Real-Time Human Physical Activity Recognition with Low Latency Prediction Feedback Using Raw IMU Data.

Area of Science:

• Human Activity Recognition (HAR)
• Machine Learning
• Wearable Sensor Technology

Background:

• Supervised and deep learning methods commonly train on extracted time and frequency features from sensor data.
• Raw inertial and gyroscopic data are underexplored in Human Activity Recognition (HAR) model training.
• Custom wireless sensors with embedded IMU and sEMG offer potential for direct raw data utilization.

Purpose of the Study:

• To investigate the effectiveness of using raw sensor data for Human Activity Recognition (HAR).
• To develop and evaluate a real-time HAR system using custom wireless motion sensors.
• To achieve high motion classification accuracy with low latency.

Main Methods:

• Recording a dataset from able-bodied participants using 3 custom wireless motion sensors with IMU and sEMG.
• Implementing a Support Vector Machine with Radius Basis Function Kernel (RBF-SVM) classifier on a Raspberry Pi 3 base station.
• Augmenting the RBF-SVM classifier with Spherical Normalization for enhanced motion classification.

Main Results:

• Achieved a motion classification accuracy of 97.35% for 8 distinct body motions.
• Demonstrated the feasibility of using raw sensor data for HAR without extensive feature extraction.
• The proposed classifier provided real-time prediction callbacks with low latency output.

Conclusions:

• Raw sensor data, when processed with advanced algorithms like augmented RBF-SVM, can yield high accuracy in HAR.
• Custom wireless sensors integrated with embedded processing offer a viable solution for real-time HAR applications.
• The developed system presents a promising low-latency, high-accuracy solution for Human Activity Recognition.