Thermal comfort assessment based on occupants’ action recognition using computer vision - A review

• A comprehensive review of the integration of action recognition with thermal comfort assessment. • Multimodal fusion enhances the robustness of action recognition in complex scenarios. • Lightweight deep learning models allow efficient edge-computing action recognition. • Limited dataset diversity reduces generalizability. • Privacy-preserving methods reduce the risk of identity disclosure but adversely affect accuracy. Indoor thermal comfort has significant impacts on occupants’ physiological health, psychological state, and productivity. Conventional assessment methods, such as questionnaires or physiological measurements, tend to be intrusive and suffer from feedback latency. With advances in computer vision, action-recognition–based thermal-comfort assessment has emerged as a promising approach that is both non-intrusive and real time. However, existing reviews remain unclear or incomplete on key aspects of this approach, for example, the mapping between thermal-adaptive behaviours and thermal sensation, the available datasets, and the underlying techniques and end-to-end workflow. Accordingly, this article systematically traces the development of thermal-comfort assessment and synthesizes recent studies on action-recognition–based approaches, providing a comprehensive account of their methods and pipelines. We conducted a systematic search of the relevant literature and then performed in-depth analyses of action-recognition models for adaptive behaviours, real-world deployment, and privacy issues. We critically examine the literature on visual data acquisition, feature extraction and behaviour-to-comfort mapping, classification algorithms, and the integration of recognized actions with HVAC control strategies. Some studies indicate that integrating action recognition into HVAC control system can reduce thermal comfort prediction error, increase occupant satisfaction and achieve energy savings. We further discuss current deployment challenges, dataset construction, and privacy concerns, outlining both the limitations of the state of the art and directions for future research. Taken together, this review provides a theoretical foundation and practical guidance for integrating action recognition into intelligent building systems, laying the groundwork for occupant-centric, energy-efficient environmental control.