MODERN COMPUTER VISION AND DEEP LEARNING APPROACHES FOR FACE MASK AND SOCIAL DISTANCING DETECTION

Abstract

This thesis proposes a new automated pipeline to detect and localize ob- jects in a given video scene using state-of-the-art computer vision models and machine learning. The inspiration of this pipeline was the ongoing global pandemic of COVID- 19, which the World Health Organization (WHO) declared as a global pandemic on January 30, 2020. The objective for this pipeline was to employ modern technologies, such as machine learning and deep learning, to help solve a real-life problem: face-mask wearing and social distancing in indoor and outdoor places. Employing state-of-the- art object detection algorithms along with a deep learning-based approach to estimate distance between people, we have created a pipeline to receive a CCTV live camera feed and output heat-maps showing whether people are adhering to face mask-wearing and social distancing. Training over 39K images on the-sate-of-art object detection models at the time, YOLOv4 and Faster R-CNN with Focal Loss, we have achieved over 83% Mean Average Precision (mAP) on a custom dataset. In addition, we imple- mented classical and modern approaches to measure distances between people in public spaces, such a projective transformation and deep learning-based models, to estimate camera calibration parameters.