AutoTherm: A Dataset and Ablation Study for Thermal Comfort Prediction in Vehicles

Abstract

State recognition in well-known and customizable environments such as vehicles enables novel insights into users and potentially their intentions. Besides safety-relevant insights into, for example, fatigue, user experience-related assessments become increasingly relevant. As thermal comfort is vital for overall comfort, we introduce a dataset for its prediction in vehicles incorporating 31 input signals and self-labeled user ratings based on a 7-point Likert scale (-3 to +3) by 21 subjects. An importance ranking of such signals indicates higher impact on prediction for signals like ambient temperature, ambient humidity, radiation temperature, and skin temperature. Leveraging modern machine learning architectures enables us to not only automatically recognize human thermal comfort state but also predict future states. We provide details on how we train a recurrent network-based classifier and, thus, perform an initial performance benchmark of our proposed thermal comfort dataset. Ultimately, we compare our collected dataset to publicly available datasets.