GhostNet

Video Demo

About this project

Falls are the leading cause of injury-related death in adults over 65, and the risk is highest for those living alone, where no can help them up. Existing solutions force a choice between cameras or compliance (wearables that elderly people simply don't wear). A 2022 Carnegie Mellon paper showed WiFi radio waves, absorbed differently by human bodies, could reconstruct human pose but they used industrial antenna arrays costing thousands. We asked: what if you could do this with a $9 chip? GhostNet turns two ESP32-S3 boards into an invisible, always-on safety net. By reading WiFi Channel State Information how radio signals change as a person moves through a room. It detects position, identifies falls in real time, and monitors heart rate in real time. When a fall is detected and the person doesn't get back up within a grace period, GhostNet automatically calls their caregiver via Twilio. No camera. No wearable. Every event streams into Snowflake for historical analysis and longitudinal health tracking. No public CSI dataset matched our hardware. So we made one. Each team member fell, lay down, and sat 50 times each, across all three actions generating 6,032 labeled CSI samples directly from our ESP32-S3 chips in real conditions. Every sample streamed into Snowflake via Snowpipe as it was collected. We then trained a CNN fall detector on that data using Snowpark ML. The model learned to distinguish a real fall from a slow lie-down using WiFi signals. No camera supervision, no wearable ground truth. For heart rate, a teammate collected his own CSI breathing and pulse data and used it to normalize a public dataset to our hardware, extracting heart rate from raw WiFi signal fluctuations with no wearable required. What's next: expanding to 4-6 nodes for full house coverage and Snowflake Cortex for detecting declining mobility weeks before a fall ever happens.

Gallery