pierrelouis-c/aporia-bends

Aporia corridor bends

Steady-state pedestrian density maps for 1,466 simulated one-way corridor bends. Each map comes with the geometry rasterised as network inputs. This is the training data of CFD-U.

Simulator: JuPedSim 1.4.2, collision-free speed model, configuration calibrated on experiments.

Design

1,500 bends drawn over four parameters, each simulated with two seeds. 1,466 passed the acceptance checks.

Parameter	Range
theta_deg: bend angle	30–150°
width: corridor width	1.5–4 m
cut: inner-corner chamfer	0–3.9 m
specific_flow: inflow	0.1–0.9 ped/(m·s)

Files

dataset.npz (NumPy). Every map is a 96 × 96 grid at 0.25 m/pixel, a 24 m window centred on the turn.

Key	Shape	Content
inputs	(1466, 5, 96, 96) float32	walkable, wall_distance (m), inlet, exit, load (ped/(m·s))
target	(1466, 96, 96) float32	density (ped/m²), mean of the two seeds
mask	(1466, 96, 96) bool	pixels that are walkable and valid in both seeds
seed_maps	(1466, 2, 96, 96) float32	each seed's map (noise floor)
geometry	(1466,) int64	geometry id
split	(1466,) str	train, validation, test_interpolation, test_extrapolation
channel_names	(5,) str	names of the input channels

manifest.jsonl: one line per geometry, with its shape parameters, split, the status of each run and its seed-to-seed RMSE.

dataset.json: versions and counts.

Splits

Split	Maps	Angles
train	896	30–120°
validation	112	30–120°
test_interpolation	111	30–120°
test_extrapolation	347	120–150°, never seen in training

import numpy as np
z = np.load("dataset.npz")
train = z["split"] == "train"
x, y, m = z["inputs"][train], z["target"][train], z["mask"][train]

Limits

These are simulations, not measurements: single bends, one-way flow, everyday loads, and no pushing or panic.