A Comparative Analysis of Pose Estimation Models and a Proposed MediaPipe-Based Framework for Vision-Based Fall Detection

Abstract

Vision-based fall detection increasingly relies on human pose estimation to convert raw video into structured skeletal data for posture and motion analysis. This paper presents a literature-based comparative study of five widely used pose estimation models — MediaPipe Pose Landmarker, OpenPose, AlphaPose, MoveNet, and HRNet — examined across architecture, accuracy, speed, computational cost, edge-deployment readiness, multi-person capability, and healthcare suitability. The comparison surfaces a research gap: existing fall-detection literature rarely justifies its choice of pose-estimation backbone through systematic comparison, and almost no published study evaluates these backbones together under fall-specific, resource-constrained conditions. Motivated by this gap, and by published evidence favoring lightweight, CPU-feasible architectures for continuous healthcare monitoring, this paper proposes a conceptual prototype framework for fall detection built around MediaPipe Pose Landmarker, combining pose-based posture analysis with temporal inactivity reasoning to suppressfalsealarms.Theproposedframeworkispresentedstrictlyatthearchitecturalandworkflowlevel;noexperimentalimplementation, datasetevaluation,orperformanceresultsarereportedhere.Thispaperestablishestheliterature-grounded foundation and prototype design that will be implemented and experimentally validated in a subsequent study.

Introduction

1. Falls are a major cause of accidental injuries among older adults, making early and accurate fall detection essential for timely medical assistance. Traditional fall detection methods include wearable sensors, ambient sensors, and vision-based systems. Vision-based approaches are particularly useful because they do not require users to wear additional devices and can protect privacy by processing body-skeleton keypoints instead of transmitting raw video.

   This paper reviews five human pose-estimation models used in vision-based fall detection: MediaPipe Pose Landmarker, OpenPose, AlphaPose, MoveNet, and HRNet. These models differ in accuracy, processing speed, computational requirements, multi-person detection, and suitability for edge devices. The comparison suggests that MediaPipe Pose Landmarker provides a good balance of real-time performance, low computational cost, CPU compatibility, and healthcare-related validation, making it suitable for the proposed system.

   The paper identifies three major research gaps: the lack of systematic comparison when selecting pose-estimation models, limited evaluation using actual fall-related postures, and inconsistent reporting of computational performance across different hardware platforms.

   To address these gaps, a conceptual AI-based fall detection framework is proposed using MediaPipe Pose Landmarker. The system captures continuous video through a camera, detects human body keypoints, and extracts features such as joint angles, body orientation, and body-centroid movement. It combines posture analysis with temporal inactivity monitoring to distinguish actual falls from normal activities such as sitting, bending, or lying down. Once a fall is confirmed, the system generates an alert for a caregiver or notification service.

   The proposed framework is designed to operate locally on a CPU-based laptop or embedded device without requiring a dedicated GPU. This reduces computational costs and improves privacy by avoiding continuous transmission of video.

   The main contribution of this research is a comparative analysis of five pose-estimation models, justification for selecting MediaPipe, identification of existing research gaps, and the design of a privacy-conscious, real-time fall detection architecture.

   An important limitation: This paper presents only the literature review and conceptual framework. The system has not yet been implemented or experimentally tested, so no actual accuracy or processing-time results are claimed. Future work will involve implementation, testing on fall-detection datasets, and comparison of the five models using accuracy, precision, recall, and latency.

   Key points

   • Aim: Develop an AI-based vision system for detecting falls among older adults.

   • Selected model: MediaPipe Pose Landmarker.

   • Technology: Computer Vision, Human Pose Estimation, Python, and edge computing.

   • Detection approach: Skeleton-based posture analysis combined with inactivity monitoring.

   • Privacy: Local video processing without continuous video transmission.

   • Expected outcome: Faster fall detection and automatic caregiver alerts.

   • Future scope: Practical implementation, real-time testing, and performance comparison.

    

Conclusion

This paper compared five pose-estimation architectures for vision-based fall detection — MediaPipe Pose Landmarker, OpenPose, AlphaPose, MoveNet, and HRNet — and identified aresearchgapconcerningsystematicbackbonejustificationand fall-specific, same-hardware evaluation. Building on this analysis and on our project synopsis, a conceptual prototype framework centered on MediaPipe Pose Landmarker was proposed, combining posture analysis with temporal inactivity reasoning within a CPU-only, edge-deployable architecture. This conclusion is intentionally bounded: the proposed framework has not been implemented or experimentally validated,andnoaccuracy,latency,orotherperformancefigures areclaimedinthispaper.Futurework,tobereportedinasecond paper, involves implementing the framework described in Section V, evaluating it experimentally on fall-detection benchmarks,andreportingprecision,recall,andlatencyresults— together with a same-hardware comparison against the alternative backbones discussed in Section III.

References

[1] V.Bazarevsky,I.Grishchenko,K.Raveendran,T.Zhu,F.Zhang,andM.Grundmann,\"BlazePose:On-devicereal timebodypose tracking,\"arXiv:2006.10204,2020. [2] Z.Cao,G.Hidalgo,T.Simon,S.E.Wei,andY.Sheikh,\"OpenPose:Realtimemultiperson2Dposeestimationusingpartaffinityfields,\"IEEE Trans. Pattern Anal. Mach. Intell., vol. 43, no. 1, pp. 172-186,2021. [3] H.-S.Fang,J.Li,H.Tang,C.Xu,H.Zhu,Y.Xiu,Y.-L.Li,andC.Lu,\"AlphaPose: Whole-body regional multi-person pose estimation andtrackinginrealtime,\"IEEETrans.PatternAnal.Mach.Intell.,vol.45,no. 6, pp. 7157-7173, 2023. [4] J. Shin, A. S. M. Miah, R. Egawa, K. Hirooka, M. A. M. Hasan, Y.Tomioka, and Y. S. Hwang, \"Fall recognition using a three-streamspatio-temporalGCNmodelwithadaptivefeatureaggregation,\"Sci.Rep.,2025. [5] K. Sun, B. Xiao, D. Liu, and J. Wang, \"Deep high-resolutionrepresentationlearningforhumanposeestimation,\"inProc.IEEECVPR, 2019, pp. 5693-5703. [6] B. Jo and S. Kim, \"Comparative analysis of OpenPose, PoseNet, andMoveNetmodelsforposeestimationinmobiledevices,\"TraitementduSignal, vol. 39, no. 1, pp. 119-124, 2022. [7] Google,\"PoseestimationandclassificationonedgedeviceswithMoveNet and TensorFlow Lite,\" TensorFlow Blog, 2021. [8] H.-S. Fang, S. Xie, Y.-W. Tai, and C. Lu, \"RMPE: Regional multi-personposeestimation,\"inProc.IEEEICCV,2017,pp.2334-2343. [9] Z.Cao,T.Simon,S.-E.Wei,andY.Sheikh,\"Realtimemulti-person2Dpose estimation using part affinity fields,\" in Proc. IEEE CVPR, 2017,pp.1302-1310. [10] C.Zheng,W.Wu,C. Chen,T.Yang,S.Zhu,J.Shen,N.Kehtarnavaz,andM.Shah,\"Deeplearning-basedhumanposeestimation:Asurvey,\"ACM Comput. Surv., vol. 56, no. 1, 2023. [11] E.Samkari,M.Arif,M.Alghamdi,andM.A.AlGhamdi,\"Humanposeestimationusingdeeplearning:Asystematicliteraturereview,\"Mach. Learn. Knowl. Extr., vol. 5, no. 4, pp. 1612-1659, 2023. [12] Z. Huang, Y. Liu, Y. Fang, and B. K. P. Horn, \"Video-based falldetectionforseniorswithhumanposeestimation,\"inProc.4thInt.Conf. Universal Village (UV), 2018, pp. 1-4. [13] Y. Wang and T. Deng, \"Enhancing elderly care: Efficient and reliablereal-timefalldetectionalgorithm,\"DIGITALHEALTH,vol.10,2024. [14] T.Riahi,M.A.H.Bappy,andM.M.Islam,\"ElderFallGuard:Real-timeIoTandcomputervision-basedfalldetectionsystemforelderlysafety,\" arXiv:2505.11845, 2025. [15] E.R.Sykes,\"Next-generationfalldetection:Harnessinghumanposeestimation and transformer technology,\" Health Systems, 2024. [16] V.MaliandS.Jaiswal,\"Pose-basedfalldetectionsystem:Efficientmonitoring on standard CPUs,\" arXiv:2503.19501, 2025. [17] H.ZhengandY.Liu,\"LightweightfalldetectionalgorithmbasedonAlphaPoseoptimizationmodelandST-GCN,\"Math.Probl.Eng.,vol.2022,2022. [18] V. Kong, S. Soeng, M. Thon, W.-S. Cho, A. Nayyar, and T.-K. Kim,\"PIFR: A novel approach for analyzing pose angle-based humanactivitytoautomatefalldetectioninvideos,\"PLoSOne,vol.20,no.6,2025. [19] E.Alam,\"GMDCSA-24:Adatasetforhumanfalldetectioninvideos,\"2024. [20] T.S.Kavya,\"Visionbasedreal-timemonitoringsystemforelderlyfall detection,\" Comput. Syst. Sci. Eng., vol. 42, no. 1, 2023. [21] P.Jesudhas,\"Computationallyefficientnon-intrusivepre-impactfalldetection system,\" arXiv:2507.03705, 2025. [22] C.S.T.Hii,K.B.Gan,N.Zainal,N.MohamedIbrahim,S.Azmin,S. [23] H.MatDesa,B.vandeWarrenburg,andH.W.You,\"Automatedgaitanalysisbasedonamarker-freeposeestimationmodel,\"Sensors,vol.23, no. 14, 2023. [24] V.Wagh,M.W.Scott,J.W.Andrushko,C.B.Jones,B.C. Larssen, L. A. Boyd, and S. N. Kraeutner, \"Using MediaPipe to track upper-limbreachingmovementsafterstroke:Aproof-of-principlestudy,\"J.NeuroEng. Rehabil., vol. 22, no. 1, 2025.

Copyright

Copyright © 2026 Rohan Singh, Nupur Thakkar, Vedant Lade, Siddhesh Bhiwapurkar. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.