Functional and Shape Data Analysis under the Frenet-Serret Framework: Application to Sign Language Motion Trajectories Analysis

Perrine Chassat

Abstract: This thesis, conducted in collaboration with MocapLab, a company specializing in motion capture, aims to determine the optimal mathematical framework and relevant descriptors for analyzing sign language motion trajectories. Drawing on principles of motor control, we identified the framework defined by the Frenet-Serret formulas, including curvature, torsion, and velocity parameters, as particularly suitable for this task. By introducing new curve analysis approaches based on the Frenet framework, this thesis contributes to developing novel methods in functional data analysis and shape analysis. The first part of this thesis addresses the challenge of smoothly estimating Frenet curvature parameters, treating the problem as parameter estimation of differential equation in $SO(d)$, $(d \ge 1)$. We introduce a functional Expectation-Maximization algorithm that defines a unified variable estimation method in the $SE(3)$ group, providing smoother estimators that are more reliable and robust than existing methods. In the second part, two new curve representations are introduced: unparametrized Frenet curvatures and the Square Root Curvatures (SRC) transform, establishing new Riemannian geometric frameworks for smooth curves in $\mathbb{R}^d$, $(d \ge 1)$. Leveraging higher-order geometric information and parametrization dependence, the Square Root Curvatures transform outperforms the state-of-the-art Square-Root Velocity Function (SRVF) representation on synthetic results. Given a collection of curves, this type of geometry allows us to define efficient statistical criteria for estimating Karcher mean shapes on the associated Riemannian shape spaces, proving particularly effective on noisy data. Finally, this developed framework opens the door to more practical applications in sign language processing, including the study of power laws on our data and the development of a generative model for a point motion in sign language.