Public safety & security analytics
Security & law enforcement agency
DeepSight — video intelligence platform
Measured results
−90%
human review labour
Source: Epsilon AI Success Stories & Project Highlights, Ver. 3 — 15 October 2025, page 12
The problem
CCTV review was manual and therefore retrospective: footage was examined after an incident, not while one was developing.
What was built
Frame-level object recognition with anomaly detection, turning footage into searchable indexed metadata.
Technical approach
- Object detection
- Real-time YOLO detection of people, vehicles and objects.
- Metadata extraction
- Footage converted into timestamped, searchable records.
- VMS integration
- Connects to third-party video management systems already in place.
Measurement framework
Deployment evidence is drawn from the published engagement. The delivery and measurement plans are case-specific recommendations grounded in the documented solution and results.
- Evidence status
- The published engagement documents an operational video-indexing and search capability, supported by a reported reduction in human review effort.
- Indicative delivery plan
- Onboard representative camera feeds, agree event categories, index a controlled footage set, validate detection and retrieval with reviewers, then expand by site and use case.
- Recommended baseline
- Recommended: reviewer minutes per hour of footage and per investigated incident before deployment, using the same camera mix, retention window and event categories.
- Recommended measurement method
- Compare matched footage and incident volumes before and after indexed search. Record human review minutes, false positives and missed events, and validate a blinded sample with authorised reviewers before operational reliance.
Recommended items are not presented as the original study method. They should be replaced with the confirmed project methodology when the underlying records are approved for publication.