ActiSight: Wearer Foreground Extraction Using a Practical RGB-Thermal Wearable

Area of Science:

• Computer Vision
• Wearable Technology
• Human-Computer Interaction

Background:

• Wearable cameras are crucial for life-logging and human activity recognition.
• Current foreground extraction methods using only RGB data are computationally expensive and energy-inefficient.
• Accurate wearer segmentation is fundamental for many body-worn video applications.

Purpose of the Study:

• To develop a practical solution for wearer segmentation in body-worn video.
• To leverage thermal information to improve the efficiency and accuracy of foreground extraction.
• To reduce the computational and energy demands of wearable camera applications.

Main Methods:

• Development of ActiSight, a wearable RGB-Thermal video camera system.
• Collection of 59 hours of video data from 6 participants in natural settings.
• Evaluation of wearer foreground extraction performance using the Dice similarity score.

Main Results:

• ActiSight successfully extracts wearer foreground with high accuracy (high Dice similarity score).
• The RGB-Thermal approach significantly reduces execution time compared to RGB-only methods.
• Energy consumption is substantially lowered when using ActiSight for wearer segmentation.

Conclusions:

• ActiSight offers a practical and efficient solution for wearer segmentation in body-worn video.
• Integrating thermal data enhances the performance of wearable camera systems.
• This technology enables more feasible implementation of complex tasks like activity recognition on resource-constrained devices.