Human Motion Models for Automated Prediction and Assessment of Motor Disorders by Computer Vision

ABSTRACT: The observation of human movement is a fundamental tool in the prediction, evaluation and monitoring of a lot of different pathologies, from rheumatology to neurodegenerative diseases. To face the diversity of symptoms and the need for early detection of subtle signs, several types of movement are now currently used: gait, writing, eye movement, facial expressions, etc. While deep learning algorithms currently represent the state of the art for video classification in general, the best way to merge data of different types remains an open problem. Furthermore, building a reliable model from a small amount of annotated data is a major challenge for medical applications. We present in this plenary some approaches based on machine learning and computer vision to address the following research objectives: (1) Propose a non intrusive method (i.e. without markers or worn device) for automatic evaluation of patients' stage from standard RGB videos. (2) Build generic human motion models that can be learned from a reduced quantity of data. (3) Design deep learning fusion models that can combine efficiently data from different movement modalities.

Professor of Computer Science at ENSTA Paris and Researcher in Computer Vision within the Autononous Systems and Robotics Laboratory.