reubenrouse/crowd-whatif

Crowd What-If

Simulated crowds where you know what everyone wants, and "what if?" pairs: the same crowd moment replayed under different gate plans, with the true outcome of each.

Use it to train and test models that watch a crowd, guess where people are heading, and predict what happens if you close or open a gate.

One test moment. First the 2 s a model gets to watch, then what really happens under each gate plan. Everyone here wants exit B (orange). Closing B sends them to exit A, which breaks up the queue but gets fewer people out in 10 s.

• 15,000 scenes: 6,000 halls, 5,000 train platforms, 4,000 bottlenecks. 1.85M people, 5 frames per second.
• Hidden labels: each person's intended exit, current target and desired speed.
• Actions: gates closing and opening, train arrivals.
• 9,145 what-if moments, each replayed under 2–3 gate plans ("keep", "close B", "open C").
• Realistic simulator: JuPedSim, tuned to real crowd experiments (10.6% flow error).

Real recordings only show what actually happened. Here you also get what would have happened.

How it was made

1. Scenes: random rooms with 1–3 exits, waiting crowds, steady streams and train doors. Each crowd has its own hurry level (0.8–1.6 m/s). Rushed crowds also stand closer together.
2. Gates close and reopen during scenes. People heading for a closed exit walk to the nearest open one.
3. What-if moments: the simulator restarts from everyone's exact position and replays 10 s once per plan: keep the gates as they are, or change one gate 1 s in. Nobody new enters during the replay.

Splits are by scene: 13,500 train, 750 validation, 750 test. What-if moments: 7,072 train, 1,039 validation, 1,034 test.

Files

Config	One row per	Main columns
scenes	scene	scene_id, split, family, layout as polygons (walkable_wkt, exits_wkt, gates_wkt), gate_events, hidden hurry
trajectories	person per frame	scene_id, person_id, t (s), x, y (m), target (current exit, hidden)
people	person	scene_id, person_id, intended exit, desired speed (hidden)
whatif	moment	moment_id, scene_id, k0 (last watched frame), plans, exit_counts (people out per exit in 10 s, per plan)
whatif_truth	person per frame per plan	moment_id, plan, person_id, n (frames after k0), x, y

Frame = round(t × 5), counted from the start of the scene. The 2 s a model "watches" are frames k0-9 … k0.

from datasets import load_dataset
import json

moments = load_dataset("reubenrouse/crowd-whatif", "whatif", split="test").to_pandas()
traj = load_dataset("reubenrouse/crowd-whatif", "trajectories", split="test").to_pandas()

m = moments.iloc[0]
watched = traj[(traj.scene_id == m.scene_id) & (traj.t.mul(5).round().between(m.k0 - 9, m.k0))]
for j, plan in enumerate(json.loads(m.plans)):
    print(plan["name"], "-> people out per exit:", json.loads(m.exit_counts)[j])

Benchmark (test set)

Every method sees only 2 s of positions, never the hidden labels.

Method	Exit-count error (people)	Gets the effect's direction right	Finds the right gate change when one helps	Picks the least crowded plan
Random	–	–	–	48%
Never touch the gates	–	–	0%	72%
Simulator, everyone to the nearest exit	6.0	83%	47%	77%
Simulator, exit guessed from walking direction	4.7	79%	58%	74%
Crowd World Model	3.7	93%	80%	80%

Crowding = people in the busiest 1 m² after 10 s.

Limitations

• It's simulated: no pushing, simple route choice, rooms up to 16 × 16 m.
• Nobody new enters during a what-if replay.
• The simulator is deterministic, so each plan has one true outcome.

Licence

CC BY 4.0. Made with JuPedSim. Code to rebuild everything: github.com/reubenjrouse/crowd-world-model.