Mobile Thermography-based Physiological Computing for Automatic Recognition of a Person’s Mental Stress
This thesis explores the use of Mobile Thermography1, a significantly less investigated sensing capability, with the aim of reliably extracting a person's multiple physiological signatures and recognising mental stress in an automatic, contactless manner. Mobile thermography has greater potentials for real-world applications because of its light-weight, low computation-cost characteristics. In addition, thermography itself does not necessarily require the sensors to be worn directly on the skin. It raises less privacy concerns and is less sensitive to ambient lighting conditions.
The work presented in this thesis is structured through a three-stage approach that aims to address the following challenges: i) thermal image processing for mobile thermography in variable thermal range scenes; ii) creation of rich and robust physiology measurements; and iii) automated stress recognition based on such measurements. Through the first stage (Chapter 4), this thesis contributes new processing techniques to address negative effects of environmental temperature changes upon automatic tracking of regions-of-interest and measuring of surface temperature patterns. In the second stage (Chapters 5,6,7), the main contributions are: robustness in tracking respiratory and cardiovascular thermal signatures both in constrained and unconstrained settings (e.g. respiration: strong correlation with ground truth, r=0.9987), and investigation of novel cortical thermal signatures associated with mental stress. The final stage (Chapters 8,9) contributes automatic stress inference systems that focus on capturing richer dynamic information of physiological variability: firstly, a novel respiration representation-based system (which has achieved state-of-the-art performance: 84.59% accuracy, two stress levels), and secondly, a novel cardiovascular representation-based system using short-term measurements of nasal thermal variability and heartrate variability from another sensing channel (78.33% accuracy achieved from 20seconds measurements). Finally, this thesis contributes software libraries and incrementally built labelled datasets of thermal images in both constrained and everyday ubiquitous settings. These are used to evaluate performance of our proposed computational methods across the three-stages.