Guide
Can I add AI to my existing CCTV?
Usually, yes. If your recorder or IP cameras can share their video over your local network, AI software can analyse that video without replacing a single camera. How well it works depends on where each camera points, how clear the picture is and what you want it to look for.
5 min read
The short answer
If your CCTV is reasonably modern, it can very likely take AI. The cameras do not need to be "smart". The analysis runs in separate software that reads the same video your recorder already receives, and your recording carries on as before.
Connecting is rarely the hard part. The harder part is deciding what you want the system to notice and checking that your cameras can see it clearly. A camera that shows a whole car park from a high corner is good for spotting people and vehicles after hours. The same camera will not read a number plate at the far end. Being clear about the jobs you want done saves a lot of disappointment later.
What your cameras and recorder need
A way to share video on the network
Most modern recorders (NVRs, and many newer DVRs) and IP cameras offer RTSP, a standard way for another device on the network to receive a live stream. Many also support ONVIF, a shared standard that lets software find cameras and read their settings. If your system offers either, AI software can usually connect without rewiring. Our guide to RTSP and ONVIF explains both in plain English.
Enough resolution for the job
Detecting people and vehicles works well on ordinary 720p cameras when the subject takes up a reasonable part of the frame. Fine detail is different: reading text, recognising small objects or telling similar items apart needs more pixels on the subject. That usually means a closer camera or a higher-resolution one.
A sensible angle
Cameras mounted at a moderate height and looking across an area give AI the clearest view of people. Very steep, straight-down views and heavily distorted fisheye images are harder. So are cameras facing bright windows or oncoming headlights.
Steady light, or decent night vision
At night most cameras switch to infrared and show a greyscale picture. People and vehicles are still detectable in a good infrared image; colours are not.
A short setup check on your own system will show which cameras suit which jobs. Expect a mix: some cameras ready for everything, some suited to simpler tasks, and a few best left recording as they are.
What changes and what doesn't
What stays the same:
- Your cameras, cabling and recorder.
- Recording, retention and playback in the CCTV app your team already uses.
- Who owns the footage.
What changes:
- Software reads the video from the cameras you choose and turns what it sees into events: a person in a restricted area after hours, a vehicle parked too long, a bag left unattended.
- Rules decide which events become alerts, and who receives them.
- A person reviews each alert and decides what to do. A good system logs every decision with who made it and when.
- Depending on the product, you need either computing hardware on site or an internet connection to a cloud service. Our guide to on-premise and cloud video analytics covers that choice.
What AI can and can't reliably detect today
AI on CCTV is good at measurable things that happen in the frame. It is reliable at:
- detecting people and vehicles, and where they are;
- noticing when someone enters a zone or crosses a line you have drawn, at the times you choose;
- counting people, measuring queues and estimating how full an area is;
- spotting long dwell, a person on the floor, crowding or an item left behind;
- following a person from one camera to the next, when the cameras overlap or sit close together.
It is less reliable, or not reliable at all, at:
- anything the camera cannot see clearly: tiny objects, far-off detail, heavy glare or deep shadow;
- judging why something happened. A system can say a person was in the store room at 23:42. It cannot say whether they were meant to be there. That is a decision for a person;
- being right every time. Every system raises some false alerts and misses some events. The useful questions are how often, and how easily your team can tune it.
A good rule of thumb: if a guard watching that screen could say what happened from the picture alone, AI can usually flag it. If the guard would need more information than the picture shows, the AI will too.
How to test it on your own cameras first
Demos on sample footage always look good, so test on your own site before you commit:
- Pick three to five cameras that cover your most common problems: an entrance, a quiet area after hours, a busy floor.
- Write down what you want flagged on each, in plain words. For example: "anyone in the yard after 22:00".
- Run a live trial for a few days. Count useful alerts, false alerts and anything it missed.
- Check the review flow. Who gets the alert, on which channel, and how quickly can they confirm or dismiss it?
- Ask where the footage goes during the trial and afterwards, and get the answer in writing.
If a supplier cannot connect to a few of your cameras for a trial, ask why before you sign anything.
Where SurveyLensys fits
SurveyLensys works this way. It connects to the recorders and cameras you already have, runs detection on the cameras you choose and sends factual alerts for a person to review. We can run a live demo on a few of your own cameras, usually in a single visit.