Skip to content
IEMI Labs

FlowCount AI

Live queue length and wait times, without a single image leaving

Concept · not for sale
Engineering design: FlowCount Lane
Target price from$899per sensor

Software: $39–$79 per month per sensor

Concept, not for sale yet. Tell us if you want it; that decides what we build.

Rate this idea and tell us what to change
FlowCount AI concept: the idea and where each part goes

Drag to rotate · right-drag to move · scroll to zoom · click a part for its size and placeSwipe to rotate · pinch to zoom · two fingers to move · tap a part for its size

How it fits together

FlowCount AI, part by part

Scroll or tap a part. Press “See the 3D CAD design” to watch the model, built from our CAD, light each part up and show how it connects.

FlowCount AI concept: the idea and where each part goes

Real-time 3D render · drag to rotate · final design may varySwipe sideways to rotate

  1. 01

    The whole FlowCount AI

    Ceiling cameras count people at a queue mouth and a service point; counts alone give the queue length and a posted wait time.

  2. 02Brain

    Raspberry Pi Zero 2 W

    The brain: it runs the flowcount-node program

  3. 03

    AI Camera (IMX500)

    Finds people in each frame on the sensor itself

    Raspberry Pi Zero 2 W · CSI-2 ribbon cable · Detection boxes, no images

  4. 04

    Lens window

    Gives the camera a clean, replaceable view down

    Raspberry Pi Zero 2 W · Optical path · Light onto the sensor

  5. 05

    Status light

    Steady means counting; blinking means no uplink

    Raspberry Pi Zero 2 W · GPIO 17 pin · LED on and off

  6. 06

    USB Ethernet adapter

    Sends counts over a wired link

    Raspberry Pi Zero 2 W · USB OTG (micro-USB) · Counts every 5 s

  7. 07

    PoE splitter

    Powers the node from the one ceiling cable

    Raspberry Pi Zero 2 W · 5 V micro-USB lead · 5 V power

  8. 08

    All connected

    6 key parts, one program, and your phone

What it is

A small white ceiling sensor with two lenses that counts people without sending images.

What it does

  • Measures queue length in real time
  • Predicts wait times per lane
  • Counts entries, exits and occupancy
  • Alerts staff before queues build

Who it's for

Airports, transit stations, venues, stores and campuses with queues or crowds.

The idea

Two ceiling counters at a concession line show the real wait and tell the supervisor when to open another register

The problem today

At a ballpark, zoo or museum stand, nobody knows how long the line really is until fans walk away. Airport-grade queue systems need many stitched sensors and an integrator, and a stereo counter costs about $899 a sensor.

With FlowCount AI

One counter at the line's mouth and one at the registers count people in and out, with the AI on the camera chip and only numbers sent. The wait engine reads each person's real wait from the two counts, the TV shows it, and the supervisor gets 'open register 3' before the line spills onto the concourse.

Stand 104 kitchenRegistersLine between stanchionsConcourse walkwaySupervisor1234
  1. 1Counter over the registersCounts each person leaving the line at the registers, so the system knows the real service rate
  2. 2Counter over the queue mouthCounts each person joining the line; joins minus leaves is the line length, and the two counts give each wait
  3. 3Wait TV by the lineShows 'Wait about 6 min', so fans pick a shorter stand before they give up
  4. 4Supervisor's phoneGets 'Stand 104: open register 3' a few minutes ahead, and a Monday report of waits by game and hour
Concession stand 104 on a ballpark concourse, seen from above: the counter and registers, the line between stanchions, the queue mouth and the walkway, drawn to show where each part goes (not to scale).

How the parts connect

FlowCount AI, wired up

Every part, what links it to the brain, and how the data reaches your phone. A solid line is a cable; a dashed line is a radio link.

AI Camera (IMX500)Detection boxes, no imagesCSI-2 ribbon cableLens windowLight onto the sensorOptical pathStatus lightLED on and offGPIO 17 pinUSB Ethernet adapterCounts every 5 sUSB OTG (micro-USB)PoE splitter5 V power5 V micro-USB leadTHE BRAINRaspberry Pi Zero 2 Wruns flowcount-nodePython 3 with Picamera2 and Num…Edge hub or cloudvia PoE Ethernet · MQTT/TLSPhone, laptop or TVvia HTTPS web page
Drawn from the FlowCount AI parts list. The 3D view above is built separately, from the CAD.

The program

What controls FlowCount AI

Runs on
Raspberry Pi Zero 2 W
Written in
Python 3 with Picamera2 and NumPy
System
Raspberry Pi OS Lite (64-bit), read-only root
AI model
Person or head detector running on the IMX500
  1. 1Read each frame's detector output; keep boxes scoring 0.45 or more
  2. 2Track people within this camera's view only
  3. 3Count a line crossing after 5 frames on the far side
  4. 4Every 5 s publish in, out and zone counts
  5. 5Reset track IDs every 10 minutes
flowcount-node.pymain loop
while True:
    req = picam2.capture_metadata()          # no image saved, tensor only
    boxes = imx500.get_outputs(req)          # x, y, w, h, score
    people = [b for b in boxes if b.score >= 0.45]
    tracks = tracker.update(people)          # IDs live in memory only
    for line in lines:
        d = line.crossings(tracks)           # needs 5 frames past the line
        counts[line.id].add(d)
    if now() - last_pub >= 5:
        mqtt.publish(f'{base}/line/{line.id}', json.dumps(counts.flush()))
        last_pub = now()
    if now() - last_reset >= 600:
        tracker.reset()

To your phone

How the data reaches FlowCount console

Manager's laptop or phone browser, venue TV page

  1. Node
  2. Edge hub or cloud
  3. Phone, laptop or TV

What you see in the app

One screen for each job FlowCount AI does.

  • Wait alertBetween innings at a ballpark stand the line grows fast. The supervisor's phone buzzes to open another register before it reaches the concourse.

  • Posted waitA screen by the museum ticket desk shows the measured wait, so visitors can decide before they join the line.

  • All linesBefore the evening rush, the ops manager checks every stand on one map and moves staff to the lines that are building.

App screens with sample data. The interface is in development.

What the app gets

Data FlowCount AI sends to the app
Queue length14 peopleEvery 5 s
Predicted waitAbout 7 minEvery 5 s
Service rate2.1 per minEvery 5 s
Wait alertStand 104 over 6 minOver threshold
Node health12.4 fps, 58 °CEvery 60 s

No images are stored or sent. Only counts leave the node; track IDs reset every 10 minutes.

CAD design · Design complete (stages 1–6), not built

Designed in code, checked in 3D

  • 4 printed parts, each a single closed solid
  • Every bought part fitted inside and checked for collisions
  • Assembled exactly as shown in the 3D CAD view (the button on the picture at the top)
See the CAD parts renderFlowCount Lane printable parts rendered from the CAD

Key specs

FlowCount AI hardware
AI processor2× Sony IMX500 intelligent vision sensors (on-sensor AI) + low-power ARM host for stereo fusion
CameraDual-lens stereo, 2× 12.3 MP IMX500; metadata-only output (no images transmitted)
SensorsAmbient light sensor · Tamper / tilt accelerometer
ConnectivityEthernet (PoE 802.3af) · Optional Wi-Fi 6 · MQTT, REST API and webhooks
PowerPoE (802.3af), under 6 W (target)
EnclosureWhite rounded-square polycarbonate ceiling housing with two lenses, flush or surface mount
Dimensions13 × 13 × 4 cm
Weight~380 g
Ingress ratingIP42 (indoor)
All target specifications

Sensing

Sensing target specifications
Sensors2× Sony IMX500, stereo baseline
OutputMetadata only (counts, positions)
Mount height2.5–6 m
CoverageUp to ~6 × 6 m per sensor at 4 m

Analytics

Analytics target specifications
Counting accuracy≥95% at line crossings (target)
Update rateWait time every 5 s
Multi-sensorStitch sensors across long queues
OutputsDashboard, displays, API, alerts

Power and physical

Power and physical target specifications
PowerPoE 802.3af, < 6 W
Size13 × 13 × 4 cm
RatingIP42 indoor
Operating temp.0 °C to 45 °C

Data and privacy

Data and privacy target specifications
ImagesNever stored or transmitted
BiometricsNo face templates, no identity
Data residencyEU / US hosting options
UpdatesSigned OTA updates

Targets for production units; they may change during development.

Pricing

Hardware: $899 per sensor

Software: $39–$79 per month per sensor

Target pricing. Large airport projects are quoted per terminal and may be delivered with an integrator.

Count

$39/mo per sensor

People counting and occupancy for doors and rooms.

  • Line-crossing counts
  • Occupancy per zone
  • Daily reports
  • API access

Queue

$79/mo per sensor

Queue length and predicted wait times.

  • Live wait times
  • Threshold alerts
  • Public display feed
  • Throughput analytics

Terminal

Custom

For airports, rail hubs and multi-site operators.

  • Multi-zone stitching
  • Integrator support
  • SLA and data residency
  • Custom reporting

Indicative launch pricing in USD, excluding taxes and shipping. Final pricing is confirmed before any balance is charged.

Community

What do you think of FlowCount AI?

FlowCount AI is in development. Rate the idea, tell us what to change about the features, specs or price, and we'll build it with your feedback.

How much people want it

–

No ratings yet

  1. 50
  2. 40
  3. 30
  4. 20
  5. 10

Follow FlowCount AI

Get build updates and first access when pre-orders open.

Leave feedback

Your rating of the idea

Feedback on a product concept in development, not a review of a shipped product. Posts may be moderated; links and contact details are held for review. See our community guidelines and privacy policy.

Comments

    No comments yet. Tell us what you'd want from FlowCount AI.

    Questions

    Ask about FlowCount AI

    Tell us what this would save you, or ask anything. We reply within two business days.

    We use your details only to respond to this request. See our privacy policy.

    From the product lab · ideation phase

    Lab updates for FlowCount AI

    All lab updates
    1. LatestA ballpark, zoo or museum food stand needs to know the real wait before fans walk away; airport queue systems need many stitched sensors and an integrator, and a stereo counter is about $899 a sensor (sources in DESIGN.md).
      • Drawn idea plan (idea.json): concession stand 104 on a ballpark concourse from above with the kitchen, the registers, the line snaking between stanchions and the walkway; pins for the counter over the registers, the counter over the queue mouth, the wait TV and the supervisor's phone
      • The product picture is now this drawn plan instead of the plain fallback card; drawn from the design notes, not the CAD
      • What we kept from the market work: sold to mid-size venues that buy without an integrator, not airports or small shops; the plan shows why two count lines are enough whatever shape the line takes
      • Next: drawn plans for more catalogue products that still show the fallback card

      See it in 3DFull research and design

    2. Independent grocers have four jobs in every aisle: shelves run empty between shelf walks (independents ran 6.6% out of stock in fiscal 2025, FMS/NGA study), spills wait to be noticed and a slip claim commonly settles at $15–75k (law-firm figures), paper floor-walk logs are easy to back-fill, and owners lose buyers to the till queue without knowing it (Dor counters cost $59–150 a month). Our own four products for it cost about $4,476 and $466 a month for a 6-aisle store.
      • Combined into StoreEye (https://iemilabs.com/products/storeeye/); FlowCount AI itself is unchanged

      See it in 3DFull research and design

    3. A builders’ merchant, timber yard or garden centre lets customers drive in, park, queue at the will-call counter and walk among forklifts. WorkSafe NZ fined a timber yard after a forklift knocked a pack onto a customer and asked yards to control people on foot and show customers where to park. Today each job is a separate tool: about $950 a truck for an AI reversing camera, $35 a month per camera for plate reading, paid processing for traffic counts and $859-919 per people counter.
      • Combined into SafeYard (https://iemilabs.com/products/safeyard/); FlowCount AI itself is unchanged

      See it in 3DFull research and design

    FlowCount AI FAQ

    General FAQ
    Does FlowCount record or send images?

    No. The Sony IMX500 sensors run the person detector on the sensor chip and output only metadata such as counts and positions. Images are not stored or transmitted, and no faces or identities are processed. A one-time calibration view during setup is shown locally and not saved.

    How is it installed?

    Mount the sensor on the ceiling above the queue at 2.5–6 m and run one PoE cable. Zones and count lines are drawn in the setup app. Long queues use several sensors stitched together.

    What connectivity does it need?

    A PoE Ethernet port and an internet connection for the dashboard, or a local MQTT/API connection if you prefer to keep data on-site. Wi-Fi is optional.

    Can it work with our existing airport systems?

    Wait times and counts are available through a REST API, webhooks and MQTT, so they can feed displays, apps and operational tools. Large projects may be delivered with an integrator.

    Is FlowCount AI available now?

    No. FlowCount AI is a new concept in development. Specifications are target specifications. It is a concept: we are not taking orders, reservations or deposits, and it has no ship date yet.

    What warranty will it come with?

    We are planning a 2-year limited hardware warranty covering manufacturing defects. Final terms will be published before shipping begins.

    Related products

    ParkSense AI concept: the idea and where each part goes
    VisionConcept

    ParkSense AI

    A palm-sized edge box that plugs into your existing parking cameras.

    • Counts free bays in real time
    • Flags overstays and wrong parking
    • Guides drivers to open spaces

    From $499 per edge box

    Details
    WasteSort Coach concept: the idea and where each part goes
    VisionConcept

    WasteSort Coach

    A clip-on camera coach for the recycling bins you already have.

    • Lights the right bin for each item
    • Says "not sure" instead of guessing
    • Estimates each station's contamination

    From $299 per 3-bin station

    Details
    AirfieldWatch AI concept: the idea and where each part goes
    VisionConcept

    AirfieldWatch AI

    A thermal and visible camera turret on a mast that spots birds and wildlife near runways.

    • Detects birds and flocks day and night
    • Classifies hazard-relevant species groups
    • Triggers dispersal devices automatically

    From $8,900 per turret

    Details

    FlowCount AI

    Concept · target $899

    I want this