FLASH: Efficient Impact Fall Detection with Unified Hypergraph State-Space Model

Researchers introduced FLASH, a skeleton-based fall-impact detection model that uses biomechanically informed hyperedges plus a Mamba state-space network.

FLASH is designed to pinpoint the exact moment of impact during a fall while keeping inference efficient enough for real-time use.

On UP-Fall and UMAFall, it delivered strong accuracy, lower compute cost, and better cross-dataset generalization than graph-, dual-representation-, and transformer-based methods.

The approach could improve safety monitoring for older adults and support deployment on edge devices.