Most building energy-saving strategies are based on thermal comfort standards defined by regulatory bodies (e.g., ASHRAE), which do not account for individual differences or preferences of the building occupants. To address this gap, this paper proposes a non-intrusive automated method to capture the individual thermal discomfort states of occupants over time based on their thermal discomfort cues. The method aims to support the identification of personalized energy-saving strategies toward improved occupant thermal comfort and reduced building energy consumption. It uses a deep learning model to recognize occupant cues from videos. To implement and test the proposed approach, participants were video recorded in uncontrolled office settings while performing work-related activities. The deep learning model was used to recognize the thermal discomfort cues, and a thermal comfort profile that captures the discomfort states of occupants over time was developed. The method achieved a weighted recall and precision of 84% and 96%, respectively, in recognizing building occupant thermal discomfort cues.