Embedded Activity Recognition

An embedded machine learning project for real-time human activity recognition on a Raspberry Pi. This project focused on improving model accuracy, managing inference tradeoffs, and deploying an optimized neural network in an edge-computing environment.

Overview

This project involved optimizing a neural network for activity classification and deploying the resulting model to a Raspberry Pi. The work emphasized embedded AI constraints such as model size, inference speed, and real-world usability, while improving classification performance through architecture tuning and preprocessing improvements.

Project Goals

• Improve the accuracy of an existing embedded neural network model
• Deploy an optimized model for edge inference on Raspberry Pi
• Evaluate tradeoffs between accuracy, model size, and inference time
• Strengthen experience with embedded ML workflows

Tech Stack

• Python
• TensorFlow / Keras
• TensorFlow Lite
• NumPy
• scikit-learn
• Raspberry Pi

Key Improvements

• Applied standard scaling to reduce noise in the data
• Increased training set size from 60% to 80%
• Expanded the architecture with additional Conv1D and MaxPooling1D layers
• Increased dropout from 0.2 to 0.3
• Increased training epochs from 15 to 50
• Evaluated optimized and non-optimized TensorFlow Lite models

Results

• Improved baseline model accuracy from approximately 0.80 to a significantly higher optimized result
• Balanced classification performance with edge-device constraints
• Produced deployable TensorFlow Lite models for embedded use
• Strengthened understanding of the tradeoff between responsiveness and model complexity

Repository Structure

• Week 10/, Week 11/, Week 12/, Week 13/ – project development over time
• README.md – project report and summary

What This Project Demonstrates

• Embedded machine learning on real hardware
• Model optimization under deployment constraints
• Neural network architecture tuning for practical use
• End-to-end experience from preprocessing to edge deployment

Future Improvements

• Add dataset and label documentation
• Include plots for training/validation accuracy and loss
• Benchmark inference latency more systematically
• Explore quantization for smaller and faster TFLite deployment
• Add a demo video or sample output workflow

Why This Project Matters

This project sits at the intersection of machine learning, embedded systems, and real-world deployment. It reflects my interest in intelligent systems that move beyond model training and into practical, resource-constrained applications.