Computer Vision Systems to Analyse Face and Body Movements, to Automated the Assessment of Physical Health, Mobility, and Safety in Natural Settings and Over Time

点击翻译按钮获取中文摘要

My area of research is the development of computer vision systems and algorithms for intelligent health monitoring and rehabilitation technologies. Over the next five years I will focus my work on two important topics in this area: 1) using and advancing recent developments in machine learning to improve how temporal and longitudinal information is incorporated in computer vision health assessment, and 2) fairness and bias in computer vision systems, particularly with respect to old age and disability. 1) Longitudinal health data often contains large amounts of missing information and is almost always irregularly sampled. Principled approaches to model missing data and irregular sampling include Gaussian Processes and, more recently, Neural Ordinary Differential Equations. My research will augment and advance these approaches to take advantage of additional information, including the patterns of missing information and the temporal distribution of samples. I will also investigate how Transformer Networks could be applied to process long-term trends in healthcare image/video data and in relation to other parameters of health or relevant information. 2) My work has identified fairness and bias issues related to the performance of state-of-the-art computer vision facial analysis models on faces of older adults with a physical or cognitive disability, e.g. stroke or dementia. This is an important limiting factor in employing computer systems in healthcare solutions. Deep learning models used to obtain best performing results on standard benchmarks are typically trained on hundreds of thousands to millions of training examples. It is not practical to collect representative examples of this size from various clinical populations to include in the training data. Patient data and identifiable health records (e.g. face images) are highly sensitive and access is often restricted. A more realistic solution is to collect targeted represented training sets and to adapt state-of-the-art models trained on standard benchmarks datasets to the target population via transfer learning. In my work so far, I have examined commonly used transfer techniques, and have also developed and evaluated novel transfer techniques for this purpose. Using these methods, the performance of facial analysis models on target populations improves, but the gap in performance between clinical and healthy populations persists. Over the next few years, I plan to expand my work in this area and, specifically, explore the use of counterfactual inference and uncertainty quantification to improve the performance of transfer learning techniques, particularly as related to computer vision systems used in facial and body movement analysis. This program of research will train seven HQPs and will result in innovative solutions to reduce bias and to better incorporate long-term temporal information in computer vision systems.