reliable. Some of the marker denotes the label ‘1’ that is, potholes are detected at
features are explained below. those points and the other markers meaning the same as earlier.

1) Using the Accelerometer sensor in mobiles: Training Classifier Model: The dataset was divided into
Dataset: The smartphone was mounted on the dashboard train and test data such that 30% of the data was used in
of the vehicle which recorded the time and accelerometer test data in classification models to calculate accuracy. It
and gyroscope readings for training. The data was collected is trained on the training data to detect potholes. Because
 camera setup or through pedestrian participation. Citizens
can participate by drawing the authorities’ attention to such
problems. Generally, it is over social media platforms, which
often gets overlooked in the heap of new information. This is
where a dedicated complaint platform can come in handy.

Fig. 3: Downsampled data

Fig. 5: (a) shows the feature extraction pyramid. Part (b) shows
the two subnetworks - one for classification of the classes and
the other for regression of the boundary boxes.

RetinaNet: In this manuscript, pothole detection is performed

using RetinaNet, a one-stage detector. It is subject to class
imbalance since the dataset only contains labelled data for the
potholes and not the smooth roads. However, RetinaNet takes
care of it by using Focal Loss allows biased learning towards
the training samples [12]. The implemented RetinaNet
architecture uses a ResNet backbone for feature extraction.
A Feature Pyramid Network is used on top of ResNet
to construct a rich multi-scale feature pyramid from one
Fig. 4: Distinguishing between potholes and not potholes on single resolution input image. This is accompanied by two
downsampled data subnetworks, one for the classification of classes and the
other for bounding box regression. The classification subnet
predicts the probability of object presence at each spatial
the training data is a labelled data set, supervised machine position for each of the anchors and object classes. The subnet
learning algorithms are used to infer a model from the labelled is a fully convolutional network (FCN) that applies four 3×3
sample data. Unlike the training data, the test data does not convolution layers, each with 256 filters and followed by
contain the label feature. Five classification algorithms were ReLU activation after each layer, followed by another 3×3
worked upon, on the testing data; Random Forest, Logistic convolution layer. The box regression subnet is a similar FCN
Regression, Decision Tree, Naive Bayes and K-Nearest that regresses the offset from each anchor box to the nearest
Neighbors and compared their accuracies. The model that ground-truth object if it exists. Figure 5 shows the RetinaNet
gave the highest accuracy was used to detect potholes on roads. architecture.

Prediction: The system records the real-time data from Dataset: The dataset used for this work is obtained from
the sensors and preprocesses it to remove the unnecessary Google Images of bad roads. It consists of 254 images and a
high-frequency data for prediction. Afterwards, the prediction total of 1757 samples of potholes. The potholes are manually
of potholes on this data is done based on the trained classifier annotated in the XML format, containing the vertices, Xmin ,
model. Alerts of the locations of the predicted potholes are Ymin , Xmax and Ymax of the bounding boxes. The dataset is
sent to the users on their smartphones. Finally, these locations trained for 150 epochs with a batch size of 4 and 400 steps
are sent to the server which maintains the database of all per epoch.
recorded pothole locations.
Detection: For detection, the input image is provided in the
2) Image-based pothole identification: BGR format. It is normalized by subtracting the ImageNet
Using the accelerometer sensors to detect potholes provides mean of all three channels. It was found that a threshold of
automatic complaint redressal by the passengers of the 0.4 detects the maximum true positive boundary boxes for
vehicle. However, it requires driving into the pothole, which the potholes, throughout multiple tests.
causes great inconvenience to the passengers. Hence, an
image-based pothole detector is also considered, which 3) Geotagging locations of the complaints:
can be utilized in two ways – as a dash-mounted phone The complaints of the users are tagged with the GPS location.
For getting the GPS location, the fused location provider API by the authorities.
is used. The API uses technologies like mobile GPS, Wi-Fi and
cellular connection to accurately determine the exact position
of the user. Hence, to achieve the best results in determining
the location, the API is set to the highest accuracy. When
the system detects a pothole or a user clicks a photo of it,
the GPS coordinates of the last known location of the device
are recorded. These values are sent to the server with the
complaint, hence marking its location on the map.
4) Complaint Feed and Upvoting similar complaints:
The feed is added to the system to allow the users to view
complaints raised by other users within their proximity. The
default minimum range is 100 meter which can be increased
using the slider by the users. This helps to display all the posts
within that range for the user. The complaints recorded by the
system are geofenced with a radius of fifty meters around each Fig. 6: Complaints per Ward
pothole location using Google Services API. Whenever a user
passes over this region, a notification is sent to them which Dataset: During the testing of the application, the data
warns them about the nearby potholes beforehand. reported potholes in Mumbai was collected. The data consists
In case a user finds any complaint critical, he can upvote of the timestamp, latitudes and longitudes of various locations
that complaint to give it a higher priority, so that immediate reported by the users, the number of people facing the
action is taken on it. Similarly, upvoting is also used to problem and the date when the issue got resolved. The data
consolidate repetitive complaints from the same location. is from the time period of 1st October 2019 to 28th February
The priority of complaints increases as the votes increase. 2020 consisting of 1995 complaints in total.
Hence, a complaint with a higher number of votes is given
the maximum priority which is considered as one of the SARIMA Model: As one of the important functionalities of
important parameters for setting the priority of the complaints. the system is forecasting the number of potholes for the future
on the dashboard, the Seasonal Autoregressive Integrated
Moving Average (SARIMA) model was used on our time
C. Web Portal series dataset which consisted of five months of data. To
make the data stationary for efficient prediction, an Augmented
The solution also deals with the redressal of pothole
Dickey-Fuller test was performed. While using univariate
complaints by the users. This feature is solely handled by
time data, some aspects like autoregression, differencing and
civic authorities using a web portal which helps them get
moving average were considered. To smoothen the data, in the
information on the complaints and forecasts the expected
moving average model, three different window sizes 5 days, 30
number of complaints arising in the future. The portal is
days and 90 days were used to check for a trend in the curve.
only accessible by the authorities who can perform several
By increasing the window size, finding a trend was easier.
tasks for the resolution of the complaints. For making it a
While dealing with this model, seven important parameters
competent system, several novel functionalities have been
need to be considered: trend autoregression order(p), trend dif-
added to the system which is explained below.
ference order(d), the trend moving average order(q), seasonal
autoregressive order(P), seasonal difference order(D), seasonal
1) Dashboard:
moving average order(Q) and the number of time steps for
The civic authorities have access to the web portal with a
a single seasonal period(s). The equation for the SARIMA
dashboard that serves the authorities to view and analyze all
model is as follows:
the complaints posted by the mobile users.
The dashboard helps in maintaining a count of active SARIM A(p, d, q)(P, D, Q)s (1)
complaints, being worked on or those which are resolved.
By using this data, we analyze and generate various insights The parameter values were set as follows: p = range(0, 5),
which can be used by the Government authorities to plan d = 1, q = range(0, 5), P = range(0, 5), D = 1, Q = range(0,
better strategies for the resolution of the pothole problem. 5), s = 5. By fitting the model on the data, the duration of
The data shows various wards which are facing the issue as predictive maintenance of the predicted potholes prone roads
shown in Figure 6. Information of the authority which was is suggested to the authorities.
responsible for maintaining the roads in a particular ward can
be found out. The statistics also show the time taken by the 2) Prioritization:
authorities to resolve the complaints and those areas where Existing systems lack the insight the officials have over the
the problem kept on reappearing even after getting resolved complaints which are of utmost importance. It so happens
that the most important complaints are spammed by those
which can be dealt with later. A unique way to deal with
such a problem is by setting the priority to all the complaints
reported by the users and sort them accordingly. Four key
parameters significant to road conditions are identified and for
setting the priority the sum of these parameters is considered.

Upvotes By Users: All the users can upvote any complaint

they are facing. By these upvotes, the number of users facing
a particular problem is derived and hence this is the most Fig. 7: Area Wise Complaints Clustered on Map
important parameter in the calculation of the priority.
It was found out that people tend to upvote complaints rather
than uploading new ones. Priority will rapidly increase with wise to get a better view. A map view is also displayed on
each upvote and this can be a bias to this parameter. Therefore, the portal which shows the number of complaints in different
a range for the number of upvotes is decided as mentioned in areas. The complaints are ordered according to their priority.
I to keep the value within a calculable range. As shown in Figure 7, the complaints are clustered area wise
such that all the complaints in a particular region are grouped
TABLE I: PRIORITY ACCORDING TO VOTES with the count of those complaints printed above the marker.
Number of Votes Priority The blue region in the figure is one such identified location
<20 0.5 with 15 unresolved complaints in it. This gives a broader view
20-50 1
>50 2
to the officials which helps them resolve complaints according
to a region.
Location of The Complaint: A complaint’s priority should
be dependent on the type of people facing the issue. A
pothole near a hospital, a school, an administrative building
or highways are of more importance than any other place.
The Locations API provided by the Google Cloud Platform
provides all the nearby places of a location whose coordinates
are provided to it. By setting The RADIUS to 50 metres and
with the help of TYPE SEARCH, if the nearby locations
include the above-mentioned buildings, the priority is
increased by a value of 1.
Fig. 8: Complaints Marked on Map
Traffic In The Area: Uneven roads filled with potholes is
sure to restrict the movement of automobiles and cause hour- The government officials can view individual complaints
long traffic jams. The use of Roads API in conjunction with marked on the map. The numbers on the marker in Figure
Directions API gives the traffic layer data. The authorities 8 denotes the priority of the complaint. It can be clicked to
get to know the reason for these traffic jams which can be get details like the coordinates, address, timestamp and the
improved by fixing the road. The API provides us with the description of the complaints.
JSON value of traffic in the area. Heavy traffic areas increase
the priority of that road by a value of 1. 4) Status Update on maintenance:
There is a need for a system that facilitates the redressal of
Repetitive Complaints: It is meaningless if potholes pothole complaints by the civic authorities. Hence, status is
are reported again even after the road has been reconstructed assigned to each complaint based on the activity performed
by the authorities. More attention is given to such complaints on them. After the complaints are arranged in the order of
by the officials to get rid of the potholes altogether. The priority, they are viewed by the authorities on the portal with
priority value of such complaints is increased by 1. a ‘pending acknowledgement’ status set as default.
Once the authority has reviewed the complaint, its status is
3) Viewing all complaints: changed to ‘acknowledged’. This can be viewed on the user’s
The proposed solution maintains and displays the records of feed. A resolution task is assigned for the complaint, which is
all the complaints in a feasible manner. The complaints are viewed as ‘work-in-progress’ and finally, when the resolution
fetched from the server and viewed by the civic authorities is completed, the status of that complaint becomes ‘resolved’.
on the portal. They have the option of viewing a specific This lets the users know that the complaint has been resolved
complaint, its address and location on the map. and after 7 days of the resolution, that complaint is deleted
Google Maps API has been used to view the location of from the users’ feed.
the complaints and those can be consolidated together, ward
 5) Contacting higher authorities: as compared to the other classifiers for the system and is best
The complaints can go unnoticed by the civic authorities due suited for our system to detect potholes.
to which the complaint is not resolved which does not fulfil
the purpose of the system. B. Image-based Pothole Detection
Maintaining one portal for the different needs of the people The manually-annotated dataset was trained with the Reti-
is a challenging task for higher authority. Therefore, if any naNet model for 150 epochs. The model obtained was used
complaint registered by users is not acknowledged or is not to detect potholes of different texture and different angles, as
worked upon by the concerned authority for a long period, shown in Figure 10, with some of the sample images that were
then the system automatically makes a detailed report about tested.
the complaint to send to the higher officials of the state to
maintain the functionality of the system. The report can also
be published on Twitter or other social media platforms to
increase awareness.

IV. R ESULTS
A. Pothole Detection Using Sensors
Sample accelerometer and gyroscope sensor readings are (a) Potholes with patches of (b) Dry-textured pothole with-
smooth road out puddles
divided into train and test data. Cross-validation is performed
on the test data that has been scrutinized from the sample Fig. 10: Results of tested images
dataset. Several evaluation measures have been taken into
consideration for choosing the most appropriate model. The progression of precision and recall over the epochs are
given in Figure 11. After testing with around 25 images, the
final model gives a precision of 0.83 and a recall of 0.72 which
gives realistic performance for the model.

Fig. 9: ROC Curve

To evaluate the performance of the five models, accuracy

was calculated. From table II, it was observed that the accuracy Fig. 11: Precision and Recall over training for 150 epochs
of the classifiers was very close, so it was more important to
optimize precision and recall than to focus on a single factor.
C. Predictive Analysis of Potholes
TABLE II: CLASSIFICATION REPORT Dickey-Fuller test was performed on the time series dataset
Algorithm Accuracy Precision Recall f1-score to stabilize the data. It was found that the p-value was not less
Random Forest 89.63 0.71 0.45 0.55 than 0.5, therefore, the data for forecasting was not stationary.
Decision Tree 83.95 0.62 0.42 0.50 The SARIMA model is a combination of several models which
Naive Bayes 84.52 0.57 0.37 0.45
Logistic Regression 82.61 0.61 0.47 0.53 works well with non-stationary data, therefore, it fits well with
KNN 85.86 0.82 0.21 0.33 our time series data.
After running the model for 625 epochs, the Mean Absolute
The ROC-AUC curve was plotted based on precision and Percentage Error (MAPE) was found to be 11%. By predicting
recall to check how well a model fitted in the system. From the values of the last 25 days in the dataset, the predicted
Figure 9, it is observed that the AUC for Random Forest is values were compared with the actual values for the month of
0.841, which is the closest to 1 which means that this classifier February as shown in Figure 12 and the trend looks pretty
was able to distinguish between the two classes quite well. close. The model gave an accuracy of 85.7% against the
Therefore, the performance of Random Forest works the best training data.
 It provides a better knowledge of routes to be used by users to
avoid road accidents and damages. Further, it helps the civic
authorities keep a track of the number of potholes and improve
the road conditions by the contribution of users.
In the future, the system can be scaled to report grievances
like garbage, debris, graffiti, and so on. Furthermore, classi-
fication of roads into good and bad roads can also help in
finding optimal routes for travel. A single integrated platform
can be made for citizens to interact with the authorities, and
express their issues with ease.
R EFERENCES
Fig. 12: Actual vs Predicted Number of Potholes
[1] U. Sharma and A. Kanoungo, “Study of causes of potholes on bitu-
minous roads – a case study,” in Journal of Civil Engineering and
Environment Technology, pp. 345–349, 04 2015.
Figure 13 gives the prediction of the total number of [2] A. Mednis, G. Strazdins, R. Zviedris, G. Kanonirs, and L. Selavo, “Real
potholes in the month March. time pothole detection using android smartphones with accelerometers,”
in 2011 International Conference on Distributed Computing in Sensor
Systems and Workshops (DCOSS), pp. 1–6, 2011.
[3] F. Seraj, B. van der Zwaag, A. Dilo, T. Luarasi, and P. Havinga, “Roads:
A road pavement monitoring system for anomaly detection using smart
phones,” in Proceedings of the 1st International Workshop on Machine
Learning for Urban Sensor Data, SenseML 2014, pp. 1–16, Springer, 9
2014.
[4] M. Ghadge, D. Pandey, and D. Kalbande, “Machine learning approach
for predicting bumps on road,” in 2015 International Conference on
Applied and Theoretical Computing and Communication Technology
(iCATccT), pp. 481–485, Oct 2015.
[5] H.-W. Wang, C.-H. Chen, D.-Y. Cheng, C.-H. Lin, and C.-C. Lo, “A real-
time pothole detection approach for intelligent transportation system,”
Mathematical Problems in Engineering, vol. 2015, pp. 1–7, 08 2015.
[6] T. Tanprasert, C. Saiprasert, and S. Thajchayapong, “Combining un-
supervised anomaly detection and neural networks for driver identifi-
cation,” Journal of Advanced Transportation, vol. 2017, pp. 1–13, 10
Fig. 13: Predicted Number of Potholes in March 2017.
[7] A. Allouch, A. Koubâa, T. Abbes, and A. Ammar, “Roadsense: Smart-
phone application to estimate road conditions using accelerometer and
V. C ONCLUSION gyroscope,” IEEE Sensors Journal, vol. 17, no. 13, pp. 4231–4238, 2017.
[8] S. Nienaber, M. T. Booysen, and R. Kroon, “Detecting potholes using
This paper identifies and addresses two important issues- the simple image processing techniques and real-world footage,” in Southern
detection of potholes, and redressal of these complaints by the African Transport Conference, 07 2015.
[9] A. Dhiman and R. Klette, “Pothole detection using computer vision
civic authorities. Using data from accelerometer and gyroscope and learning,” IEEE Transactions on Intelligent Transportation Systems,
sensors classification was performed and compared with five pp. 1–15, 2019.
different algorithms: Naive Bayes, Decision Tree, Random [10] S. Omanovic, E. Buza, and A. Huseinovic, “Pothole detection with
image processing and spectral clustering,” 10 2013.
Forest, Logistic Regression and K-Nearest Neighbor Classifier. [11] R. Fan, U. Ozgunalp, B. Hosking, M. Liu, and I. Pitas, “Pothole
Upon testing, it was revealed that the Random Forest algorithm detection based on disparity transformation and road surface modeling,”
performed the best out of the five, with an accuracy of 89.63%. IEEE Transactions on Image Processing, vol. 29, pp. 897–908, 2020.
[12] T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Dollár, “Focal loss for
For image-based pothole detection, the RetinaNet model was dense object detection,” CoRR, vol. abs/1708.02002, 2017.
able to record a precision rate of 83%. [13] A. Vora, L. Reznik, and I. Khokhlov, “Mobile road pothole classification
Secondly, to ensure timely redressal of the road safety and reporting with data quality estimates,” in 2018 Fourth International
Conference on Mobile and Secure Services (MobiSecServ), pp. 1–6,
grievances, a web portal is designed for the authorities to 2018.
improve communication with the citizens. The dashboard [14] E. Afify and M. Kadry, “Electronic-customer complaint management
maintains a list of all the complaints and prioritizes them based system (e-ccms) – a generic approach,” International Journal of Ad-
vanced Networking and Applications, vol. 11, pp. 4125–4141, 07 2019.
on the criticality. It also allows the authorities to update the [15] V. Mareeswari and V. Gopalakrishnan, “Complaint go: an online com-
status of complaints to help the users get information on the plaint registration system using web services and android,” IOP Confer-
resolution of complaints. Prediction of the number of potholes ence Series: Materials Science and Engineering, vol. 263, p. 042073,
11 2017.
is another important feature that helps the authorities, as well [16] D. Radhakrishnan, N. Gandhewar, R. Narnaware, P. Pagade, A. Tiwari,
as the users, be aware of the areas populated by how many and P. Vijaywargi, “Smart complaint management system,” in Interna-
potholes. This is performed using the SARIMA model which tional Journal of Trend in Research and Development, 2016.
gave an accuracy of 85.7%.
Therefore, the combination of the two solutions provides an
efficient and an effective system which takes care of all the
detected potholes with the involvement of higher authorities.