Evaluating Pose Estimation as a Solution to the Fall Detection Problem

Abstract

In this age of evolving technological capabilities, assisted living has proven useful to ease the frailty that comes with aging. Within this context, the detection and prevention of accidents are paramount to ensure a longer life expectancy for the elderly. Over the years, many approaches for fall detection have been proposed, such as ambient sensors, wearable devices, and automated camera monitoring. A recent approach is to use pose estimation software to identify humans and pinpoint the location of their most important joints. This pose information can be later used as features for an effective fall detection system. This scenario begs the question: Can pose estimation methods be as effective as the sensor or other camera-based ones? To answer this question, we analyzed three pose estimation frameworks, totalizing eleven models, paired with a simple neural network classifier. In our experiments, we have obtained competitive results among the state-of-the-art on the UR Fall Detection dataset, a multi-modal fall detection benchmark, comprised of RGB, depth, and acceleration data. More specifically, our best model achieved a sensitivity rate of 94.5% and a specificity rate of 99.9%, in line with the best camera and sensor-based solutions.