📚 https://doi.org/10.1038/s41598-024-83475-4

✅ Day 2 – Literature & Method Review: Foundations Behind the Framework

Today’s reading focused on the foundation of the proposed scoring framework —
including how previous motion analysis methods compare, and how this paper constructs a full pipeline from skeleton data to explainable output.

📚 What I Read – From Motion Capture to Alignment Logic

📌 Key Literature Threads

• Motion Capture Types:
  • Wearable: Infrared (lab-accurate), IMU (portable but sparse)
  • Markerless: RGB-D (e.g., Kinect), 2D/3D Pose Estimation (OpenPose, HRNet, BlazePose)
• Evaluation Approaches:
  1. Rule-based: Hard thresholds from experts (simple but not generalizable)
  2. Similarity metrics: DTW, cosine, Euclidean (scoring but not interpretable)
  3. Model-based: Regressors/classifiers for learning performance → higher accuracy, but often black-box

⚙️ Method Overview

🔹 Skeleton Feature Extraction

• MediaPipe is used to extract 3D skeleton coordinates (33 joints)
• 18 angles are computed from body parts (shoulder, elbow, hip, knee, torso)
• Each angle is calculated as:

🔹 Feature Alignment

• Uses Dynamic Time Warping (DTW) to align motions temporally
• A 32-frame template is created from the top-performing sequence
• All others are aligned to this reference → compensates for speed/rhythm differences

🔹 Regression & Ensemble

• Base models: Linear, Lasso, SVM, KNN, Decision Tree, Random Forest, Bagging
• Uses adaptive weighting → models with lower RMSE get higher weights
• Final score = weighted average of all predictions

🔹 Explainability (SHAP)

• Global SHAP: shows which features consistently influence scores
• Local SHAP: explains why a specific motion got its score
• Makes feedback joint-specific and interpretable

💡 Why This Matters

This section gave me insight into:

• Why alignment isn’t just a preprocessing step — it’s the backbone of fair comparison
• How classic models (DT, SVM) can be repurposed for scoring if aligned features are strong
• The value of temporal normalization in movement analysis
• How interpretability doesn’t require deep nets — just good features and clear scoring logic

A solid skeleton feature design + fair alignment + transparent scoring = usable real-world evaluation system.

🛠️ What I’ll Implement Next

• Define 2D skeleton angles using MediaPipe keypoints
• Build a 32-frame DTW-based alignment system
• Train multiple regressors (e.g., Ridge, KNN, RF)
• Weight model outputs by validation RMSE
• Try SHAP (or cosine error maps) to highlight important joints

🔭 What’s Coming on Day 3

• Evaluate aligned vs. non-aligned features
• Compare model-based vs. similarity-based scoring
• Try visualizing SHAP impact over sequence heatmaps
• See if scoring outperforms average human inter-rater agreement

📝 Reflection

Today’s section helped ground the project:
alignment and angle feature design may actually matter more than model choice.
Also, I realized that explainability doesn’t have to wait until the end — it can shape how we choose and evaluate input features from the start.

The idea of using scoring models not just for labels, but for feedback is now clearer than ever.