SOCIAL DISTANCE ANALYZER

“Whatever You Want To Say, Say It From 6 Feet Away”

A Project Report submitted

for the partial fulfilment of the degree of

Bachelor of Technology
In
Information Technology

By

Swarnim Yadav 1805213059

Priya Adarsh 1805213041
Alankrita Singh 1805213007

Under the guidance of

Prof. Girish Chandra Dr. Tulika Narang

Department of Computer Science and Engineering

Institute of Engineering and Technology, Lucknow
Dr. A.P.J. Abdul Kalam Technical University, Lucknow, Uttar Pradesh
 Contents
S.No Content Page No.

1 Declaration i

2 Certificate ii

3 Acknowledgement iii

4 Abstract iv

5 List of figures v-vi

6 List of tables vii

7 List of abbreviations viii

8 Introduction 1-2

9 Literature review 3-4

• Related works
• State of the art
10 Methodology 5-9
• Workflow
• Technology stack
11 Experimental results 10-25

12 Conclusions 26

13 Future works 27

14 References 28-30

15 Annexure 31-33
 Declaration

We hereby declare that this submission is our own work and that, to the best of our belief and

knowledge, it contains no material previously published or written by another person or

material which to a substantial error has been accepted for the award of any degree or diploma

of university or other institute of higher learning, except where the acknowledgement has been

made in the text. The project has not been submitted by us at any other institute for requirement

of any other degree.

Submitted by: - Date:

1. Swarnim Yadav
1805213059
Information Technology

2. Priya Adarsh
1805213041
Information Technology

3. Alankrita Singh
1805213007
Information Technology

i
 Certificate

This is to certify that the project report entitled “Social Distance Analyser” presented by

Swarnim Yadav, Priya Adarsh, Alankrita Singh in the partial fulfilment for the award of

Bachelor of Technology in Computer Science and Engineering, is a record of work carried out

by them under my supervision and guidance at the Department of Computer Science and

Engineering at Institute of Engineering and Technology, Lucknow. It is also certified that this

project has not been submitted at any other Institute for the award of any other degrees to the

best of my knowledge.

Prof. Girish Chandra

Department of Computer Science and Engineering

Institute of Engineering and Technology, Lucknow

Dr. Tulika Narang

Department of Computer Science and Engineering

Institute of Engineering and Technology, Lucknow

ii
 Acknowledgement

On the very outset of this report, we would like to extend our sincere & heartfelt gratitude

towards all luminaries who have helped us directly or indirectly in this project.

We are ineffably thankful and pay our sincere gratitude to our supervisors Prof. Girish

Chandra, CSE department and Dr. Tulika Narang, CSE department for their helpful

information, keen interest, valuable guidance, constructive criticism, and insightful advice at

various stages of our training period.

We would like to thank Dr. D. S. Yadav, the head of department, Dr. Promila Bahadur, the

project coordinator and the project monitoring committee members for delivering the

guidelines and organizing the online presentations composedly with time and ease.

Finally, we would like to wind up by paying our heartfelt thanks to our supportive family and

friends for motivating us and putting out their idea for the project.

iii
 Abstract

In this report, a model is proposed that aims to track social distancing among people by

calculating the physical distance among every pair of pedestrians. Based on this calculated

distance, the model classifies them as violator that in turn helps to mitigate the impact of the

pandemic in the area of interest. The proposed model uses a deep learning object detection

algorithm, YOLOv3 that helps in detecting the pedestrians and the detection results helps to

evaluate the distance between the pedestrians. The distance between any pair of pedestrians in

the display is estimated by using Euclidean distance and the pair which is non-compliant with

the set parameters is indicated with a red frame and the pair who is on the verge of violating

the parameters of social distancing is indicated with a yellow frame and the pair which is well

within the set parameters of the same is indicated with a green frame. The proposed model was

corroborated on a pre-recorded video. The result exhibits that the model can be used for

monitoring whether the social distancing is being followed or not. The model can be further

developed for real time monitoring of the pedestrians by providing camera URL.

iv
 List of Figures

Fig 1.1 Social distance

Fig 3.1 YOLO algorithm depiction

Fig 3.2 Flowchart of SDA

Fig 4.1 Threshold limit

Fig 4.2 Code snippet for threshold value

Fig 4.3 Code snippet for parameters

Fig 4.4 Use Case for Human count

Fig 4.5 Code snippet for turning on human count feature

Fig 4.6.1 Code snippet for human count function

Fig 4.6.2 Code snippet for human count function

Fig 4.7 Alert function and abnormal violations

Fig 4.8 Code snippet for calculation of violations

Fig 4.9 Overlapping camera view

Fig 4.10 Region of interest

Fig 4.11 Unit length for ROI

Fig 4.12 Bird eye view

Fig 4.13 Output frames per second

Fig 4.14 Safe and unsafe distance

Fig 4.15 Code snippet for email alert feature

v
Fig 4.16 Email alert function

Fig 4.17 Email receiver function

Fig 4.18 Code snippet for providing camera URL

Fig 4.19 Code snippet for grabbing camera URL/video file

Fig 4.20 Threading parameter

Fig 4.21 Threading function

Fig 4.22 Analysis in college campus

vi
 List Of Tables

Table 1 Task categorization

vii
 List Of Abbreviations

YOLO You only look once

ROI Region of Interest

SDA Social Distance Analyzer

AI Artificial Intelligence

DL Deep Learning

viii
 CHAPTER-1

Introduction
In today’s era, digital technologies such as surveillance of population, contact tracking

and evaluation of interventions based on mobility data, case identification etc are being

utilized to support the public-health response to COVID-19 worldwide [1]. In future,

the public health is likely to shift towards these digital technologies and these will be

used to strengthen pandemic management and help in preparing for COVID-19 and

other contagious diseases.

In the time of outbreaks such as COVID-19, research and development on new

methods and technologies or updating existing methods is done to empower the core

capacities [2]. A deeper understanding of the epidemic suggests that in order to reduce

the risk of virus spread can be minimized by increasing the physical distance or

avoiding the physical contact among the people. Effective technologies and

comprehensive tools can be used to implement social distancing. If social distancing

is enforced at the initial stages, it can perform a vital role in overcoming the spread of

the virus and preventing the peak of the pandemic.

Artificial Intelligence, a core technology that came into existence during the fourth

industrial revolution, plays an important non-medical intervention in overcoming the

current global health crisis. AI helps in building next-generation epidemic

preparedness, and in moving towards a flexible recovery. In the area of computer

vision and machine learning, different algorithms have been developed for object

1
detection. The same methods can also be used to detect the physical distance between

people.

The following points summarize the main components of this approach:

1. For object detection, there has been a shift towards using deep learning technologies.
the same technique is used for the purpose of human detection.

2. A social distancing detection model is developed that will detect the distance between
the pedestrians to analyse safe distance.

3. Evaluating of the result of classification by analysing real-time video streams by

providing the camera [3].

The main purpose of this project is to track the social distance by providing a deep

learning platform. To accomplish this objective of social distance monitoring, we will

be using a framework, YOLOv3 to identify humans in video sequences. The proposed

system is favourable because it is automated and helps in reducing the manpower

required to physically inspect the areas where the people may not be following social

distancing norms.

Fig 1.1 Social distance

2
 CHAPTER-2

Literature Review
This chapter mentions about the sources from where the ideas and references have been drawn

to come up with this project.

2.1 RELATED WORKS

2.1.1 Works on social distance analyser

Numerous amounts of work have been done in the area of detection for objects and person

using various techniques of deep learning. Real-time object detection algorithms that use model

of Convolutional Neural Networks (CNN) like You Only Look Once (YOLO) and Region

based Convolutional Neural Networks (RCNN) for detecting multiple classes in several

regions was developed by G.V. Shalini et al[2].

Latest improvements in the area of deep learning facilitates the object detection in a more

effective way. Researchers frequently utilize these methods to track physical distancing among

pedestrians in the dynamic frames.

2.2 STATE OF THE ART

2.2.1 Real time object detection using deep learning and OpenCV

The work proposed by Chandan G, Ayush Jain, Harsh Jain, Mohana, aimed at achieving

detection of objects in real time. A real time video that was taken from a webcam was given as

an input for the object detection. The input video was pre- processed. The output of pre-

processing was then sent to the module for detection where a series of operations were carried

out for generating the result. The work will help in the future aspect for real time object

detection [5].
3
2.2.2 Person detection for social distancing and safety violation alert

The work by Afiq Harith Ahamad, Norliza Zaini, Mohd Fuad Abdul Latip, proposed a way for

the implementation of two features using python language and OpenCV library. The first

feature was to detect social distancing violations and the second feature was to detect violations

if a person enters in restricted areas. When different tests were performed on this approach, the

results suggested that the object detection model was not having high accuracy and was facing

issues in detecting people correctly in the public areas. The work helped us in getting to know

about the YOLOv3 algorithm and with this; it also helped us in defining the region of interest

parameter [6].

2.2.3 Deep learning based safe social distancing and face mask detection in public

areas

The work by Shashi Yadav proposed a system that made use of raspberry pi4, computer vision

and Mobile Net V2 that helped the local police officials in the surveillance of public areas. In

this system the camera feed real time videos and detects whether a pedestrian is equipped with

a mask or not. The work will help in implementing our proposed model using raspberry pi in

future [7].

2.2.4 Object detection algorithms for video surveillance

The work proposed by Apoorva Raghunandan; Mohana; Pakala Raghav; H.V. Ravish Aradhya,

implements object detection, colour detection and skin detection in a video running in real time.

It also comprises of background and foreground image processing model for object detection

in real time. The work helped in determing the method to video as an input for the detection

algorithm [8].

4
 CHAPTER-3

Methodology
This chapter describes the steps that have been taken to construct and implement the model and
analyse the final result.

3.1 WORKFLOW

In this work, we have used a Deep Learning model for object detection to detect people along
with an algorithm for classifying the social distance based on the provided input video. Initially,
we have used an open-source object detection algorithm, YOLO for the detection of the
pedestrians in the input video frame. Detector may detect several objects like pedestrians, trees,
cars, etc but we have used only the pedestrian class and ignored the rest classes in this proposed
model. The bounding boxes are fitted for each pedestrian that is detected and the data of these
detected pedestrian is used for measuring the distance between them.

The tasks were divided in the following categories-

1. Environment Configuration

2. Data Acquisition

3. Applying YOLO algorithm

4. Generating Result

Table 1 Task Categorization

5
3.1.1 Environment Configuration

The program is developed using python3 language. The model is implemented

in Jupyter notebook.

3.1.2 Data Acquisition

Pre-recorded videos were downloaded from the internet in order to train the

model. We have also recorded some videos with the help of smart phone to test

the model.

3.1.3 Applying YOLO Algorithm

On providing the input, the YOLO algorithm divides the in input into grids, say SxS.

The features are extracted from each grid. The output contains the bounding boxes

along with the confidence scores for the classes that are predicted in the bounding

boxes. The model identifies the people using the obtained bounding box information.

The distance between each pair of bounding boxes is determined by measuring the

Euclidean distance between the bounding box’s centroid. In order to estimate the social

distance violations between the people, an approximate threshold limit has been

determined. The threshold limit has been set to check whether there has been a breach

in the minimum social distance threshold.

6
 The steps taken in YOLO algorithm are –

• Video is taken as an input

• The input frame is divided into grids

• Image classification in applied on each grid.

• Bounding boxes and class probabilities are predicted to identify the type of

object.

Fig. 3.1 YOLO algorithm depiction

3.1.4 Generating Result

To get the output videos, we have run the code in Window PowerShell.

7
 3.2 TECHNOLOGY STACK

To implement this application, following technologies have been used

3.2.1 Python
We have used Python version3 as the centre programming language for this project as
it has support for Artificial Intelligence and various supporting libraries.

3.2.2 OpenCV

OpenCV is an open–source library which is used for machine learning, image

processing and computer vision. By using this, we can easily detect objects, humans,
etc. We have used cv2 to grab the video file if reference path is provided otherwise, to
grab reference to the camera.

3.2.3 NumPy

This library consists of multi-dimensional arrays and matrix data structures. With the
help of Numpy, we are able to achieve array operations in less space complexity. In this
project, we have used Numpy library to extract the centroid details.

3.2.4 Argparse

With the help of argparse, we can create a command-line environment in order to

improve the interaction. It generates help methods and issues if wrong arguments are
provided.

3.2.5 Scipy

With the help of this library, we have imported the distance computations.

8
Program flowchart for social distancing detection in each video frame:

Fig. 3.2 Flowchart of SDA

9
 CHAPTER-4

Experimental Results
In this Chapter, we have provided the detailed description of each output snapshot and code
snippet along with all the use cases of the project.

Fig 4.1 Threshold limit

• The above figure shows the snapshot of output video result on the pre-recorded input
video. In the bottom right corner of the snippet (Highlighted with a border) are the
parameters on which we are calculating distance between pedestrians and displaying
them with a red frame when they are under the set parameter of safe distance i.e., 80px,
and if they are beyond the set parameter of safe distance i.e., 80px they are represented
with a green frame.

10
 Fig 4.2 Code snippet for threshold value

• The above code snippet has a code for the parameters that has been written in config
module which here states about the parameters of threshold that has been set to 12 here
which signifies that atmost 12 violations can be tolerated with a parameter of maximum
distance of 80px which has been written at the bottom in the above snippet.

11
 Fig 4.3 Code snippet for parameters

• The above code snippet accepts the arguments for safe distance in the main python
file that has been sent from config module as MAX_DISTANCE. This code snippet
also accepts the threshold value whose value can be set by the user and is also sent
from config module as a argument to the above function.

Fig 4.4 UseCase for human count

• As one of the usecase of this project is to calculate the human count in the video. In the
above snapshot of the video as we can see total human count is 5 which is perfectly
accurate as per the video snapshot provided above. This feature can be used to count
no. of people in a gathering, to avoid stampede, to be in compliance of the curfew act,
to take attendance of no. of people and many more such cases.

12
 Fig 4.5 Code snippet for turning on human count feature

• The Human count feature is completely optional and can be turn off/on in the config
module under People counter variable as true or false. The has value has been imported
to the detection python file under mylib folder.

13
 Fig 4.6.1 Count snippet for human count function

Fig 4.6.2 Code snippet for human count function

• The above code is written under mylib -> detection python file to detect people and
count the no. of people.
14
 Fig 4.7 Alert function and abnormal violations

• The above snippet of the output shows the serious violations (as red frame) , abnormal
violations(as yellow frame)and the safe zone(as green frame).The set parameters of the
serious and abnormal violations are mentioned in the bottom left corner of the output
snippet.

Fig 4.8 Code snippet for calculation of violations

15
• The above code snippet states the code for the total no. of social distancing violations
on the output frame.

Fig 4.9 Overlapping camera view

• One of the use cases of our project is to identify the violations on a prerecorded video
to defeat the aspect of overlapping video angles. We enhance the project by restricting
the area of visualization (Region of interest) which is set by the user itself as per its
discretion.

16
 Fig 4.10 Region of interest

• To achieve the restricted field of visualization (Region of interest) which is set by the
user, the user draws four points(red) which is termed as region of interest (ROI) here.

Fig 4.11 Unit length for ROI

17
• In the above snapshot the three blue dots which are parallel to the lines drawn for the
region of interest indicates unit length in the real world.

Fig 4.12 Bird eye view

• The above snapshot represents the bird eye view which is actually the output video
frame received by providing the ROI and unit length points.

18
 Fig 4.13 Output frames per second

• The above snapshot is of the output folder of the project which captures and store the
output frames per second, and it continues to store the output frames until the process
is killed.

19
 Fig 4.14 Safe and unsafe distance

• In the above snapshot is captured from the output folder of project, the line determines
that the person 1 and person 3 are at a reasonably safe distance from each other. The
red lines represent that person 1, person 2 and person 2, person 3 are at an unsafe
distance from each other.

20
 Fig 4.15 Code snippet for email alert feature

• The above code snippet is taken from the config file in which we have provided the
option for turning on the alert function to true/false, whose value is further exported to
the mailer file and used as per the requirements.

21
 Fig 4.16 Email alert function

• In the above code snippet, the script initiates the email alert function. This contains two
functions ,one which contain the sender email along with its authorization provided 2
step verification is not turned on .

Fig 4.17 Email receiver function

• The above code snippet has a function which ensures that the administration will
receive an email in case if the violations are beyond the set parameters of threshold.

22
 Fig 4.18 Code snippet for providing camera URL

• As a future aspect of our project, we aim to bring much more of the scalability. As the
result of which we aspire this project to be a real time analyser this we have put forward
our step and provided the camera URL variable in which we can provide our camera
URL for the inference.

Fig 4.19 Code snippet for grabbing camera URL/Video fie

• In the above code snippet, we have used if-else statement in which if the video path
reference is not supplied then the code will grab the reference to the camera otherwise
it will grab a reference to the video file.

23
 Fig 4.20 Threading parameter
• The above code snippet provides user the flexibility to turn on the threading, if user
experience lag or delay in the stream of the output video frame, by assigning value to
thread parameter as true or false.

Fig 4.21 Threading function

24
• The above code snippet initiates threading class, which read the frames as soon as they
are available and discard any unprocessed frames. This approach removes OpenCV’s
internal buffer and reduces the frame lag. This function stores the frame in queue
instead of the buffer and fetch frames from the queue one by one.

Fig 4.22 Analysis in college campus

• The above snapshot was taken around January 2022 in our college campus during the
third wave of covid-19. At this time, we were in our initial phase of the project and we
took the opportunity to capture this video and later we performed an analysis on this
video so that we can closely relate and associate with the situation of how serious and
deadly surrounding we create for ourselves as well as for others.

25
 CHAPTER-5

CONCLUSION

This project presents a deep learning based intelligent system that will track people and

identify whether social distancing is being followed as per the predefined norms. We have

used the pretrained YOLOv3 exemplar for the detection of human. Our model of

detection gives the information of violation through bounding box, containing centroid

coordinates information. The specific algorithm implemented on bounding boxes,

distinguished between the safe and unsafe conditions, respectively, marking as green and

red the bounding box for detected persons. We have used Euclidean distance to measure

the distances between two pairs of detected bounding boxes through their centroid. To

classify the social distance violations, we have used numerous parameters, such as

max_distance between a pair of pedestrians, threshold value, according to which model

assigns serious and abnormal violation flags and trigger an alert function if threshold

value is hit. We have furthermore added various usecases to our model such as human

count which is useful for preventing stampede, taking attendance, tackling contagious

diseases, alert mail which is sent to admin whenever the serious violation has crosses the

parameter of threshold value, modifying ROI as a usecase is useful whenever we have

two overlapping video angles input , we can manually set out our region of interest and

run our analyser on that particular ROI only. The proposed technique has achieved

promising results in identifying people walking too close and violating social distancing

norms.

26
 FUTURE WORKS

The proposed model may be executed in a distributed video surveillance system. It will

definitely be an effective solution for the administration to analyse the compliance of

pedestrians with physical distancing. In the coming times, this model can also be put into effect

on mobile cameras. For eg.,established on drone system and it can simply operate and capture

fast and dynamic actions of the captured objects from various angles.

In our project we have made space for real time monitoring of the pedestrians by providing

camera URL variable in config file, in which if we provide secure url for the camera the main

function would run inference on the reference of the camera otherwise it would run inference

on the reference of the stored video file.

27
 References

[1] Jobie Budd, Benjamin S. Miller, Erin M. Manning, Vasileios Lampos, Mengdie
Zhuang, Michael Edelstein, Geraint Rees, Vincent C. Emery, Molly
M.Stevens, NeilKeegan, Michael J. Short, Deenan Pillay, Ed Manley, Ingemar J.
Cox, David Heymann, Anne M. Johnson & Rachel A. McKendry, “Digital technologies in
the public-health response to COVID-19”

https://www.nature.com/articles/s41591-020-1011-
4#:~:text=Digital%20technologies%20are%20being%20harnessed,and%20communicatio
n%20with%20the%20public

[Accessed 7 August 2020]

[2] G V Shalini, “Social Distancing Analyzer Using Computer Vision and Deep Learning”

https://iopscience.iop.org/article/10.1088/1742-6596/1916/1/012039/pdf

[Accessed 23 February 2022]

[3] Yew Cheong Hou; Mohd Zafri Baharuddin; Salman Yussof; Sumayyah Dzulkifly, “Social
Distancing Detection with Deep Learning Model”

https://ieeexplore.ieee.org/abstract/document/9243478

[Accessed 11 November 2020]

28
[4] Riya Chaudhri; Lokesh M. Girpunje; Arpita Patra; Aniket Sagar “A review on Social
Distance and Face mask detection”

https://www.jetir.org/view?paper=JETIR2104023
[Accessed 8 April 2021]

[5] G Chandan; Ayush Jain; Harsh Jain; Mohana, “Real time object detection and
tracking using deep learning and openCV”

https://ieeexplore.ieee.org/document/8597266?denied=
[Accessed 2018]

[6] Afiq harith Ahamad; Norliza Zaini; M F Abdul Latip, “Person Detection for Social
Distancing and Safety Violation Alert based on Segmented ROI”

https://www.researchgate.net/publication/347021016_Person_Detection_for_Social_Dista
ncing_and_Safety_Violation_Alert_based_on_Segmented_ROI
[Accessed August 2020]

[7] Shashi Yadav, “Deep Learning based Safe Social Distancing and Face Mask Detection in
Public Areas for COVID-19 Safety Guidelines Adherence”

https://www.researchgate.net/publication/343346690_Deep_Learning_based_Safe_Social
_Distancing_and_Face_Mask_Detection_in_Public_Areas_for_COVID-
19_Safety_Guidelines_Adherence
[Accessed July 2020]

29
[8] Apoorva Raghunandan; Mohana; Pakala Raghav; H. V. Ravish Aradhya, “Object
Detection Algorithms for Video Surveillance Applications”

https://ieeexplore.ieee.org/document/8524461
[Accessed 8 November 2018]

30
 Annexure

Other Experimental Results

This section consists of the snapshots of the experimental results obtained from the video
captured in the campus of Institute of Engineering and Technology, Lucknow.

31
32
33