Human Activity Recognition by Combining a Small Number of Classifiers.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) systems often rely on numerous sensors.
• Systems with limited sensors face challenges in accuracy and robustness.
• Combining outputs from multiple low-resource sensors is crucial for effective HAR.

Purpose of the Study:

• To develop advanced Bayesian models for HAR using a minimal set of wireless inertial sensors.
• To enhance the accuracy and reliability of HAR systems with few sensors.
• To address the challenge of combining soft outputs from individual classifiers.

Main Methods:

• Proposed novel Bayesian models integrating soft outputs from individual classifiers.
• Incorporated the dynamic nature of human activities using a first-order homogeneous Markov chain.
• Developed inductive and transductive inference methods for supervised and semi-supervised learning.

Main Results:

• Demonstrated consistent reduction in error rates compared to single-classifier approaches.
• Showcased increased robustness against sensor failures.
• Outperformed existing combination models lacking soft outputs and Markovian structure.

Conclusions:

• The proposed Bayesian models offer a superior approach for HAR with limited sensors.
• These models enhance system accuracy, robustness, and adaptability to different learning scenarios.
• The integration of soft outputs and Markovian dynamics is key to improved HAR performance.