Skip to the main content
FlockOpen Flock

Why we study crowds, and what we found.

The first study is the crowd forecast, the one behind the number on a venue card wherever Flock knows that venue's weekly pattern. A venue without one gets an estimate from what is typical for its kind. The second, a sensor that counts the people actually there, has only been tested on research footage so far.

Research summary · October 2026 · 3 pages

Forecasting How Busy a Venue Will Be, and Counting How Many People Are There

Jayden Bansal, Flock Social LLC

Read the summary (PDF)Full paper (PDF, 30 pages)

Why we did this research

A group deciding where to go wants to know one thing: how busy will this place be when we get there? Map apps answer with a popular-times chart, each venue's usual level for that weekday and hour. That chart can't know tonight is different, like when it rains or there's a game two blocks away.

So we asked two questions. Can AI tell how crowded a venue will be before you go? And can a credit-card-sized computer model be designed to count the number of people inside a venue using a thermal camera without identifying individuals?

The numbers side by side

Flock against the popular-times chart

FlockChart

Within one crowd level
Flock 78.4%
Chart 64.8%
Exact crowd level
Flock 54.9%
Chart 30.2%
Within 10 points
Flock 55.2%
Chart 30.9%
Share of the same 119,939 live readings, one hour ahead, every one the forecast was served for at 1,754 venues in Philadelphia, the Lehigh Valley and Miami from 9 September to 8 October 2026. None of them was used to set the forecast. Higher is better.

Owl against the heat rule it replaced

OwlRule

Flux's camera modelSDT (TU Wien), indoor rooms, 2,000 frames
Owl 89.7%
Rule 77.1%
Ceiling camera2,347 frames
Owl 92.9%
Rule 9.2%
Rooms with hot laptops265 frames
Owl 95.5%
Rule 71.3%
Wall-mounted meeting rooms813 frames
Owl 93.0%
Rule 51.5%
Side view2,151 frames
Owl 82.1%
Rule 36.6%
Frames counted exactly, on six benchmark sets that no model trained on. The sealed tests, including the sets where the rule did better, are in the paper.

What it covers so far

Flock's crowd numbers are scores from 0 to 100 for each venue, not headcounts, and the live test covers one month. Its findings were tested on 6.5 million public hourly counts from 31 sources in 15 countries, including people counted inside rooms and libraries. The counter has been tested on six research datasets and in fifteen sealed tests on public footage, including a café and a restaurant, and not yet inside a working venue.

What's next

A retrain of the crowd model is scored in early November against thresholds and held-out days written down before their data existed, and it ships only if it beats today's numbers on days it never saw. After that, the plan is to put a Flux sensor in a venue Flock already reads and compare its counts with the forecast hour by hour.

Thank you

Both studies were tested on data that other people collected and chose to share. These are the labs, cities and libraries behind it, each with the credit its licence asks for.

Thermal images, for Owl

  • SDT, real and rendered, Computer Vision Lab, TU Wien (Pramerdorfer, Strohmayer and Kampel) (by permission)
  • Thermal Vision, classrooms, Arijit Samal and Haroon R. Lone, IISER Bhopal (by permission)
  • Thermo Presence, PUT Vision Lab, Poznań University of Technology (MIT)
  • SenSys 2021 low-resolution thermal set, Zhu, Voigt, Perez-Ramirez and Eriksson (CC BY 4.0)
  • VAP Trimodal, Aalborg University (Palmero et al.) (CC BY 4.0)
  • OpenThermalPose2 and OpenThermalAction, ISSAI, Nazarbayev University (MIT)
  • InfoLab MLX90640 set, InfoLab, Sungkyunkwan University (MIT)
  • Thermal Dogs and People, Roboflow Universe (CC BY 4.0)
  • Multi-person localization, HTPA 32x32, Klir et al. (GPL-3.0)
  • BU-TIV, Atrium and Lab, Wu, Fuller, Theriault and Betke, Boston University (by permission)
  • TIDOS, Cokbas, Ishwar and Konrad, Boston University (research use, by permission)
  • Thermal soccer, Gade and Moeslund, Aalborg University (CC BY-SA 4.0)
  • Long-term Thermal Drift (LTD), Nikolov et al., Aalborg University (CC BY 4.0)
  • VILNIUS TECH thermal pose, Vdoviak and Sledevic (CC BY 4.0)
  • TPD, thermal pose, Dhillon, Chaturvedi, Sharma and Kumar, Bennett University (CC BY 4.0)
  • FLC-Rural, Dhillon, Sharma and Kumar (CC BY 4.0)
  • Thermopile occupancy set, HTPA 32x32, Poudel et al. (CC BY 4.0)
  • UPPET TPE images, Cormier, Specker and Beyerer, Fraunhofer IOSB (CC BY-NC-SA 3.0, evaluation only)
  • WLRI-HRC, Süme, Ponomarjova, Wendt and Rupitsch, University of Offenburg (CC BY 4.0)
  • LITIV 2012 and 2014 thermal-visible sets, Bilodeau, Torabi, Massé, St-Charles and Riahi, Polytechnique Montréal (by permission)
  • MlDetection, Vandersteegen, Reusen, Van Beeck and Goedemé, KU Leuven EAVISE, with Melexis (by permission)
  • Thermal-IM, Tang, Ye, Ma and Zhao, Tsinghua University and MIT (BSD-3-Clause)
  • Multi-View Visible-Thermal images for Human Pose Estimation, Lupión, Polo-Rodríguez, Medina-Quero, Sanjuan and Ortigosa, Universidad de Almería (CC BY-NC 3.0, by permission)

From Roboflow Universe

  • CIIS Thermal, ciis
  • Thermal People, drone-work
  • Thermal Human 4, football-3arjv
  • mlx90640, mlx90640
  • MLX90640 Camera Person Detection, thermalobjectdetection
  • Thermal Camera Person Detection, Mondragon Unibertsitatea
  • Outcomex Thermal Person Detection, development-outcomex
  • Thermal Human Detection, overhead, pnusafetynet
  • Thermal Human Detection, a Roboflow Universe publisher
  • thermla-mlx90640, muhammad-abdul-latif
  • Thermal Human, toinonen, KTH
  • yolo-thermal-pedestrian-indoor, yolovthermalindoor
  • Thermal People Cafe, uon101
  • Thermal Occupancy Counter, QuietBites Project
  • Occupancy Estimation, 32x32 array, IISER Bhopal

Software

  • XGBoost, Tianqi Chen and Carlos Guestrin, and its contributors
  • ONNX Runtime, the ONNX Runtime developers
  • matplotlib, the Matplotlib development team

Public counts, for the forecast

  • Brisbane City Council, Pedestrian movement counts, data.brisbane.qld.gov.au, licensed under CC BY 4.0.
  • Contains data from the City of Melbourne Pedestrian Counting System (data.melbourne.vic.gov.au), CC BY 4.0.
  • Source: University of Alberta Library, Borealis, doi:10.5683/SP3/TCY5XY (CC0).
  • Contains information licensed under the Open Government Licence – Toronto (Toronto Island Ferry Ticket Counts, City of Toronto Fleet Services).
  • Source: Aarhus Kommune, Borgerservice og Biblioteker (ITK), opendata.dk (CC BY 4.0).
  • Source: Montpellier Méditerranée Métropole, data.montpellier3m.fr (Open Database License, ODbL).
  • Source: Orléans Métropole, data.orleans-metropole.fr, Fréquentation par heure (Licence Ouverte v2.0), data as of 2026-10-15.
  • KIT Library, Karlsruhe Institute of Technology, Seatfinder (seatfinder.bibliothek.kit.edu), used with permission.
  • Contains data from Cork City Council (data.corkcity.ie), licensed under CC BY 4.0.
  • Contains data from Dublin City Council and the National Transport Authority (data.smartdublin.ie), licensed under CC BY 4.0.
  • Contains data from Dún Laoghaire-Rathdown County Council (data.smartdublin.ie), licensed under CC BY 4.0.
  • Source: Comune di Bologna, opendata.comune.bologna.it, BolognaWifi Affollamento (CC BY 4.0).
  • Source: Saga City (佐賀市), Saga City Library (佐賀市立図書館), BODIK open data (CC BY 4.0).
  • Pedestrian count data: Heart of the City, Auckland (hotcity.co.nz), CC BY 4.0.
  • Tekler, Z.D., Ono, E., Peng, Y., Zhan, S., Lasternas, B. and Chong, A. ROBOD, room-level occupancy and building operation dataset. Building Simulation (2022). Data: figshare, https://doi.org/10.6084/m9.figshare.19234530, CC BY 4.0.
  • Source: Boryeong City (충청남도 보령시), Public Data Portal (data.go.kr), dataset 15134612.
  • Space and Time User Distribution in a University Campus, Universitat Politecnica de Catalunya. Mendeley Data, V1, https://doi.org/10.17632/55vx86j8wf.1, CC BY 4.0.
  • Source: Kanton Basel-Stadt, data.bs.ch (dataset 100013), CC BY 4.0.
  • Source: University of St. Gallen Library, Seatfinder seat estimates, used with the library's permission.
  • Source: Stadt Zürich, Open Data Zürich (Stadtentwicklung, Präsidialdepartement), measurements by hystreet.com, CC BY.
  • Source: Stadt Zürich, Verkehrsbetriebe Zürich (VBZ, Departement der Industriellen Betriebe), Open Data Zürich (CC0).
  • Source: Stadt Zürich, Tiefbauamt (Tiefbau- und Entsorgungsdepartement), Open Data Zürich (CC0).
  • Contains public sector information licensed under the Open Government Licence v3.0 (Leeds City Council, Data Mill North).
  • Contains public sector information licensed under the Open Government Licence v3.0 (City of York Council footfall).
  • Source: Cambridge Water Department, Fresh Pond Reservation Usership, data.cambridgema.gov (PDDL).
  • Source: Metropolitan Transportation Authority, MTA Subway Hourly Ridership, data.ny.gov.
  • Source: Seattle Department of Transportation, data.seattle.gov (public domain).
  • Crowdedness at the Campus Gym, Nick Rose, Kaggle, made available under the ODbL v1.0; contents under the DbCL v1.0.

Read the whole thing.

Start with the 3-page summary. The full paper covers both studies in detail, including what did not work.

Read the summary (PDF, 3 pages)Full paper (PDF, 30 pages)