AI
Computer Vision & Visual Intelligence
SERVICE ENGINEERING GUIDE
Turn images or video into actionable events for a clearly defined task. We build visual-inspection, counting and scene-understanding workflows around camera conditions, labelled examples and an operational response, rather than treating every detection as a confirmed incident.
What we deliver in practice
Visual inspection
Define observable defect classes and acceptable variation. Prepare representative examples, train or adapt a suitable model, and distinguish uncertain results from confirmed inspection findings.
Labelled dataset and defect-detection evaluation.
Counting & event detection
Configure regions, object classes and event rules for agreed operational use cases. Test occlusion, lighting, camera angle and density to understand where counts or alerts become unreliable.
Scene-specific rules and error analysis.
Edge or central inference
Compare local processing with server-based inference using bandwidth, latency and privacy constraints. Validate camera streams, frame sampling, hardware capacity and outage behaviour before extending coverage.
Deployment topology and capacity measurements.
Operational review & feedback
Deliver events with time, location and reviewable evidence. Route uncertain events to a person, track false alarms and use reviewed feedback to improve later versions.
Event dashboard, review queue and feedback process.
The technology, explained
Detection & segmentation
Detection locates objects; segmentation labels image regions or pixels. The required output determines annotation effort, model choice and acceptance tests.
Precision / recall
Precision measures how many reported detections are correct; recall measures how many true events are found. Their tradeoff depends on missed-event and false-alarm cost.
Edge inference
Run the model close to the sensor instead of sending all data to a central server. Hardware capacity, updates and local failure handling still need planning.
Reference architectures & design practices
Illustrative patterns for project discussions. Final topology, sizing, products and acceptance criteria are developed against your requirements. These are BrightGate concept diagrams, not vendor-certified designs.
Camera-to-reviewed-event architecture+
On smaller screens, scroll the diagram sideways to see all components.
A model proposes detections; event rules add time and scene context before an operator or approved workflow acts. Model metrics must be measured on the intended environment. Safety-critical decisions need separately approved controls and specialist review.
Design practices to apply
- Test day/night, occlusion and camera movement.
- Measure false alarms as well as missed events.
- Minimise retained imagery and control reviewer access.
Systems, products & official references
Examples to evaluate during scoping. Selection, supported features, hardware and licences depend on the agreed solution; a product mention does not imply a certification or an included purchase.
IP cameras & video streams
Camera and lens selection affect the quality of model input.
Official reference Axis network camerasDetection / segmentation models
Models locate objects or regions; suitability is established on task-specific examples.
Edge compute / GPU servers
Processing placement follows measured throughput, latency, connectivity and privacy requirements.
A result your team can accept
Acceptance evidence
Detection quality
Measure precision/recall by event class and scene.
Operational usefulness
Review false alerts and response ownership with operators.
Capacity
Test sustained streams, dropped frames and connection recovery.
View the complete capability checklist
- Manufacturing defect and quality inspection
- Object detection and inventory counting
- Asset condition and visual anomaly detection
- Site progress and image comparison
- PPE and defined safety-event detection
- Document image classification
- Edge or private-cloud inference deployment
- Alert workflows and human review dashboards