A Compact Dual-IMU Fall Detection Model for TinyML Deployment at Low Sampling Rates

Falls in the elderly pose a significant health risk, necessitating reliable, real-time detection systems suitable for efficient operation on wearable devices. This paper introduces a lightweight fall detection solution employing dual IMUs at the waist and thigh, operating at a low 20 Hz sampling rate for TinyML platform deployment. The system extracts angular velocity and Root Mean Square (RMS) features, processed by a compact Convolutional Neural Network (CNN) with integrated 1D and 2D layers to discern temporal and inter-sensor motion patterns. On a public dataset of 21,847 windows, the dual-IMU configuration achieved a compelling <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{F 1}$</tex>-score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 9 5 2 2}$</tex> and a recall of 0.9733, significantly outperforming single-sensor approaches. Crucially, the model demonstrates high efficiency, requiring less than 130 KB of flash memory (97 KB on STM32F407) and performing inference in under <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$60 ~\text{ms}(36 ~\text{ms}$</tex> on STM32F407) on typical microcontrollers. These findings affirm the feasibility of accurate, responsive fall detection on resourceconstrained embedded systems without demanding high sampling frequencies or overly complex architectures.