SAFER-Activities: A Dataset for Smart Assessment of Fall Events and Routine Activities
Diwas Lamsal, Pramod Wickramatilake, Jednipat Moonrinta, Mongkol Ekpanyapong, Matthew N. Dailey
cs.CV · cs.AI
Quick verdict
Claims largest frame-level fall/ADL dataset incl. wheelchair subset; benchmarks back skeleton-over-RGB OOD findings, but young actors and no stats limit reach.
Summary
The authors introduce SAFER-Activities, a video dataset for fall detection and activity monitoring with frame-level annotations: over 66 hours from 46 participants, 85,310 action instances across 30 classes, including 5,406 falls and a dedicated wheelchair subset (840 wheelchair falls), plus a complementary wheelchair pose-estimation set (6,846 images). Annotation reliability is quantified (Cohen's κ=0.89; boundary F1 0.86 at 500 ms). The paper benchmarks 2D/3D skeleton models, frozen RGB backbones (CLIP, DINOv3, VideoMAE), and five fusion strategies on in-lab, out-of-distribution home-environment, wheelchair, and cross-dataset (ImViA) test sets. Best in-lab macro F1 reaches 87.2% (2D MS-G3D, non-wheelchair); under domain shift, skeleton models degrade least (79.3% OOD) while frozen RGB collapses (DINOv3 23.8%; fall F1 0.0), and no fusion strategy beats the skeleton-only CNN1D baseline (73.8% OOD). On ImViA without fine-tuning, PoseC3D reaches 96.8% fall-detection F1. Main limitations: falls are simulated by healthy adults aged 18–58 rather than elderly or regular wheelchair users; the OOD set has only 5 participants and 30 videos; results lack variance estimates or significance tests; RGB conclusions rest solely on frozen linear-probe models; dataset access is granted upon request.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SAFER-Activities is the largest vision-based fall/ADL dataset with frame-level annotations (85,310 instances, 5,406 falls, >66 h, 46 subjects) and the only one with a wheelchair subset (840 wheelchair falls).
- Skeleton-based models generalize best under domain shift, while frozen RGB features collapse on the OOD home-environment set and on external fall data.
- Fusing frozen RGB features with the CNN1D skeleton stream improves in-domain recognition (most clearly on the wheelchair subset: 68.2%→78.6%) but degrades out of distribution, and no tested fusion strategy beats the skeleton-only OOD baseline (73.8%).
- Models trained on SAFER-Activities transfer to unseen environments and external fall data without fine-tuning, with PoseC3D reaching 96.8% fall-detection F1 on ImViA.
Strengths
Scale and annotation density decisively exceed prior fall datasets
Table 1: 5,406 falls across camera views vs. 275 in the next-largest (Baldewijns et al.), 46 subjects aged 18–58, >66 h, and frame-level labels for all 30 classes — only the second resource after Auvinet et al. to do so; 840 wheelchair falls are unique.
Quantified annotation reliability, rare in HAR datasets
Sec. 3.2: second-annotator verification plus 5% blind re-annotation yields Cohen's κ=0.89 (0.91 wheelchair), boundary F1 0.86 at 500 ms tolerance, median boundary deviation 145 ms — this makes the frame-level labels trustworthy rather than merely claimed precise.
Generalization-first evaluation design
Honest negative results and candid limitations
Released reproducibility assets
Weaknesses
Ecological validity gap for the intended population
No uncertainty quantification for model results
RGB modality conclusions rest only on frozen linear probes
Fusion evaluated only against the weakest skeleton encoder
Request-gated dataset access
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a dataset-and-benchmark paper making four testable claims: (1) SAFER-Activities is the largest vision-based fall/ADL dataset with frame-level annotations, including a wheelchair subset; (2) skeleton models generalize best under domain shift; (3) frozen-RGB fusion improves in-domain recognition but degrades out of distribution; (4) models transfer cross-dataset without fine-tuning. The text supports all four at their stated scope. Table 1 substantiates claim (1) concretely: 5,406 falls across camera views vs. 275 in the next-largest prior dataset (Baldewijns et al.), 46 subjects spanning ages 18–58, >66 hours, and frame-level labels for all 30 classes — only the second resource after Auvinet et al. to provide these. Claims (2)–(4) each carry dedicated experiments (Tables 2–6). Wording is mostly careful — "frozen RGB", "over the baseline CNN1D" — though the abstract's modality narrative risks being read beyond the frozen-linear-probe evidence.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the SAFER-Activities dataset?
A public video dataset for fall detection and physical activity monitoring: over 66 hours of multi-camera video from 46 participants, with frame-level annotations for 85,310 action instances across 30 classes, including 5,406 falls and a dedicated wheelchair subset (840 wheelchair falls), plus a complementary pose-estimation subset of 6,846 wheelchair images.
How does SAFER-Activities compare to existing fall detection datasets?
Per its Table 1 comparison, prior vision-based fall datasets top out around 275 falls (Baldewijns et al.) with mostly clip-level or fall-only labels; SAFER-Activities offers 5,406 falls, 30 labeled activity classes, frame-level labels for all actions (previously only Auvinet et al. did this), untrimmed multi-camera recordings, and a home-environment out-of-distribution test set.
Which models perform best on SAFER-Activities?
Skeleton-based models: 2D MS-G3D is best in-lab (87.2% macro F1 on the non-wheelchair subset). Under domain shift (home environment), skeleton models degrade least (best 79.3%), while frozen RGB backbones collapse (DINOv3 falls from 72.1% to 23.8%; 0.0 fall F1). Fusion helps in-domain (78.6% on wheelchair) but no fusion strategy beats the skeleton-only CNN1D baseline (73.8%) out of distribution.
Does the dataset contain real falls by elderly or wheelchair users?
No. All falls are simulated by healthy adult actors aged 18–58, including able-bodied actors using wheelchairs. The authors state this explicitly in the limitations: the dataset is positioned as a pre-deployment benchmark for fall dynamics, and validation with elderly people, regular wheelchair users, and naturalistic falls remains future work.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Fig. 1, 2, 5, sample images) cannot be evaluated from text extraction; the qualitative claims of Sec. 4.4 (Fig. 5) rest on images that are not verifiable here.
- Quantitative results were transcribed from dense multi-column tables; column alignment was cross-checked against the in-text narrative (e.g., Sec. 4.2 summaries match Tables 2–3), but individual cell values could not be verified against source data.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.