Tezislər / Theses

REAL-TIME SOCIAL DISTANCE SURVEILLANCE SYSTEM 
Laman Hasanli, Parvana Guliyeva, Sona Mehdizada 
Baku Higher Oil School 
Baku, Azerbaijan 
[e-mail] 
Supervisor: Ph.D Associate Professor Kamala Pashayeva 
Keywords:
YOLOv3, Convolutional Neural Network, video processing, object detection 
With its devastating spread, the ongoing COVID-19 coronavirus 
outbreak has caused a global calamity. Because there are no vaccines 

---

THE 3
rd
 INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS 
dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
164
which would cure the virus available in the current circumstances, 
social separation is considered an adequate precaution against the 
spread of the pandemic virus. This paper describes a method for 
detecting social distancing using deep learning to assess the distance 
between people in order to reduce the impact of the coronavirus 
epidemic. By analyzing a video feed, the detecting tool was created to 
warn people to keep a safe distance from one another. A pre-recorded 
video of pedestrians walking was used to validate the proposed 
strategy. With the accuracy of 95%, the outcome demonstrates that 
the suggested method is capable of determining the social distancing 
measures between many participants in a video.
Social distancing refers to strategies for preventing the spread of 
a virus by limiting human physical contact in public locations (e.g., 
shopping malls, parks, schools, universities, airports, and workplaces), 
avoiding crowds, and maintaining a safe distance between individuals. 
If applied early on, social distance can play a critical role in halting 
virus propagation and preventing the pandemic from reaching its peak. 
Social distance has been shown to minimize the number of infected 
patients and the strain on healthcare institutions. Therefore, in today’s 
world an accurate social distance detection system is a necessity 
rather than being preferred. 
The pre-trained model based on the YOLOv3 method was utilized 
to identify pedestrians using the video frames from the real-time 
camera as an input. Later, the video frame was transformed to a top-
down view for determining distance values in the 2D plane. Any pair 
of persons having distance less than predefined threshold value in the 
display will be depicted with a red box on views. An approximation of 
physical distance to pixel is utilized to approximately estimate social 
distance violations between people. To determine whether the 
distance value exceeds the minimal social distance, a violation 
threshold is determined.
The model is created using the image and video captured by the 
camera. The camera is set up to capture at a fixed angle, and the video 
frame's view has been converted to a 2D bird's eye view to precisely 
estimate the distance between each object detected. The people in the 
picture are assumed to be leveled on the horizontal plane for simplicity. 
Then four locations on the horizontal plane are determined, and the 
scene is converted into a bird's eye view. Then, using the bird's eye 
view, each person's position can be estimated. The CNN model was 

---