A diffusion model for inertial based time series generation on scarce data availability to improve human activity...

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Human Activity Recognition (HAR) relies on sensory systems like Inertial Measurement Units (IMUs) to differentiate human movements.
• Current methods for generating synthetic HAR data struggle with generalization to new subjects and limited sensor types.
• Generating realistic synthetic data is crucial to overcome the time and cost constraints of real-world data collection.

Purpose of the Study:

• To develop a novel method for generating multi-IMU synthetic human motion sequences.
• To enhance the generalization capability of HAR models to unseen participants.
• To improve classification performance in HAR by augmenting training datasets with synthetic data.

Main Methods:

• Adapted a denoising diffusion probabilistic model, originally from the vision domain, for synthetic human motion generation.
• Generated synthetic data from multiple IMUs, focusing on meaningful human motion sequences.
• Evaluated synthetic data quality through visual analysis using a novel clustering approach and by assessing classifier improvement.

Main Results:

• The proposed model successfully generated meaningful multi-IMU human motion sequences.
• Adding synthetic samples to training data led to significant improvements in HAR classification tasks.
• Performance gains were observed even with very limited real data (as few as 2 samples per subject).

Conclusions:

• Diffusion models can effectively generate high-quality synthetic multi-IMU data for HAR.
• Synthetic data generation offers a viable solution to reduce data collection burdens in HAR research.
• The approach demonstrates strong generalization capabilities, benefiting HAR applications with scarce datasets.