Vision at the edge

Edge AI & Computer Vision

Computer vision that runs beside the camera: on-device inference, model optimisation, camera pipelines for inspection, safety and analytics, private by design.

On-device inferencePrivate by design
Illustrative camera view with sample detections and counts
Inference beside the camera, not a network round trip away
On-device
Frames processed locally; only events leave the site
Private by design
Decisions continue when the connection drops
Works offline
Accuracy reported on your own footage, not a benchmark
Measured
Why the edge

Decide beside the camera, learn in the cloud

Vision models that run on a device beside the camera, so decisions are fast, keep working when the network does not, and footage need not leave the site.

On the device

Latency
No network round trip per frame
Bandwidth
Events and counts only
Privacy
Footage stays on site
Link down
Keeps deciding

In the cloud

Latency
A round trip on every decision
Bandwidth
Continuous video upload
Privacy
Footage leaves the site
Link down
Decisions stop

The cloud still has a role: training, and fleet-wide reporting.

Use cases

Where vision earns its keep

Quality inspection

Spot defects, missing parts and misalignment on the line, with an image of every rejection kept for review.

Workplace safety

Detect missing PPE, entry into restricted zones and near misses between people and vehicles, and alert supervisors as it happens.

Operational analytics

Counts, queue lengths and dwell times for sites and stores, kept as numbers rather than stored as video.

Drawings use sample detections and values

On the device

Capture, infer, decide, report

Every stage runs on the edge device itself; only the last one talks to the network.

  1. Capture

    RTSP, USB or MIPI cameras decoded with hardware acceleration on the device.

  2. Infer

    An optimised detection, classification or segmentation model runs on the edge accelerator.

  3. Decide

    Rules turn detections into actions: reject a part, raise an alert, update a count.

  4. Report

    Events, metrics and chosen frames go to your systems; raw video stays local.

Model optimisation

Made to fit the device, measured after

FP3232 bits per weight
FP1616 bits per weight
INT88 bits per weight

Storage per weight at each precision; accuracy is measured on your footage after each step

  • QuantisationFP16 or INT8 models calibrated on your own images, so accuracy is measured after compression, not assumed.
  • Pruning and distillationSmaller models trained to match a larger one where the hardware needs it.
  • Runtime compilationExported to ONNX and compiled with TensorRT, OpenVINO or TensorFlow Lite for the target accelerator.
  • Hardware fitNVIDIA Jetson, an Intel or ARM box with an accelerator, or a smart camera, chosen against your frame rate, conditions and budget.
Illustrative frame and event with sample people, plate and values
Privacy by design

Events leave the site, footage does not

  • Frames are processed in memory on the device and discarded by default
  • Faces and number plates are blurred before any image is stored or sent
  • Only events and aggregate counts leave the site unless you choose otherwise
  • No facial recognition or identity tracking in our standard pipelines
How we work

Accuracy known before you commit

  1. Feasibility

    Footage from your cameras, a labelled test set and a baseline model, so accuracy is known before you commit.

  2. Model and device

    Training, optimisation and a hardware choice tested against your frame rate and conditions.

  3. Pilot

    One line, zone or site live, with results checked against manual inspection.

  4. Roll out and improve

    More cameras, staged over-the-air model updates, and retraining fed by reviewed edge cases.

Stack

Models, runtimes and hardware

  • PyTorch
  • ONNX Runtime
  • TensorRT
  • OpenVINO
  • OpenCV
  • GStreamer
  • NVIDIA Jetson
  • Rust
Questions

Frequently asked

Why run vision on the device rather than in the cloud?
Latency, bandwidth and privacy. A decision made beside the camera does not wait on a network round trip, it keeps working when the link drops, and video does not have to be streamed off site. The cloud still has a role: training, and fleet-wide reporting.
How accurate will it be?
We do not quote accuracy before seeing your footage. The feasibility phase measures it on a labelled test set from your own cameras, and that figure, not a public benchmark, is what you decide on.
Do we need new cameras?
Often not. Most IP cameras that provide an RTSP stream can be used. Where the task needs better images, we advise on resolution, placement and lighting.
What happens to the video?
By default frames are processed in memory on the device and then discarded. Only events, counts and any images you explicitly choose to keep are stored, with faces and number plates blurred.
How do the models improve after launch?
Low-confidence and disputed detections are queued for review, labelled, and added to the next training run. Updated models reach devices over the air, in stages, with the previous model kept for rollback.

Find out what your cameras can see

Share a few minutes of footage and the decision you want made from it. We will tell you whether it is feasible on the edge, and how we would measure it.