Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Unsupervised Keypoints for Real-Time Fall Detection: Comparative Analysis Under Real-world Conditions with Predictive Bandwidth Reduction

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper shows that when bodies are partially occluded, fall detection built on unsupervised motion keypoints recalls 89% of falls versus 49% for supervised anatomical keypoints, because the learned landmarks adapt to whatever body structu

desk verdict A useful split-level comparison of learned vs anatomical keypoints for fall detection, but the occlusion-robustness claim conflates task supervision with representation type; worth a serious revision. read the letter →

arxiv 2607.15400 v1 pith:CLDC3RND submitted 2026-07-16 cs.CV cs.LG

Tasmiah Haque , Jacob Kosinski , Sumit Mohan , Srinjoy Das , Mohammad Abdullah Al-Mamun This is my paper
classification cs.CVcs.LG
keywords falldetectionunsupervisedkeypointssupervisedposeestimationocclusionrobustnessout-of-distributionevaluationvideomonitoringpredictivebandwidthreductionprivacy-preserving
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that the right motion representation for fall detection depends on the visual conditions a monitoring camera will actually face, and that under occlusion an unsupervised keypoint representation is substantially more reliable than anatomically supervised pose. Its central evidence: with bodies partially hidden, the unsupervised pipeline recalls 88.9% of falls versus 48.9% for the supervised pose pipeline (F1 76.3 vs 56.3), and the gap persists when half the keypoint frames are replaced by forecasted values. Under random train/test splits neither representation consistently dominates, so the paper argues that standard benchmarking hides the meaningful difference. A sympathetic reader would care because real home and hospital monitoring regularly has occlusion and partial body visibility, and a system that misses nearly half of falls is clinically unacceptable.

What carries the argument

The mechanism carrying the argument is a keypoint-based representational bottleneck coupled to a temporal forecasting objective. The unsupervised detector turns each video frame into K spatial coordinates by taking expectation over learned heatmaps, then the whole pipeline—detector, variational recurrent predictor, and recurrent classifier—is optimized jointly with reconstruction, separation, sparsity, forecast, and classification losses. This joint task-aligned training is what lets the learned landmarks behave as motion anchors free of anatomical constraints; supervised pose, by contrast, is a fixed pretrained detector whose predefined joints must remain visible. The same temporal model se

What would settle it

Train the unsupervised keypoint detector without the classification and forecasting losses (reconstruction-only), retrain the downstream classifier, and rerun the occlusion split; if recall drops to supervised levels, the claimed anatomical-independence advantage is not causal. Alternatively, fine-tune the supervised detector with the same classification objective and see whether the gap closes.

Watch

Extended reading notes

Core claim

The paper's central claim is that the anatomical independence of unsupervised keypoints—learned motion landmarks that are not tied to predefined body joints—is what makes them robust in exactly the conditions where supervised pose fails. In the occlusion-based test, supervised keypoints miss roughly half of all falls (recall 48.9±11.3) because the detector is obligated to localize joints that are absent or occluded, while unsupervised keypoints retain high sensitivity (recall 88.9±9.9) by anchoring to whatever salient body structure remains visible. The paper further claims this advantage comes with a trade-off: under clean full-body visibility, supervised keypoints generalize better to new

Load-bearing premise

The load-bearing premise is that the two pipelines differ only in representation type; in practice the unsupervised detector is optimized with the classification objective while the supervised detector is frozen, so task supervision rather than anatomical independence could explain the result.

Editorial extensions

If this is right

  • Under occlusion or partial body visibility, using unsupervised keypoints instead of supervised pose raises fall recall by roughly 40 percentage points without prediction, and by about 33 points when only half the keypoint frames are transmitted.
  • A monitor that transmits only half the keypoint frames and forecasts the rest locally keeps the unsupervised advantage while cutting bandwidth by roughly 35–52x relative to video transmission.
  • Random train/test splits are not sufficient for choosing a representation; subject-disjoint and occlusion-based splits reveal that supervised pose wins under clean visibility while unsupervised keypoints win under degraded visibility.
  • The same forecasting module that enables bandwidth reduction can fill short gaps when keypoint frames are lost in transmission, preserving temporal continuity for classification.
  • For clinical deployment, recall behavior and interpretability both matter: supervised keypoints retain value for biomechanical or explainability needs, while unsupervised keypoints are preferable when missed falls are the dominant risk.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the mechanism is truly anatomical independence rather than task supervision, the same robustness should appear in other occlusion-prone vision tasks—patient mobility monitoring, activity recognition, or gesture understanding—when supervised pose is replaced by unsupervised motion anchors.
  • The comparison is not fully symmetric: the unsupervised detector is trained jointly with the classification objective while the supervised detector is frozen, so an alternative explanation for the occlusion advantage is task-aligned training; a matched experiment training both detectors under the same optimization would settle which factor is responsible.
  • Adding an auxiliary loss enforcing some anatomical consistency to the unsupervised detector could reduce false positives on varied daily activities while preserving the occlusion advantage, addressing the paper's admitted trade-off.
  • The forecasting module's ability to fill missing keypoints suggests it could double as a dropout-recovery mechanism in real deployments, something the paper mentions but does not evaluate explicitly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a privacy-preserving fall-detection pipeline that extracts unsupervised keypoints locally, transmits them to a central system, uses a VRNN to predict future keypoints for bandwidth reduction, and classifies with an LSTM. It compares this pipeline against an OpenPose-based supervised keypoint pipeline on UR Fall and Human Fall datasets under random, subject-disjoint, and occlusion-based splits. The central claim is that unsupervised keypoints are substantially more robust to occlusion because their anatomical independence allows spatial anchors to adapt to visible body structure, and that supervised keypoints miss nearly half of all falls under occlusion. The paper also reports bandwidth and compute advantages of the unsupervised front-end.

Significance. If the central comparison were clean, the finding that unsupervised keypoints outperform supervised pose under occlusion would be clinically relevant and of interest to both the computer vision and healthcare monitoring communities. The paper's strengths include a thoughtful split protocol (random, subject-disjoint, occlusion), the use of paired statistical tests across seeds, a predictive bandwidth-reduction mechanism with explicit bitrate estimates, and an assessment of computational cost. However, the main result is currently confounded, and the mechanistic interpretation is not directly evidenced. The work is potentially valuable, but the central claim needs additional controlled experiments before the conclusion can be accepted.

major comments (4)
  1. [§3.6, Eq. (21); §3.3.2] The central comparison is not representation-only. The 'unsupervised' keypoint detector is trained jointly with L_cls, L_VRNN, L_future, and other losses, while OpenPose is 'pretrained and fixed' with only the predictor and classifier trained. Thus the occlusion advantage in Table 7 could stem from task-level supervision or end-to-end temporal/classification feedback shaping the detector, not from anatomical independence. To support the claim in §5.1 and §6, please add a controlled experiment: e.g., fine-tune the supervised detector with L_cls and VRNN losses (or train a supervised keypoint detector on the target data), or evaluate a fixed unsupervised detector without L_cls and VRNN. This is necessary to isolate representation type.
  2. [§4.4.3, Table 7] The occlusion-based evaluation is built on a manually selected test set of only 18 videos. The split is not a clean manipulation of body visibility: identity, clothing, camera angle, and ADL composition differ between train and test, and ADL identities are not separated. The Wilcoxon test uses only 6 F1 pairs (3 seeds × 2 prediction settings), which is low-powered for the strong claim. Please report the specific composition of the occlusion test set, provide a more principled occlusion protocol (e.g., synthetic occlusion or a larger sample), and include per-video results and confidence intervals.
  3. [§5.1, Fig. 3] The mechanistic claim that unsupervised keypoints 'adapt to visible body structure' is not directly measured. The paper reports only downstream F1/recall; Figure 3 is qualitative. Add quantitative evidence on keypoint behavior under occlusion, such as localization error on visible body parts, keypoint stability across occluded frames, or an ablation that explicitly compares keypoint tracking on partially visible bodies. Without this, the mechanism remains speculative.
  4. [§3.3.1 vs §3.3.2] The two pipelines differ in feature preprocessing beyond the detector: OpenPose uses MidHip/Neck-to-MidHip normalization, drops confidence scores, and appends frame-to-frame displacements, while the unsupervised pipeline uses global centering, retains confidence, and does not use displacement features (Eqs. 8–12). These are additional asymmetries that may affect the comparison independent of representation type. Please control for these by applying a common preprocessing scheme or explicitly discussing their influence.
minor comments (6)
  1. [Algorithm 1, line 7] The indexing for the predicted segment is written as P(j) = {kpt,k}_{t=M}^{M+N}, but Eq. (20) defines prediction for t = M+1 to M+N. This is inconsistent and should be fixed.
  2. [§4.5.2, Table 6] The Wilcoxon signed-rank test description is ambiguous about which pairs are included (all seeds and groups, with/without prediction? n=24 or n=12?). Please specify the number of pairs and whether the test pools both prediction settings.
  3. [Tables 1, 5, 7] The column headers contain a formatting artifact: 'Unsup. 1 Sup.1' appears in several tables. Please correct these headers.
  4. [§5.4] There is a typo: 'multiple multiple residents' should be 'multiple residents'.
  5. [Abstract / §1] The paper emphasizes 'real-time' fall detection, but no end-to-end latency measurement is reported; only TOPS estimates are given. Either add latency experiments or soften the real-time claim.
  6. [§2.3 / Contributions] The statement 'first work to apply unsupervised keypoint learning ... to fall detection' is difficult to verify and may be overclaimed. Consider softening with 'to the best of our knowledge' and a more detailed related-work comparison.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the core empirical comparison is self-contained. Minor self-citations in the VRNN motivation are not load-bearing, and the main occlusion-robustness claim is a genuine held-out finding, though a supervision confound is a correctness risk rather than circularity.

full rationale

The central claim—that unsupervised keypoints retain higher recall than supervised OpenPose keypoints under occlusion (Table 7: 88.9 ± 9.9 vs 48.9 ± 11.3 without prediction)—is an empirical result on held-out test videos, not a quantity forced by the paper's own definitions. The with-prediction setting is a true forward test: Eq. 20 defines the observed/predicted split, but the VRNN must forecast frames M+1..M+N from frames 1..M, and the classifier is evaluated on fixed held-out test data; no equation inserts the test labels into the input. The preprocessing steps (interpolation, centering, scale normalization) are applied uniformly to both pipelines and do not encode the target class. The paper's only self-citations are in the motivation for choosing a VRNN (Section 3.4, refs [44,45]); those references support the architectural rationale, not the occlusion-robustness conclusion, and the VRNN itself is standard (Chung et al., ref [43]). A real concern is that Section 3.6 and Eq. 21 train the 'unsupervised' keypoint detector jointly with L_cls, while OpenPose is 'pretrained and fixed,' so task supervision may contribute to the occlusion advantage; this is an attribution/confound problem, not circularity, because minimizing L_cls on training data does not by construction determine held-out occlusion recall. The methodology is therefore not equivalent to its own inputs, and no prediction reduces to a fitted parameter by construction.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper's central comparison assumes that both keypoint front ends are equivalent except for representation type. That assumption is strained because the unsupervised detector is jointly optimized with the fall classifier while OpenPose is fixed; several hyperparameters (K, lambda_cls, tau, M/N, augmentation counts) are chosen by hand and affect the comparison, though none is fitted specifically to the occlusion result. No new physical entities are introduced.

free parameters (6)
  • Number of keypoints K = 20
    Chosen for both datasets (Table 9); sets representation dimensionality for both pipelines and affects downstream classifier capacity.
  • Classifier loss scale lambda_cls = 0.1 - 0.5
    Tuned range in Table 9; controls how strongly fall/non-fall labels shape the 'unsupervised' keypoint detector through Eq. 21.
  • Confidence thresholds tau = 0.05 unsupervised, 0.15 supervised
    Hand-chosen thresholds for marking keypoints as missing before interpolation (Sections 3.3.1, 3.3.2).
  • Observed/predicted steps M/N = 12/12
    Defines the half-transmission bandwidth setting in Algorithm 1; not derived from data.
  • Augmentation variants per video = 16 for UR Fall, 8 for Human Fall
    Augmentation counts differ by dataset; augmentation is excluded from OOD test sets, changing train set size and diversity.
  • VRNN KL loss scale beta = 0.0001
    Hyperparameter in Eq. 17; affects how strongly latent-space regularization shapes keypoint forecasting.
assumptions (6)
  • domain assumption The unsupervised keypoint detector architecture of Minderer et al. transfers to fall-detection videos when trained with the losses in Eq. 21.
    The paper defers architecture details to Section S1.1.1 of reference 19 and assumes the detector produces motion-relevant anchors on fall and ADL clips.
  • domain assumption OpenPose, used as a fixed pretrained detector, provides reliable anatomical joints on clean full-body frames.
    The supervised pipeline does not fine-tune OpenPose; if OpenPose is systematically inaccurate even in clean frames, the comparison is biased.
  • domain assumption YOLOv8-m person segmentation isolates the human without introducing representation-dependent bias.
    Segmentation is shared by both pipelines, but segmentation errors in occluded frames could interact differently with the two keypoint detectors.
  • domain assumption Manual identity and visibility annotations for the UR Fall occlusion split are correct.
    Occlusion-based evaluation is built by manual inspection: five fall identities for identity separation and 18 test videos chosen as partially occluded (Section 4.4.3).
  • standard math VRNN/ELBO formulation (Eqs. 13-17) is a valid latent-variable model for keypoint coordinate dynamics.
    The paper uses the standard variational recurrent derivation from reference 43 without re-deriving it.
  • domain assumption Dataset fall/non-fall labels in UR Fall and Human Fall are accurate.
    All classification metrics depend on ground-truth labels; no label audit is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Keypoints for Real-Time Fall Detection: Comparative Analysis Under Real-world Conditions with Predictive Bandwidth Reduction." pith.science (2026). https://pith.science/paper/CLDC3RND

@misc{pith2026260715400,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Keypoints for Real-Time Fall Detection: Comparative Analysis Under Real-world Conditions with Predictive Bandwidth Reduction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CLDC3RND}},
  note         = {Machine review of arXiv:2607.15400}
}
read the original abstract

Falls among older adults are a major safety challenge, but continuous monitoring is difficult to sustain. Video captures fall-related posture and motion, yet deployment is limited by privacy, computation, and bandwidth. Supervised pose estimation is anatomically interpretable but vulnerable to occlusion and partial body visibility. We propose a privacy-preserving framework that replaces RGB transmission with compact motion representations based on unsupervised keypoints and predictive temporal modeling. Local processing performs segmentation and keypoint extraction; variational recurrent prediction and sequence classification then detect falls from observed and forecasted motion. We evaluate the framework on the UR Fall Detection and Human Fall datasets using random, subject-disjoint, and occlusion-based splits. Under random splits, neither representation consistently dominates, suggesting that standard protocols may hide meaningful differences. Under subject-disjoint evaluation, supervised keypoints show a statistically significant advantage, but performance varies by subject: they perform better when anatomical landmarks are visible, whereas unsupervised keypoints are more robust to occlusion and partial visibility, though they produce more false positives for complex activities. Under occlusion-based evaluation, supervised keypoints miss nearly half of all falls, while unsupervised keypoints retain strong sensitivity and substantially outperform them. Their anatomical independence allows spatial anchors to adapt to visible body structure rather than fail on absent landmarks. The gap widens under bandwidth constraints, where supervised localization errors compound through the temporal model. These findings show that representation choice should reflect expected visual conditions and that unsupervised keypoints offer an advantage when body visibility is compromised.

Figures

Figures reproduced from arXiv: 2607.15400 by the authors.

Figure 1
Figure 1. Overview of our proposed privacy-preserving fall-monitoring pipeline [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The modular design of the proposed pipeline [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Supervised and unsupervised keypoints overlaid on video frames from the UR Fall dataset. For visualization, [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Remote patient monitoring using the proposed pipeline [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 3 linked inside Pith

  1. [1]

    Get the facts on falls prevention

    National Council on Aging. Get the facts on falls prevention. Online, 2025. Accessed: 2026-04-22

    2025

  2. [2]

    Osonuga, Ayokunle Osonuga, Deborah Omeni, Gloria C

    Adewoyin A. Osonuga, Ayokunle Osonuga, Deborah Omeni, Gloria C. Okoye, Eghosasere Egbon, and David B. Olawade. Artificial intelligence in hospital fall prevention: Current applications, challenges, and future directions. Safety Science, 196:107104, 2026

    2026

  3. [3]

    Risk factors for hip fracture in very old people: a population-based study

    Robert Wiklund, Annika Toots, Mia Conradsson, Birgitta Olofsson, Henrik Holmberg, Erik Rosendahl, Yngve Gustafson, and Håkan Littbrand. Risk factors for hip fracture in very old people: a population-based study. Osteoporosis International, 27(3):923–931, 2016

    2016

  4. [4]

    Fall-related brain injuries and the risk of dementia in elderly people: a population-based study.European journal of neurology, 12(2):86–92, 2005

    Heikki Luukinen, Petteri Viramo, Mika Herala, Kari Kervinen, YA Kesäniemi, Olli Savola, Satu Winqvist, Jari Jokelainen, and Matti Hillbom. Fall-related brain injuries and the risk of dementia in elderly people: a population-based study.European journal of neurology, 12(2):86–92, 2005

    2005

  5. [5]

    Factors related to fear of falling among community-dwelling older adults.Journal of clinical nursing, 26(1-2):68–76, 2017

    Oanh Tran Thi Hoang, Pornchai Jullamate, Naiyana Piphatvanitcha, and Edwin Rosenberg. Factors related to fear of falling among community-dwelling older adults.Journal of clinical nursing, 26(1-2):68–76, 2017

    2017

  6. [6]

    Schneider, and Bernd Reuschenbach

    Jenny Kubitza, Iris T. Schneider, and Bernd Reuschenbach. Concept of the term long lie: A scoping review. European Review of Aging and Physical Activity, 20(1):16, 2023

    2023

  7. [7]

    Exploring the impact of experiencing a long lie fall on physical and clinical outcomes in older people requiring an ambulance: A systematic review

    Joanna Blackburn, Karen Ousey, John Stephenson, and Sun Chong Lui. Exploring the impact of experiencing a long lie fall on physical and clinical outcomes in older people requiring an ambulance: A systematic review. International Emergency Nursing, 62:101148, 2022

    2022

  8. [8]

    The CDC injury center’s response to the growing public health problem of falls among older adults.American Journal of Lifestyle Medicine, 10(1):74–77, 2016

    Debra Houry, Curtis Florence, Grant Baldwin, Judy Stevens, and Rod McClure. The CDC injury center’s response to the growing public health problem of falls among older adults.American Journal of Lifestyle Medicine, 10(1):74–77, 2016

    2016

Show all 52 references
  1. [9]

    Haddad, Gabrielle F

    Yara K. Haddad, Gabrielle F. Miller, Ramakrishna Kakara, Curtis Florence, Gwen Bergen, Elizabeth R. Burns, and Adam Atherly. Healthcare spending for non-fatal falls among older adults, USA.Injury Prevention, 30(4):272–276, 2024

    2024

  2. [10]

    Falls in the community-dwelling older adult: a review for primary-care providers.Clinical interventions in aging, 2(4):545–553, 2007

    Theresa A Soriano, Linda V DeCherrie, and David C Thomas. Falls in the community-dwelling older adult: a review for primary-care providers.Clinical interventions in aging, 2(4):545–553, 2007

    2007

  3. [11]

    Inquiry concerning the deployment of advanced telecommunications capability to all americans in a reasonable and timely fashion: 2024 section 706 report

    Federal Communications Commission. Inquiry concerning the deployment of advanced telecommunications capability to all americans in a reasonable and timely fashion: 2024 section 706 report. Technical Report FCC 24-27, GN Docket No. 22-270, Federal Communications Commission, Mar...

    2024

  4. [12]

    K. L. Rush, S. Singh, C. L. Seaton, L. Burton, E. Li, C. Jones, J. C. Davis, K. Hasan, B. Kern, and R. Janke. Telehealth use for enhancing the health of rural older adults: A systematic mixed studies review.The Gerontologist, 62(10):e564–e577, 2022

    2022

  5. [13]

    Warrington, Elizabeth J

    Daniel J. Warrington, Elizabeth J. Shortis, and Paula J. Whittaker. Are wearable devices effective for preventing and detecting falls: An umbrella review (a review of systematic reviews).BMC Public Health, 21(1):2091, 2021

  6. [14]

    Rajendra Acharya

    Thanveer Shaik, Xiaohui Tao, Niall Higgins, Lin Li, Raj Gururajan, Xujuan Zhou, and U. Rajendra Acharya. Remote patient monitoring using artificial intelligence: Current state, applications, and challenges.WIREs Data Mining and Knowledge Discovery, 13(2):e1485, 2023

    2023

  7. [15]

    Continuous patient monitoring with AI: Real-time analysis of video in hospital care settings.Frontiers in Imaging, 4:1547166, 2025

    Paolo Gabriel, Peter Rehani, Tyler Troy, Tiffany Wyatt, Michael Choma, and Narinder Singh. Continuous patient monitoring with AI: Real-time analysis of video in hospital care settings.Frontiers in Imaging, 4:1547166, 2025

    2025

  8. [16]

    Sullivan

    Mostafa Habibi, Zeinab Delaram, Mehrdad Nourani, and Dennis H. Sullivan. Video-based human-object interaction analysis for patient behavioral monitoring. In2025 IEEE 13th International Conference on Healthcare Informatics (ICHI), pages 414–422, Rende, Italy, 2025

    2025

  9. [17]

    Acceptance and privacy perceptions toward video-based active and assisted living technologies: Scoping review.Journal of Medical Internet Research, 25:e45297, 2023

    Tamara Mujirishvili, Caterina Maidhof, Francisco Flórez-Revuelta, Martina Ziefle, Miguel Richart-Martínez, and Julio Cabrero-García. Acceptance and privacy perceptions toward video-based active and assisted living technologies: Scoping review.Journal of Medical Internet Resear...

    2023

  10. [18]

    Unsupervised learning of object landmarks through conditional image generation.Advances in neural information processing systems, 31, 2018

    Tomas Jakab, Ankush Gupta, Hakan Bilen, and Andrea Vedaldi. Unsupervised learning of object landmarks through conditional image generation.Advances in neural information processing systems, 31, 2018

    2018

  11. [19]

    Murphy, and Honglak Lee

    Matthias Minderer, Chen Sun, Ruben Villegas, Forrester Cole, Kevin P. Murphy, and Honglak Lee. Unsupervised learning of object structure and dynamics from videos. InAdvances in Neural Information Processing Systems 32 (NeurIPS 2019), pages 92–102, 2019

    2019

  12. [20]

    Alharbi, Naif S

    Imran Ali Abro, Sultan S. Alharbi, Naif S. Alshammari, Abdullah Algarni, Nasser A. Almujally, Asim Jalal, and Hui Liu. Multimodal intelligent biosensors framework for fall disease detection and healthcare monitoring. Frontiers in Bioengineering and Biotechnology, 13:1544968, 2025

    2025

  13. [21]

    Transformer-based fall detection in videos.Engineering Applications of Artificial Intelligence, 132:107937, 2024

    Adrián Núñez-Marcos and Ignacio Arganda-Carreras. Transformer-based fall detection in videos.Engineering Applications of Artificial Intelligence, 132:107937, 2024

    2024

  14. [22]

    Khan, and Alex Mihailidis

    Jacob Nogas, Shehroz S. Khan, and Alex Mihailidis. DeepFall: Non-invasive fall detection with deep spatio- temporal convolutional autoencoders.Journal of Healthcare Informatics Research, 4(1):50–70, 2020

    2020

  15. [23]

    Fall recognition using a three stream spatio-temporal GCN model with adaptive feature aggregation.Scientific Reports, 15:10635, 2025

    Jungpil Shin, Abu Saleh Musa Miah, Rei Egawa, Koki Hirooka, Md Al Mehedi Hasan, Yoichi Tomioka, and Yong Seok Hwang. Fall recognition using a three stream spatio-temporal GCN model with adaptive feature aggregation.Scientific Reports, 15:10635, 2025

    2025

  16. [24]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  17. [25]

    2d human pose estimation: A survey.Multimedia systems, 29(5):3115–3138, 2023

    Haoming Chen, Runyang Feng, Sifan Wu, Hao Xu, Fengcheng Zhou, and Zhenguang Liu. 2d human pose estimation: A survey.Multimedia systems, 29(5):3115–3138, 2023

    2023

  18. [26]

    A survey on deep learning-based 2d human pose estimation models.Computers, Materials, & Continua, 76(2):2385, 2023

    Sani Salisu, M Jaafar, S Ainun, Hussain Younis, et al. A survey on deep learning-based 2d human pose estimation models.Computers, Materials, & Continua, 76(2):2385, 2023

    2023

  19. [27]

    Edward R. Sykes. Next-generation fall detection: Harnessing human pose estimation and transformer technology. Health Systems, 14(2):85–103, 2024

    2024

  20. [28]

    Privacy preserving human fall detection using video data

    Umar Asif, Benjamin Mashford, Stefan V on Cavallar, Shivanthan Yohanandan, Subhrajit Roy, Jianbin Tang, and Stefan Harrer. Privacy preserving human fall detection using video data. InMachine Learning for Health Workshop, pages 39–51. PMLR, 2020

    2020

  21. [29]

    Skeleton-based privacy-preserving smart activity sensor for senior care and patient monitoring

    Jie Liang, Andrew Au, Minghua Chen, Cyrus Chan, Jiannan Zheng, Zachary DeVries, Ying Xiao, and Paeton Dhesi. Skeleton-based privacy-preserving smart activity sensor for senior care and patient monitoring. 2024

    2024

  22. [30]

    A real-time skeleton-based fall detection algorithm based on temporal convolutional networks and transformer encoder.Pervasive and Mobile Computing, 107:102016, 2025

    Xiaoqun Yu, Chenfeng Wang, Wenyu Wu, and Shuping Xiong. A real-time skeleton-based fall detection algorithm based on temporal convolutional networks and transformer encoder.Pervasive and Mobile Computing, 107:102016, 2025

    2025

  23. [31]

    Privacy-preserving approach for early detection of long-lie incidents: A pilot study with healthy subjects.Sensors, 25(12):3836, 2025

    Riska Analia, Anne Forster, Sheng-Quan Xie, and Zhiqiang Zhang. Privacy-preserving approach for early detection of long-lie incidents: A pilot study with healthy subjects.Sensors, 25(12):3836, 2025

    2025

  24. [32]

    A framework for fall detection based on openpose skeleton and lstm/gru models.Applied Sciences, 11(1):329, 2020

    Chuan-Bi Lin, Ziqian Dong, Wei-Kai Kuan, and Yung-Fa Huang. A framework for fall detection based on openpose skeleton and lstm/gru models.Applied Sciences, 11(1):329, 2020. 24 APREPRINT- JULY20, 2026

    2020

  25. [33]

    OpenPose: Realtime multi-person 2d pose estimation using part affinity fields.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172–186, 2021

    Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. OpenPose: Realtime multi-person 2d pose estimation using part affinity fields.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172–186, 2021

    2021

  26. [34]

    Posenet: A convolutional network for real-time 6-dof camera relocalization

    Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. InProceedings of the IEEE international conference on computer vision, pages 2938–2946, 2015

    2015

  27. [35]

    Mediapipe solutions guide

    Google. Mediapipe solutions guide. https://ai.google.dev/edge/mediapipe/solutions/guide, 2024. Accessed: 2026-04-28

    2024

  28. [36]

    HEViT- Pose: Towards high-accuracy and efficient 2d human pose estimation with cascaded group spatial reduction attention.Scientific Reports, 16(1):5637, 2026

    Chengpeng Wu, Zhidong Chen, Beihua Ying, Guangxing Tan, Bing Hu, Chunyu Li, and Haifeng Chen. HEViT- Pose: Towards high-accuracy and efficient 2d human pose estimation with cascaded group spatial reduction attention.Scientific Reports, 16(1):5637, 2026

    2026

  29. [37]

    Explore ultralytics yolov8, 2023

    Ultralytics. Explore ultralytics yolov8, 2023

    2023

  30. [38]

    Sina Mehdizadeh, Hoda Nabavi, Andrea Sabo, Twinkle Arora, Andrea Iaboni, and Babak Taati. Concurrent validity of human pose tracking in video for measuring gait parameters in older adults: a preliminary analysis with multiple trackers, viewing angles, and walking directions.Jo...

    2021

  31. [39]

    Skeleton-based human action recognition using spatial temporal 3d convolutional neural networks

    Juanhui Tu, Mengyuan Liu, and Hong Liu. Skeleton-based human action recognition using spatial temporal 3d convolutional neural networks. In2018 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6, 2018

    2018

  32. [40]

    Sensor data augmentation from skeleton pose sequences for improving human activity recognition

    Parham Zolfaghari, Vitor Fortes Rey, Lala Ray, Hyun Kim, Sungho Suh, and Paul Lukowicz. Sensor data augmentation from skeleton pose sequences for improving human activity recognition. In2024 International Conference on Activity and Behavior Computing (ABC), pages 1–8. IEEE, 2024

    2024

  33. [41]

    Towards the extraction of robust sign embeddings for low resource sign language recognition.arXiv preprint arXiv:2306.17558, 2023

    Mathieu De Coster, Ellen Rushe, Ruth Holmes, Anthony Ventresque, and Joni Dambre. Towards the extraction of robust sign embeddings for low resource sign language recognition.arXiv preprint arXiv:2306.17558, 2023

    2023 arXiv

  34. [42]

    Co-occurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation.arXiv preprint arXiv:1804.06055, 2018

    Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. Co-occurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation.arXiv preprint arXiv:1804.06055, 2018

    2018 arXiv

  35. [43]

    Courville, and Yoshua Bengio

    Junyoung Chung, Kyle Kastner, Laurent Dinh, Kratarth Goel, Aaron C. Courville, and Yoshua Bengio. A recurrent latent variable model for sequential data. InAdvances in Neural Information Processing Systems 28 (NeurIPS 2015), pages 2980–2988, 2015

    2015

  36. [44]

    Towards efficient real-time video motion transfer via generative time series modeling.Multimedia Tools and Applications, 85(2):51, 2026

    Tasmiah Haque, Md Asif Bin Syed, Byungheon Jeong, Xue Bai, Sumit Mohan, Somdyuti Paul, Imtiaz Ahmed, and Srinjoy Das. Towards efficient real-time video motion transfer via generative time series modeling.Multimedia Tools and Applications, 85(2):51, 2026

    2026

  37. [45]

    Inference-time stochastic refinement of gru-normalizing flow for real-time video motion transfer.arXiv preprint arXiv:2512.04282, 2025

    Tasmiah Haque and Srinjoy Das. Inference-time stochastic refinement of gru-normalizing flow for real-time video motion transfer.arXiv preprint arXiv:2512.04282, 2025

    2025

  38. [46]

    Long short-term memory.Neural computation, 9(8):1735–1780, 1997

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory.Neural computation, 9(8):1735–1780, 1997

    1997

  39. [47]

    Human fall detection on embedded platform using depth maps and wireless accelerometer.Computer methods and programs in biomedicine, 117(3):489–501, 2014

    Bogdan Kwolek and Michal Kepski. Human fall detection on embedded platform using depth maps and wireless accelerometer.Computer methods and programs in biomedicine, 117(3):489–501, 2014

    2014

  40. [48]

    Gmdcsa-24: A dataset for human fall detection in videos.Data in Brief, 57:110892, 2024

    Ekram Alam, Abu Sufian, Paramartha Dutta, Marco Leo, and Ibrahim A Hameed. Gmdcsa-24: A dataset for human fall detection in videos.Data in Brief, 57:110892, 2024

    2024

  41. [49]

    Individual comparisons by ranking methods

    Frank Wilcoxon. Individual comparisons by ranking methods. InBreakthroughs in statistics: Methodology and distribution, pages 196–202. Springer, 1992

    1992

  42. [50]

    An empirical study on low gpu utilization of deep learning jobs

    Yanjie Gao, Yichen He, Xinze Li, Bo Zhao, Haoxiang Lin, Yoyo Liang, Jing Zhong, Hongyu Zhang, Jingzhou Wang, Yonghua Zeng, et al. An empirical study on low gpu utilization of deep learning jobs. InProceedings of the IEEE/ACM 46th International Conference on Software Engineerin...

    2024

  43. [51]

    Real-time 2d multi-person pose estimation on cpu: Lightweight openpose.arXiv preprint arXiv:1811.12004, 2018

    Daniil Osokin. Real-time 2d multi-person pose estimation on cpu: Lightweight openpose.arXiv preprint arXiv:1811.12004, 2018

    2018 arXiv

  44. [52]

    Cisco Webex. What are the minimum bandwidth requirements for sending and re- ceiving video in cisco webex meetings? https://help.webex.com/en-us/article/ WBX22158/What-are-the-Minimum-Bandwidth-Requirements-for-Sending-and-Receiving\ -Video-inCisco-Webex-Meetings , 2023. Webex...

    2023

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.