Fall prediction algorithm with built-in instability metrics.

Abstract

This article introduces an artificial intelligence (AI) platform that uses computer vision to monitor human body posture and predict falls. Typically, camera-based systems for "fall prediction" are rarely investigated compared to "fall detection" systems because falls represent a nonlinear, time-dependent challenge. This complexity is especially prominent when attempts are made to predict falls in uncontrolled environments. In addition, privacy concerns and the requirement to install expensive cameras or sensors are considered key limitations for a camera-based system. To address this gap, we introduce the extraction of new features independent of the camera type that help eliminate the need for high-cost cameras, and improve prediction accuracy without requiring extensive room modification or the addition of body markers or wearable sensors for the user. In this study, we used a 4 K AKASO camera to record different fall scenarios, and the features were extracted from these videos. These features include key landmarks, the centroid's locations, and the angular position of body segments. The results show that this system can accurately predict falls with an approximate accuracy of 91 %. Additionally, the feature importance analysis highlights the significance of the extracted features, indicating that these features have a significant effect on improving the prediction of a fall up to two seconds before it happens, which is three times faster than today's single camera systems.