Reliable head pose estimation is critical for human activity monitoring, yet machine-learning (ML) approaches typically provide no explicit measures of uncertainty or reliability. We propose a hybrid strategy that embeds ML-based pose estimation within a rigorous photogrammetric framework. A calibrated multi-camera system observes both static reference markers and dynamic markers attached to the human head. Through bundle adjustment, 3D marker positions together with uncertainty, redundancy, and correlation measures are estimated. By modelling the head-mounted markers as an inner sub-network, stability and reliability indicators are derived and reprojected into image space, directly informing markerless ML-based keypoint detection. We validate the approach on synthetic data and a real test measurement in the context of radiotherapy patient monitoring. Results show that the framework reliably separates rigid head motion from local facial deformation, achieving a signal-to-noise ratio above 6 for face deformation detection, and that studentized per-point stress indicators effectively partition the marker network into deformation-active and noise-dominated subsets, providing a principled basis for spatially adaptive keypoint weighting.