General movement assessment (GMA, Prechtl’s method) relies on visual evaluation of spontaneous neonatal motion and benefits from objective 3D quantification. However, clinical videos are often monocular and lack metric (true) depth, limiting biomechanical inference. We process clinical videos with a pose-landmark detector (33 points) pipeline and a monocular depth estimator. Inverse depth is converted to true depth via a frame-wise scale recovery fixed by anatomical points (constant bone lengths) and temporal regularization (Gaussian, bilateral, and sliding window smoothing). From the shoulder–elbow–wrist segment, we calculate elbow angles with four estimators (law of cosines, cross product, rotation matrix, and quaternion) and a robust ensemble method. Wrist trajectories reconstructed from angles are compared against a Denavit–Hartenberg (DH) kinematic model. The data consist of nonidentifiable, publicly available neonatal videos (split 70/20/10 train/validation/test); pose thresholds were set to 0.52 for detection/tracking. The ensemble method consistently outperformed individual estimators for trajectory tracking (MAPE range 0.65–6.68% across axes and wrist). The ensemble model obtained a similarity with the DH model of 81.24% (left) and 87.13% (right). The framework converts monocular video into true depth and clinically interpretable kinematics in a lightweight pipeline, providing a reproducible complement to GMA and a foundation for bias-aware neonatal movement analysis.