A movement quality assessment system combining 30 biomechanical features, a validated scoring pipeline (ρ=0.958), and a 100-video annotated dataset, applying kinesiology principles to real-time pose analysis.
Modern pose estimation frameworks like MediaPipe can track 33 body landmarks in real time from a single webcam feed. But landmark detection answers a narrow question: where is the body? It does not address the question a movement practitioner needs answered: how well is the body moving?
The distinction is significant. Two people can hold what appears to be the same yoga pose while exhibiting fundamentally different movement quality. One has proper joint alignment, distributed weight, and engaged stabilizing musculature. The other has compensatory patterns: a hyperextended knee masking hip inflexibility, a rotated pelvis creating the illusion of depth in a forward fold. A pose classifier labels both as “Warrior II.” A kinesiologist observes two distinct movement profiles.
ZENith began with a question at the intersection of computer vision and human movement science: can a computational system evaluate movement quality, not just pose identity, using the same biomechanical principles a human practitioner would apply?
The central idea is biomechanically-grounded feature engineering: rather than feeding raw landmark coordinates into a model, the system encodes the same anatomical measurements a kinesiologist would use to evaluate form.
30 biomechanical features. The features span four categories: 16 joint angles (bilateral shoulder flexion/abduction, elbow flexion, hip flexion/abduction, knee flexion, ankle dorsiflexion, spinal lateral flexion, trunk forward lean), 6 segment ratios (torso-leg ratio, arm span symmetry, stance width, shoulder-hip alignment, center-of-mass over base-of-support, head-spine alignment), 4 symmetry metrics (bilateral comparison of shoulder, elbow, hip, and knee angles), and 4 stability indicators (velocity variance, CoM oscillation, base-of-support area, weight distribution). These features are hand-designed from domain knowledge rather than learned from data, encoding clinical assessment principles directly into the representation.
Classification tells you what a pose is. Biomechanical features tell you how well it’s being performed. That shift from identity to quality is the difference between a pose library and a coaching system.
Data-calibrated pose profiles. Each of the 10 poses has a pose-specific quality profile defining which features are critical and their ideal ranges. These profiles were calibrated from correct-form video distributions: the system records what “good” looks like for each pose across the annotated dataset, then measures deviations from those empirical baselines.
A 100-video annotated dataset. Existing yoga pose datasets (Yoga-82, others) support classification but lack ground-truth annotations for biomechanical correctness, making them unsuitable for quality assessment. This dataset contains 100 video sequences across 10 poses with deliberate variation in form quality, from textbook alignment to common compensatory patterns. Each video was annotated using a semi-automated protocol: the system proposed feature-level quality assessments, and a kinesiologist reviewed and corrected them. The dataset follows the Gebru et al. (2021) datasheet standard with a full dataset card and annotation protocol.
The pipeline has four stages, each translating between different representational levels.
Stage 1: Pose Estimation (MediaPipe → 33 Landmarks). MediaPipe BlazePose processes each video frame and returns 33 3D landmark coordinates. This provides the spatial foundation for downstream analysis.
Stage 2: Biomechanical Feature Extraction (33 Landmarks → 30 Features). Raw landmark coordinates lack the anatomical context needed for movement assessment. The feature extraction layer computes the 30 biomechanical features described above, with pose-specific weighting. A Warrior II evaluation emphasizes front knee flexion angle (target: 90°), rear leg extension, hip abduction symmetry, and shoulder-line horizontality. A Tree Pose evaluation emphasizes standing-leg knee extension, hip external rotation of the lifted leg, and center-of-mass alignment over the base of support.
Stage 3: Quality Scoring (Pose Profiles + VAE). Quality is assessed through two complementary mechanisms. Pose-specific profiles compare the 30 features against calibrated ideal ranges, generating per-feature deviation scores. Simultaneously, a Variational Autoencoder trained on correct-form biomechanical features learns a latent distribution representing the space of “correct” execution. Reconstruction error provides a continuous quality score; the pattern of error localizes which aspects of the pose are problematic.
Stage 4: Classification & Coaching (Random Forest + Gemini). The Random Forest classifies pose identity from the 30 biomechanical features, while quality scores and specific angle deviations feed into a Gemini-powered coaching module that generates natural-language feedback. The system can say: “Your Warrior II front knee is at 72°. Try to deepen to 90°. Your hip abduction looks good. Watch your rear foot; it’s rotating inward, which often compensates for tight hip flexors.”
Four evaluation studies validate different aspects of the pipeline.
Ablation study. Stratified 5-fold cross-validation across 4,587 frames and 10 pose classes compared four feature conditions (RAW landmarks, BIO features, RAW+BIO, BIO+TEMPORAL) with two classifiers (Random Forest, MLP). The 30 biomechanical features achieved equivalent classification accuracy to 132 raw landmarks (a 4.4× dimensionality reduction) while preserving the interpretability needed for quality feedback.
Quality validation. The composite quality score was evaluated against expert biomechanical ratings across the annotated dataset. Spearman rank correlation: ρ=0.958 (p < 10-53), indicating strong agreement between the system’s assessments and practitioner ratings.
VAE discrimination. The Bio-VAE (30→8 latent dimensions), trained only on correct-form examples, produced a 9.84× reconstruction error ratio between correct and incorrect form, demonstrating reliable quality discrimination without requiring labeled negative examples.
Latency benchmarks. The full biomechanical pipeline executes in 0.29ms per frame, less than 1% of the 30fps frame budget, indicating that the feature extraction layer does not introduce meaningful latency overhead.
The central methodological tension in this project is the use of the researcher as both practitioner and dataset subject. In quantitative computer science, this is typically considered a limitation (N=1, no inter-rater reliability, potential bias in ground-truth labeling). In the phenomenological and kinesiology traditions, however, the trained practitioner’s body serves as a calibrated instrument. The semi-automated annotation protocol was designed to bridge these perspectives: the system proposes assessments, the practitioner corrects them, and the quality validation (ρ=0.958) provides empirical evidence that the pipeline produces clinically meaningful output.
The semi-automated annotation protocol treats embodied expertise as ground truth, with the practitioner’s trained proprioception as the labeling function, while maintaining systematic oversight through structured review.
An unexpected finding emerged from the VAE latent space. Poses that would typically be grouped by difficulty did not cluster that way in the learned representation. Instead, the latent space organized primarily by mechanical chain—poses that load similar joint systems clustered together regardless of subjective difficulty. Tree Pose and Crescent Lunge appeared near each other (both load the single-leg stance chain with hip stabilization demands), while Warrior II and Triangle were more distant than expected despite similar stance widths (the trunk orientation difference creates distinct kinetic chains).
Two pose profiles (Extended Side Angle and High Lunge) showed poor discrimination between correct and incorrect form. This likely reflects either insufficient profile specificity or narrower quality spectra that the current feature set does not capture well. The next step is targeted data collection and profile recalibration for these poses.
The single-performer limitation is the most significant constraint. While the quality validation demonstrates strong correlation with expert judgment, the system has been calibrated to one body’s biomechanics. Extending to diverse body types would require personalized biomechanical baselines, an open research direction that connects to the broader challenge of adaptive human-AI systems.