Posture Prediction with Neural Networks

Overview

This project aims to develop a sensor-agnostic posture classification system that uses an IMU sensor unit embedded in an Arduino board. The system will gather sensor data, including 3-axis accelerometer, 3-axis gyroscope, and 3-axis magnetometer, and implement a machine learning algorithm for real-time posture detection. The final model will be deployed on the microcontroller for live predictions and communicate the results to a base station (laptop or smartphone).

Project Phases

Phase 1: IMU Sensor Data Collection

Run readData.ino code to read IMU sensor data and store signal readings on your computer.

Phase 2: Data Collection

Run readData.py Collect data for different postures (supine, prone, side, sitting, and unknown). Ensure to gather data for various sensor orientations to ensure robustness.

Phase 3: Dataset Construction

Construct data.csv with all 3 sensors data for all 5 postures and split it into training, validation, and test sets. Train the model with only 3 input channels (x, y, and z).

Phase 4: Neural Network Architecture

Run Relu_datamodel.py a custom neural network architecture and train your model offline. Obtain model_pred.tflite and model_pred_quant.tflite

Phase 5: Model Performance Assessment

Assess the performance of your model. Make adjustments to the architecture and dataset to prevent overfitting or underfitting.

Phase 6: Model Testing

Test your model on the test dataset.

Phase 8: Deployment and rediction Interface

Convert model_pred.tflite or model_pred_quant.tflite into a model.cc source file for Arduino BLE Sense.

Robustness Considerations

• Ensure the model is insensitive to sensor orientation changes.
• Collect data representing the same posture in different orientations and label them consistently.
• Discuss assumptions about sensor positioning, operating points, and corner cases in the report.

Results

• The following has a prediction value of about 81% because of the gyroscope sensor.
• The gyroscope measured the angular velocity of the system and when we were collecting data we kept it at rest so the angular velocity values were kind of gibberish.
• The value of the magnetometer affects the magnetic field of the thighs around it. So we made sure that during data collection and inference, we kept the same setup.

Video Demonstration

For a visual demonstration of this project, please refer to the video linked below:

Project Video Demonstration