Key takeaways
- Most modern IP cameras that support ONVIF or RTSP can run AI detection. You rarely need to rip and replace.
- Placement and lighting matter more than megapixels.
- Choose the three detections that matter most and prove them on one zone before scaling.
- Where there is no power or network, add solar 4G units that report into the same platform.
Do you need new cameras?
Usually not. AI video analytics reads the same video stream your recorder already receives. If your cameras are IP cameras that support ONVIF or RTSP, which most installed in the last decade do, they can normally feed an AI platform without being replaced. Older analogue cameras can sometimes be connected through an encoder, but that is where results start to vary.
01
Existing camera
ONVIF / RTSP stream
02
AI detection
Edge device or unit
03
Scopis platform
Zones, rules, schedules
04
Action
Warning, alert, clip, log
What makes a camera good enough for AI
| Factor | What good looks like | Why it matters |
|---|---|---|
| Field of view | The subject fills a reasonable part of the frame | A person a few pixels tall can't be checked for a helmet |
| Angle | Looking across the area, not straight down or into the sun | Overhead views hide PPE; glare hides everything |
| Lighting | Lit areas, IR or full-colour night cameras | Most security events happen after dark |
| Mounting | Rigid, out of reach, clear of obstructions | Shake and blocked views cause missed detections |
| Stream | Steady stream, sensible frame rate | Dropped frames mean dropped events |
Edge, cloud or both?
Edge processing runs the AI next to the camera or on the unit itself. It reacts fastest, keeps working when the connection drops and sends only events and clips over the network. Cloud processing is easier to scale across many sites but depends on bandwidth. Most real deployments mix the two: detection at the edge, with dashboards, reporting and multi-site management in the cloud.
Network and IT: what to ask
- Bandwidth: sending events and short clips needs far less than streaming everything to the cloud.
- Segregation: will the analytics sit on the CCTV network, or need a route to it? Agree this with IT early.
- Remote sites: units on their own encrypted cellular link avoid touching the customer network at all.
- Access: who can see live views and clips, and how is that logged?
Where existing CCTV falls short
Existing cameras cover where cameras were installed, which is usually entrances, car parks and building perimeters. Temporary works, pipe laying, remote compounds and yard edges often have no power, no network and no camera at all. That is where self-contained solar 4G units fill the gap, reporting into the same dashboard as the fixed cameras.
Start with three detections, not thirty
- 1
Pick the problems
Choose the three events that cost you most: for example PPE at the gate, people in plant zones, and after-hours intrusion.
- 2
Map them to cameras
For each, name the cameras that can see it and check them by day and night.
- 3
Pilot one zone
Run the detections on one zone of one site and agree up front what success looks like.
- 4
Tune, then scale
Adjust zones, schedules and confidence thresholds, then roll out site by site.
The question isn't whether the AI can see it. It's whether this camera, in this spot, at 2am, can.
Data protection
Adding analytics changes what your CCTV does, so revisit your data protection paperwork. Update the DPIA, especially for number-plate recognition and anything that monitors staff, confirm retention periods, and update signage and privacy notices.
Before you start
0/6 doneCommon questions
Often, yes. Placement, angle and lighting usually matter more than resolution. Very wide views or distant subjects may need a closer or higher-resolution camera.
Sources
Niall Dorr
Founder of Scopis AI. More than 25 years in computer vision, three computer vision patents, and a career building machine vision and remote monitoring systems.
LinkedIn





