Human Activity Recognition for RISC-V

This repository contains the complete pipeline for training, quantizing, and deploying Human Activity Recognition (HAR) models on RISC-V architectures. The project utilizes deep learning architectures (CNN, GRU, LSTM) and TensorFlow Lite for Microcontrollers (TFLite Micro) to enable efficient, edge-based inference.

Overview

The goal of this project is to create highly optimized HAR models that can run on resource-constrained RISC-V microcontrollers. The pipeline covers:

1. Data Preprocessing: Handling and scaling raw sensor data (e.g., accelerometer, gyroscope).
2. Model Training & Evaluation: Training models like CNN and CNN-GRU, including Leave-One-Subject-Out (LOSO) cross-validation for robustness.
3. Quantization: Converting models to INT8 format via TFLite to dramatically reduce memory footprint and latency.
4. RISC-V Simulation & Execution: Compiling the quantized models with TFLite Micro and running them in a RISC-V environment (using simulators like Spike).
5. Real-time Demonstrations: Providing websocket-based streaming tools to visualize activity recognition in real-time.

Repository Structure

• train_*.py / evaluate_*.py: Scripts used to train the Keras models, evaluate their accuracy, and convert/quantize them to TFLite formats.
• main.cc, model_data.cc, test_data.h: C++ source files for initializing TFLite Micro, loading the quantized model arrays, and performing inference.
• models/: Contains the generated .keras and .tflite model files along with their training history and performance graphs.
• websocket_har_demo/: A real-time web application to stream sensor data (e.g., via Phyphox) and visualize live predictions.
• exhibition_showcase/: Front-end components for demonstrating the project's capabilities.

Getting Started

1. Training Models

To train the CNN model with quantization-aware strategies or standard pipelines, use the provided Python scripts:

python train_cnn_tinyml.py
python train_realworld_cnn_loso.py

Models will be evaluated and exported as .keras and .tflite artifacts in the models/ directory.

2. C++ Simulation

Export the quantized TFLite model to a C-byte array (e.g., model_data.cc). Then, compile the main.cc with the TFLite Micro library and run the executable in your RISC-V simulator.

3. Real-time Demo

To launch the real-time websocket demo, navigate to the websocket_har_demo directory to start the backend server and frontend dashboard.

Dependencies

• Python 3.8+ (TensorFlow, Keras, NumPy, Pandas, Scikit-Learn, Matplotlib)
• RISC-V Toolchain (e.g., riscv64-unknown-elf-gcc)
• Spike RISC-V Simulator
• TensorFlow Lite for Microcontrollers (TFLite Micro)