Deep Learning architectures for generating rehabilitation human motion

Abstract

Exercise-based rehabilitation programs play a vital role in helping patients recover from injuries or surgeries, significantly enhancing their quality of life while reducing mortality rates and the likelihood of re-hospitalization. However, traditional rehabilitation methods are often resource-intensive, requiring continuous supervision by therapists. This creates a substantial barrier for patients who need to attend rehabilitation sessions multiple times a week. One way to resolve this is by providing technological support for home-based rehabilitation. Recent advancements in deep learning offer promising new avenues for automating and improving these rehabilitation practices by generating human motion patterns. Previous works in the domain of motion generation have rarely explored the generation of rehabilitation movements. In this project, we propose a deep learning-based framework for generating rehabilitation human motion. Given a prescribed action type and a quality score, we aim to generate plausible 3D rehabilitation human motion sequences. We propose a conditional Variational Auto-Encoder (VAE) that encourages a diverse sampling of the motion space according to a specified performance score.