Feature-Free Activity Classification of Inertial Sensor Data With Machine Vision Techniques: Method, Development, and...

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Inertial sensors are crucial for human activity recognition (HAR) and exercise detection (ED).
• Traditional methods rely on complex digital signal processing (DSP) and feature engineering, requiring specialized expertise.
• Many researchers lack the necessary DSP background for effective feature-set development.

Purpose of the Study:

• To present a novel application of machine vision for HAR and ED.
• To simplify entry into HAR and ED research by reducing the need for deep DSP skills.
• To leverage transfer learning with a pretrained convolutional neural network (CNN) for exercise classification.

Main Methods:

• A CNN, a machine vision technique, was applied to exercise detection (ED).
• Time series plots from accelerometer and gyroscope signals were used to retrain an Inception neural network.
• Data from 82 volunteers performing 5 exercises using an inertial measurement unit (IMU) were collected and analyzed.

Main Results:

• The proposed method achieved 95.89% accuracy in classifying exercises.
• This accuracy is competitive with current state-of-the-art techniques in ED.
• Classification performance was compared to conventional feature-extraction and random forest methods.

Conclusions:

• Distinct waveform morphologies in time-series plots enable effective machine vision approaches for ED.
• The novel use of machine vision and high-level machine learning frameworks simplifies HAR research.
• This approach facilitates access for researchers without extensive DSP or machine learning backgrounds.