Real-time Fall Detection Based on Vision for Low-Power Edge Platforms

This paper reframes visual fall detection as a stability-loss event in a coupled human-support dynamical system. Its dual-LTC architecture models the Center of Mass (CoM) and Base of Support (BoS) with adaptive time constants, links them through a learnable coupling module, and classifies boundary crossing in a joint Stability Manifold space using Lyapunov-inspired metrics. Counterfactual trajectory projection and Time-to-Collision (TTC) estimation are intended to assess irreversibility and provide early warnings. The reported network has fewer than 50K parameters and targets real-time inference on constrained edge hardware. The preliminary validation covers only Normal-versus-Falling classification; the proposed three-state Normal/Falling/Fallen transition remains future work.