Physical rehabilitation is central to the recovery process after many type of musculoskeletal and neurological injuries. Considering limited resources in the healthcare system, delivering effective rehabilitation at home is crucial to achieving optimal outcomes. For this reason, tools to track rehabilitation activities in different environments are needed to generate data that will assist with individualized treatment planning, performance monitoring, and the development of evidence to support the most effective approaches.

Systems that combine video data with deep learning are achieving impressive performance in tracking posture and recognizing activities. However, individuals with disabilities may perform movement exercises in varied ways. In order to support effective tracking of rehabilitation, systems are needed that recognize activities without requiring large numbers of examples of the same activity being performed in exactly the same way. As an important step towards this goal, the objective of this project will be to develop a video-based deep learning approach that can perform few-shot action recognition.

The student will be responsible for:

• Developing a deep learning system that integrates existing neural networks for encoding motion data into a few-shot learning architecture.
• Evaluating the performance of the system on public action recognition datasets, using different motion encoders.