First-person Vision-based Assessment of Fall Risks in The Wild, Towards Fall Prevention in Older Adults

Abstract

Falls in older adults is one of the most important public health problems world-wide. In our previous works, we showed that first-person
vision (FPV) data acquired by chest- and waist-mounted cameras
have the potential to be utilized to (A) develop novel markerless
deep models to estimate spatiotemporal gait parameters over time
(e.g., step width) by localizing feet in 2D coordinate system of RGB
frames (using optical flow and RGB streams) and (B) automatically
identify environmental hazards (e.g., curbs, stairs, different terrains)
that may lead to falling. In this paper, a summary of our recent FPV-
based approaches for fall risk assessment in the wild are being discussed. These approaches aimed to eventually inform clinical decisions on the most appropriate prevention interventions to reduce
fall incidence in older populations.