Human Activity Recognition with an HMM-Based Generative Model.

Area of Science:

• Computer Science
• Machine Learning
• Healthcare Informatics

Background:

• Human Activity Recognition (HAR) is crucial for healthcare applications like elder support and disease diagnosis.
• Smart devices generate vast amounts of sequential data daily.
• Existing models may not fully capture the proportional nature of activity data.

Purpose of the Study:

• To propose an unsupervised, scaled, Dirichlet-based Hidden Markov Model (HMM) for HAR.
• To model sequential patterns in human activities using a novel statistical approach.
• To leverage the strengths of HMMs for continuous data flow analysis.

Main Methods:

• Developed unsupervised, scaled, Dirichlet-based Hidden Markov Models (HMMs).
• Assumed emission probabilities follow a bounded-scaled Dirichlet distribution for proportional data.
• Employed variational inference for model parameter learning.
• Evaluated the model using a publicly available dataset.

Main Results:

• The proposed model demonstrates effectiveness in analyzing human activities from sequential data.
• The scaled Dirichlet distribution proved suitable for modeling proportional emission probabilities.
• Variational inference provided a viable method for learning the complex model.

Conclusions:

• The novel Dirichlet-based HMM offers a robust method for unsupervised HAR.
• This approach enhances capabilities in health monitoring and disease diagnosis through activity analysis.
• The model's performance validates its potential in leveraging smart device data for healthcare.