
Smartphone motion analysis: what 2D pose estimation measures, and with what error
An iPhone can now locate joints on video without markers. The clinical question is a different one: for which angles, in which planes and with what error do those numbers come close to a 3D motion analysis laboratory?
- With the camera in the sagittal plane, 2D pose estimation typically falls within 4–6° of marker-based capture at the hip and knee during gait 1,2.
- The ankle is the sagittal joint with the largest error: 7.4° during gait 1.
- FPPA measured in 2D agrees well across 2D methods but correlates poorly with knee abduction measured in 3D (r = 0.20) 3.
- Transverse-plane rotations remain the weak point, even in multi-camera systems 4.
- OpenCap achieves a mean error of 4.5° with two calibrated iPhones; this figure does not transfer to a single camera 5.
- Keep the camera perpendicular to the plane of movement and level: oblique views produced biases of up to 26° 6,7.
Why validity matters
Motion analysis with reflective markers and multiple infrared cameras remains the laboratory reference, but it requires equipment, time and specialist staff. Markerless pose estimation promises to bring part of that information into the gym and the clinic through a smartphone camera 8. A 2022 review summarised the state of the art cautiously: spatiotemporal parameters appear equivalent to those of marker-based systems, but joint centre locations and joint angles are not yet accurate enough for clinical applications 9.
For the clinician, the practical question is not whether the smartphone ‘works’, but which variables can be trusted, under which capture conditions and for which decision: comparing sides, tracking progress or classifying risk.
How pose estimation works on a smartphone
Pose estimation models are neural networks that locate anatomical landmarks in each frame; angles are then calculated from those landmarks. BlazePose, the model family behind MediaPipe, detects 33 body landmarks and runs at more than 30 frames per second on a mobile phone 10. It was designed for fitness and real-time use, not as a biomechanics instrument, so its accuracy has to be tested separately.
Two problems recur in the literature. The first is systematic offset: when OpenPose, AlphaPose and DeepLabCut were compared with marker-based capture, hip and knee joint centres differed by about 30–50 mm, most likely because of annotation errors in the training data 11. The second is occasional gross failure: in a multi-camera OpenPose system, 80% of mean errors were below 30 mm, but 10% exceeded 40 mm, mainly when the algorithm mistook objects for body segments or swapped segments with each other 12.
A constant offset has little effect when a person is compared with themselves; an occasional failure, if it goes undetected, can spoil an entire repetition.
Accuracy by joint and plane
The evidence closest to a single-iPhone application comes from studies using one camera and 2D analysis. The pattern is consistent: the sagittal plane is the most reliable, the frontal plane is acceptable as a measure of 2D alignment, and the transverse plane cannot be measured with one camera.
| Measure | What the evidence shows | Confidence |
|---|---|---|
| Hip flexion/extension (sagittal) | Mean absolute error of 4.0° during gait 1; 3.7 ± 1.3° with OpenPose 2; r = 0.68–0.94 against 3D in the single-leg squat 3 | High |
| Knee flexion/extension (sagittal) | 5.6° 1 and 5.1 ± 2.5° 2 during gait; MediaPipe accurate when the movement is large and the joint is visible 13 | High, with frame-level quality control |
| Ankle (sagittal) | 7.4° during gait 1; misidentification of anatomical landmarks 13 | Moderate to low |
| Knee in the frontal plane (FPPA) | r = 0.95 between 2D methods, but r = 0.20 against 3D abduction 3; r = 0.308 in the single-leg squat 14 | Good as 2D alignment; not equivalent to 3D valgus |
| Rotations (transverse plane) | Excluded from the sub-5.5° error even in an 8-camera system 4 | Do not report |
| Shoulder (Apple Vision, iPhone) | R² > 0.93 against 3D capture, with overestimation of range; SEM of 9.2° for abduction 15 | Moderate |
Studies mainly in healthy adults or adolescents; error depends on the pose model, the device and the capture conditions.
Models are not interchangeable. In a comparison of four open-source models, the lightest ones, designed for mobile phones, were not the most accurate 2. In a sample of older adults, MediaPipe performed worst of the methods tested for gait, although it was accurate at the knee 13. In a recent study using BlazePose and lateral smartphone capture, the ICC for the squat was 0.773 when low-confidence frames were excluded, but fell to 0.428, with an RMSE of 25.43°, without that filter; 15.7% of dynamic observations did not meet the confidence criterion 16. Quality control of each frame determines whether the number is usable.
The iPhone’s native 3D tools also have limits. ARKit 3D body tracking showed a weighted mean error of 18.80° ± 12.12° against Vicon, ranging from 3.75° to 47.06° depending on joint and exercise 17. Apple Vision 2D pose, tested at the shoulder, followed the movement pattern well but overestimated range of motion 15.
OpenCap and other iPhone precedents
OpenCap is the most robust precedent for smartphone motion analysis. With two calibrated iPhones, it estimated joint angles with a mean absolute error of 4.5° against marker-based capture (4.1° in gait and squatting, 4.7° in sit-to-stand and 5.1° in the drop jump), ground reaction forces with an error of 6.2% of body weight and joint moments with an error of 1.2% of body weight × height 5. In a field study with 100 participants, it was 25 times faster than the laboratory approach, at less than 1% of the cost 5. The essential condition is geometry: two or more cameras allow movement to be reconstructed in 3D; one camera does not.
For other measures, the iPhone already has solid validation. The My Jump app measured countermovement jump height with an ICC of 0.997 against a force plate, with a bias of 1.1 ± 0.5 cm 18, and MySprint measured sprint split times with an ICC of 1.0 against photocells 19. Both measure well-defined temporal events, which are easier to validate than angles. A review of sports science apps offers a useful warning: many were tested on a single device 20.
How to film: good capture practice
- Camera perpendicular to the plane of movement. Kinovea was valid between 90° and 45° of perspective, up to 5 m, but 90° is recommended for optimal results 6. With oblique cameras, biases ranged from −0.35° to −26.07° depending on joint and view 7.
- Phone level and untilted. Tilting the phone downwards or upwards changed the measured range in opposite directions; in the reference study, the iPhone was 0.9 m high and about 3 m from the participant, levelled with a spirit-level app 15.
- Whole body in frame, one person only and an uncluttered background. Errors above 40 mm resulted from objects mistaken for segments and from swapped segments 12.
- Several repetitions or steps. In running, at least 7 steps are needed for a stable average 21; in gait, per-participant averages were more accurate than single steps 1.
- Close-fitting clothing and even lighting. Error increased with different clothing (below 11°) and lighting (below 9°), although without simultaneous comparison with 3D capture 22.
- Frame rate. The iPhone jump and sprint studies used high-speed video 18,19, and the shoulder study used 30 fps for slow movements 15. We found no studies directly comparing 30, 60 and 240 fps for joint angles: preferring higher frame rates for jumps, landings and sprints is a practical recommendation, not a published finding.
At reassessment, the most important rule is to repeat exactly the same set-up: distance, height, view, footwear and clothing. Changing the view changes the number 7.
What this means for clinical practice
2D video measures angles projected onto a plane. FPPA, for example, was only weakly to moderately associated with hip adduction (r = 0.32–0.38) and knee external rotation (r = 0.48–0.55) measured in 3D, and the authors concluded that it should not be used to quantify 3D rotations 23. Even so, 2D ranks athletes reasonably well: between subjects, the 2D frontal-plane angle correlated with 3D valgus with r² of 0.58–0.64, but within the same subject only with r² of 0.25–0.36 24. 2D is therefore better suited to identifying extreme values than to detecting small changes.
Automation itself does not appear to be the problem: in the frontal plane, automatic FPPA agreed well with manual video measurement (CCC of 0.89 with BlazePose and 0.95 with OpenPose) 25. The limit lies in the 2D projection, not in the algorithm. In addition, most validation studies included healthy adults; evidence in clinical populations, children and pathological gait is scarce.
In summary, with a well-positioned camera, sagittal-plane hip and knee flexion is accurate enough to compare sides and track progress; the frontal plane is useful as a measure of 2D alignment; rotations should not be reported. Any change should be compared with the test’s measurement error before it is credited, and diagnosis remains with the clinician.
Frequently asked questions
Is smartphone motion analysis valid?
It depends on the variable. For sagittal-plane hip and knee flexion, single-camera error is typically 4–6° during gait 1,2; in the frontal plane and for rotations, agreement with 3D analysis is much lower 3.
What is the difference between 2D video and 3D motion analysis?
2D video measures angles projected onto the camera plane; 3D analysis reconstructs the true position of the segments. 2D FPPA should not be used to quantify 3D rotations 23.
Is MediaPipe accurate for measuring joint angles?
At the knee, with large movements and good visibility, it can be accurate 13; without frame-level quality control, RMSE reached 25.43° in the squat 16. The ankle is the most problematic joint.
How should I position my iPhone to film a movement test?
Perpendicular to the plane of movement, level, with the whole body in frame, and always with the same set-up at reassessment 6,15.
Is OpenCap comparable to a single-iPhone app?
No. OpenCap uses at least two calibrated iOS devices to reconstruct movement in 3D, with a mean joint-angle error of 4.5° 5; a single camera measures projected 2D angles.
References
- Stenum J, Rossi C, Roemmich RT. Two-dimensional video-based analysis of human gait using pose estimation. PLoS Comput Biol. 2021;17(4):e1008935. doi.org/10.1371/journal.pcbi.1008935
- Washabaugh EP, Shanmugam TA, Ranganathan R, Krishnan C. Comparing the accuracy of open-source pose estimation methods for measuring gait kinematics. Gait Posture. 2022;97:188–195. doi.org/10.1016/j.gaitpost.2022.08.008
- Haberkamp LD, Garcia MC, Bazett-Jones DM. Validity of an artificial intelligence, human pose estimation model for measuring single-leg squat kinematics. J Biomech. 2022;144:111333. doi.org/10.1016/j.jbiomech.2022.111333
- Kanko RM, Laende EK, Davis EM, Selbie WS, Deluzio KJ. Concurrent assessment of gait kinematics using marker-based and markerless motion capture. J Biomech. 2021;127:110665. doi.org/10.1016/j.jbiomech.2021.110665
- Uhlrich SD, Falisse A, Kidziński Ł, et al. OpenCap: human movement dynamics from smartphone videos. PLoS Comput Biol. 2023;19(10):e1011462. doi.org/10.1371/journal.pcbi.1011462
- Puig-Diví A, Escalona-Marfil C, Padullés-Riu JM, Busquets A, Padullés-Chando X, Marcos-Ruiz D. Validity and reliability of the Kinovea program in obtaining angles and distances using coordinates in 4 perspectives. PLoS One. 2019;14(6):e0216448. doi.org/10.1371/journal.pone.0216448
- Baldinger M, Reimer LM, Senner V. Influence of the camera viewing angle on OpenPose validity in motion analysis. Sensors (Basel). 2025;25(3):799. doi.org/10.3390/s25030799
- Colyer SL, Evans M, Cosker DP, Salo AIT. A review of the evolution of vision-based motion analysis and the integration of advanced computer vision methods towards developing a markerless system. Sports Med Open. 2018;4(1):24. doi.org/10.1186/s40798-018-0139-y
- Wade L, Needham L, McGuigan P, Bilzon J. Applications and limitations of current markerless motion capture methods for clinical gait biomechanics. PeerJ. 2022;10:e12995. doi.org/10.7717/peerj.12995
- Bazarevsky V, Grishchenko I, Raveendran K, Zhu T, Zhang F, Grundmann M. BlazePose: on-device real-time body pose tracking. arXiv:2006.10204 [preprint]. 2020. arxiv.org/abs/2006.10204
- Needham L, Evans M, Cosker DP, et al. The accuracy of several pose estimation methods for 3D joint centre localisation. Sci Rep. 2021;11(1):20673. doi.org/10.1038/s41598-021-00212-x
- Nakano N, Sakura T, Ueda K, et al. Evaluation of 3D markerless motion capture accuracy using OpenPose with multiple video cameras. Front Sports Act Living. 2020;2:50. doi.org/10.3389/fspor.2020.00050
- Menychtas D, Petrou N, Kansizoglou I, et al. Gait analysis comparison between manual marking, 2D pose estimation algorithms, and 3D marker-based system. Front Rehabil Sci. 2023;4:1238134. doi.org/10.3389/fresc.2023.1238134
- Schurr SA, Marshall AN, Resch JE, Saliba SA. Two-dimensional video analysis is comparable to 3D motion capture in lower extremity movement assessment. Int J Sports Phys Ther. 2017;12(2):163–172. pubmed.ncbi.nlm.nih.gov/28515970/
- van den Hoorn W, Lavaill M, Cutbush K, Gupta A, Kerr G. Comparison of shoulder range of motion quantified with mobile phone video-based skeletal tracking and 3D motion capture—preliminary study. Sensors (Basel). 2024;24(2):534. doi.org/10.3390/s24020534
- Yue X, Gu Y, Yu B. Validation of single-camera MediaPipe BlazePose for knee joint angle measurement across multiple functional tasks: concurrent validity, interrater reliability, and exploratory test-retest reliability study. JMIR Rehabil Assist Technol. 2026;13:e102399. doi.org/10.2196/102399
- Reimer LM, Kapsecker M, Fukushima T, Jonas SM. Evaluating 3D human motion capture on mobile devices. Appl Sci. 2022;12(10):4806. doi.org/10.3390/app12104806
- Balsalobre-Fernández C, Glaister M, Lockey RA. The validity and reliability of an iPhone app for measuring vertical jump performance. J Sports Sci. 2015;33(15):1574–1579. doi.org/10.1080/02640414.2014.996184
- Romero-Franco N, Jiménez-Reyes P, Castaño-Zambudio A, et al. Sprint performance and mechanical outputs computed with an iPhone app: comparison with existing reference methods. Eur J Sport Sci. 2017;17(4):386–392. doi.org/10.1080/17461391.2016.1249031
- Peart DJ, Balsalobre-Fernández C, Shaw MP. Use of mobile applications to collect data in sport, health, and exercise science: a narrative review. J Strength Cond Res. 2019;33(4):1167–1177. doi.org/10.1519/JSC.0000000000002344
- Dingenen B, Barton C, Janssen T, Benoit A, Malliaras P. Test-retest reliability of two-dimensional video analysis during running. Phys Ther Sport. 2018;33:40–47. doi.org/10.1016/j.ptsp.2018.06.009
- Viswakumar A, Rajagopalan V, Ray T, Gottipati P, Parimi C. Development of a robust, simple, and affordable human gait analysis system using bottom-up pose estimation with a smartphone camera. Front Physiol. 2021;12:784865. doi.org/10.3389/fphys.2021.784865
- Willson JD, Davis IS. Utility of the frontal plane projection angle in females with patellofemoral pain. J Orthop Sports Phys Ther. 2008;38(10):606–615. doi.org/10.2519/jospt.2008.2706
- McLean SG, Walker K, Ford KR, Myer GD, Hewett TE, van den Bogert AJ. Evaluation of a two dimensional analysis method as a screening and evaluation tool for anterior cruciate ligament injury. Br J Sports Med. 2005;39(6):355–362. doi.org/10.1136/bjsm.2005.018598
- Janisch L, Maack LM, Fohrmann D, Marques CJ. Assessment of dynamic knee angle deviations in the frontal plane in physiotherapy clinical practice: intra- and inter-rater reliability of an application and agreement with two AI-models. BMC Musculoskelet Disord. 2026;27(1):407. doi.org/10.1186/s12891-026-09958-9
Educational content for health professionals. It does not replace clinical assessment or reading the original studies.



