A Deep Learning Approach for Basic Human Activity Recognition with YOLOv4-tiny

The regression of the elderly body condition over time causes the elderly to be more susceptible to illness and other accidents. When the elderly experience illness or an accident such as a fall, it is necessary for the family to realize this and take prompt treatment immediately. However, with relatively many elderly people in Indonesia choosing to live independently, it will be difficult for families to find out and provide help instantly. Therefore, in this study, the authors form a model for recognizing basic human activities with deep learning-based computer vision that can be implemented in a supervisory system in a room. A deep-learning approach is needed because of the complexity and variance of body postures and forms of human activities. However, the deep learning approach requires extensive resources and computational capabilities. Therefore, the model is formed by the YOLOv4-tiny method, one of the tiny versions of YOLO. Model training using the author's laptop and model inference testing was carried out on the Jetson Nano and the author's laptop to compare the inference time between the two devices. We investigate the performance of the YOLOv4-tiny model application on the Jetson Nano and a laptop, as well as the accuracy of recognizing human activities. This study shows that this particular vision-based human activity recognition model formed using YOLOv4-tiny as a deep learning method can be applied using Jetson Nano as an embedded device in real-time, with a speed of about 20 frames per second, mAP@0.50 of 99.04%, and an average F1-Score of 94.18%.