Sensors cost a small car.
Regulatory-grade monitors run $15,000 to $50,000 each. Cities buy a handful. Hundreds of square kilometres are then "covered" by ten data points.
No. 44 · Research · Public Health · Hacksagon 2026
4.2M
deaths a year · outdoor air pollution · WHO 2023Most cities have fewer than ten sensors.
A hardware air-quality monitor costs fifteen to fifty thousand dollars and takes months to install. There are over a million traffic cameras already on poles in US cities. ECO-LENS turns each one into a virtual air-quality sensor using vision-counted vehicles, EPA AP-42 emission factors and Gaussian plume dispersion. No new hardware. Hyper-local. Near-zero cost.
4.2MAnnual deaths · WHO 2023
1M+US traffic cameras already deployed
$15-50KCost of a single hardware sensor
50×50Virtual-sensor mesh per area
5 sWebSocket update interval
Act I · The Problem
The air the sensor reads is not the air you breathe.
A reference monitor on a rooftop two kilometres away does not know that you live next to a six-lane intersection. The communities living near busy roads, often low-income, breathe air two to three times more polluted than the nearest station ever reports.
2-3× under-reported, near roads.
The pollution gradient between an intersection and the nearest fixed monitor is steep. Vulnerable communities are systematically misrepresented on official maps.
Procurement takes months.
Every new pole, every new sensor, every new calibration. Meanwhile the camera infrastructure already on every other corner watches traffic and watches nothing else.
Act II · The Promise
No new hardware.
Only the physics the existing cameras don't know.
Three peer-reviewed pieces, multiplied. The output is a continuous, hyper-local AQI surface · at the cost of the compute alone.
The ECO-LENS Identity
YOLOv8 vehicle counts × EPA AP-42 emission factors × Gaussian plume dispersion = hyper-local AQI
01 · Virtual sensor mesh
Kriging interpolates the gaps.
Ordinary Kriging with variogram modelling builds a continuous pollution surface from a handful of camera-derived points, with confidence intervals at every cell.
A 50×50 virtual sensor mesh from 3-5 camera feeds.
02 · Equivalent cigarettes
μg/m³ becomes a feeling.
Berkeley Earth: 22 μg/m³ PM2.5 over 24 hours equals one cigarette. ECO-LENS streams a live counter, dose accrued, in plain English.
"You have breathed 0.3 cigarettes in the last hour."
03 · Plume particle simulation
Watch the invisible drift.
Pasquill-Gifford stability classes, real-time wind from OpenWeatherMap, thousands of particles colour-coded by PM2.5. The dashboard makes the chemistry visible.
Plumes drift downwind from intersections, in real time.
04 · Green corridor routing
The cleanest route, not the fastest.
A modified A* with edge weights = distance × PM2.5. Compares fastest vs cleanest with an exposure-savings percentage. Same source, same vehicles · different physics.
"2 minutes longer. 40% less PM2.5 inhaled."
Act III · The Science
Every number is traceable to a paper.
Nothing is hand-waved. Every layer of the pipeline points back to a peer-reviewed source · the same sources EPA, WHO and FHWA already trust.
EPA AP-42 · Chapter 13
Per-vehicle-class emission factors for PM2.5, PM10, NOx, CO, CO2, VOC, SO2 · the US gold standard for emission estimation.
Gaussian plume · Pasquill-Gifford A-F
Same model framework as EPA's AERMOD and SCREEN3 regulatory dispersion models, with real-time wind and atmospheric stability.
WHO Air Quality Guidelines 2021
Annual PM2.5 below 5 μg/m³, 24-hour below 15 μg/m³. The thresholds the dashboard surfaces against your live readings.
Global Burden of Disease · dose-response
All-cause mortality RR 1.06 per 10 μg/m³ PM2.5 · respiratory 1.10 · cardiovascular 1.08. The relative risks behind the cigarette equivalence and the health score.
FHWA Traffic Noise Model (TNM)
Reference sound levels per vehicle class, log-distance attenuation, energy-sum aggregation. The same source · vehicles · drives both pollution and noise.
Berkeley Earth · Muller 2015
22 μg/m³ over 24 hours = 1 cigarette. The one number the public actually feels.
The Stack
Async API. WebSocket stream. One docker compose up.
- Python 3.11+
- FastAPI
- Next.js 16 · React 18
- YOLOv8 (Ultralytics)
- NumPy + SciPy
- Leaflet + Deck.gl
- WebSockets · 5s
- SQLite / PostgreSQL
- OpenWeatherMap
- Docker Compose
Act IV · Proof
Hacksagon 2026, shipping form.
One-command deploy
docker compose up --build · backend, frontend, database, all networked. Or autoconfig.sh on Linux/macOS, autoconfig.bat on Windows, with key rotation built in.
9 REST endpoints + WebSocket
/api/sensors, /api/grid, /api/forecast, /api/health-impact, /api/routing/green-path. Five-second WebSocket stream of every sensor, particle and statistic.
Eight scientific engines
Detection, emission, dispersion, kriging, dosimetry, noise, routing, forecasting (Holt-Winters). Each isolated, each test-covered, each documented.
Demo mode included
No camera feed needed for the demo. Simulation mode generates realistic vehicle counts so judges and city procurement officers can run the dashboard immediately.
The cameras are on the poles. The science is in the papers.
I build hackathon-grade research that ships as a real system. Async APIs, WebSocket streams, regulatory-grade physics, accessible dashboards. The kind of work cities can actually pilot.