ActiveTrack

About this project

ActiveTrack uses motion sensor data to recognize barbell exercises and estimate repetitions from the movement itself. The dataset was collected with a Meta Motion sensor across bench press, deadlift, overhead press, rowing, and squats, with both medium and heavy sets recorded through a Bluetooth-connected phone.

The pipeline begins with raw accelerometer and gyroscope data, which is cleaned, merged, resampled, and visualized before model development. Several approaches to anomalous data detection are explored, including IQR, Chauvenet's Criterion, and Local Outlier Factor, followed by missing-value interpolation and the creation of numerical, temporal, frequency, and cluster-based features.

The resulting features are used for model selection and hyperparameter tuning through train/test evaluation, forward feature selection, and grid search. Signal filtering is then applied to support automated repetition counting, connecting the machine learning pipeline back to the original sensor measurements.