Adaptive Fall Detection Using WiFi CSI for Unseen Environments and New Individuals
Israa Bayad, Sandy Mahfouz, Khouloud Samrouth, Farah Mourad-Chehade, Hassan Amoud · 2025
Quick detection of falls is crucial for elderly individuals to prevent severe injuries, reduce hospitalization risks, and ensure timely medical intervention. This work introduces a scalable, non-invasive fall detection system that leverages WiFi Channel State Information (CSI), eliminating the need for wearable devices or intrusive cameras, thus providing a privacy-preserving solution. Our lightweight 2D-CNN architecture consists of only 202,705 parameters, making it computationally efficient while achieving up to 100% accuracy in controlled settings, where training and testing data are collected from the same indoor environments with different sets of activities. It also demonstrates strong generalization, detecting new falls in unseen environments with 87% accuracy and falls from entirely new individuals with 79% accuracy. To enhance adaptability to real-world conditions, various data augmentation techniques such as shadowing and time shifting are applied, leading to a 13% improvement in generalization accuracy compared to baseline models up to 97% larger. Designed for real-time operation, our system is highly effective in practical fall scenarios. It also contributes to human activity recognition (HAR), demonstrating broader applicability beyond fall detection. These findings establish our model as a robust, efficient, and practical solution for fall detection and beyond.