Drone based face mask detection system

Abstract

Over the past year and a half, Malta just like most countries in the world has been affected due to the Coronavirus outbreak. With over 160 million confirmed cases and a death toll of over three million worldwide, people around the world are trying to contain the spread of this deadly virus. The most effective way to prevent the spread of the virus is by simply wearing a face mask. Most countries have made wearing a mask both inside and outside crowded establishments mandatory through legal notices. However, it requires a lot of manpower to constantly go to different premises and check that people are wearing their masks. Therefore, there is a need for an automatic and real-time image processing face mask detection system helping out with the difficult task of surveilling the obedience of mask wearing in public places. The biggest advantage this system would have is that it is detecting a person’s face in real time which speeds up the process. In recent years, we have seen a great advancement to image processing Neural Networks, hence many more projects such as this one, have become feasible. In enclosed public spaces such as airports, a camera can easily be placed, however in public spaces this is more difficult since there is a larger area to cover and people are constantly moving. In these types of situations unmanned aerial vehicles (UAV) such as drones come in place as they offer various advantages. One of their biggest advantages is that they have a great range of movement which allows them to navigate through hard to access areas. The aim of this project is to investigate, design and develop a drone-based real-time face detection system. Different deep learning algorithms which can be used in this were evaluated in order to find the best performing object detection algorithm. From the research and tests carried out, the You Only Look Once (YOLO) algorithm was found to be one of the best convolutional neural network architecture models due to its high accuracy and speed, especially in real time detection. The system was implemented using different versions of this algorithm such as YOLOv3 and the Tiny-yoloV3 model. A custom dataset was created and used to train and evaluate each model. In conclusion, both algorithms were proven to give good results.