Source-linked AI summary
Markerless Pose Estimation for Resistance Training Technique Assessment
Joseph Turner, Jeff Clark, Nawid Keshtmand
TL;DR
Resistance-training technique is often assessed subjectively or with laboratory systems that are difficult to access, motivating practical video-based alternatives. The paper develops a BlazePose framework that converts squat, bench press, and deadlift footage into joint-angle trajectories compared with reference repetitions using RMSE. The framework recovers meaningful squat and deadlift kinematic patterns for quantitative repetition comparison, while viewpoint sensitivity and occlusion constrain deployment.
Problem
Accessible and interpretable biomechanical assessment of resistance-training technique outside laboratory environments remains limited, particularly regarding viewpoint, occlusion, and intra-set variability.
Method
The framework uses BlazePose landmarks from squat, bench press, and deadlift videos, converts them into joint-angle trajectories, and compares repetitions with reference trajectories using RMSE.
Results
The framework successfully reconstructed squat and deadlift joint-angle trajectories, with mean usable-frame rates of 99.7% and 99.0% respectively.
Takeaways & Limitations
Markerless pose estimation can provide accessible, interpretable biomechanical assessment outside laboratory environments and support data-driven movement feedback for everyday training.
Takeaways & Limitations
2D joint-angle estimation is sensitive to camera orientation because single-view landmarks remain subject to perspective ambiguity.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper develops a markerless pose-estimation framework to assess resistance-training technique from standard video, targeting a more accessible alternative to laboratory analysis. It covers squat, bench press, and deadlift recordings while examining camera viewpoint effects.
- Laboratory motion capture provides accurate kinematic measurements but requires expensive equipment and controlled environments.
- BlazePose infers anatomical landmarks from ordinary video, enabling low-cost extraction of kinematic information from smartphone recordings.
- The framework extracts landmarks from squat, bench press, and deadlift videos and converts them into joint-angle trajectories.
- Squat analysis uses knee and trunk angles compared quantitatively against gold-standard reference repetitions.
- The contributions include framework development, curation of reference repetitions, and investigation of camera viewpoint effects on joint-angle estimation.
2 Related Work
Prior work establishes markerless pose estimation as an accessible source of movement measures, but its robustness for resistance-training technique remains incompletely evaluated. This paper focuses on viewpoint, occlusion, and intra-set variability in complex powerlifting movements.
- Markerless approaches offer accessible video-based landmark estimation, but accuracy remains sensitive to camera viewpoint, occlusion, and image quality.
- Lightweight models such as BlazePose are attractive for practical applications because they provide real-time full-body tracking from standard video.
- Earlier sports applications show that meaningful biomechanical measures can be extracted outside laboratory conditions from athletic video.
- Resistance-training computer vision has estimated external variables such as barbell velocity, while movement quality requires joint-level biomechanical analysis.
- The effects of camera viewpoint, occlusion, and intra-set variability remain underexplored for markerless assessment of complex powerlifting movements.
3 Method
The method filters resistance-training videos, extracts 2D landmarks with BlazePose, derives exercise-specific joint angles, temporally normalizes repetitions, and compares them with reference trajectories using RMSE-based similarity.
- Data and reference construction: The dataset combines two publicly available Kaggle datasets containing 2618 videos of squats, bench presses, and deadlifts.
- Data and reference construction: Videos with poor lighting, severe occlusion, incomplete body visibility, or extreme camera viewpoints were excluded for reliable biomechanical analysis.
- Data and reference construction: Reference trajectories came from curated instructional repetitions selected for technical soundness against widely accepted coaching standards.
- Pose and joint-angle analysis: BlazePose returns M = 33 anatomical landmarks per frame, which are used as the foundation of the analysis pipeline.
- Pose and joint-angle analysis: Joint angles are derived from landmark triplets, including hip-knee-ankle for squat knee angle and shoulder-hip posture for trunk angle.
- Temporal normalization: Signals are linearly interpolated onto a common normalized time axis, yielding 101 samples per repetition regardless of original frame count.
- Temporal normalization: Multi-repetition recordings are segmented into descent, hold, and ascent phases assigned fixed 40%, 20%, and 40% proportions before concatenation.
- Trajectory comparison: RMSE is computed independently for knee-angle and trunk-angle trajectories and converted into a 0–100 similarity score for repetition comparison.
4 Results and Discussion
The framework performed most reliably for upright squat and deadlift recordings, producing interpretable joint-angle trajectories and identifying repetition variability. Camera orientation and occlusion substantially constrained 2D angle estimates, making sagittal recording preferable for deployment.
- Pose Estimation Performance Across Exercises: 99.7% and 99.0% mean usable-frame rates were achieved for squats and deadlifts, compared with 73.6% for bench press.Only 77.2% of bench-press videos were successfully processed, reflecting barbell occlusion and supine orientation.
- Squat Trajectories and Reference Trajectory Comparison: Most of ten squat repetitions exhibited the expected U-shaped knee trajectory after temporal normalisation, with minimum knee angles of approximately 25°–50°.The trajectories captured the characteristic descent and ascent structure of the squat.
- Squat Trajectories and Reference Trajectory Comparison: 17.56° mean RMSE and 64.9/100 mean similarity showed that repetitions could be quantitatively compared with a reference trajectory.Rep 1 had the largest deviation at 37.50° RMSE, while Rep 8 had the closest agreement at 13.18° RMSE.
- Camera Sensitivity and Occlusion Effects: The camera-rotation experiment found substantial knee-angle changes despite a fixed squat posture, with sagittal side-on view used as the accuracy reference.The experiment evaluated camera positions across approximately 270° around the subject.
- Camera Sensitivity and Occlusion Effects: Power-rack occlusion produced the largest deviations from the side-on knee-angle estimate, indicating that apparent angle changes can reflect projection distortion.The shaded occlusion regions in Figure 4 identify where lower-limb landmarks were partially blocked.
- Deadlift Generalisation: The deadlift showed smooth hip-hinge and trunk-angle trajectories, with progressive hip extension and controlled trunk extension during ascent.Its monotonic structure contrasted with the squat’s trajectory pattern and reflected the hinge-dominant movement.
- Limitations and Future Work: Camera orientation is the primary limitation because single-view 2D estimates distort joint angles even when posture remains unchanged.More robust depth estimation or stricter camera-placement guidance is needed for uncontrolled environments.
- Limitations and Future Work: RMSE weights all movement phases equally, although deviations at biomechanically critical points such as the squat bottom may matter more in practice.A phase-weighted metric could better reflect technique evaluation.
5 Conclusion
The framework reconstructed meaningful squat and deadlift joint-angle trajectories and captured variation across squat repetitions. Camera viewpoint was the strongest practical constraint, while the overall approach supports accessible biomechanical assessment outside laboratory environments.
- 99.7% and 99.0% mean usable-frame rates were achieved for squat and deadlift recordings, respectively.A usable frame contained all BlazePose landmarks required for joint-angle calculation with valid coordinates.
- Squat trajectories showed consistent biomechanical structure and agreement with a reference repetition under sagittal viewing conditions.Intra-set analysis also captured movement-quality variation across repetitions.
- Camera viewpoint strongly constrained 2D joint-angle estimates, with non-sagittal recordings producing misleading trajectories despite technically sound movement.The authors identify camera placement guidance or more robust depth estimation as requirements for reliable uncontrolled deployment.
- Markerless pose estimation can provide accessible, interpretable biomechanical assessment outside laboratory environments.The framework is presented as a practical step toward data-driven movement feedback for everyday training settings.
6 Code Availability
The implementation code will be released upon publication. The paper also records funding support and reports no relevant competing interests.
- The proposed framework code will be released upon publication at the stated GitHub repository.
- NK received support from the EPSRC LEAP Digital Health Hub grant EP/X031349/1.
- The authors report no competing interests relevant to the article.