COVID-19 SOP's Violations Detection in Terms of Face Mask Using Deep Learning

Facemask detection is a need of time as we are suffering in a pandemic situation of COVID-19, and facemask is considered the best preventive measure to stop the rapid spread. The vast majority of the world population is still unvaccinated, especially young and kids. Moreover, despite the vaccination, people are still getting Covid positive, and the majority are due to the Delta variant. So, we still need to have strict SOP implementation. The best way is to have some autonomous system to monitor SOP compliance and alert the authority to take countermeasures. Many people wear the mask, but the mask is usually on the chin and does not serve the purpose because the facemask must cover the mouth and nose to stop the spread. This study has proposed the improved version of the YOLOv4 model for the robust detection of face masks and checks whether the mask is worn in the recommended way. 2D convolutions of Yolov4 are replaced with the spatially separable convolutional in YOLOv4 to reduce the parameters so that the model can work in real-time. We have achieved an accuracy of 86.61% in terms of proper mask-wearing. Unlike other proposed approaches, our model is not only detecting the mask but also determines that whether the mask is worn in the recommended manner.

Multiple Parallel Activity Detection and Recognition to Avoid COVID-19 Spread-Out

COVID-19 has been affecting people around the globe. It is affecting almost every country currently, according to the World Health Organization (WHO). This virus is transmitted to another person if an asymptomatic person makes close contact with the everyday person. There is no cure for this virus, and the only solution is social distancing and avoids the people doing these activities. In this paper, we proposed a system for detecting and recognizing the activities that violate social distancing. These activities involve handshakes and hugging. We implement a system that is capable of detecting and identifying multiple parallel activities. Temporal features are extracted for similar activities in 16 frames. We use the two models for this purpose: Faster RCNN for the detection and ResNet-50 to recognize the activities. First, Faster RCNN detects the group of people and that region of interest ROI saved and passes to the ResNet-50 to recognize the activities. We also generated our dataset on the local environment in multiple parallel activities. This system achieves the accuracy of 95.03% for the detection of testing dataset and recognition of multiple parallel activities 92.88% accuracy accomplished. The system used different public datasets and generated some local datasets for handshake and hugging activities. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.