Gated-CNN: Lightweight Dual-Stream Architecture for Watch-Based Fall Detection

A novel lightweight dual-stream architecture for fall detection in wearables, named Gated-CNN, has been introduced by researchers. This model independently processes accelerometer and gyroscope data using one-dimensional convolutional feature extractors. It overcomes the challenges posed by self-attention mechanisms, which can lead to quadratic computational demands and hinder accurate localization of the brief impact signatures associated with falls. The design features a sigmoid gating module to filter out irrelevant background signals while enhancing features that distinguish falls, a global average pooling layer to condense each stream into a fixed-length descriptor, and a shared classification head that integrates both descriptors for binary fall prediction. This research is available on arXiv under ID 2605.20275.