Falls present a significant global public health challenge, especially in today's aging society, underscoring the importance of developing an effective fall detection system. Non-invasive radio-frequency (RF) based fall detection has garnered substantial attention due to its wide coverage and privacy-preserving nature. Existing RF-based fall detection systems approach falls as an activity classification problem, assuming that human falls introduce reproducible patterns to the RF signals. However, we argue that falls are inherently accidental, making their impact uncontrollable and unforeseeable. We propose a fundamentally different approach to fall detection by shifting the focus from directly identifying hard-to-quantify falls to recognizing normal, repeatable human activities, thus treating falls as abnormal activities outside the normal activity distribution. We introduce a self-supervised incremental learning system incorporating FallNet, a deep neural network that employs unsupervised learning techniques. Our real-time fall detection system prototype leverages WiFi Channel State Information (CSI) sensing data and has been extensively tested with 16 human subjects.
View on arXiv