Dynamic Segmentation of Sensor Events for Real-Time Human Activity Recognition in a Smart Home Context

Area of Science:

• Computer Science
• Artificial Intelligence
• Ubiquitous Computing

Background:

• Human Activity Recognition (HAR) is crucial for smart building services.
• Robust HAR systems for real-world deployment face significant challenges.
• Existing research often relies on pre-segmented sensor data, limiting real-time application.

Purpose of the Study:

• To investigate real-time human activity recognition using streaming sensor data.
• To develop a methodology for dynamic event windowing based on spatio-temporal correlations.
• To accurately identify the start of new activities from continuous sensor streams.

Main Methods:

• Proposed a dynamic event windowing technique incorporating spatio-temporal correlation.
• Developed an algorithm with three steps: sensor correlation (SC) verification, temporal correlation (TC) verification, and activity trigger determination.
• Applied the approach to the 'Aruba' dataset from the CASAS database for real-world validation.

Main Results:

• The proposed method achieved high segmentation quality for multiple activities, including sleeping, meal preparation, and housekeeping.
• Segmentation performance was assessed using the F1 score, yielding results between 0.63 and 0.99.
• Demonstrated the effectiveness of using sensor correlation and temporal patterns for real-time activity recognition.

Conclusions:

• The developed methodology provides a robust solution for real-time human activity recognition in smart environments.
• Dynamic event windowing based on spatio-temporal correlation enhances the accuracy of activity segmentation.
• This approach overcomes limitations of pre-segmented data, enabling more confident deployment in ordinary real environments.