Graph dictionary learning for the study of human motion.

In this article, we introduce a method inspired by Graph Signal Processing (GSP) for the analysis of human motion based on the 3D positions of skeletal joints. Our approach uses a graph dictionary learning technique, in which each velocity sample is decomposed into a linear combination of a limited set of atoms acquired directly from the data. The efficacy of this methodology is evaluated using a dataset focused on upper limb elevations. We present features and visualizations, and validate the robustness of the approach through the construction of inter-and intra-subject distances. The features are also used as inputs for Human Activity Recognition with competitive results. The interpretability of the features and visualizations obtained from this method make it suitable for applications such as inter-individual comparisons or longitudinal follow-up of patients.