Human Activity Detection Using YOLOV4

Abstract

A simple activity recognition model can allow a single human person to monitor all our surrounding with the purpose to ensure safety and privacy while preserving maintenance cost and efficiency with the soaring level of precision. This monitoring system with realtime video surveillance could be deployed for patients and the elderly in a hospital or old age home along with various human activity on important area such as the airport. For speedy analyze of action and accurate result while working with complex human behaviour, we decided to use YOLOv4 (You Only Look Once) algorithm which is the latest and the fastest among them all. This technique uses bounding boxes to highlight the action. In this case, we have collected 4,674 number of different data from the hospital or different condition ourselves for fastest accuracy with the one of the largest data-set ever used in such kind of project. During our research, we had divided our action into three different class which are standing, sitting and walking. This model was able to detect and recognize multiple patients or other regular person activity and multiple human activities tracing support at once. After completing our project, this model manages an average accuracy of 94.6667% while recognizing image and about 63.00% while recognizing activity from video file. We also work on two other different projects with TensorFlow and OpenPose while the YOLOv4 perform better than those two. In future, more complex data can be added and prediction can be implemented which will improve and add utility to this project.