Empirical Mode Decomposition Based Multi-Modal Activity Recognition.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Traditional electroencephalogram (EEG) analysis using infinite impulse response filters introduces nonlinear phase distortions.
• Limited information is available when using only energy and entropy as EEG features.
• Conventional activity recognition relies on single signal types, hindering performance.

Discussion:

• This paper proposes empirical mode decomposition (EMD) to decompose EEG into intrinsic mode functions (IMFs), categorizing them into four groups.
• Eleven distinct physical quantities are extracted from IMFs as features, expanding analytical information.
• A multi-modal system integrating EEG, image sequences, and motion signals is developed for enhanced activity recognition.

Key Insights:

• The multi-modal approach significantly outperforms single-signal methods, with three signal types yielding higher accuracy than two or individual signals.
• The proposed EMD-based method surpasses conventional filtering-based techniques, highlighting the benefits of nonlinear, adaptive time-frequency analysis.
• This research offers a more robust and accurate solution for at-home child activity monitoring.

Outlook:

• Future work could explore additional signal modalities or advanced feature extraction techniques.
• Further research into real-world deployment and scalability of the multi-modal system is warranted.
• Investigating the application of this EMD-based multi-modal approach in other domains could yield valuable insights.