HyperFallNet: Human Fall Detection in Real-Time Using YOLOv13 With Hypergraph-Enhanced Correlation Learning

Falls remain a leading cause of injury-related morbidity among elderly populations and vulnerable individuals in smart living environments. Despite advances in deep learning-based fall detection systems, existing methods often suffer from trade-offs between real-time efficiency, detection accuracy, and semantic interpretability. Most state-of-the-art models either rely on computationally expensive architectures that are impractical for real-time edge deployment or cannot model complex fall-related spatial cues. This paper introduces HyperFallNet, a novel lightweight yet semantically rich fall detection framework designed for fast and accurate inference in real-world surveillance scenarios. The core motivation behind this work stems from the persistent gap between high-speed models and context-aware reasoning in fall detection. The objective of this paper is to bridge this gap by proposing an enhanced YOLOv13-based pipeline augmented with two architectural innovations: 1) HyperACE, a hypergraph-assisted attention module that captures high-order spatial dependencies among human joints and body posture: and 2) FullPAD, a pipeline-aware semantic injection mechanism that ensures refined feature propagation across the network hierarchy. We evaluate HyperFallNet on two benchmark datasets, CUCAFall and DiverseFALL10500, and conduct rigorous ablation studies and statistical tests. Results show that HyperFallNet achieves a mean average precision (mAP@0.5) of 0.982 while operating at 131.2 FPS—surpassing recent fall detection models in both accuracy and speed. Additionally, our model demonstrates statistically significant improvements and stable confidence intervals across key metrics. These findings highlight HyperFallNet’s suitability for real-time, privacy-aware, and resource-constrained fall detection applications.