Skip to content
A row of security camera lenses on a dark wall, one catching a reflection of light.

Service 05 · Computer Vision

Vision that holds up on your cameras

A vision model can look good on demo footage. Yours has to hold up on live feeds, poor light and the camera someone mounted crooked.

Book a readiness reviewSee the cases
  • Detection, tracking, recognition and OCR
  • Tested on your footage before the build starts
  • Edge, cloud or hybrid
  • Connected to your cameras and systems

Example record

Each alert, with its frame

So you can check an alert against its evidence. Each detection is logged with its confidence, its frame and what happened next.

The model flags. Your operations team sets the thresholds and decides what an alert means.

stream · loading dock camera 04Sample log

  • Person entered restricted zone

    confidence above threshold, frame kept

    Alert

  • Forklift tracked across 3 cameras

    one track, handed between views

    Tracked

  • Pallet label read

    OCR, matched to the shipment record

    Matched

  • Low light after 22:00

    confidence dropping, flagged for review

    Review
A machine vision camera on an industrial arm above a conveyor of identical parts in a dark factory.

On your cameras

Your footage is the hard part

Live feeds, factory floors, exam sessions, public safety video. We help your model hold up where light, angle and motion work against it.

Book a consultation

Where vision projects stall

A model that scores well on a benchmark often fails in its first week on live footage. So these get checked on your footage first.

  • Training data that does not look like your cameras

  • Accuracy measured on the wrong scenes

  • Latency left unmeasured until the stream was live

  • A model with no path into the systems that act on it

Capabilities

What goes into your build

So one team carries your project, dataset audit through inference on your hardware.

01Object detection and trackingDetection, segmentation and tracking that follow an object across your frames and cameras.
02Video analyticsContent classified, sessions monitored and anomalies flagged in your live and recorded video.
03Text in imagesDocuments, number plates and signage read by OCR into structured data you can query.
04RecognitionFace and person-attribute recognition, where your use case and the law allow it.
05Edge inferenceModels optimised to run on the device, so inference happens next to the camera.
06System integrationStream reliability, frame-level latency and connections to your SCADA, ERP and cloud.

Where you can start

oxagile.com publishes 2 of the systems behind this practice: a proctoring platform running 5,000+ exams a day across 129 countries, and a law-enforcement video platform with 95%+ detection accuracy.

  • Consulting: architecture review, dataset audit, success criteria

  • Proof of concept: one hypothesis on your data, in 2 to 6 weeks

  • Full build: data strategy through deployment and retraining

  • Integration: trained models connected to your cameras and pipelines

Model

What we fix, and what you decide

What we fix

  1. F1Success criteria agreed on your footage before model work
  2. F2Accuracy and latency reported scene by scene
  3. F3A retraining cadence planned with the build

What you decide

  1. C1Edge, cloud or hybrid
  2. C2Which cameras and systems it connects to
  3. C3Where recognition is allowed, and where it is not

Related

Where to go next

Next step

Send us an hour of your footage.

You learn what a model can see in it and what it will miss. That gives you a go or no-go before you fund a build.

Book a readiness review →