Activity Recognition from Daily-Life Sounds Using Unsupervised Learning with Dirichlet Multinomial Mixture Models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Ambient assisted living is crucial for the elderly living alone.
• Activity recognition systems often require extensive labeled data, increasing costs.
• Household sounds contain rich information about daily activities.

Purpose of the Study:

• To develop a cost-effective method for recognizing daily activities of the elderly using household sounds.
• To reduce the reliance on labeled data for activity recognition model development.
• To support ambient assisted living initiatives.

Main Methods:

• An unsupervised learning approach using a Dirichlet multinomial mixture model.
• Modeling the generative process of neural audio codec codes conditioned on latent activities.
• Extending the model to handle multiple streams of audio codes from different sound directions.

Main Results:

• The proposed unsupervised learning method effectively clusters daily activities from household sounds.
• Handling multiple audio code streams improves activity cluster accuracy.
• The model leverages the burstiness of code occurrence patterns for better recognition.

Conclusions:

• The developed Dirichlet multinomial mixture model offers a viable unsupervised approach for activity recognition.
• This method significantly reduces the need for labeled data and user inquiries.
• The approach is a key component for scalable ambient assisted living systems.