Deep Learning Approaches to Social Distancing Compliance and Mask Detection in Dining Environment

Social distancing is an effort to maintain distance and reduce physical contact between individuals or groups to prevent the transmission of diseases such as the COVID-19 virus. The simulated research was carried out on a camera installed in the simulation room of a restaurant. The camera will detect the implementation of social distancing and the use of masks. The camera will detect objects in the form of persons using the You Only Look Once (YOLO) algorithm and calculate the distance between the detected individuals using the Euclidean Distance method. The system can detect each individual's face using Haar Cascade and classify the use of masks or not using the Convolutional Neural Network (CNN) algorithm. The system can detect violations of social distancing and the use of masks between person objects. The results of the YOLOv4 training model at a ratio of 90%:10%, max batches of 6000, and a learning rate of 0.001 received a mAP of 49.02%. As for CNN, was obtained with a ratio of 90%:10%, batch size 256, and learning rate of 0.0001 got an accuracy of 98%. In the condition that the camera is parallel to the object has an accuracy rate of 100%, in the state where the camera is located above the object, an accuracy of 77.8% is obtained. While CNN with a range of distance between faces with a camera of 100 cm to 300 cm with an accuracy rate of 80.36%.