Visual anomaly detection for operations and quality

The problem

How it works

1. 1Normal footage or images are collected for a process or product.
2. 2A model learns the normal visual pattern.
3. 3Deviations are scored and flagged above a threshold.
4. 4Flagged cases are reviewed and used to refine the model or create dedicated detectors.

What can be detected

• Unusual product appearance
• Activity in a zone at an unusual time or pattern
• Process steps skipped or out of order
• Objects present where they should not be

Typical scenarios

• Quality inspection for rare defects
• Process compliance in workshops
• After-hours activity in sensitive areas
• Cash-handling and back-office procedures

Working with existing cameras

• Stable camera positions and consistent lighting make "normal" easier to learn.
• Most IP cameras and NVRs that provide RTSP or ONVIF streams can be connected, regardless of brand.
• Before a pilot, sample footage from each candidate camera is reviewed for resolution, angle, distance and lighting.
• Where a view is not suitable, the recommendation is usually to reposition the camera or add one — not to replace the whole system.

Integration

• Review queue in the dashboard
• REST API / webhook for follow-up workflows

Privacy and security

• No face recognition is used in safety and operations analytics; events describe what happened, not who.
• On-premise processing keeps video inside your network.
• Evidence clips, retention periods and user access are configurable and should follow your GDPR/KVKK policies.
• Workplace monitoring should be introduced with clear employee information and, where applicable, consultation with employee representatives.

Limitations

• Anomaly scores need human review in early stages; thresholds are tuned to balance alerts and misses.
• Not every anomaly is a problem — context rules are usually needed.