FluidFlow
Real-time crowd crush prevention using physics-informed neural networks and computer vision detecting dangerous crowd states 30-120 seconds before they become visible.
FluidFlow - AI-Powered Crowd Crush Prevention System
FluidFlow is a real-time early warning system that predicts crowd crush events before they become visible to the human eye. It treats dense crowds as physical fluids and applies computational fluid dynamics principles to detect dangerous crowd states 30 to 120 seconds before critical pressure develops.
The system works entirely on existing CCTV infrastructure no new cameras, no new sensors, no additional hardware. Video footage is processed through a four-stage pipeline: YOLOv8-nano detects persons per frame and maps crowd density zone by zone, Farneback Optical Flow computes a dense velocity field across every pixel, a NumPy physics engine derives Turbulence Intensity, Reynolds Number, Pressure Proxy and Shockwave Flag by treating the crowd as a non-Newtonian fluid governed by Navier-Stokes dynamics, and a PyTorch LSTM forecaster reads the last 30 seconds of metric data to predict crowd state 30 seconds ahead.
When predicted Turbulence Intensity exceeds 0.30 or Reynolds Number crosses 3000, a Level 2 alert fires before the crush is visible giving authorities time to open exits, redirect movement, and deploy ground staff.
The core insight is grounded in peer-reviewed research. Helbing et al. (2007) documented that real crowd disasters transition through measurable laminar to turbulent flow phases before visible panic providing a detection window that existing systems completely ignore. Current solutions count people. FluidFlow measures how they are moving, how pressure is building, and where the shockwave is forming.
Deployed across 50 high-risk locations railway stations, religious gatherings, stadiums FluidFlow projects a 40% reduction in stampede fatalities, translating to 50 to 100 lives saved per year across India.