Fine-Grained Motion Recognition in At-Home Fitness Monitoring with Smartwatch: A Comparative Analysis of Explainable...

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• The squat is a fundamental exercise for overall fitness, but variations in form can impact effectiveness and safety.
• Accurate classification of squat execution is crucial for personalized fitness guidance and injury prevention.
• Existing methods may struggle to differentiate subtle variations in squat technique.

Purpose of the Study:

• To develop and evaluate a smartwatch-based system for fine-grained classification of squat motions.
• To compare the performance of deep neural network models against a conventional machine learning baseline.
• To investigate the feature learning capabilities of attention-based models for complex motion analysis.

Main Methods:

• Utilized a smartwatch to collect motion data from 52 participants performing correct and incorrect squats with varied arm postures.
• Implemented and compared deep neural network models (bidirectional GRU/LSTMs with attention) and a Random Forest baseline.
• Analyzed high-dimensional embeddings and attention patterns to understand model decision-making processes.

Main Results:

• The bidirectional GRU/LSTMs with an attention mechanism achieved the highest test accuracy (F1-score of 0.856) for squat classification.
• Attention-based models demonstrated more efficient learning from complex multivariate time-series motion signals, evidenced by clustered latent space embeddings.
• Bidirectional GRU/LSTMs showed consistent attention patterns across squat classes, focusing on distinct kinematic events like descending and ascending phases.

Conclusions:

• A smartwatch-based system with attention-enhanced deep learning models can effectively recognize subtle differences in squat execution.
• Attention mechanisms improve the efficiency of learning from complex motion data, leading to superior classification performance.
• Further analysis of attention patterns can provide insights into the kinematic features critical for accurate squat classification.