Computer vision / How it works

How computer
vision works.

Cameras supply the view. An on-site device turns it into counts and events. Use the results in dashboards, reports and alerts.

On-site

Camera

Provides video frames

Edge device

Detects, counts & stores

Cloud sync

Events & statistics

Your results

Dashboards, reports,
API & alerts

Video processing stays on-site.The standard cloud connection carries event data, not the camera stream.

An illustrative example

How a count is created.

A configured line turns movement through a camera view into a recorded crossing.

A car approaches a counting line. Illustrative overhead view of a single-lane entry. COUNT LINE
01 / Detect

Recognise the object

The local model identifies a car in the camera view.

The same car has crossed the counting line. Illustrative overhead view of a single-lane entry. COUNT LINE
02 / Track

Measure the crossing

The same object is followed across frames. Crossing the configured line creates an event.

03 / Record

Store an event

Object class, direction, timestamp and location form a useful record, without reading the vehicle's registration number.

Illustrative sequence, not live data. A crossing count is not a count of unique people or vehicles.

Complete processing sequence7 steps, from camera capture to results
  1. Capture the frame

    A suitable camera sends frames to the edge device. Supported input options include USB, RTSP and ONVIF, with PTZ control where compatible.

  2. Detect and classify

    A locally running model detects supported objects. Traffic classes include car, utility, motorcycle, bicycle, bus and person.

  3. Connect detections across frames

    The tracker assigns an object ID across successive frames so movement can be measured without repeatedly counting the same detection. This is a temporary object track, not a vehicle registration number or personal identity.

  4. Apply lines and zones

    Crossing a configured line creates an event with direction. Calibrated speed zones and dwell-time rules support additional measurements where configured.

  5. Store the event locally

    Timestamped event metadata is stored in an encrypted local database. Local buffering supports operation when the cloud connection is unavailable.

  6. Sync anonymous statistics

    Batches of events and statistics are sent over HTTPS when connectivity allows. The camera stream does not follow that route.

  7. Evaluate rules and deliver the output

    Configured inactivity, volume or loiter rules run on their evaluation cadence. Access results in dashboards, the API, CSV exports and alert notifications, with custom rule types scoped as needed.

Privacy & data

Video and event data.

Traffic analytics records vehicle types and movements, not vehicle registration numbers.

On-site processing

Video frames

Analysed locally and discarded in the standard pipeline.

Standard sync

Counts & event data

Object classes, directions, times and configured activity measurements.

No number-plate reading (ANPR)

Vehicle registration numbers are not extracted or stored as analytics records.

See the data included and excluded

Collected for analytics

  • Vehicle counts by class, including the unified utility category.
  • Direction: entry, exit and passing through a monitored line.
  • Timestamps and location context for trend analysis.
  • Pedestrian counts when the relevant model is enabled.
  • Calibrated speed measurements and accuracy ratings where configured.
  • Zone activity, dwell-time events and configured alert notifications.

Not part of standard analytics

  • Continuous video uploaded to the cloud or retained as an analytics recording.
  • Number plate reading or plate records.
  • Facial recognition, names or personal identities.
  • Individual visitor histories or identity tracking between sites.
  • Audio capture for the analytics pipeline.

Frames are processed and discarded in the standard pipeline. Traffic deployments can separately enable optional snapshots for speed-threshold events; agree their evidence handling and retention in the deployment scope. A snapshot option is not continuous recording.

Read the full privacy & security principles

Access & reporting

Using your results.

Access is separated by customer. Choose the format that suits your work.

Web dashboard

Current device status, historical trend charts, filterable event tables and report generation. Role-based access controls govern who can see the data.

REST API

Programmatic access to paginated events and hourly or daily aggregates. Tenant API keys and IP allowlists support controlled integrations and automation.

Near real-time alerts

Built-in or custom rules cover inactivity, thresholds and loiter time. Review alert history and resolution tracking; timing depends on the configured evaluation cadence.

Bulk CSV export

Download data filtered by date and location for Excel, Python or other analysis tools. Background processing supports exports of up to 365 days of available data.

For the technical brief

Technical details.

These reference specifications are not a performance guarantee. Hardware, rates, available history and deployment requirements must be confirmed for your site.

Edge hardware & camera inputs
  • NVIDIA Jetson Orin series with Ampere GPU architecture.
  • NVDEC hardware decoding for H.264 and H.265.
  • Reference processing rate: 12-15 frames per second at 1080p; camera capture can be up to 30 frames per second.
  • USB, RTSP and ONVIF inputs, with compatible PTZ control.
  • Encrypted SQLite event storage, local buffering and offline-resilient processing.
Detection model & tracking
  • Custom-trained UK traffic model with TensorRT FP16 acceleration.
  • Six supported classes: car, utility, motorcycle, bicycle, bus and person.
  • Training includes elevated and aerial views, night conditions and more than 50,000 road images in the reference model.
  • Multi-object tracking reconnects detections across frames; these object IDs do not identify people.
  • Reference inference time is under 100 milliseconds per frame, with local capture-to-event storage under 200 milliseconds. Confirm performance against the deployment configuration.
  • Bespoke detection model integration can be scoped for other use cases.
Data protection & access controls
  • Full-disk encryption and SQLCipher AES-256 local database encryption.
  • Password-protected devices without default credentials.
  • HTTPS/TLS 1.3 transport and per-device API keys in the reference configuration.
  • Tenant isolation, role-based portal access and configurable retention.
  • The client portal supports JWT-based access and MFA.
Cloud, sync & operational health
  • FastAPI (Python) backend, PostgreSQL database and a Next.js web portal.
  • Railway hosting, with an EU-region option; confirm the deployment region.
  • REST data API with tenant API keys and IP allowlist controls.
  • Reference statistics sync interval: 15 seconds; reference device heartbeat: 30 seconds.
  • Alert evaluation follows its own configured cadence after data arrives, so a sync interval is not an alert-delivery promise.
  • Health monitoring, automatic recovery and transparent device status support operations.
  • Bulk CSV exports support up to 365 days, subject to the data available and configured retention.

A practical starting point

Planning a deployment.

Camera placement, lighting, calibration and the question being asked all influence a useful deployment.

Confirm the conditions

Review the camera view, occlusion, indoor or outdoor setting, power and connectivity. Define which areas are and are not monitored.

Agree the success criteria

Specify the required counts, zones, rules, reports and handling of any optional evidence snapshots.

Review a site pilot

Check the resulting data against the operational question before expanding. Confirm that local processing, cloud sync and reporting work together for your site.

Start with a conversation

Discuss your camera setup.

Bring the camera setup, the data you need and the systems that need to receive it. We can work through the fit.

Discuss your project
Direct to Percepta

Tell us about your site

Describe your site, your cameras and the question you need answered. We can discuss the right approach for your project.

EDINBURGH, SCOTLAND
UK & INTERNATIONAL ENQUIRIES

Send a secure enquiry

All fields marked * are required.

Please do not include passwords, payment-card details or sensitive personal information.

We use the details you provide to review and respond to your enquiry and protect the form from abuse. See our privacy notice.

Your choice

Optional video settings

These demonstration videos are hosted by Google/YouTube, separately from Percepta's on-device camera analytics. Loading a player sends your IP address to Google/YouTube and may let it use cookies or similar storage for measurement and advertising.

No YouTube player connects until you allow YouTube and explicitly load a video. Allowing here does not load any videos. Each player needs its own Load action, then Play; nothing starts automatically.

Your choice stays in memory while you browse these public pages and resets on a full page reload. We do not save a preference cookie or use browser storage.

Turning videos off immediately removes all loaded players. It cannot erase information YouTube has already received or cookies it has already set.

YouTube videos: off.