An attempt at fall detection on an embedded device based on YOLOv8n-pose

Human-action recognition aims to identify the actions performed by individuals. Due to the broad spectrum of human activities, action recognition covers a wide range. Among all, fall detection is a critical aspect of surveillance, particularly in environments where individuals are at risk. Throughout the years, several sensors, data types, and classification techniques have been investigated to address this issue. This paper proposes a lightweight fall detection designed to process sequences of images in real-time. This system is deployed on an embedded device, specifically the Jetson nano. Our goal is to construct a comprehensive dataset that accurately detects falls in various lighting conditions. The proposed system is constructed using YOLOv8n-pose, which have been trained to identify people using widely recognized dataset. Our methodology includes the design and implementation of the YOLOv8n-pose, data collection, and rigorous testing to ensure the accuracy of fall detection in real-time using surveillance camera. The experimental results show that high detection accuracy and acceptable timing capabilities are achieved.