Hybrid Graph-Recurrent Architectures for Spatio-Temporal Motion Analysis and Fall Forecasting

Abstract

Falls constitute a predominant etiology of injury across demographic spectra and represent a significant contributor to morbidity, mortality, and escalating healthcare expenditures in Australia. Mitigating such incidents necessitates integrating intelligent surveillance systems capable of preemptive detection and rapid intervention. To support these goals, our study advances a deep learning model designed to anticipate and classify fall-related human motion by modeling the intricate spatiotemporal interdependencies inherent in bodily dynamics. The proposed approach is developed and evaluated exclusively using the TelUP dataset. Among our exploration on graph-recurrent models, GCN-GRU operationalizes a synergistic representation of spatial posture and temporal progression to forecast imminent falls with heightened precision. Empirical validation reveals a marked performance improvement over extant baselines, attaining an accuracy of 97.40%, an F1-score of 96.89%, a mean per-joint position error (MPJPE) of 113.4 pixels, and a mean per-joint velocity error (MPJVE) of 64.66 pixels. Supplementary analyses concerning computational complexity and comparative visual evaluation corroborate the model’s superior efficacy relative to prevailing architectures, notably outperforming the LSTM benchmark. Future work will evaluate the proposed approach across additional public datasets and real-world deployment settings to assess robustness while accounting for inference-intervention latency under variations in physical environments.

Manuscript received: 01 Feb 2026 | Revised: 27 Mac 2026 |  Accepted: 29 Apr 2026 | Published: 31 Jul 2026