SAFER‑Activities
A Dataset for Smart Assessment of Fall Events and Routine Activities
Overview
SAFER‑Activities is a dataset for fall detection and physical activity monitoring, with frame-level annotations over long, untrimmed, multi-camera videos and a dedicated subset for wheelchair users. We benchmark 2D/3D skeleton models, frozen RGB foundation-model features, and multimodal fusion for action recognition and fall detection across in-lab, out-of-distribution, wheelchair, and cross-dataset settings: skeleton models generalize best under domain shift, while RGB features help in-domain but degrade sharply out of distribution.
What's Inside
- Over 66 hours of untrimmed video from 46 participants.
- Frame-level start/end annotations for 30 action classes (85,310 action instances).
- 5,406 fall instances alongside look-alike actions such as lying down and sitting.
- A dedicated wheelchair subset and a pose-estimation set of 6,846 images / 8,324 annotated instances.
- 2D/3D pose skeletons, bounding boxes, and frozen pretrained visual features, with subject- and view-wise splits.
Detecting Falls in the Wild
Models trained on SAFER‑Activities detect falls in public videos far outside the training distribution, including wheelchair users.
Sources: non-wheelchair falls from FailArmy; wheelchair fall from regor09.