Home / Case studies / AI Object Counting for Ice Block Loading at Plant Gates
Case study · Manufacturing · Video analytics & edge AI
AI Object Counting for Ice Block Loading at Plant Gates
An Indian ice manufacturer needed AI object counting because its loading gates kept miscounting ice blocks whenever four trucks loaded at the same time. Brainy Neurals built a camera system that uses object detection & tracking to count every block crossing each gate. One machine inside the plant reads all four conveyor cameras & replaces a tally that depended on whoever stood closest. Counts reach the owner’s phone every hour, & no video leaves the site.
4
Gates counted at once
25
Frames read per second
On site
Where every count is made
Published October 2026
At a glance
Ice block counting at four gates moved from a hand tally to cameras & one machine on site.
What problem did this solve?
Ice blocks leaving on trucks at four gates were miscounted whenever the owner stepped away. One person can’t watch four conveyors at once, so the tallies drifted with every distraction.
What did Brainy Neurals build?
Brainy Neurals built an AI counting system for an Indian ice manufacturer. A custom-trained detector watches four conveyor cameras from one machine on site & counts every block.
What changed after it went live?
Counting now runs unattended at all four loading gates. The owner reads hourly & daily reports for each gate on a mobile app, with no video sent to the cloud.
Who else could use this?
Camera counting fits any site that moves identical items past a fixed point all day. Warehouse docks, bottling lines, brick yards & parcel belts all qualify.
| Fact | Detail |
|---|---|
| Industry | Manufacturing |
| Sub-industry | Ice production |
| Client | Indian ice manufacturer |
| Engagement | AI counting system |
| Timeline | Not disclosed |
| Capabilities | Computer vision & edge AI |
| Delivery | Project-based delivery |
| Video sent to the cloud | None |
Why did the count keep going wrong?
The count kept going wrong because an Indian ice manufacturer loads trucks at four gates at once, with no AI object counting in place. Each gate is fed by its own conveyor, & every block that leaves is revenue. The tally was whatever the nearest person remembered.
When the owner stood on the dock, the numbers held together well enough. When the owner traveled, they drifted, & nobody could say by how much.
The plant already had cameras at every gate, installed for security rather than video analytics. Those cameras watched everything & counted nothing.
Where it broke
- Four belts fed into one shared tally.
- Counts drifted every time the owner traveled.
- Disputes were settled on memory instead of records.
- A recount meant stopping a fully loaded truck.
- Every cloud video quote came with a monthly bill.
The drift was nobody’s fault in particular & came as no surprise. A study of nearly 370,000 records at one retailer found 65% of its inventory counts were wrong[1]. Manual counts drift, & they drift faster when nobody watches the person counting.
What do plants usually try first?
Plants usually try one of four counting routes first, & each is a reasonable place to start.
| Approach | What it gets right | Where it stops | Who it still suits |
|---|---|---|---|
| A counter at each gate | No new technology | Four gates, shifts & disputes | One gate, short loading hours |
| Beam or weight sensors | Cheap, counts every crossing | Counts hands & debris too | Single-file items, clean flow |
| Cloud video counting | No on-site hardware | Bandwidth & a monthly bill | Few cameras, strong uplink |
| On-site AI, our route | Counts only blocks, at all gates | Training footage & hardware up front | Parallel gates, high volume |
Three familiar routes each stop at a wall the fourth route clears.
The trade-off between these routes is an old one. A 2016 survey of edge computing named response time & bandwidth cost as reasons to process video on site[2].
How we designed the counting system
We designed the counting system to run entirely inside the plant, on one machine the client bought once. That machine carries a graphics processing unit, or GPU, sized for real-time object detection on four streams.
Each gate’s camera feeds the machine, & a custom-trained model finds every block in the frame. A tracker follows each block, & a fixed line turns each exit into one tick.
Video never leaves the plant, & the numbers are the only thing that does.
Every stage from detection to counting runs inside the plant, & only counts cross the wire.
Video never leaves the plant, & the numbers are the only thing that does.
Our first decision was what should cross the wire beyond the plant. We rejected streaming four cameras to cloud machines, because that bill grows with every camera added.
A count that depends on an uplink also stops when the uplink does. Only counts leave the building, a few bytes an hour in place of four video streams.
The second decision put all of the intelligence at the gate, with the cloud only storing & serving totals. A phone can ask for a report, but it can’t slow a gate down.
The technology stack we used
The technology stack we used earns its place on one machine & one bill. The same layout runs in our edge AI builds on smaller boards, with the hardware swapped.
On-site capture & inference
Cameras
We used the gate cameras already in place, so no new mounts were needed. A full camera fit-out was ruled out.
Compute
One GPU workstation, bought once, handles four streams in real time. Paying for cloud compute on every frame was ruled out.
Detection model
A custom-trained detector looks for one class of object, the ice block. General off-the-shelf detection models were ruled out.
Counting & reporting logic
Tracking & counting
Each block keeps one identity & counts once when it crosses, so nothing is counted twice. Counting every raw detection was ruled out.
Stream scheduling
All four streams share one model on a fixed rotation, so every gate keeps its full frame rate. One machine per gate was ruled out.
Local buffering
Counts are stored on site first, so a dropped link loses nothing. Fire-and-forget uploads to the cloud were ruled out.
Cloud & mobile delivery
Cloud backend
A small hosted count service stores totals & serves hourly rollups. Streaming video off site was ruled out.
Mobile app
One codebase serves both phone platforms, so the owner can read counts anywhere. Building two native apps was ruled out.
How does one block get counted?
One block gets counted in six steps, from gate to tally, with nothing waiting on a network.
- The camera at each gate feeds live frames to the plant machine, all four streams at 25 frames per second.
- The detector finds every ice block in the frame & draws a tight box around each one.
- The tracker links those boxes from frame to frame, so every block keeps one identity as it moves.
- The gate’s tally rises by exactly one when a tracked block crosses its fixed counting line.
- The count is stamped with its gate & hour, then stored on the plant machine first.
- The totals sync to a small cloud service, & the mobile app updates every gate’s numbers.
Each block is counted once, at the line, & only step six leaves the site.
Only the last of the six steps leaves the plant. Everything before it happens on the machine beside the gates, whether the internet is up or not.
What nearly broke the count?
Three problems nearly broke the count, & the first was that ice blocks travel nose to tail. Two touching blocks read as one, then two, then one again, & every flicker minted a phantom block. Those phantom blocks went straight onto the invoices.
The dock door then began making weather of its own. Cold air met warm air & fogged the lenses, while midday glare off wet ice washed whole frames out. The detector began missing blocks it had found without trouble an hour earlier.
Four streams also fought over one machine. Scheduled naively, one gate’s frames queued behind another’s, & a queued frame is a skipped frame.
Tracking benchmarks list identity switches & ghost tracks among the standard errors, which are polite names for phantom blocks[3]. These are the weeks when a plant brings in specialist engineers instead of losing a season.
How we fixed each problem
We fixed each problem with a change that reads short on the page, though none of them was short to find.
One identity per block & one tick per crossing replaced counting raw detections.
Phantom blocks
We stopped counting detections & started counting crossings instead. The tracker gives every block one identity, & the tally moves only once, when that identity crosses the line. The phantom blocks then vanished from the invoices for good.
The dock door
We collected fog & glare footage from the worst hours, added night loads to it & retrained on those frames. We also moved each counting line clear of the mist.
Four streams, one machine
We batch the four streams through the model on a fixed rotation. Every gate gets its slice of every cycle, so all four keep their full frame rate & no gate starves.
Conveyor belts, as it happens, are older than the light bulb itself.
Finding these three problems took most of the build, so an AI proof of concept now runs this gauntlet first.
What changed once it went live?
Once it went live, the count stopped depending on whoever stood closest to the belt.
| What | Before | After |
|---|---|---|
| Who keeps the count | Whoever stood closest | The system, at every gate |
| When the owner travels | The tallies drifted | Nothing changes |
| Video sent to the cloud | Quoted as the fix | None |
| Cost that grows per camera | Cloud streaming & bandwidth | Neither |
| The loading report | End-of-day memory | Hourly & daily, per gate |
We don’t publish model accuracy figures, on this page or any other. What we will publish is what runs on the counts, which is every hourly & daily report at every gate.
Day to day, the owner opens the app from anywhere & reads what left each gate that hour. Because the machine was bought once, the cost of counting stopped scaling with the camera count.
A dispute now ends with a report instead of a recount at the gate. The recount, with a loaded truck idling beside it, is the ritual this build retired.
Counting identical items at your own gates?
If products leave your site faster than anyone can count them, tell us about the project. We’ll look at your cameras & say whether camera counting fits.
What runs at the plant today
Today the AI object counting system runs in production at the plant, on the same machine the first pilot used. All four gates report every hour & every day, & the owner reads those reports from anywhere.
Nothing about the loading day changed for the crews on the belts. The belts keep running & the trucks keep filling while the count happens above everyone’s head.
For AI in manufacturing work, that is usually the right amount of AI to ship. Plant staff watch the system far more often than they operate it.
What would we change next time?
Next time we would change four things, each one learned the slow way on this build.
A tally that moves on every detection is a random number generator with confidence.
Film the worst hour first
We trained on clean daylight footage & met the fog later, at the door. A site’s ugliest frames are the ones to start with.
Count each crossing once
A tally that moves on every detection is a random number generator with confidence. We knew the rule going in & still spent a week relearning it.
Load all four gates early
The one-gate build ran beautifully for a week & taught us nothing. Four parallel streams make a different machine, so test the full load from day one.
Decide what crosses the wire first
Sending counts instead of video cost almost nothing to decide on this project. It still shaped the hardware & the backend. It also set the monthly bill, so make it before anything else.
Where else does camera counting fit?
Camera counting fits wherever identical products pass a fixed point faster than anyone can count. It keeps a running tally of items passing a camera by detecting & tracking every one of them.
The same counting pattern on a parcel belt, with counts leaving & video staying.
| Industry | The equivalent problem | What changes in the build |
|---|---|---|
| Logistics | Warehouse dock cameras counting pallets & parcels out | Wider classes, one line per door |
| Retail | Back-room cases counted onto the shop floor | Smaller items, counts in both directions |
| Agriculture | Produce crates from the packhouse onto trucks | Outdoor light, seasonal retraining |
| Construction materials | Cement bags & bricks loaded onto flatbeds | Dust-tolerant training, rugged mounts |
| Beverage bottling | Cases leaving the line faster than sight | Higher frame rates, tighter timing |
| Food processing | Frozen cartons from freezer to dispatch | Frost-proof retraining, cold-rated housings |
Porting the system takes fresh footage of the new item & a retrained detector. Hand tallies then run beside it until the two counts agree.
What buyers ask about AI object counting
Can AI count objects from existing CCTV cameras?
Usually it can, because counting needs a clear view & steady framing more than it needs new cameras. We test your current footage first & only ask for new mounts where a view fails.
How does AI object counting work?
A detection model finds each item in every frame the camera produces. A tracker then links those detections across frames into one identity per item. The count rises exactly once when that identity crosses a set line.
How accurate is camera-based counting?
Accuracy depends on the item & the camera view. The training footage matters just as much, so no honest vendor quotes one number. A serious counting system is checked against hand tallies before go-live, on your gates & in your light.
How long does an AI counting system take to build?
A proof of concept on your own cameras usually takes a few weeks. Production follows once the counts hold against manual checks across shifts & weather, through a full loading day.
How much does an AI object counting system cost?
The cost depends on the camera count & the item, as well as the hardware already on site. Brainy Neurals scopes it from sample footage & a short conversation, then quotes a fixed price. An AI readiness assessment tells you first whether your cameras & site can carry it.
Is edge AI cheaper than cloud AI for video?
For always-on video, edge AI usually costs less. Edge hardware is bought once per site, while cloud inference sends a bill for every hour of video streamed. A few cameras running short hours can still favor the cloud, but four gates running all day don’t.
What else can the same cameras count?
Anything that moves past the lens as a distinct object can be counted this way. The same pipeline counts pallets, crates, bags, cases & vehicles from the same camera mounts. Each new item needs its own footage & a check against hand tallies.
Services behind this case study
Six Brainy Neurals services sit behind this case study, from video analytics to AI in manufacturing.
Video analytics
Monitoring & counting from the cameras a site already owns, with alerts on events.
Computer vision development
Detectors trained on your product & your belts, under your worst lighting.
Edge AI & embedded services
The same counting pattern on smaller boards, when a workstation is too much.
Robotics & hardware automation
For sites where a count must stop a belt or open a gate itself.
Hire AI developers
Vision & edge engineers who join your team for the hard weeks.
AI in manufacturing
Counting & inspection for plants that can’t wait on a network.
An AI proof of concept is a quick, low-cost way to test counting on your own footage. AI consulting helps choose the hardware before anything is bought.
An AI readiness assessment checks whether your cameras are enough, & the industries hub shows where this pattern already runs.
Similar case studies
Three similar case studies show Brainy Neurals builds shipped into working sites instead of left as demos.
Overhead Line Geometry Measurement
Stereo cameras on a moving train measure wire geometry, with inference on the train.
AI Diet Assistant for Gastroenterology
Clinical dietary guidance generated under review gates, live in a healthcare setting.
Personalised AI Meal Planning for Chronic Care
Structured meal plans built from messy personal health data, grounded & reviewable.
Cite this case study
Mori, Prasiddh & Shah, Rushabh. AI Object Counting for Ice Block Loading at Plant Gates. Brainy Neurals, August 2026. https://brainyneurals.com/case-studies/ai-object-counting-ice-blocks/
Sources cited on this page
- DeHoratius N, Raman A. Inventory Record Inaccuracy: An Empirical Analysis. Management Science. 2008;54(4):627-641. DOI 10.1287/mnsc.1070.0789.
- Shi W, Cao J, Zhang Q, Li Y, Xu L. Edge Computing: Vision and Challenges. IEEE Internet of Things Journal. 2016;3(5):637-646. DOI 10.1109/JIOT.2016.2579198.
- Milan A, Leal-Taixe L, Reid I, Roth S, Schindler K. MOT16: A Benchmark for Multi-Object Tracking. arXiv. 2016. arXiv:1603.00831.