Real-time fall detection based on vision for low-power edge platforms

2026-07-14

AI Summary

This work addresses the common oversight in existing vision-based fall detection methods—the dynamic instability of the human support system—by modeling falls as stability loss events in a coupled dynamical system. The authors propose a dual Liquid Time-Constant (LTC) neural network architecture that separately captures the continuous evolution of the center-of-mass trajectory and the base of support, while a learnable coupling module emulates their physical interaction. Within a shared latent space, the framework integrates a Lyapunov-like stability metric, stability manifold classification, counterfactual trajectory projection, and time-to-collision estimation to enable physically interpretable early warnings. The resulting model, with fewer than 50,000 parameters, achieves state-of-the-art accuracy in binary fall detection while remaining lightweight enough for real-time deployment on edge devices.

Abstract

Falling detection is vital for elderly care and intelligent surveillance; however, prevailing vision-based approaches predominantly frame it as static pose classification or discrete temporal pattern matching, fundamentally overlooking the instability dynamics of the human support system. This paper proposes a physics-informed falling detection framework that recasts falling as a stability-loss event in a coupled dynamical system. We introduce a novel dual-LTC architecture comprising a Center-of-Mass (CoM) subsystem and a Base-of-Support (BoS) subsystem, both instantiated as Liquid Time-Constant (LTC) neural networks to continuously model inertial trajectory evolution and ground-contact adjustment through adaptive time constants, Physical interpretability of falling motion. A learnable coupling module emulates physical interaction between the two subsystems, while a Stability Manifold classifier operates in the joint latent space to detect boundary crossing via Lyapunov-inspired stability metrics. Complementary counterfactual trajectory projection and Time-to-Collision (TTC) estimation further enable irreversibility assessment and early warning. The architecture is designed to support a three-state prediction paradigm (Normal, Falling, Fallen); in this preliminary study, we validate the core stability discrimination capability on a two-class dataset (Normal vs. Falling), leaving the full three-state temporal transition to future work. Unlike conventional CNN--RNN pipelines, the proposed formulation encodes continuous-time mechanical inertia, yielding a sub-50K-parameter network capable of real-time inference on resource-constrained edge devices. Extensive experiments demonstrate competitive accuracy with superior physical interpretability, validating its efficacy for low-compute visual fall detection.

Problem

• fall detection
• vision-based
• stability dynamics
• edge computing
• human support system

Innovation

• physics-informed
• Liquid Time-Constant networks
• stability manifold
• fall detection
• edge computing