Open-source crowd-safety AI · built and tested for Indian crowds

See a crowd crush coming, before it happens.

Crowd Shield counts every head in CCTV, drone and phone images, turns the count into people per square metre on a satellite map, watches camera feeds for zones filling up or surging, and plans where to move people. These are the warning signs crowd-safety research says appear minutes before a crush.

Try the live demo Technical report (PDF) Source code
86–91%counting accuracy on 10 hand-counted
Indian photos (pilot)
86.1%counting accuracy, packed crowds
(ShanghaiTech A)
92.1%counting accuracy, streets
(ShanghaiTech B)
11 / 11staged escapes caught (UMN);
1 of 3 real panics
0false alarms on the same test
105freely licensed Indian crowd photos in the test set
Live demo

Use the real app, right here

This is the working system, not a recording. Count a crowd photo, run the Kumbh and Madurai site maps, or see how accurate it is.

crowd-shield-india.streamlit.app Open full screen ↗

Runs on free hosting. If nobody has used it for a while it is asleep: press Yes, get this app back up if you see it, and give it about a minute.

  • Count people in a photo
  • Analyse a video
  • Watch a live camera
  • Site maps
  • Evacuation plan and alerts
  • How accurate is it?
Inside the app

What you see when you use it

Analyse a video
Analyse a video: Every head marked, zones turning amber and red as they fill, and each zone over time.

Every head marked, zones turning amber and red as they fill, and each zone over time.

Site map
Site map: Zones on satellite imagery of the Sangam, people per m², and walking routes to move people.

Zones on satellite imagery of the Sangam, people per m², and walking routes to move people.

Evacuation plan and alerts
Evacuation plan and alerts: Each zone against its limits, and orders for every zone's team (example numbers).

Each zone against its limits, and orders for every zone's team (example numbers).

How accurate is it?
How accurate is it?: Every hand-counted Indian photo, with the model's count and the true one.

Every hand-counted Indian photo, with the model's count and the true one.

Why this matters

Crushes don't come from nowhere

In a crush, people die standing up, squeezed until they can't breathe. Density and pressure build for minutes before that happens. Crowd Shield measures both, using cameras that are already there.

Maha Kumbh, Prayagraj · 29 January 2025
30+ killed

On the Mauni Amavasya bathing day, barricades gave way near the Sangam Nose and the crowd surged (Uttar Pradesh Police).

Chithirai festival, Madurai · 16 April 2022
2 killed

More than five lakh people gathered for Lord Kallazhagar's entry into the Vaigai river, and a crush broke out on the riverbank (police reports).

The number that matters is people per square metre

A head count alone can't tell you if a crowd is dangerous. You also need the ground area.

0–2free movement
≈ 2.2normal walking breaks down (Fruin)
5+serious risk (Still)
≈ 7crowd moves like a fluid
How it works

From a camera frame to an evacuation plan

1

Count every head

A point model (PET) puts a dot on each head, even where bodies overlap. A density model and YOLO are the fallbacks.

2

Measure the ground

Zones drawn on satellite imagery give the real area in m², and a ground calibration does the same for a camera view.

3

Read the danger

Density sets the risk level. On video, optical flow tracks speed and crowd pressure, and an alarm fires on a sudden surge.

4

Act

An optimiser plans where to move people, only along routes they can walk and without overfilling any zone, and alerts each zone's team by Telegram or email.

Crowd Shield architecture: input, measure, understand, act
Site maps

Density on the real ground

Two of India's biggest gatherings, with zones drawn on satellite imagery. The colours show a simulated surge using the Fruin "standard" thresholds. In the app, live counts fill these zones and animated arrows show where to move people.

Results

Every number has its data and code

Tested on held-out data only: no model was picked using the test images. Accuracy = 1 − total counting error ÷ total true count.

Counting · ShanghaiTech test sets

From 7% to 86% on packed crowds

MethodPacked (A)Streets (B)
Accuracy by crowd size

PET holds up on large crowds of over 1,000 people. It is weaker on small groups: two of the 10 pilot photos were 21% and 32% off. When the app can't tell how much ground a photo covers, it says Unknown instead of Safe.

Surge alarm · UMN escape videos

11 of 11 escapes caught, no false alarms

The alarm fires a median of 0.7 s before the labelled start of the panic. It held up when each scene was left out in turn while tuning: lawn 2/2, indoor 6/6, plaza 3/3.

Alarm timeline on UMN clips
India test set

105 photos of melas, festivals, stations and temples

All freely licensed from Wikimedia Commons, with credits. On the first 10 hand-checked photos the first model scored 66.7%, because it took railway gravel for heads. The second, also trained on 1,200 dense crowd photos and on empty Indian tracks, scores 91% (86% on the free online server).

What I learned

Where it fails, and what I did about it

A safety tool that hides its failures is dangerous. These are published in full.

Gravel looks like heads

From above, railway ballast reads as a sea of heads: 1,267 counted where there were 415. Retraining fixed that photo but hurt the others, so I kept the better model and documented the failure.

Phones move

All 15 freely licensed Indian crowd clips I found were filmed handheld, so speed readings measured the camera, not the crowd. The app now detects a moving camera and hides motion results instead of guessing.

No GPU, still useful

On a plain 4-core CPU, PET takes about 16 s per photo. The free server has 3.2 GB, so it counts big photos in overlapping pieces: same accuracy, about 30 seconds each. A live feed is recounted every 30 s or more; video-rate analysis would need a GPU.

Field test · Indian festival video
Head counts on Indian festival video frames

Counting works on handheld Indian clips (Holi in Howrah, Bhagoriya, Gangasagar). Motion analysis needs a fixed camera, such as CCTV.

No facesIt counts heads as points. It never recognises anyone.
No identitiesMovement is measured per zone, never per person.
Nothing storedPhotos and clips stay in memory for your session only. No database.
Open sourceCode, models, data credits and results are all public.