Fall detection paradigm for embedded devices based on YOLOv8.

Human-action recognition attempts to determine which deed occurs by people. Owing to the wide spectrum of human activities, action recognition spans a wide range. Among all, falls the conduct gathering special interest, especially for the elderly, as 37.3 million accidents happen annually, with 1.83% being fatal. Over the years, several sensors, data types, and classification techniques have been explored to cope with this problem. Nowadays, ambient assisting living environments employ cameras and companion robots to monitor human activity. The need for low computing power platforms in such environments is essential. To this end, we propose here a lightweight fall detection pipeline for processing sequences-of-images in real-time, which is deployed on an embedded device, namely the Raspberry Pi 4. The proposed system is built upon YOLOv8, having been trained to locate people on the well-established COCO dataset, and a binary classification support vector machine trained on E-FDPS. Through an extensive experimental protocol on LE2I, we show that high detection accuracy and fast timing capabilities are achieved. Lastly, aiming to facilitate the community, our framework is open-sourced at https://github.com/smoutsis/real_time_fall_detection.