Feature Representation and Data Augmentation for Human Activity Classification Based on Wearable IMU Sensor Data...

Area of Science:

• * Wearable sensor technology and machine learning applications.
• * Biomedical engineering and human-computer interaction.

Background:

• * Wearable inertial measurement unit (IMU) sensors are crucial for collecting human motion data for activity recognition.
• * Effective human activity recognition (HAR) relies on appropriate feature representations and classifiers, but is hindered by limited labeled data.
• * Data-driven learning models require sufficient labeled data to accurately assess performance capabilities.

Purpose of the Study:

• * To propose a spectrogram-based feature extraction method using raw IMU sensor data.
• * To introduce an ensemble of data augmentation techniques in feature space to mitigate data scarcity.
• * To evaluate the impact of these methods on HAR accuracy using a deep long short-term memory (LSTM) neural network.

Main Methods:

• * Development of a spectrogram-based feature extraction approach from raw IMU sensor data.
• * Implementation of an ensemble of data augmentation techniques applied in the feature space.
• * Utilizing a deep long short-term memory (LSTM) neural network for performance testing and evaluation.

Main Results:

• * The proposed feature extraction and data augmentation ensemble achieved state-of-the-art accuracy in HAR.
• * Performance evaluations demonstrated the significant influence of each augmentation approach on classification accuracy.
• * The technique yielded state-of-the-art results on both a custom dataset and the UCI HAR dataset.

Conclusions:

• * Spectrogram-based feature extraction combined with feature-space data augmentation effectively enhances HAR accuracy.
• * The proposed methods successfully address the challenge of limited labeled data in HAR.
• * The approach demonstrates robust performance and achieves state-of-the-art results across different datasets and learning rates.