Skeleton-based Human Action Recognition in a Thermal Comfort Context
John Martins, Katherine A. Flanigan, Christopher McComb · 2023
Thermal comfort optimization is key to ensuring the well-being of building occupants and promoting intelligent energy use in the built environment. In the current space of thermal comfort analysis, many techniques are used ranging from more intrusive wearable sensors and qualitative occupant surveys to less intrusive infrared thermal monitoring and human action recognition (HAR). However, as these methods increase surveillance of building occupants in often complex environments, privacy preservation and accurate analysis are essential for optimal thermal control. This paper focuses on uplifting the ability of skeleton-based HAR for use in thermal comfort-related action recognition as this method has been shown to have promising action recognition accuracy while maintaining user privacy. While we focus here on thermal comfort, our analysis and emphasis on the use and advancement of privacy-preserving technologies naturally extend to other systems (e.g., cyber-physical-social infrastructure systems) underpinned by human-centered, or “social,” objectives and interactions. We benchmark several fundamental deep learning models in the skeleton-based HAR space and compare their performance on a new dataset of thermal comfort-related actions.