A Tiny Transformer for Physiotherapy Exercise Recognition based on Pose Landmark Time Series
Rita Meziati, Mehdi Roudesli, N. Ducrocq · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Human pose estimation is one of the most powerful applications of machine learning and computer vision. It consists of identifying (and tracking in the case of input videos) body key landmarks such as the head, shoulders, hands or legs. It has various use cases such as pose and action recognition, motion tracking and 3D character animation. In our study, we use pose estimation to extract x and y body landmark coordinates from input videos of human participants performing physiotherapy functional exercises. We developed a transformer-based model and trained it on a small dataset, in order to recognize these exercises (14 exercises in total). We achieved a 66% accuracy on our test data. In this paper, we present the dataset we used and the methodology we followed, including the data processing steps and the classification model conception. The pose landmark time series extracted from our dataset will be publicly shared.