Object detection and counting
Locate and count items, vehicles or people in image and video streams, with tracking across frames.
Cameras produce far more footage than anyone reviews. Computer vision turns that into counts, alerts and measurements — reliably, provided you are honest about lighting, angles and how often the model is allowed to be wrong.
The same detection task can be straightforward or impossible depending on lighting, camera placement and how much the scene varies. We assess the real environment before promising accuracy, and often the highest-value change is moving a camera rather than improving a model.
We work from your actual footage, not a public benchmark, before committing to a target.
Building a representative dataset including the awkward edge cases — normally the largest task.
Driven by bandwidth, latency and privacy. On-site processing is frequently the right answer.
Precision and recall agreed explicitly, because the cost of a false positive and a false negative are rarely equal.
Locate and count items, vehicles or people in image and video streams, with tracking across frames.
Detect defects, missing components and tolerance breaches on a line, faster and more consistently than sampling by eye.
Read printed and handwritten text from photographs, forms, labels and plates under real-world conditions.
Event detection across continuous streams — zone entry, dwell time, process timing — summarised rather than stored in full.
PPE checks, restricted-zone breaches and process adherence, with alerting rather than passive recording.
Optimised models running on-site so video never leaves the premises and results survive a connectivity outage.
Video is heavy. Moving it to the cloud for inference is often the most expensive design available. We push processing toward the camera and send results rather than frames wherever that is viable.
Camera selection, placement and lighting — where most accuracy problems are created or solved.
Quantised models on on-site hardware, so bandwidth and privacy stop being constraints.
Only detections and clips of interest are transmitted and retained, not continuous footage.
Misclassifications flagged by staff feed back into the next training round rather than being lost.
Capability in isolation changes nothing. These are the systems this normally has to reach in order to be useful.
RTSP streams from the CCTV and IP cameras already installed, in most cases.
Detections raised as ERP, MES or ticketing events rather than as another dashboard.
Notifications to the people who can act, with the supporting clip attached.
Aggregated counts and timings written to the warehouse for trend analysis.
A defect detected on the line and raised as an actionable event, without a person watching a monitor.
The same six stages apply to every AI build we run. The evaluation set is created before development starts — without it there is no way to tell an improvement from a regression.
Is a model the right instrument here? A deterministic rule is cheaper, faster and easier to audit when the problem allows one. We say so when it does.
Sources identified, access agreed, content parsed and indexed. Data quality sets the ceiling on everything after this point.
Real examples with known-good outcomes, collected before any development. This is what makes later decisions measurable.
A working slice scored against the evaluation set — accuracy, latency and cost per request all measured rather than estimated.
Wired into the real system with permissions, rate limits, spend caps, fallbacks and human approval wherever the action is irreversible.
Quality, cost and drift tracked after launch, with the evaluation set re-run whenever a prompt, model or data source changes.
These are not add-ons quoted separately. They are the conditions under which a system like this is allowed near production data.
On-device processing with faces and identifiers blurred or discarded where identity is not the purpose.
Only clips around a detected event are kept, on a defined schedule enforced by the system.
Where monitoring involves employees, we help you scope it to the operational purpose and communicate it honestly.
Defined behaviour when a camera fails or a model is uncertain — never a silent gap in coverage.
Catch defects at the point of production rather than at final inspection, when rework is far cheaper.
Automated progress capture from site photography, plus PPE and exclusion-zone monitoring.
Confirm the right goods, quantities and labels before a vehicle departs.
Detect out-of-stock facings and queue build-up in time to act on either.
Extract data from handwritten forms and printed records where digital capture is not yet in place.
Assess condition and count from drone or fixed imagery across areas too large to inspect manually.
Usually. Most IP and CCTV systems expose RTSP streams that we can consume directly. The limits are resolution, frame rate and position — a camera placed for general surveillance is often at the wrong angle for a specific detection task. We review your footage first and will tell you when moving a camera beats any model improvement.
It depends entirely on the task and the environment, which is why we will not quote a number before seeing real footage. Controlled conditions such as a production line can reach very high reliability; variable outdoor scenes with occlusion are harder. We agree precision and recall targets explicitly, because the cost of a false positive and a false negative are rarely the same.
No, and for most deployments it should not. Edge inference runs the model on-site and transmits only detections and short clips, which resolves bandwidth cost and privacy exposure at the same time. Cloud processing makes sense mainly for periodic batch analysis of already-stored imagery.
Typically a few hundred to a few thousand labelled examples per class, weighted toward difficult and edge cases rather than easy ones. We often start from a pre-trained model and fine-tune, which reduces the requirement substantially. Labelling is usually the largest single line in a vision project and we plan for it explicitly.
Send us a representative sample. We will assess feasibility honestly, including when the answer is to move a camera instead.