From 2D video markerless acquisition of infant motion to anatomy constrained trajectories using a Kalman-Filtering approach

Preterm birth represents a major risk factor for neurodevelopmental disorders (NDDs). Assessment of General Movements is a well-established diagnostic tool; however traditional visual evaluations are subjective and time-consuming. Recent 2D markerless video tracking provides a low-cost, automated alternative, yet it is intrinsically limited to the image plane and lacks 3-dimensional evaluations. The aim of this thesis was to bridge this gap by developing a 3D biomechanical model of the infant integrated into an Unscented Kalman Filter (UKF), that estimates anatomy-constrained 3D trajectories from single-camera 2D video recordings. The UKF was tested on 21 clinical videos of preterm infants at 40 weeks and 3 months corrected age. It integrates anatomical constraints, including segment-length consistency, physical boundaries (such as crib surface), and physiological joint Range of Motion (ROM). Although this is a preliminary work and the absolute 3D accuracy requires further validation against a multi-camera ground truth, the framework demonstrates the feasibility of extracting physically plausible, constrained 3D poses from conventional monocular 2D videos. This represents a promising step toward non-invasive 3D computational tools for the early clinical detection of NDDs.