Real-Time Sensor-Embedded Neural Network for Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for healthcare and human-computer interaction.
• Existing HAR systems often rely on external processing, limiting real-time applications.
• There is a need for efficient, on-device HAR solutions.

Purpose of the Study:

• To introduce a novel, self-contained sensor for real-time human activity recognition.
• To demonstrate the feasibility of implementing a Convolutional Neural Network (CNN) on a low-cost microcontroller for HAR.
• To evaluate the performance of the embedded system for predicting human behaviors.

Main Methods:

• Development of a wearable sensor integrating an Inertial Measurement Unit (IMU) and a low-cost microcontroller.
• Implementation of a Convolutional Neural Network (CNN) directly onto the microcontroller for real-time data processing.
• Real-time data acquisition from the IMU and activity prediction using the embedded CNN.

Main Results:

• The proposed sensor accurately detects and predicts human activities in real-time.
• The embedded CNN achieves high inference performance on the low-cost microcontroller.
• The system successfully eliminates the need for external processing devices.

Conclusions:

• The developed sensor offers an accurate and efficient solution for embedded, real-time human activity recognition.
• This approach enables practical, on-device HAR applications without external computational resources.
• The findings highlight the potential of edge computing for advanced activity monitoring.