RainTrace
Hyperlocal rainfall detection & urban flood mesh. MobileNetV2 + GRU spatiotemporal network turns municipal traffic cameras into optical rain gauges feeding a 100 m micro-lattice road-weather simulation over Amritsar.
Municipalities need real-time, street-by-street flood awareness during monsoon storms without deploying expensive physical rain gauge hardware on every intersection.
Municipalities need real-time, street-by-street flood awareness during monsoon storms without deploying expensive physical rain gauge hardware on every intersection. Centralizing raw CCTV surveillance video creates massive bandwidth saturation and severe citizen privacy violations.
Trained a MobileNetV2 + GRU spatiotemporal deep learning model on 8-frame (5-second) sliding windows across 25–30 fps traffic camera video to quantify optical rainfall signals without physical sensors.
Engineered a 100 m square micro-lattice road-weather mesh simulating 9,119 virtual surveillance towers over a 5 km urban radius in Amritsar on an ESRI Dark Canvas.
Built pure delta-encoded WebSocket streaming on a 1 Hz FastAPI tick loop, achieving 99.7% bandwidth reduction over full grid state broadcasts.
Implemented privacy by construction: camera streams are processed at the edge and discarded; the central server renders road flood states across 81,557 OpenStreetMap (OSM) road segments purely from sparse float reports.