New physics-inspired AI detects falls in real-time on edge devices

Current vision-based fall detection often treats falling as a static pose or a discrete temporal pattern, missing the underlying physics of instability. To address this, researchers introduce a physics-informed framework that reimagines falling as a stability-loss event in a coupled dynamical system. The core innovation is a dual-LTC (Liquid Time-Constant) architecture split into a Center-of-Mass (CoM) subsystem and a Base-of-Support (BoS) subsystem. These continuous-time neural networks model inertial trajectory evolution and ground-contact adjustment using adaptive time constants, providing physical interpretability of falling motion. A learnable coupling module emulates the interaction between CoM and BoS, while a Stability Manifold classifier detects boundary crossings via Lyapunov-inspired stability metrics. The system also includes counterfactual trajectory projection and Time-to-Collision (TTC) estimation for early warning and irreversibility assessment.

The architecture is designed to support a three-state prediction paradigm (Normal, Falling, Fallen). In this preliminary study, the team validated core stability discrimination on a two-class dataset (Normal vs. Falling), achieving competitive accuracy while keeping the network under 50,000 parameters. This tiny footprint allows real-time inference on low-power edge platforms such as surveillance cameras or wearable devices, unlike heavier CNN-RNN pipelines. The framework's physical interpretability offers a significant advantage in safety-critical applications, as it can explain why a fall is predicted. Future work will extend to full three-state temporal transitions and real-world deployment in elderly care and intelligent surveillance.