Bridging the visual-to-physical gap: physically aligned representations for fall risk analysis

Introduction – Vision-based fall analysis has advanced rapidly, yet a key bottleneck remains: visually similar motions can correspond to markedly different physical outcomes, as subtle differences in contact mechanics and protective responses are difficult to infer from appearance alone. Existing approaches typically rely on supervised injury prediction, which requires reliable clinical labels. In practice, such labels are difficult to obtain due to ambiguity in video evidence (e.g. occlusion and viewpoint limitations) and the rarity and ethical constraints of real injury events, leading to noisy supervision. Methods – We propose PHARL (PHysics-aware Alignment Representation Learning), a framework that learns physically meaningful fall representations without requiring clinical outcome labels. PHARL introduces two complementary regularization mechanisms: (1) trajectory-level temporal consistency to stabilize motion representations, and (2) physics-aware alignment, where simulation-derived multi-class contact outcomes are used to structure the embedding space. By associating video windows with temporally aligned simulation descriptors, the model captures local impact-relevant dynamics while maintaining a purely feed-forward inference process. Notably, physics information is only used during training as a structural regularizer and does not define an explicit outcome predictor. Results – Experiments on four public datasets demonstrate that PHARL consistently improves risk-aligned representation quality compared to visual-only baselines, while maintaining strong performance on fall detection tasks. Discussion – Beyond quantitative improvements, PHARL exhibits an emergent ordinal structure in the learned representation space, where an interpretable severity ordering (Head > Trunk > Supported) arises without explicit ordinal supervision. This suggests that physics-aware regularization can induce meaningful structural priors in representation learning, offering a promising direction for label-efficient fall analysis.