Robust 3D Skeletal Joint Fall Detection in Occluded and Rotated Views Using Data Augmentation and Inference–Time Aggregation
Open accessAbstract
Fall detection systems are a critical application of human pose estimation, frequently struggle with achieving real-world robustness due to their reliance on domain-specific datasets and a limited capacity for generalization to novel conditions. Models trained on controlled, canonical camera views often fail when subjects are viewed from new perspectives or are partially occluded, resulting in missed detections or false positives. This study tackles these limitations by proposing the Viewpoint Invariant Robust Aggregation Graph Convolutional Network (VIRA-GCN), an adaptation of the Richly Activated GCN for fall detection. The VIRA-GCN introduces a novel dual-strategy solution: a synthetic viewpoint generation process to augment training data and an efficient inference-time aggregation method to form consensus-based predictions. We demonstrate that augmenting the Le2i dataset with simulated rotations and occlusions allows a standard pose estimation model to achieve a significant increase in its fall detection capabilities. The VIRA-GCN achieved 99.81% accuracy on the Le2i dataset, confirming its enhanced robustness. Furthermore, the model is suitable for low-resource deployment, utilizing only 4.06 M parameters and achieving a real-time inference latency of 7.50 ms. This work presents a practical and efficient solution for developing a single-camera fall detection system robust to viewpoint variations, and introduces a reusable mapping function to convert Kinect data to the MMPose format, ensuring consistent comparison with state-of-the-art models.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
A rapid transition of trunk inclination from 0° to 90° served as a primary fall trigger.
“A rapid transition from 0∘ (vertical) to 90∘ (horizontal) served as a primary fall trigger [6].”
From Safe Fall: Use of Predictive Modeling and Machine Vision Techniques for Fall Analysis and Fall Quality (2026) · cites this paper as [6]
The Methodology (§3.3; page unavailable) identifies rapid changes in vertical displacement as fall indicators. Related Work (§2; page unavailable) describes trunk angle as a verification measure in an earlier system, but says that system first checks an accelerometer threshold. Neither passage supports a rapid 0°-to-90° trunk transition as the primary fall trigger. (LLM, conf=0.82)
Experimental — accuracy not yet measuredThe viewpoint was specifically selected to maximise the visibility of critical anatomical landmarks and minimise self-occlusion during the rotational phase of backward rolling falls.
“This viewpoint was specifically selected to maximise the visibility of critical anatomical landmarks (head, hips, knees) and minimise self-occlusion during the rotational phase of backward rolling falls (Ushiro-Ukemi) (see Figure 1), a challenge noted in recent pose estimation benchmarks [6].”
From Safe Fall: Use of Predictive Modeling and Machine Vision Techniques for Fall Analysis and Fall Quality (2026) · cites this paper as [6]
Standard pose estimation frameworks, such as OpenPose or MediaPipe, frequently suffer from tracking degradation in high-velocity, non-linear scenarios, resulting in lost keypoints and identity switching.
“Standard pose estimation frameworks, such as OpenPose or MediaPipe, frequently suffer from tracking degradation in these high-velocity, non-linear scenarios, resulting in lost keypoints and identity switching [5,6].”
From Safe Fall: Use of Predictive Modeling and Machine Vision Techniques for Fall Analysis and Fall Quality (2026) · cites this paper as [6]
Section 3.2 (page unavailable) discusses OpenPose’s computational cost and the choice of MMPose, but not tracking failures during high-velocity, non-linear motion. Searches for lost keypoints, identity switching, and fast-motion tracking found no passage addressing those claims for OpenPose or MediaPipe. (LLM, conf=0.73)
Experimental — accuracy not yet measuredThe ceiling effect confirms that the vision pipeline (YOLOv8 + SAM 2) is robust enough to track subjects perfectly in both chaotic and controlled scenarios, validating the tool for real-world usage.
“This ceiling effect confirms that our vision pipeline (YOLOv8 + SAM 2) is robust enough to track subjects perfectly in both chaotic and controlled scenarios, validating the tool for real-world usage [6].”
From Safe Fall: Use of Predictive Modeling and Machine Vision Techniques for Fall Analysis and Fall Quality (2026) · cites this paper as [6]
The Conclusions report strong fall-detection performance for VIRA-GCN and discuss its readiness for real-world deployment. The Methodology identifies MMPose as its pose-estimation framework; neither those passages nor searches for YOLOv8, SAM 2, a ceiling effect, and perfect subject tracking address the citing paper’s pipeline or its claimed tracking results. (LLM, conf=0.78)
Experimental — accuracy not yet measuredThis validates the robustness of the YOLOv8 + SAM 2 pipeline, demonstrating that tracking remains reliable even during severe self-occlusions.
“This validates the robustness of our YOLOv8 + SAM 2 pipeline, demonstrating that tracking remains reliable even during the severe self-occlusions typical of a backward roll (where legs obscure the torso) [6].”
From Safe Fall: Use of Predictive Modeling and Machine Vision Techniques for Fall Analysis and Fall Quality (2026) · cites this paper as [6]
Section 3.1 describes simulated occlusions used to train a skeletal fall-detection model, and Section 6 reports robustness of VIRA-GCN. Neither passage evaluates YOLOv8 + SAM 2 tracking or a backward roll. Searches for those terms and for self-occlusion found no passage establishing the attributed claim. (LLM, conf=0.78)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 23 · Cited by 6
Research works printed but not linked: 9
- [3]Nazar M. Ahsan A. Khan M.I. A review of vision-based fall detection systems for elderly people: A performance analysis J. King Saud-Univ. Comput. Inf. Sci. 2020 32 808 819
- [6]Sannino M.M. De Masi L. D’Angelo V. Sannino L. D’Agnese M. Sannino F. Fall Detection with Multiple Cameras: An Occlusion-Resistant Method Based on 3-D Silhouette Vertical Distribution IEEE J. Biomed. Health Inform. 2020 24 262 273 20952341
- [10]Cao Z. Hidalgo G. Simon T. Wei S.-E. Sheikh Y. OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields IEEE Trans. Pattern Anal. Mach. Intell. 2019 43 150 166 10.1109/TPAMI.2019.2929257 31331883 · DOI · Paper page
- [26]Kepski M. Kwolek B. A New Fall Detection Dataset and the Comparison of Fall Detection Methods Proceedings of the 2019 International Conference on Computer Vision and Graphics Warsaw, Poland 16–18 September 2019 Springer Cham, Switzerland 2019 11 22
- [29]TST Fall Detection Dataset v2. IEEE Dataport, 2016 Available online: https://ieee-dataport.org/documents/tst-fall-detection-dataset-v2 (accessed on 19 October 2025)
- [37]Li W. Wang Y. Wang Y. Li B. Zhang H. Wu Y. Lightweight Spatio-Temporal Graph Convolutional Networks for Skeleton-based Action Recognition Proceedings of the 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) Virtual 6–12 June 2021 1605 1609
- [38]Alaoui A. El-Yaagoubi M. Chehri A. V2V-PoseNet: A View-Invariant Pose Estimation Network for Action Recognition Proceedings of the International Conference on Advanced Communication Technologies and Information Systems (ICACTIS) Coimbatore, India 19–20 March 2021 1 6
- [41]Alanazi S. Alshahrani S. Alqahtani A. Alghamdi A. 3D Multi-Stream CNNs for Real-Time Fall Detection IEEE Access 2023 11 15993 16004
- [42]Shin J. Kim S. Kim D. Three-Stream Spatio-Temporal Graph Convolutional Networks for Human Fall Detection Proceedings of the 2024 International Conference on Information and Communication Technology Convergence (ICTC) Jeju Island, Republic of Korea 16–18 October 2024 1205 1210
- Authors
- Zobi, M., Bolzani, L., Tabii, Y., & Oulad Haj Thami, R.
- Affiliations
- Mohammed V University (MA)
- flAIght, Viale dell’Innovazione 13, 20126 Milan, Italy
- ADMIR Laboratory, National School of Computer Science and Systems Analysis, Mohammed V University, Rabat 10000, Morocco
- Corresponding author
- Zobi, Maryem
- Venue
- Sensors
- Publisher
- MDPI AG
- Published
- 2025-11-06
- Volume
- 25
- Issue
- 21
- Pages
- 6783
- ISSN
- 1424-8220 (ISSN-L)
- PMC
- PMC12609388
- OpenAlex
- W4415952832
- Cited by
- 6
- References
- 23
- Citations by year
- 2026: 6
- FWCI
- 2.09
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Sensors (publishedVersion, cc-by)
- Europe PMC (PubMed Central) (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Human Pose and Action Recognition
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
- Health Sciences › Health Professions › Physical Therapy, Sports Therapy and Rehabilitation › Balance, Gait, and Falls Prevention
- Keywords
- fall detection, human pose estimation, viewpoint invariance, occlusion robustness, data augmentation, consensus forecasts, graph convolutional networks, low-resource deployment, real-time inference
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
