Smart Activity Tracker using RT-Spark Development Board

Project Overview

This project is a full-stack IoT system designed for the real-time classification of user activities. Using an RT-Thread RT-Spark Development board equipped with an accelerometer and gyroscope, motion data is captured and transmitted over Wi-Fi. The system bridges the gap between raw hardware signals and meaningful user insights by leveraging cloud-based analytics and mobile visualization.

Software Architecture

The project is divided into three distinct layers:

• The Edge: An RT-Spark board collecting 6-axis accelerometer and gyroscope readings at an optimized sampling rate.
• The Brain (Back-end): A Flask web server hosting a pre-trained Machine Learning model. It receives raw data, extracts temporal features, and predicts activities (Walking, Running, or Stationary).
• The Interface: An Android mobile application that provides a real-time dashboard and a historical log of user activity retrieved from a Firebase database.

Technical Implementation

Phase 1: Hardware Configuration

The RT-Thread RT-Spark Development board served as the edge sensing device. Due to the integration of the ICM20608 sensor suite, high-fidelity motion data was accessible without external peripheral wiring. The device was configured for high-speed Wi-Fi transmission to maintain a continuous data stream to the back-end.

Phase 2: Embedded Firmware Development (RT-Thread Studio)

Firmware development was conducted within RT-Thread Studio.

• Driver Initialization: Onboard sensor drivers were enabled to capture raw data from the 3-axis accelerometer and 3-axis gyroscope.
• Network Management: The WLAN management framework was utilized to establish a stable TCP/IP connection.
• Data Serialization: Sensor readings were aggregated into 30-sample windows, serialized into JSON format, and transmitted via HTTP POST requests to the server’s REST API endpoint.

Phase 3: Machine Learning Model Development

Activity recognition was powered by a supervised learning model trained on a comprehensive motion dataset.

• Feature Extraction: Statistical features, specifically the mean and standard deviation of each axis, were calculated to represent the temporal rhythm of movement.
• Model Training: A lightweight Scikit-Learn classifier was selected to ensure rapid inference times. The model was optimized to distinguish between "Stationary, " "Walking, " and "Running" states.
• Deployment: The finalized model was serialized into a .pkl format for integration into the Python-based Flask environment.

Phase 4: Cloud Back-end and Database Integration

The Flask server functioned as the central processing unit of the architecture.

• Inference Engine: The server received JSON packets, extracted real-time features, and passed them through the ML model to generate predictions and confidence scores.
• Firebase Realtime Database: Predictions were pushed to a cloud database. A Throttling Mechanism was implemented to update the "Current Activity" node instantly while limiting "History" log entries to every 10 seconds (or upon activity change) to prevent database congestion.

Phase 5: Mobile Application Development

The user interface was developed using Android Studio and Java.

• Real-time Synchronization: Firebase listeners were utilized to ensure the UI responded to cloud changes with negligible latency.
• User Experience (UX): The dashboard was designed to provide visual cues (color-coded backgrounds) corresponding to the detected activity. A RecyclerView was implemented to render the historical logs efficiently.

Challenges and Solutions

Conclusion

The completed system demonstrates a successful integration of embedded systems, cloud computing, and machine learning. By offloading complex inference to a Flask back-end and utilizing a real-time database, the prototype provides accurate, low-latency activity tracking suitable for modern wearable applications.Project Demo:
https://drive.google.com/file/d/1wDI3Qmi6LiGicCj13V2mVDGC7vzYfNl_/view?usp=sharing