Human Activity Recognition (HAR)

Project Overview

This project presents a deep learning approach to sensor-based Human Activity Recognition (HAR). The model processes temporal signals gathered from smartphone accelerometers and gyroscopes to classify user physical actions (e.g. walking, climbing stairs, sitting, laying down), enabling smart health tracking integrations.

Key Features & Objectives

• Temporal Classification: Categorizes physical actions based on signal changes over time.
• Multi-Channel Signal Inputs: Processes 3-axis acceleration and angular velocity signals concurrently.
• Noise Filtering: Cleans raw signal streams to extract clean activity metrics.

Technical Stack & Technologies

• Language: Python
• Deep Learning: TensorFlow, Keras (LSTM layers)
• Signal Processing: SciPy, NumPy, Pandas
• Dataset: UCI HAR Dataset

Core Techniques & Deep Architectures

• LSTM Recurrent Networks: Built stacked Long Short-Term Memory (LSTM) network architectures capable of mapping temporal coordinate dependencies across sliding time windows.
• Signal Noise Filtering: Applied Butterworth low-pass filters to segregate gravity components from body acceleration signals.
• Spectral Feature Mapping: Utilized Fourier Transforms (FFT) to convert signal periods into frequency-domain power distributions, extracting additional feature arrays for the model.