Pitfree: Pot-holes detection on Indian Roads using

Mobile Sensors
Gaurav Singal, Anurag Goswami, Suneet Gupta, Tejalal Choudhary
Department of Computer Science & Engineering
Bennett University, India
gauravsingal789@gmail.com, anurag.goswami@bennett.edu.in, suneet.gupta@bennett.edu.in, tejalal.choudhary@gmail.com

Abstract— Pot-holes on road will make transportation Among different road anomalies, Potholes, shown in Fig.
slower and costly. India has a big network of roads to connect 1. remains as a major cause of road accidents. According to
the villages and cities, the authority persons cannot travel Road accidents statistics (2013-16) by Ministry of road
across the region for identification of holes. As per transport and Highways [5], more than 11,386 people were
advancement in machine learning in recent time, we can use killed due to Pothole related accidents. More than 45% of
this technology for the identification and patching the pot- road accidents are caused by potholes. The major reason
holes. As per the recent survey around 400millions, people behind these accidents is due to ignorance and poor
have a smartphone in India. We can use smartphone sensors maintenance of roadways by concerned authorities.
(such as Accelerometer and gyroscope) to identify the pot-holes
on road and GPS for the location of the pit. The major task of Hence, there’s a need for infrastructure/system that can
this problem is to capture the data and annotation. We have detect defective roads (especially potholes in this case) and
developed an android app for capturing the value of inform the authorities to take quick action for its repair. The
displacement while travelling on road. We have applied overall result will lead to a reduction of road accidents and
different classification algorithms to sensor raw-data. SVM is save human lives on Indian roads along with the reduction of
the most suitable classification technique for this problem. The the maintenance cost. We were motivated to develop a
android app will sound an alarm when a pothole is detected. mobile app that will identify potholes on the road by utilizing
accelerometer and GPS of the Smartphone placed on the
Keywords— Machine Learning, Mobile Sensors, Road
Safety, Classification
vehicle. The app will process this data using the designed
algorithm and transfer the collected data (in some time
intervals) to the database via the Internet. The data then will
I. INTRODUCTION be utilized to classify pits into a different degree (low,
In the developing country like India, various medium, high) which can be used by authorities to identify
transportation mediums are present (e.g., road, airlines, and to provide necessary maintenance to the roads.
railways). From all the available options, road transportation We have run a number of classification algorithms on
is the preferred mode as it covers 75% of the total captured sensor data to check the accuracy, SVM is
transportation system. Since 2001-2015, 2098624.255 providing 99.6%. We have shown the distribution of raw
kilometers (km) were added towards total road length in data by K-Means algorithm for clustering into potholes and
India with the Compound Annual Growth Rate (CAGR) of others. The current results show that our approach can
3.52% in total road length in India [1]. Data shows a differentiate between potholes and bump.
significant increase in road traffic injuries in India [2]. Due
to the lack of proper infrastructure to perform maintenance, Section II describes the background work and brief
in the year 2016, 150,785 people died in 480,652 road description of the terms used in this paper. Experiment
accidents. It was also found that over the past 10 years, road design is detailed in Section III. Section IV elucidates the
traffic injuries deaths increased by 43% [3]. It is also results of the experiment, and Section V describes the
observed from the recent data that as many as 1.3 lakh conclusion and future work of the application developed.
people have died and 3.8 million are seriously disabled for
their entire life [4]. Timely road maintenance leads to a II. BACKGROUND
reduction of costs and increases the lifetime along with the
reduction of accidents. This section describes the overview of machine learning
and classification techniques along with the related work that
utilizes a mobile sensor for detecting pot-holes.

A. Machine learning and classification

Machine Learning has been evolving at a spectacular
rate. With the help of high computation and memory
capacity, it has been possible to solve problems with ease.
E.g., Google Search Engine heavily depends on Machine
Learning, which makes it faster, reliable and makes search
prediction based on our use. It can be applied to solve many
problems in the society.
Machine Learning is a part of Artificial Intelligence. It is
a process by which the computer learns to identify by itself,
Fig. 1. Pot-holes with help of respective data, without explicitly writing the
code. It has high statistical computations which use
prediction models to find out after identifying patterns in the

978-1-5386-6678-4/18/$31.00 2018
c IEEE 185
 data. Thus, it’s is now used in every task. Data Scientists, C. Related Work
who deal with data and their manipulation heavily use From the research paper “The Pothole Patrol: Using a
Machine Learning algorithms. Mobile Sensor Network for Road Surface Monitoring” by
Classification technique is an approach through which we Jakob Eriksson, et al. [9] we were able to understand
can identify the category/class of the data [6]. For example: parameters like speed, Z-peak, x-z ration and speed vs. z-
Man/Woman, Animal/Plant, Car/Bike (binary classification), ratio which are used to classify potholes from other road
Bus/Train/Truck/Tractor (multiclass classification) etc. This anomalies.
kind of algorithms are widely used in our daily lives where The research paper “Pothole Detection and Warning
humans can classify one from the other. This same process is System using Wireless Sensor Networks” by Sudarshan S
followed by the computers, where they are intelligently able Rode, et al. [10] also proposes a similar system of using
to classify it. accelerometer and GPS data to detect potholes. The sensor
readings from accelerometer and GPS are sent to a server
B. Machine Learning Classifiers and this is further processed to classify a pothole from other
There are number of classifiers that exist for the road anomalies. The research paper “Pothole Detection
classification problem, we have experimented with the System using Machine Learning on Android” by Aniket
following classifier to examine the best classifier on our Kulkarni, et al. [11] detects potholes in real time using
sensor data: accelerometer sensors and GPS and these detected potholes
x Logistic Regression [7]: It is a statistical model, that are plotted on Google Map.
works with binary identifiers for calculating the Various models that may be utilized to detect road
parameters of a logical model. conditions via mobile sensors are described briefly below:
x K Nearest Neighbour’s [7]: It is the basic x Cartel System
classification and regression algorithm among all the This system proposed by Bychkovsky, et al. [12] which
machine learning algorithms that work on local uses object detection sensors and GPS to detect the road
approximation. anomalies and send the corresponding data to a server.
x Random Forest [8]: It is the most useful and accurate x Pothole Patrol System
Machine Learning algorithm without carrying out The system [9] which uses a tri-axial accelerometer and a
any hyper-parameters tuning. Random forest first GPS mounted on the dashboard and using this accelerometer
constructs multiple decision trees and then combine data we can identify potholes and other road anomalies using
for better accuracy and prediction. machine learning algorithms.
x Naïve Bias [7]: It is an algorithm of probabilistic
classifiers group that depends on the Bayes theorem
Reading of Accelerometer Sensor Data
and works on independent parameters. S. No. X-Axis Y-Axis Z-Axis
x Support Vector Machine (SVM) [7] 1 0.17 4.383 6.828
These classifier algorithms are well known. It’s a 2 0.755 4.446 9.573
supervised learning model associated that analyze the data 3 -0.575 3.435 8.573
used for classification and regression. Given a set of training 4 -1.351 3.007 8.619
examples, each marked as belonging to one or the other of
two categories, an SVM training algorithm builds a model 5 -5.319 3.667 7.751
that assigns new examples to one category or the other, 6 -1.951 3.834 10.086
making it a non-probabilistic binary linear classification.
7 -2.667 0.965 9.829
An SVM model is a representation of the examples as
points in space, mapped so that the examples of the separate 8 -3.513 4.648 7.269
categories are divided by a clear gap that is as wide as 9 -1.604 6.18 8.538
possible. New examples are then mapped into that same 10 -0.55 5.213 7.795
space and predicted to belong to a category based on which
side of the gap they fall. 11 -1.386 5.781 7.085
Thus, for the problem we identified, we see that the 12 -0.8 6.949 5.679
Machine Learning approach can help us in detecting the 13 -2.117 10.502 5.422
potholes intelligently. Machine Learning techniques can help
us finding a pattern and making the computer learn by itself.
14 0.657 7.673 7.673
This pattern will be analyzed by the model and will be able 15 1.319 8.999 5.851
to predict if the Road Anomaly is a pothole. In turn, this can 16 0.696 7.695 3.683
help us in identifying the locations where a pothole has
occurred and thus, it can automatically store the Geo 17 -1.785 7.684 0.441
Location of that spot. Real-Time data can be tracked by the 18 2.372 10.735 4.824
authorities without and manual inspections and road
conditions. It can be deployed on smartphones, with the help 19 2.849 9.423 5.818
of TensorFlow libraries supported in Android OS. 20 3.341 9.58 4.178
Fig. 2. Dataset captured from the sensor.

186 8th International Advance Computing Conference (IACC).

 Fig. 3. Graphs (taken as a screenshot) from accelerometer representing (A) Normal Anomalies, (B) Speed
Breaker, (C) Pot-hole

x RCM-GPS System
A. Data Collection
Proposed by Chen, et al. [13] which also uses an
accelerometer and a GPS which calculates current time, With an objective to detect pot-holes via mobile sensors,
location, velocity, and three direction accelerations. This data the data was collected through an Android App. To achieve
is processed and can be used to detect road anomalies. this objective, an android app was developed with
functionalities to start and stop the data collection and save
x Nericell system the data in the Android Device. For this, we save it in a file
This system, as Proposed by Mohan, et al. [14] detects in ‘CSV’ format. The readings are captured across X-axis,
potholes, braking, bumps, and honks using the Y-axis and Z-axis of the accelerometer sensor. Depending on
accelerometer, microphone, GSM radio and GPS sensors of the orientation, we would get the gravitational force
smartphones. It uses triggered sensing mechanism where a acceleration on either of these three axes. Along with the
high energy consuming sensor (GPS, microphone) is readings, we also take into account GPS readings of Latitude
activated by a low energy consuming sensor (accelerometer, and Longitude. This would help us in identifying the location
cellular radio) making the system energy-efficient. This of the potholes.
sensed data is sent to the server and is processed to detect
and classify different road anomalies. In Fig 3. we are presenting the subset of the captured
dataset with the help of accelerometer sensor using our
x Mednis System Android App on a mobile phone during the travel. The
Proposed by Mednis, et al [15] which uses the 3-axis sensor provided us with the three-dimensional motion of our
accelerometer sensors present in Smart-phones. This system device in terms of X, Y, and Z direction. The changes in the
contains four algorithms for detection of potholes. The first reading will predict the pothole on road.
two algorithms (ZTHRESH and Z-DIFF) are for real-time
detection and the other two (STDEV(Z) and G-ZERO) are B. Implementation
used for off-line post-processing of data.
Annotations is a tough task in this project after data
collection. A small script was written for labelling the data in
III. METHODOLOGY AND EXPERIMENT DESIGN the training part. K-Means clustering [16] was used to
This section elucidates the data collection, flow-chart and identify the patterns shown in Fig. 3. where the x-axis
implementation of the technique to detect pot-holes. denotes time in milli-seconds and y-axis represents the

8th International Advance Computing Conference (IACC). 187

 vertical displacement of the mobile phone. Basically, this x If the readings for the vertical axis (i.e. one with the
algorithm helps us to cluster the data into two groups of gravitational force) is between 8.2 and 12.m/s2 then
potholes and non-potholes. In Fig. 3. we are showing the the road doesn’t have any road anomalies.
graph (screenshots) of an android app that was used for
collecting the data and different type of accelerometer x If the reading suddenly dips, we check for how long
readings. (A) shows the normal anomalies in the road as the the data had a dip. Then the readings again increase.
graph shows continuous variation. From (B), it can be If the readings keep increasing in the next 10
observed that whenever there’s a speed breaker, the signal readings after the dip and are back to normal by the
(i.e. graph) will be up first and then down in the continuous 10th reading, then it’s a pothole. Otherwise, it isn’t.
signal recording. Along with that; (C) shows the graph, For measuring that, we keep taking the difference
where accelerometer readings for pot-holes have dipped all between the current and the next and check the
of sudden and then rise. We have to identify the threshold of above condition.
dipped and rise value for predicting the potholes. For x In this manner, we confirm that it’s a pothole. We
identification of potholes, we have written a script that collected the data by keeping mobile on the
examines the continuous sensors data. dashboard of a car.
In the pothole section, as we have seen in the graph, the After collected the readings, we run the script on the
accelerometer readings have dipped suddenly decrease and saved file taken from Android Application and labelled it.
then increase as shown below: We have used multiple Machine Learning models (Section
Next, a script was developed with the following criteria: II.B) to train the collected data. Before training the data set
on the models, we are applying K-Means Clustering in order

Fig. 4. Representation of the process followed to achieve maximum accuracy to detect pot-holes via mobile
application

188 8th International Advance Computing Conference (IACC).

to confirm the pattern on Road Anomalies and verify our probability of having bad road conditions. This will help in
script labelling. Fig. 4. shows a graphical representation of timely action. There are several advantages as listed below:
data in the form of clusters. We use Python’s Scikit Machine
Learning library and ran it on Jupyter Notebook. x The process is automated. Thus, it’s cost-effective
and it will enable the Government for swift actions.
C. Flow-chart x This will help in timely action. If the road
In Fig. 4. we are representing the flow of the complete conditions start deteriorating, timely action can be
project. First, we have created an application for detection taken before the pothole becomes too dangerous for
of pot, side by side collection of data. We have built two vehicles.
applications one for data collection and second for x This will ensure the safety of the public. Potholes
implementing the algorithm to get recognized the potholes may lead to accidents and such cases can be easily
and send the notification to the respective authorities. K- avoided, thus saving human lives.
Means clustering applied onto the dataset to see the
distribution of data. After that, we applied various x The government can have real-time data of the road
classification algorithm to predict the potholes on testing quality and time at which the roads start
deteriorating. This will enable the Government
data. SVM providing the best accuracy among all the
Authorities to know if the constructions standards
classification problem.
have met the specific requirements.
IV. RESULTS An efficient solution towards road safety is the need of
To detect pot-holes with maximum accuracy, we run tests the developing country like India. Thus, embracing and
using various classifiers as stated in Section II B. The overall involving new technologies which help solve problems will
implementation process can be depicted in Fig. 4. The test surely help us to lead a better and safe life. With timely
result for each classifier is stated in Table I below: actions, it can help us in saving lives. The system is limited
to classify major potholes. Minor potholes cannot be easily
identified. While our system can classify potholes with other
Table I. Accuracy Results Derived from Various Classifiers road anomalies. We were not able to classify speed breakers.

S. Confusion Future improvements include improving the efficiency of

Classifier Accuracy our system on real-time by considering the gyroscope
No. Matrix
[[244 0] readings so that the real-time tracking of potholes can work
1 Logistic Regression 88.8% in any orientation of the mobile phone. Another future work
[ 3 3]]
[[243 1] includes pushing the data from our App to server in real time
2 K-Nearest Neighbours 98% that can implement Machine and Deep Learning algorithms
[ 1 5]]
to plot them in google maps.
Support Vector [[242 2]
3 99.6%
Machine [ 2 4]]
[[237 3] ACKNOWLEDGMENT
4 Naïve Bias 98.4%
[ 1 9]] We would like to thank Maneesh Singh Bhakuni, Allen T
[[240 0] Abraham, Prashanth Duvvada, and Naman Jain who helped
5 Random Forest 89.4
[ 2 8]] us with the development of the tool.

In the table, confusion matrix with an accuracy of each REFERENCES

classifier is shown. Total 250 samples have been tested with [1] “Total road length in India from 2001 to 2015,”
each classifier. The results from various classifiers showed community.data.gov.in, 2017. [Online]. Available:
that SVM has the highest accuracy (99.6%) compare with
https://community.data.gov.in/total-road-length-in-india-from-
other techniques to detect pot-holes. Our Android app will
2001-to-2015/.
notify the potholes during the travel by a warning
sound/tones. [2] L. Dandona, Y. S. Sivan, M. N. Jyothi, V. S. U. Bhaskar, and R.
Dandona, “The lack of public health research output from India.,”
V. CONCLUSION AND FUTURE WORK BMC Public Health, vol. 4, p. 55, 2004.

We want our technology to truly improve road conditions [3] E. S, R. AB, R. A, R. M, R. S, and A. M, “An Analysis of Road
and enhance people’s lives. In the recent times, increase in Traffic Injuries in India from 2013 to 2016: A Review Article,” J.
smartphone usage has enabled us to use the mobile sensors Community Med. Health Educ., vol. 8, no. 2, 2018.
and their GPS facility to keep track of Road Conditions. [4] J. K.P. and Parashuramlu, “An epidemiological study of road
Machine Learning techniques have provided us with traffic accident (RTA) cases admitted in a tertiary care hospital -
promising results, with an average accuracy of over 90%, we a retrospective study,” Indian J. Public Heal. Res. Dev., vol. 8,
are able to achieve and keep track of Road conditions no. 3, pp. 364–368, 2017.
without any human intervention and manual inspections.
Android is the most used OS, especially in India. Thus, [5] “Statistics of Road Accidents in India From 2013 to 2016.”
without any much effort to use the app and less burden on [Online]. Available: http://morth.nic.in/.
networks, we are able to detect potholes and help the [6] R. S. Michalski, J. G. Carbonell, and T. M. Mitchell, Machine
Government officials and Road Authorities to keep track of learning: An artificial intelligence approach. Springer Science &
the data. They have Geolocation of Potholes and several user Business Media, 2013.
sensors detecting the potholes are sure enough to have a

8th International Advance Computing Conference (IACC). 189

[7] Y. Zhang et al., “Empirical study of seven data mining [12] B. Hull et al., “CarTel : A Distributed Mobile Sensor Computing
algorithms on different characteristics of datasets for biomedical System,” 4th ACM Conf. Embed. Networked Sens. Syst., pp. 125–
classification applications,” Biomed. Eng. Online, vol. 16, no. 1, 138, 2006.
2017. [13] K. Chen, M. Lu, X. Fan, M. Wei, and J. Wu, “Road condition
[8] N. Donges, “Randon Forest.” [Online]. Available: monitoring using on-board three-axis accelerometer and GPS
https://towardsdatascience.com/the-random-forest-algorithm- sensor,” in Proceedings of the 2011 6th International ICST
d457d499ffcd. [Accessed: 27-Jul-2018]. Conference on Communications and Networking in China,
[9] J. Eriksson, L. Girod, B. Hull, R. Newton, S. Madden, and H. CHINACOM 2011, 2011, pp. 1032–1037.
Balakrishnan, “The pothole patrol: using a mobile sensor network [14] P. Mohan, V. N. Padmanabhan, and R. Ramjee, “Nericell: rich
for road surface monitoring,” Proceeding 6th Int. Conf. Mob. monitoring of road and traffic conditions using mobile
Syst. Appl. Serv. - MobiSys ’08, p. 29, 2008. smartphones,” in Proceedings of the 6th ACM conference on
[10] S. S. Rode, S. Vijay, P. Goyal, P. Kulkarni, and K. Arya, Embedded network sensor systems, 2008, pp. 323–336.
“Pothole Detection and Warning System using Wireless Sensor [15] A. Mednis, G. Strazdins, R. Zviedris, G. Kanonirs, and L. Selavo,
Networks,” Embed. Real-Time Syst. Lab. Indian Inst. Technol. “Real time pothole detection using Android smartphones with
Bombay, 2007. accelerometers,” in 2011 International Conference on Distributed
[11] A. Kulkarni, N. Mhalgi, S. Gurnani, and N. Giri, “Pothole Computing in Sensor Systems and Workshops, DCOSS’11, 2011.
Detection System using Machine Learning on Android,” Int. J. [16] J. A. Hartigan and M. A. Wong, “A K-Means Clustering
Emerg. Technol. Adv. Eng., vol. 4, no. 7, pp. 360–364, 2014. Algorithm,” Appl. Stat., vol. 28, no. 1, pp. 100–108, 1979.

190 8th International Advance Computing Conference (IACC).