Improving Responsiveness of Fall Detection using Spiking Neural Networks

Fall detection systems, essential for the safety of elderly individuals, have increasingly incorporated Deep Neural Networks (DNNs) for improved accuracy. However, real-time processing, especially on resource-constrained wearable devices, remains a major challenge due to the computational intensity of DNNs. Recently, Spiking Neural Networks (SNNs) have shown promise in improving energy efficiency in fall detection. Despite these advancements, the potential of SNNs to leverage their temporal dynamics—particularly, the ability to make sequential decisions based on the timing of input spikes—has not been fully explored for enhancing detection responsiveness. This paper proposes a novel approach to exploit the temporal nature of SNNs for faster fall detection. Specifically, we introduce a novel encoding technique, Quick Spike Encoding (QSE), which prioritizes critical inputs based on signal amplitude in temporal dimension, and a Linear Weighted Mean Squared Error Count (LW-MSEC) Loss Function, which emphasizes early detection by penalizing errors more heavily at the beginning of the detection process. Evaluations on two fall detection datasets demonstrate that our method significantly enhances the responsiveness of SNNs (>60%), achieving 91% accuracy in just 10 time steps compared to 25 time steps with traditional approaches, on the SisFall dataset.