m monad-knowledge FIIT STU · open research vault
dataset

atc-shopping-mall

Also known as: ATC, ATC dataset, ATC shopping mall, ATR ATC pedestrian tracking

What it is

The ATC dataset (Brščić, Kanda et al., IEEE Trans. Human-Machine Systems 2013) is a real pedestrian-tracking corpus collected by 49 3D range sensors covering ~900 m² of the Asia and Pacific Trade Center (ATC) in Osaka — a shopping center + transportation hub + conference center. It spans 92 days (24 Oct 2012 – 29 Nov 2013), logging per-person position, velocity, motion direction and body-facing angle at high rate, plus a registered occupancy grid / floor plan of the tracked area.

Its value here is orthogonal to the floorplan corpora: it is real crowd flow in a real public venue, with a floor map — the ground-truth footfall the placement-oracle's counting and coverage-under-crowd claims can be validated against, instead of only ray-traced or jupedsim-simulated crowds. It directly attacks the sim-to-real crowd gap flagged across the CSI campaigns.

Reference paper: D. Brščić, T. Kanda, T. Ikeda, T. Miyashita, "Person Tracking in Large Public Spaces Using 3-D Range Sensors", IEEE THMS 43(6):522–534, 2013. One-year usage analysis: Brščić & Kanda, IEEE THMS 45(2):228–237, 2015.

Verified specs

Confirmed by downloading + parsing the real files (not the page).

  • Trajectory CSV — headerless, comma-separated, 8 columns: time_s (unixtime + ms), person_id, x_mm, y_mm, z_mm (height), velocity_mm_s, motion_angle_rad, facing_angle_rad. Positions in mm.
  • Sample day atc-20121114.csv (the day we mirrored): 15,606,041 rows, 18,145 distinct persons, 10.9 h (09:40–20:20), tracking extent ~90 m × 52 m (x∈[-41.5, 48.4] m, y∈[-27.8, 24.2] m). 996 MB uncompressed (208 MB upstream .7z).
  • Occupancy grid map/localization_grid.pgm + .yaml — ROS map_server format, 0.05 m/px, origin [-60, -40, 0] m, same coordinate frame as the trajectories. map/ATC-ITM_building_tracking_area-entry_exit.pptx carries the annotated floor plan + entry/exit gates.
  • docs/list_of_days.txt enumerates all 117 candidate dates (92 clean; some flagged * for partial-day tracking gaps). docs/atc-sensors.xlsx = sensor placement/calibration. code/ = upstream readers (ATCRawDataRead.py, loadallcsv.sql).

Acquisition & storage

  • Status: acquired-partial. Mirrored to s3://monad-knowledge/datasets/atc-shopping-mall/:
    • map/ — occupancy grid (.pgm+.yaml) + floor-plan .pptx
    • tracking/atc-20121114.csv — one full sample day (1.04 GB, uncompressed for byte-range loading)
    • docs/ — day list + sensor calibration · code/ — upstream readers
  • NOT mirrored: the full 9-part / ~40 GB 92-day trajectory corpus (atc-tracking-part1..9.7z) and the raw 3D-sensor sample (tar.gz, ~15 min). They stay upstream — pull more days on demand only if a campaign needs them (list_of_days.txt is the index; upstream atc-tracking-partN.7z).
  • Loading (S3-only, bounded): read datasets/atc-shopping-mall/_manifest.json first. The map is small; the day CSV is 1 GB and headerless — get_range the head or stream rows in batches, never load whole. Positions are mm; divide by 1000 for metres to overlay on the .pgm (0.05 m/px, origin [-60,-40]).
  • License: research use only, citation required; our bucket is private.

Why it matters here

  • Real footfall ground truth for the placement-oracle. The AP-placement + crowd-counting work (c-placement-oracle, coverage-meets-crowds) runs on simulated crowds; ATC provides a real public-venue occupancy field with a floor map to check whether coverage/count-informativeness heuristics hold against real pedestrian density — a real-data anchor the sim corpus lacks.
  • Public-venue crowd dynamics (arrivals, dwell, flow bottlenecks at entries) to condition or validate the jupedsim agenda engine (ip107-agenda-engine) for non-residential venues — complementing the residential-flat crowd priors.
  • Pairs with the geometry gap fill. archcad-400k / s3dis give public geometry; ATC gives public crowd behaviour on real geometry — geometry + dynamics for the same class of space.