Paper deep dive
Markerless Pose Estimation for Resistance Training Technique Assessment
Joseph Turner, Jeff Clark, Nawid Keshtmand
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 8/29/2026, 4:35:27 AM
Summary
This paper presents a markerless pose estimation framework using BlazePose to assess resistance training technique (squat, bench press, deadlift) from standard video footage. The method extracts anatomical landmarks, converts them into joint-angle trajectories, and compares them against reference repetitions using Root Mean Square Error (RMSE) and a similarity score. Results indicate high usability for squats and deadlifts but lower reliability for bench press due to occlusion. The study highlights that camera orientation (sagittal view preferred) and occlusion significantly impact 2D joint-angle accuracy.
Entities (12)
Relation Signals (9)
BlazePose → usedfor → Pose Estimation
confidence 98% · Using BlazePose, anatomical landmarks were extracted from squat, bench press, and deadlift videos
Joint-Angle Trajectories → assessedby → Root Mean Square Error
confidence 95% · Trajectories were assessed against a defined reference repetition using root mean square error (RMSE)
BlazePose → extracts → Anatomical Landmarks
confidence 95% · anatomical landmarks were extracted from squat, bench press, and deadlift videos
Squat → uses → Knee Angle
confidence 93% · For the squat, the knee angle was computed from the hip-knee-ankle triplet
Deadlift → uses → Hip-Hinge Angle
confidence 93% · For the deadlift... a hip-hinge angle defined from the shoulder-hip-knee triplet
Anatomical Landmarks → convertedto → Joint-Angle Trajectories
confidence 92% · anatomical landmarks were extracted... and converted into joint-angle trajectories
Camera Orientation → affects → Joint-Angle Accuracy
confidence 90% · Performance depended strongly on camera orientation... with non-sagittal views distorting 2D joint-angle estimates
Visual Occlusion → affects → Pose Estimation Reliability
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Resistance training can be a high risk activity, and safe form is essential to avoiding injury. Laboratory-based movement analysis provides quantitive technique assessment, yet is not easily accessible. Markerless pose estimation infers body landmarks from images or video without physical markers and could offer a feasible alternative for technique assessment. We present a pose estimation framework to evaluate resistance-training technique from ordinary video footage. Using BlazePose, anatomical landmarks were extracted from squat, bench press, and deadlift videos and converted into joint-angle trajectories, with the squat serving as the primary case study. Trajectories were assessed against a defined reference repetition using root mean square error (RMSE). Results show that the framework recovers meaningful kinematic patterns for the squat and deadlift, enabling quantitative comparison between repetitions and identification of technique variability within a set. Performance depended strongly on camera orientation and visual occlusion, with non-sagittal views distorting 2D joint-angle estimates. The findings demonstrate that markerless pose estimation can support accessible biomechanical assessment outside laboratory environments.
Tags
Links
- Source: https://arxiv.org/abs/2608.24384v1
- Canonical: https://arxiv.org/abs/2608.24384v1
Trouble viewing inline? Open PDF directly →
Full Text
26,488 characters extracted from source content.
Expand or collapse full text
Markerless Pose Estimation for Resistance Training Technique Assessment Joseph Turner, Jeff Clark [0000−0003−0118−3999] , and Nawid Keshtmand [0009−0008−5552−1395] School of Engineering Mathematics and Technology, University of Bristol, UK nawid.keshtmand@bristol.ac.uk Abstract. Resistance training can be a high risk activity, and safe form is essential to avoiding injury. Laboratory-based movement anal- ysis provides quantitative technique assessment, yet is not easily acces- sible. Markerless pose estimation infers body landmarks from images or video without physical markers and could offer a feasible alterna- tive for technique assessment. We present a pose estimation framework to evaluate resistance-training technique from ordinary video footage. Using BlazePose [4], anatomical landmarks were extracted from squat, bench press, and deadlift videos and converted into joint-angle trajecto- ries, with the squat serving as the primary case study. Trajectories were assessed against a defined reference repetition using root mean square er- ror (RMSE). Results show that the framework recovers meaningful kine- matic patterns for the squat and deadlift, enabling quantitative compari- son between repetitions and identification of technique variability within a set. Performance depended strongly on camera orientation and visual occlusion, with non-sagittal views distorting 2D joint-angle estimates. The findings demonstrate that markerless pose estimation can support accessible biomechanical assessment outside laboratory environments. Keywords: Pose estimation· Safety· Biomechanics 1 Introduction Efficient and safe movement is central to sports performance, rehabilitation, and injury prevention [2]. In resistance training, technique is typically assessed through subjective coaching or, less commonly, laboratory-based biomechanical analysis [8]. Marker-based motion capture systems such as Vicon provide highly accurate kinematic measurements but require expensive specialised equipment and controlled environments, limiting their use in everyday training settings [3,15]. Recent advances in computer vision have made markerless pose estimation a promising alternative. Models such as BlazePose infer anatomical landmarks directly from ordinary video, enabling extraction of kinematic information from smartphone recordings at low cost [4,14]. This creates opportunities for accessible movement analysis in training and rehabilitation [12]. arXiv:2608.24384v1 [cs.CV] 25 Aug 2026 2J. Turner et al. This paper aims to determine whether markerless pose estimation can provide a practical and interpretable basis for biomechanical assessment of resistance- training exercises outside laboratory environments. To investigate this, we devel- oped and evaluated a framework for analysing resistance-training technique from standard video footage using pose estimation. Anatomical landmarks were ex- tracted from recordings of the squat, bench press, and deadlift and converted into joint-angle trajectories. The squat served as the primary case study, with knee and trunk angles used for detailed biomechanical analysis, and quantitatively compared against gold standard reference repetitions. This work is intended as a feasibility study demonstrating the potential of markerless pose estimation for accessible resistance-training assessment rather than a fully validated biome- chanical assessment system. The main contributions of this paper are: – Development and evaluation of a pose-estimation framework for analysing resistance-training technique from standard video footage – Curation of a resistance-training dataset, including reference repetitions demon- strating good technique – Investigation of the effects of camera viewpoint on joint-angle estimation 2 Related Work Markerless approaches estimate anatomical landmarks directly from video, of- fering an accessible solution for coaching, rehabilitation, and performance mon- itoring [5]. Their accuracy, however, remains sensitive to camera viewpoint, oc- clusion, and image quality [16]. Human pose estimation has progressed from part-based models [18] to deep learning approaches such as HRNet [13], VideoPose3D [9], and transformer-based models such as ViTPose [17]. These methods have significantly improved land- mark detection and, in some cases, enabled monocular 3D reconstruction. For practical applications, lightweight models such as BlazePose [4] are particularly attractive because they provide real-time full-body tracking from standard video. For sports applications, OpenPose has been applied to elite long-jump com- petition footage, demonstrating that meaningful biomechanical measures can be extracted outside laboratory conditions [6]. More recently, AthletePose3D intro- duced a large-scale dataset of athletic movements to improve model performance on high-speed sports actions [19]. In resistance training, computer vision has been used to estimate external variables such as barbell velocity [1]. However, assessing movement quality re- quires joint-level biomechanical analysis. Squat biomechanics studies have shown that knee and trunk motion are key determinants of loading and technique [7]. Mercadal-Baudart et al. demonstrated that joint angles derived from single- camera pose estimation can provide interpretable metrics for strength exercises [8]. Despite these advances, limited work has evaluated how robustly markerless pose estimation can assess complex powerlifting movements under realistic gym Pose Estimation for Resistance Training3 conditions. In particular, the effects of camera viewpoint, occlusion, and intra-set variability remain under explored, which this paper seeks to address. 3 Method 3.1 Dataset Curation and Preprocessing Two publicly available Kaggle datasets [11,10] were combined (n = 2618 videos). The three target powerlifts (squat, bench press, and deadlift) were selected to fo- cus on compound movements involving multiple joints and a higher risk of injury. As the original datasets contained a wide variety of gym exercises, the majority of videos were excluded during this initial selection step. The remaining record- ings were then filtered to retain only videos suitable for reliable biomechanical analysis. Clips with poor lighting, severe occlusion, incomplete body visibility, or extreme camera viewpoints were removed, as these conditions prevent consistent landmark detection and can introduce substantial errors into 2D joint-angle es- timation. Following this filtering process, the final dataset comprised 89 videos, which were subsequently normalised to a consistent frame rate and resolution before analysis. To enable quantitative comparison between repetitions, a reference trajec- tory was established from a curated second set of instructional videos sourced from publicly available material, selected for technical soundness against widely accepted coaching standards. Each repetition was then evaluated against four biomechanical criteria. For the squat these were: (i) sufficient depth, indicated by a low minimum knee angle; (i) a smooth, continuous U-shaped descent-ascent profile; (i) stable trunk angle throughout; and (iv) minimal discontinuities at- tributable to pose estimation noise. The repetition best satisfying all four criteria was retained as the reference trajectory for subsequent comparisons. Reference repetitions could be tailored to an individual, providing quantified feedback tai- lored to the athlete’s specific anatomy and training style. 3.2 Pose-Based Biomechanical Analysis Framework BlazePose is a lightweight convolutional neural network designed for real-time 2D human pose estimation [4]. This was implemented as the foundation of the analysis pipeline. For each frame t, the model returns a set of M = 33 anatomical landmarks p (t) i = x (t) i , y (t) i , i = 1,...,M,(1) where x (t) i ,y (t) i are the 2D image-plane coordinates of the i-th keypoint. Joint angles were derived from landmark triplets using a standard vector formulation: for three points A, B, C where B is the joint of interest, θ = cos −1 (A− B)· (C− B) ∥A− B∥C− B∥ .(2) 4J. Turner et al. For the squat, the knee angle was computed from the hip-knee-ankle triplet, and trunk posture was quantified as the angular deviation of the shoulder-hip segment from the vertical axis. For the deadlift, trunk angle was derived using the same formulation, with a hip-hinge angle defined from the shoulder-hip-knee triplet to capture the degree of forward flexion at the hip. For the bench press, upper-limb kinematics were quantified using the shoulder–elbow–wrist triplet to compute elbow flexion angle, defined as the internal angle at the elbow joint. 3.3 Temporal Normalisation Since repetitions differ in duration due to differences in pacing and frame count, direct frame-by-frame comparison is not meaningful without temporal align- ment. Linear interpolation was applied to resample each joint-angle signal onto a common normalised time axis. For a repetition of T frames, the normalised coordinate was defined as: ˆτ t = t T − 1 , t = 0, 1,...,T − 1(3) Each joint-angle signal was then linearly interpolated and resampled onto a common normalised time axis τ i = i 100 , i = 0, 1,..., 100,(4) yielding 101 samples per repetition regardless of its original frame count. This process preserves the shape of each movement trajectory whilst removing vari- ability due to execution speed. The reference was normalised using the same procedure, ensuring all comparisons were made on a common movement axis. For multi-repetition analysis, a phase-aware variant was also applied. Since a multi-repetition recording contains several consecutive movement cycles, simple linear normalisation cannot distinguish between individual repetitions or align their phases meaningfully. Each repetition was therefore first segmented, then divided into descent, hold, and ascent phases, assigned fixed proportions of 40%, 20%, and 40% of the full movement cycle respectively, with each phase inde- pendently normalised to [0, 1] prior to concatenation. The 40–20–40 proportions were selected as a simple heuristic to preserve the approximate temporal struc- ture of a squat repetition, where descent and ascent typically occupy most of the movement and any pause at the bottom is comparatively brief. More adap- tive phase alignment approaches, such as dynamic time warping, represent an important direction for future work. 3.4 Joint Angle Trajectory Comparison Following reference repetition trajectory selection, deviation between each test repetition and the reference trajectory was quantified using RMSE, which was Pose Estimation for Resistance Training5 computed independently for the knee-angle and trunk-angle trajectories. To ex- press RMSE on a more interpretable scale, each value was converted to a simi- larity score ranging from 0 to 100: S = max(0, 100− 2 RMSE).(5) The scaling factor of two was selected based on the typical RMSE values observed in the dataset, allowing the resulting scores to span an interpretable 0–100 range. The similarity score is not intended as a universally validated measure of exercise technique quality, but rather as a task-specific metric for quantifying consistency between a test repetition and a reference trajectory. This provides a simple and interpretable measure of movement agreement suitable for automated assessment using markerless pose estimation. An overall similarity score is reported as the mean of the knee and trunk scores. 3.5 Intra-Set Variability Analysis Repetitions were segmented by detecting local minima of the knee-angle tra- jectory, each corresponding to the bottom position of a squat. Only complete descent-ascent cycles were retained for analysis. Each segmented repetition was independently time-normalised and per-repetition metrics were computed. Set- level summary statistics (mean RMSE, standard deviation, and best and worst repetitions by RMSE) were used to quantify intra-set technical consistency. The pipeline’s code will be released upon publication. 6J. Turner et al. 4 Results and Discussion 4.1 Pose Estimation Performance Across Exercises The pipeline was evaluated on the dataset of squat, deadlift, and bench press recordings. Squats and deadlifts performed reliably, achieving mean usable-frame rates of 99.7% and 99.0% respectively. Bench press was substantially less reli- able, with only 73.6% usable frames and 77.2% of videos successfully processed, reflecting the impact of barbell occlusion and supine body orientation on land- mark extraction. These results indicate that markerless tracking performs best for upright exercises where the limbs of interest remain visible throughout the movement cycle. Consequently, the remaining analysis focuses primarily on squat and deadlift movements. (a) Squat(b) Bench press(c) Deadlift Fig. 1: Representative examples of markerless pose estimation across the three lifts. (a) Squat: the upright side-on view provides clear lower-limb visibility. (b) Bench press: upper-limb landmarks (i.e. shoulder, elbow, wrist) are difficult to track due to supine position and barbell occlusion during the repetition. (c) Deadlift: the upright setup supports consistent tracking of main joint landmarks. 4.2 Squat Trajectories and Reference Trajectory Comparison Joint-angle trajectories extracted from squat videos demonstrated the expected biomechanical structure. When applied to multi-repetition analysis for a ten- repetition squat set, after temporal normalisation, most repetitions exhibited the characteristic U-shaped knee profile associated with squat descent and as- cent, with minimum knee angles ranging approximately between 25 ◦ and 50 ◦ (Figure 2). Most repetitions followed similar movement profiles and remained close to the reference repetition trajectory (dashed blue line, Figure 2). Quantitative re- sults are reported in Table 1, with a mean RMSE of 17.56 ◦ and mean similarity score of 64.9/100. Rep 1 showed the largest deviation from the reference repeti- tion trajectory (RMSE 37.50 ◦ ), whereas Rep 8 produced the closest agreement (RMSE 13.18 ◦ ). These findings demonstrate that the framework can identify technique variability both within individual repetitions and across an entire set. Pose Estimation for Resistance Training7 Fig. 2: Time-normalised knee trajectories for all ten repetitions within a squat set compared with the reference repetition trajectory (dashed line). Rep Knee Min ( ◦ ) Knee Max ( ◦ ) Knee RMSE ( ◦ ) Knee Similarity 135.5173.937.5025.0/100 235.4166.915.6268.8/100 334.1168.314.1571.7/100 436.3169.919.2461.5/100 537.3168.119.8060.4/100 637.6165.215.6168.8/100 736.4165.119.4561.1/100 834.8172.313.1873.6/100 938.0176.718.7462.5/100 1036.9179.322.3355.3/100 Mean17.5664.9/100 SD6.56 Table 1: Per-repetition knee-angle metrics for the analysed squat set using phase- aware normalisation. RMSE is computed relative to the reference repetition tra- jectory. 8J. Turner et al. 4.3 Camera Sensitivity and Occlusion Effects One of the strongest findings of this work was the sensitivity of joint-angle esti- mation to viewpoint, seen in the camera-rotation experiment (Figure 3). Fig. 3: Bird’s eye view of the camera rotation experiment. The lifter maintained a static squat posture at the centre whilst the camera was moved approximately 270 ◦ around the subject. The blue circle marks the sagittal side-on position (0 ◦ ), which served as the reference angle due to its superior joint angle estimation accuracy. Key intermediate viewpoints rear (∼90 ◦ ), opposite side-on (∼180 ◦ ), and frontal (∼270 ◦ ) are also marked to illustrate the range of camera orientations evaluated. It showed that estimated knee angle changed substantially even when the par- ticipant maintained a fixed squat posture. The effect became most pronounced inside a power rack where lower-limb landmarks became partially obscured (Fig- ure 4). The largest deviations occur when the rack partially blocks the lower limbs. Consequently, apparent changes in knee angle may reflect projection dis- tortion rather than genuine biomechanical differences, a limitation consistent with the occlusion and viewpoint sensitivity reported by Wade et al [16]. This Pose Estimation for Resistance Training9 explains several examples in the dataset where visually acceptable squats failed to produce the expected trajectory shape. The results therefore identify sagittal recording as the preferred viewpoint for practical deployment. Fig. 4: Estimated knee angle during camera rotation around a static squat pos- ture inside a power rack. Occlusion regions (shaded blue) produce larger de- viations from the side-on reference estimate. See Figure 3 for details on the experimental setup and corresponding camera angles. 4.4 Joint-Angle Trajectories for a Representative Deadlift Trial To demonstrate that the proposed framework generalises beyond the squat, a representative deadlift recording was analysed from a sagittal viewpoint. The same temporal normalisation procedure described for the squat was applied prior to trajectory comparison. The hip-hinge angle increased smoothly throughout the movement, reflecting progressive hip extension from the bottom position to full lockout. The trunk an- gle began in a forward-leaning position and became more upright during ascent, indicating controlled trunk extension throughout the lift. Both signals exhibited a monotonic structure compared with the squat, reflecting the hinge-dominant nature of the movement. These results demonstrate that the framework can extract meaningful joint-angle trajectories beyond the squat, supporting its gen- eralisability to compound movements with distinct biomechanical profiles. 4.5 Limitations and Future Work The primary limitation of this framework is the sensitivity of 2D joint-angle es- timation to camera orientation. As demonstrated in Section 4.3, even when the 10J. Turner et al. (a) Time-normalised hip hinge angle. (b) Time-normalised trunk angle from vertical. Fig. 5: Joint-angle trajectories for a representative deadlift repetition. underlying posture remains unchanged, viewpoint changes introduce substantial distortion in estimated joint angles. Preliminary work into 3D landmark recon- struction suggests partial improvement under non-sagittal conditions but did not fully recover the expected range of motion, as the landmarks are inferred from a single view and remain subject to perspective ambiguity. Reliable deployment in uncontrolled environments will therefore require either stricter guidance on camera placement or more robust depth-estimation techniques. The RMSE metric, whilst interpretable, weights all phases of the movement equally. Deviations at biomechanically critical points, such as the bottom posi- tion of the squat, may carry greater practical significance than deviations during transitional phases, and a phase-weighted metric could better reflect how tech- nique is evaluated in practice. The dataset used is relatively small and specific. Real-world deployment would involve a wider range of lighting conditions, camera placements, and lifter experience levels, which the current pipeline does not fully account for. Expand- ing the dataset to include a wider range of lifters, experience levels, and recording environments would improve the external validity of the framework and support more robust evaluation of the similarity metric across populations. Pose Estimation for Resistance Training11 5 Conclusion This paper presents and evaluates a markerless pose estimation framework for assessing resistance-training technique from standard video footage, and suc- cessfully reconstructed joint-angle trajectories for the squat and deadlift, with mean usable frame rates of 99.7% and 99.0% respectively. A frame was consid- ered usable when BlazePose successfully detected all landmarks required for the joint-angle calculation with valid coordinates. The usable-frame rate therefore represents the percentage of frames within a video that could be included in the biomechanical analysis. Squat trajectories demonstrated consistent biome- chanical structure and strong agreement with a reference repetition under sagit- tal viewing conditions. Intra-set analysis captured movement quality variations across a set, with early and late repetitions showing the greatest deviation from the reference repetition, consistent with movement initiation effects and fatigue. Camera viewpoint emerged as the strongest practical constraint. Rotation experiments confirmed that 2D joint-angle estimates are highly sensitive to per- spective distortion, and that non-sagittal recordings can produce misleading tra- jectories even when the underlying movement is technically sound. The extension of the pipeline to the deadlift produced meaningful hip-hinge and trunk trajec- tories consistent with expected biomechanics, supporting the generalisability of the approach to compound movements with distinct mechanical profiles. Overall, the results indicate that markerless pose estimation can provide accessible, interpretable biomechanical assessment outside laboratory environ- ments. Whilst limitations around viewpoint sensitivity, occlusion, and metric design remain, the framework represents a practical step towards data-driven movement feedback for everyday training settings. 6 Code Availability The code used to implement the proposed framework will be released upon publication at github.com/ac2771/Markerless-Pose-Resistance-Training. Acknowledgments. NK was supported by the EPSRC LEAP Digital Health Hub grant EP/X031349/1. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. References 1. Ågren, O., Palm, J.: Enhancing athletic training through AI: A comparative anal- ysis of YOLO versions for image segmentation in velocity-based training. Student thesis, DiVA portal (2024), https://w.diva-portal.org/smash/get/diva2: 1872383/FULLTEXT01.pdf 12J. Turner et al. 2. Bahr, R., Krosshaug, T.: Understanding injury mechanisms: a key component of preventing injuries in sport. British journal of sports medicine 39(6), 324–329 (2005) 3. Barris, S., Button, C.: A review of vision-based motion analysis in sport. Sports Medicine 38(12), 1025–1043 (2008). https://doi.org/10.2165/ 00007256-200838120-00006, https://link.springer.com/article/10.2165/ 00007256-200838120-00006 4. Bazarevsky, V., Kartynnik, Y., Vakunov, A., Raveendran, K., Grundmann, M.: BlazePose: On-device real-time body pose tracking. https://arxiv.org/abs/ 2006.10204 (2020), accessed: 2025-11-06 5. Colyer, S.L., Evans, M., Cosker, D.P., Salo, A.I.T.: A review of the evolution of vision-based motion analysis and the integration of advanced computer vi- sion methods towards developing a markerless system. Sports Medicine – Open 4(1), 24 (2018). https://doi.org/10.1186/s40798-018-0139-y, https://link. springer.com/article/10.1186/s40798-018-0139-y 6. Cronin, N.J., Walker, J., Tucker, C.B., Nicholson, G., Cooke, M., Merlino, S., Bis- sas, A.: Feasibility of OpenPose markerless motion analysis in a real athletics com- petition. Frontiers in Sports and Active Living 5, 1298003 (2024). https://doi. org/10.3389/fspor.2023.1298003, https://pmc.ncbi.nlm.nih.gov/articles/ PMC10796501/ 7. Escamilla, R.F.: Knee biomechanics of the dynamic squat exercise. Medicine & Science in Sports & Exercise 33(1), 127–141 (2001). https://doi.org/10.1097/ 00005768-200101000-00020, https://pubmed.ncbi.nlm.nih.gov/11194098/ 8. Mercadal-Baudart, C., Liu, C.J., Farrell, G., Boyne, M., González Es- cribano, J., Smolic, A., Simms, C.: Exercise quantification from sin- gle camera view markerless 3D pose estimation. Heliyon 10(6), e27596 (2024). https://doi.org/10.1016/j.heliyon.2024.e27596, https://w. sciencedirect.com/science/article/pii/S2405844024036272 9. Pavllo, D., Feichtenhofer, C., Grangier, D., Auli, M.: 3D human pose estimation in video with temporal convolutions and semi-supervised training. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 7753– 7762 (2019) 10. Philosopher0808: Gym workout/exercises video dataset. https://w.kaggle. com/datasets/philosopher0808/gym-workoutexercises-video/data (2023), ac- cessed: 2025-11-06 11. Sinabutar, D.: Deadlift pose estimation dataset. https://universe.roboflow. com/daniel-sinabutar/deadlift-pose-estimation (2023), accessed: 2025-11-06 12. Souaifi, M., Dhahbi, W., Jebabli, N., Ceylan, H.İ., Boujabli, M., Muntean, R.I., Dergaa, I.: Artificial intelligence in sports biomechanics: A scoping review on wear- able technology, motion analysis, and injury prevention. Bioengineering 12(8), 887 (2025) 13. Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learn- ing for human pose estimation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). p. 5693–5703 (2019). https://doi.org/10.1109/CVPR.2019.00585 14. Ultralytics:Whatisposeestimationandwhere canitbeused? https://w.ultralytics.com/blog/ what-is-pose-estimation-and-where-can-it-be-used (2024), accessed: 2025-11-06 15. Vicon Motion Systems Ltd.: Vicon: Motion capture systems for biomechanics and sports (2026), https://w.vicon.com/, accessed: 1 April 2026 Pose Estimation for Resistance Training13 16. Wade, L., Needham, L., McGuigan, P., Bilzon, J.: Applications and limitations of current markerless motion capture methods for clinical gait biomechanics. PeerJ 10, e12995 (2022). https://doi.org/10.7717/peerj.12995, https://pmc.ncbi. nlm.nih.gov/articles/PMC8884063/, published 25 February 2022 17. Xu, Y., Zhang, J., Zhang, Q., Tao, D.: ViTPose: Simple vision transformer baselines for human pose estimation. arXiv preprint arXiv:2204.12484 (2022). https://doi. org/10.48550/arXiv.2204.12484, https://arxiv.org/abs/2204.12484 18. Yang, Y., Ramanan, D.: Articulated pose estimation with flexible mixtures- of-parts. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). p. 1385–1392 (2011). https://doi.org/10. 1109/CVPR.2011.5995741, https://yangyi02.github.io/research/pose/pose_ cvpr2011.pdf 19. Yeung, C., Suzuki, T., Tanaka, R., Yin, Z., Fujii, K.: AthletePose3D: A benchmark dataset for 3D human pose estimation and kinematic validation in athletic move- ments. arXiv preprint arXiv:2503.07499 (2025), https://arxiv.org/abs/2503. 07499