CV People Counter

In a flood, fire or earthquake, two questions matter: who is still inside, and which way is safe right now? Responders rely on sign-in sheets and roll calls, and static exit signs point to the nearest exit even when it’s under water or packed.

CV People Counter answers both. A webcam above each door counts people walking in and out, and a server keeps a live count for every room, floor and building. When a hazard is reported, by staff, a drill button or a sensor, it works out the nearest safe exit for every room. Any screen showing the exit-sign page shows the route, and changes the moment the route does.

An exit sign during a flood: a large arrow and GO UPSTAIRS, with the route to the Room 202 refuge underneath.

How it works

The system in three columns: a webcam and laptop at each door, plus an offline queue and sensors; the server with occupancy, routing, hazards and SQLite; the dashboard and exit signs people see.

At the door, a laptop runs a YOLO11 pose model and a tracker on the webcam. When a person’s head crosses the door line, it sends one event: which door, which way, and when. Video never leaves the laptop.
Nothing is lost offline. Each event goes into a local SQLite queue before it’s sent, so counting carries on when the Wi-Fi drops and the backlog arrives later. Every event has its own id, so a retry is never counted twice, and late events are replayed in time order, giving the same counts as if they had arrived on time.
The server (FastAPI) keeps occupancy per room, floor and building. Counts never go below zero; a room that would is flagged as drifting, so staff can recount it.
Routing is a shortest path over the building’s map of rooms and doors (Dijkstra, with NetworkX). A door costs more the more crowded the room behind it is, and a room with a hazard is all but closed. If every exit is blocked, people are sent up to a refuge floor instead; if nothing is safe, the sign says to shelter in place.
Sensors: an MPU-6050 accelerometer on an Arduino detects shaking with an STA/LTA trigger (the short-term average of the motion against the long-term one), and thresholds turn water, gas and temperature readings into hazards. In an earthquake the signs say DROP · COVER · HOLD ON first and only show routes once the shaking stops.
The dashboard and the signs update live over a WebSocket, and the whole building (rooms, doors, exits, capacities) is one JSON file.

Counting heads, not feet

The first version counted the bottom of each person’s box: their feet. In a doorway, the person behind is often hidden, so a pair walking close together counted as one. Counting the head instead, found with pose keypoints, keeps both people visible from a camera mounted high and tilted down.

Side view: a camera above the door tilted about 45° down, and two people walking toward it, both heads in view.

Someone standing in the doorway also used to flip in, out, in. Now there’s a neutral band around the door line, and a person only counts once their head reaches the far side.

The camera’s view: an inside zone and an outside zone with a neutral band between them along the door line. A path that reaches the far zone is counted; one that turns back in the band isn’t.

How well it worked

About 95 % counting accuracy in our walk tests under good conditions, including two people walking through overlapping. The misses came from people sprinting through: a runner crosses the line in only a few, often blurred frames at 30 fps.
On a laptop with a GTX 1660 Ti, two real webcams (a Logitech Brio 100 and a C270) counted live at the same time at 25 to 27 frames a second each.

Where it could go next

Ideas from the pitch, none of them built: ranking buildings by who needs rescuing first, ESP32 and LoRa radios so it keeps working when the Wi-Fi fails, routes on phones from a QR code in every room, and depth sensors for better counts.