Sensor-Based Exercise Classifier

Overview

Group project for Machine Learning for Quantitative Self (ML4QS) at Vrije Universiteit Amsterdam (M.Sc. AI). Built a system that classifies gym exercises from smartphone sensor data strapped to the upper arm, using both classical ML and deep learning approaches.

Data Collection

• Smartphone (PhyPhox app) strapped to the upper left arm during workouts
• Sensors: 3-axis gyroscope, 3-axis accelerometer, compensated accelerometer, light, proximity, GPS, magnetometer
• Polling at ~500Hz (0.002s intervals) for accelerometer and gyroscope
• 8 exercises across 4 muscle groups:
  • Chest: Bench Press, Cable Flys
  • Back: Deadlift, Pull-ups
  • Arms: Bicep Curls, Shoulder Press
  • Core: Crunches, Russian Twists

Feature Engineering

• Time-step aggregation to 0.02s intervals to reduce sparsity
• Velocity computation from discrete acceleration integration
• Resultant acceleration: sqrt(ax² + ay² + az²)
• Butterworth low-pass filter (50Hz base, 55Hz cutoff) for noise removal
• KNN imputation for missing values after time aggregation
• Removed irrelevant features (GPS, magnetometer) based on correlation analysis

Models

• Random Forest: For feature importance analysis and baseline classification
• LSTM Network: Sequence-based classification using rolling window features with MinMaxScaler normalization
• Both approaches significantly outperformed random baselines, with accelerometer and gyroscope features dominating importance

Technologies

Python, TensorFlow/Keras (LSTM), scikit-learn (Random Forest), Pandas, NumPy, PhyPhox