Two-stage posture enhancement and stability modelling for pre-fall risk detection in athletic movements.

Background: In sports training and athletic performance, fall risk degrades performance and elevates injury incidence, making early warning of pre-fall states critical. Monocular 3D pose estimation often yields incomplete or biomechanically inconsistent skeletons due to occlusion, blur, and detector noise, limiting reliable monitoring of movement stability and quality. Method: We propose a two-stage deep learning framework: 1. RePoseNet: Refines initially estimated 3D skeleton sequences generated by a pretrained monocular pose estimator, rather than performing monocular 3D pose estimation from scratch. It repairs missing or unreliable joints, suppresses temporal jitter, and enforces biomechanical plausibility through joint symmetry, limb-length consistency, and joint-angle constraints. 2. StaFallNet: Integrates stability metrics-Step Frequency Variation Rate (SFVR), Trunk Leaning Angle (TLA), and Stride Variability (SV)-into a dynamic graph network to model spatial coordination and temporal dynamics for frame-level and event-level pre-fall risk prediction. Methodology Details: RePoseNet combines GRU-based temporal repair, cross-joint attention, and biomechanical refinement (limb symmetry, joint-angle limits). StaFallNet extracts biomechanical features (SFVR, TLA, SV), constructs dynamic joint graphs modulated by stability cues, and uses attention pooling for risk classification. Evaluation: Tested on a fall-related subset of UCF101 with manually annotated frame-level labels. Pre-fall frames were defined as frames within a fixed temporal window before the visually identified fall onset, while non-risk frames were sampled from stable movement periods. Results: RePoseNet improves temporal smoothness and biomechanical plausibility under controlled skeleton-level corruption. Reference-based Mean Per Joint Position Error and PCK evaluations indicate that the refined skeletons remain consistent with the initial pseudo-3D pose estimates while reducing temporal instability; these metrics are not interpreted as evidence of improved absolute 3D pose accuracy. More importantly, downstream experiments demonstrate that the refined skeletons provide more reliable inputs for stability-aware fall-risk prediction. Conclusion: This work offers a robust, interpretable approach for fall-risk prediction in sports. It provides actionable insights for performance evaluation, training intervention, and injury prevention, bridging computational modeling with biomechanical principles.