Where : at ENS Paris-Saclay, Centre Borelli
Keywords : change-point detection, geometric signal processing, statistics, 3D motion capture
The Centre Borelli is involved in analyzing various protocols designed to quantify human behavior. Examples include projects that use cameras to study human and animal behavior, as well as projects that use 3D markers (motion capture) to study movement. The 3D motion capture systems (Vicon, Qualisys) and vision-based pose estimation pipelines (MediaPipe, DeepLabCut) used in these projects produce, at each time step, the 3D positions of anatomical landmarks. Since the generated signals are non-stationary, one approach to processing them is to use offline change-detection algorithms, which allow the signals to be divided into homogeneous time segments.
The central question of this proposal is : how to define cost functions adapted to the geometric structure of skeleton data, and what mathematical properties must they satisfy so that standard change-point algorithms apply with their usual guarantees?
More details can be found here : https://www.laurentoudre.fr/intern/postdocCPD.pdf
