Modular Bayesian Networks with Low-Power Wearable Sensors for Recognizing Eating Activities

Area of Science:

• Human-Computer Interaction
• Wearable Computing
• Activity Recognition

Background:

• Recognizing complex daily activities using smartphones and wearable sensors is challenging due to diverse real-life contexts.
• Existing methods struggle with power, memory, and user obtrusiveness constraints in wearable environments.
• Eating is a typical complex activity that requires sophisticated recognition methods.

Purpose of the Study:

• To develop a method for recognizing complex activities, specifically eating, using only low-power mobile and wearable sensors.
• To systematically organize contextual information using a model based on activity theory and the "Five Ws".
• To propose a probabilistic approach for handling uncertain contexts in activity recognition.

Main Methods:

• Constructed a context model based on activity theory and the "Five Ws".
• Proposed a modular and tree-structured Bayesian network with 88 nodes for probabilistic context prediction.
• Collected data from 25 volunteers performing 10 different activities using mobile and wearable sensors.

Main Results:

• Achieved 79.71% accuracy in recognizing complex activities, outperforming conventional classifiers by 7.54-14.4%.
• The probabilistic approach provided approximate results even with missing or heterogeneous sensor data and contexts.
• Demonstrated the effectiveness of the Bayesian network in handling uncertainty and complexity in real-world activity recognition.

Conclusions:

• The proposed probabilistic Bayesian network effectively recognizes complex daily activities using low-power sensors.
• The method is scalable and robust, offering practical solutions for wearable activity recognition systems.
• Contextual modeling significantly enhances the accuracy and reliability of activity recognition in diverse environments.