Monitoramento de distanciamento social utilizando aprendizado de máquina e visão computacional

Abstract

Considering the history of human viral diseases, from the flu pandemic to the current COVID-19 pandemic, we are always looking for more and more innovative ways to combat these diseases. Researchers from different fields are relentlessly researching effective methods to help people protect themselves. Some important studies have been proposed in the area of computing, such as a smart cap, which can help in protection by detecting high temperatures and warnings of breaking social distance. We also have social distance detectors that use security cameras and deep neural networks and analysis of infection risks through this monitoring. This work proposes the development of a tool with machine learning and computer vision techniques to detect agglomeration in public spaces. YOLOv3 (You Only Look Once) was used to detect people and calculate the Euclidean distance between pairs and verify if there is a minimum distance violation of 1.5m, as established by OMS. Because it is a simple conversion, this minimum distance is not exactly the same in relation to the real distance, but it comes close. The tests were carried out using pre-recorded video and real-time video at a public place where people usually walk. A qualitative analysis was performed and the accuracy of each video was calculated at certain times. The average accuracy was 94.66% for pre-recorded videos, 90.28% for videos recorded in real-time and 92.47% for videos recorded in real-time, without using GPU (Graphics Processing Unit) and high FPS (Frames per Second), which is considered very efficient, in terms of video speed and detection, especially compared to other jobs where GPUs help. As a result, we have a tool capable of helping to fight not only COVID-19, but also several diseases WITH similar contagion.