Models / moi

moi/stadium-crowd-estimator

Stadium Crowd Estimator

Government-RestrictedHigh riskGov ReuseComputer VisionPyTorchN/A1.1.0

Test in Sandbox

260 8.2K/30d 14

MAE (persons/zone)

11.8

MAPE (occupancy)

7.6%

Congestion-alert precision

0.87

Latency p95 (A10G)

92 ms

About this model

Density-map crowd estimation model for stadium CCTV, producing per-zone occupancy counts and flow-rate estimates for concourses, gates, and stands. Tuned on Qatari venue geometries from the World Cup legacy stadium network.

Intended use

Safety operations in approved venue command centres: gate-load balancing, egress monitoring, and early congestion alerts. Not for identification of individuals.

Training data lineage

Trained on the stadium-crowd-annotations set of 68,000 annotated CCTV frames across eight venues and multiple event types.

Limitations & bias notes

Counts degrade under heavy flag and banner occlusion and in low-light concourse zones, with errors up to 15%. Camera repositioning after maintenance requires zone-mask recalibration or estimates silently drift.

Evaluation metrics

MAE (persons/zone)	11.8
MAPE (occupancy)	7.6%
Congestion-alert precision	0.87
Latency p95 (A10G)	92 ms

#crowd-counting#cctv#stadiums#safety#density-estimation

Try it

Live sandbox

stadium_gate_c_1840.jpg

fixture

Evaluation only: gated models run against redacted sample data until your entity's entitlement is approved.

Feedback

Owning entity

moMinistry of Interior

Updated: 2026-02-17
Latest version: 1.1.0
License: Government Internal Reuse
Access: Gated

Trained on

stadium-crowd-annotations →