Wearable Activity Recognition

1. What is the project? An end-to-end system for wearable activity recognition using raw wrist-worn accelerometer data. The goal is to automatically identify activity intensity levels and specific physical activities without relying on manual input or additional sensors. 2. How was it built? Raw acceleration signals are segmented into fixed time windows and transformed into features used to train two independent models: a Random Forest for activity intensity classification and a Multilayer Perceptron (MLP) for specific activity recognition. Model inference is exposed through a FastAPI REST API and accessed via a Streamlit front-end application. 3. What are the next steps? Planned improvements include Leave-One-Participant-Out (LOPO) evaluation, deep learning models trained on raw signals, and enhanced handling of class imbalance.