Video analytics for a small business: what it is and how to judge it

Video analytics is software that looks at the pictures from a camera and decides whether something you care about is happening. Intelligent video analytics, AI video analytics, video content analysis and VCA are different names for the same idea. A UK government brochure for its i-LIDS test library describes video analytics as automated, real-time video analysis and event detection.

For a small business, the point is simple. Your cameras record hours of video that nobody will ever watch. Video analytics watches it for you and tells you about the one moment that matters.

There are three layers, and a product may have one or all of them.

• Motion detection notices that the picture changed. It cannot tell a person from a shadow, so it is easily set off by shadows, rain, insects and headlights.
• Object detection recognises what is in the picture: a person, a vehicle, smoke, a flame.
• Rules decide when it matters: a person in this zone, at this time of night, for this long.

What it can detect depends on the software and on what the camera can see. Software that is good at one of these may be poor at another, so ask for a demonstration of each one you need.

Most small businesses do not need twenty detections. They need one or two answers they cannot get today. Start with the problem, then pick the detection.

• People or vehicles in a zone, especially after hours. For a shop, workshop or yard, that is the stockroom, the back door or the gate after closing.
• Smoke and flame in a place nobody is watching. It is an early warning that sits beside your fire alarm, never a replacement for it.
• A camera that has been covered, blocked or turned away, so you hear about the gap before the night you needed that camera.
• Safety equipment, such as helmets and vests, for sites where that matters.
• Number plates, for gates and yards. Ask which plate formats a product can read before you rely on it.
• Machines that are idle, or trucks and people too close together, in a workshop or warehouse.

Already have cameras and a recorder? Try Foreman free for 7 days, no card

Many recorders and cameras already include basic analytics. The recorder makers' own manuals describe motion detection that can be limited to people or vehicles, and line-crossing and intrusion rules that you draw on the picture. Some recorders limit these rules to a set number of channels, so check the manual for your model.

Turn them on, live with them for a week and count how many alerts were real. If they are good enough, you are done. If they are noisy, or they only tell you after the fact, that is the gap software can fill. Our business security cameras guide has a longer checklist for what to check before you buy anything.

Video analytics software has to run somewhere. That choice affects bandwidth, cost and privacy, so it is worth asking about.

Running on site, sometimes called edge processing, means the software works on a computer at the site, or inside the camera or recorder. Running in the cloud means the video is sent to a data centre to be analysed there. Some products mix the two. Whatever product you look at, ask three things. What leaves my site? How often? What happens when the internet drops?

• On site: the video does not have to leave the building to be analysed, and it needs little bandwidth. You need a computer or other equipment at the site.
• In the cloud: easier to manage from one place, and able to use very large models. It needs a steady, strong upload from every camera, and the video leaves your site.
• Mixed: the first pass happens on site and only the events that matter are sent on for a second look.

False alarms are the main reason people stop trusting video analytics. An alert that cries wolf gets muted, and then a real one is missed. Common causes are shadows, rain, insects close to the lens, headlights, reflections, steam, dust and animals.

The other side is missed events. A system tuned never to raise a false alarm will miss real things. Ask any vendor how they balance the two.

• A second look. A first model watches every camera all day. A second, larger model checks only the events the first one raises, before anyone is alerted. It is the same idea as asking a colleague to look at a photo before you wake the manager.
• Zones and times. A rule that says a person in this area after hours is much narrower than any movement at all.
• Careful camera views. A clean lens, steady mounting and good lighting help every system.
• A trial on your own cameras. It is the only honest test, because your cameras, lighting and weather are not a demo room.

Prices differ a lot by product and by country, so we will not quote a figure here. What helps is knowing how a price is built, so quotes can be compared fairly.

Ask every vendor for the total cost for your number of cameras over three years, including any hardware, and what happens to your video and settings if you leave.

• Per camera per month. Common for cloud and software products. The bill grows with every camera you add.
• Per site or per recorder. Simple, but check how many cameras are included.
• An upfront licence, or new hardware. Some systems need their own cameras or a special box, which turns a software price into an equipment bill.
• Extras. Storage days, extra detections, more users and support are sometimes charged separately.

Video analytics does not change the basic privacy steps. If people can be seen in the video, the usual good practice applies: tell people cameras are in use, keep only what you need, keep it secure and limit who can look.

Analytics adds a few points of its own. Ask how long the software keeps clips, who at the vendor can see them, and whether your video is used to train their models. Be extra careful with anything that identifies people by their face. If you do not need to identify people, choose software that does not.

Rules on cameras, recording and face matching depend on where you are and on your industry. We cannot advise on them, so ask a lawyer about yours. Here is what we do.

• Recordings stay on site, and your recorder keeps recording as usual.
• Event clips are encrypted and deleted after 30 days.
• We never train on your footage.
• We do not use facial recognition or identify people.

Demos are easy and your site is hard. These questions get past the demo.

In the UK, the government built a benchmark library of test footage, called i-LIDS, so that video analytics systems could be judged on the same scenes. The point carries over. Ask for results on footage like yours, from an independent test or from a trial on your own cameras.

• What does it detect, and can you show me on cameras like mine?
• Where does the analysis happen, and what leaves my site?
• How are false alarms handled? Is there a second check?
• Who is told, how fast, and by what?
• Does it work with the cameras and recorder I have, or do I need new ones?
• Can I set zones and times in plain words?
• What does it cost in total for my number of cameras?
• Where is video kept, for how long, and who can see it?
• Is my video used to train your models?
• Does it identify people by their face?
• Can I try it free on my own cameras?

Video analytics sees only what the camera sees. It is no use against a dark, blocked or badly aimed view. It does not replace people who can act on what they are told, and where it watches for smoke and fire, it does not replace certified fire alarms or life-safety systems.

Many sites already have cameras and a recorder (a DVR or NVR). They record. Nobody is looking at 2 a.m.

Ours is the mixed kind. Our software runs on a Windows PC at your site and watches every camera there with our own AI models, trained in-house. Your recorder keeps recording as usual, and recordings stay on site. When our models see an event, a second, larger model double-checks it, and the photo goes to you. Detection runs on site. The second check and the alert need the internet.

It works with the cameras and recorders that are already there: Hikvision and Prama, Dahua, CP Plus, TVT, Uniview, Axis, Hanwha, and any camera or recorder that supports RTSP or ONVIF. It installs in 60 seconds.

It watches every camera for three things on every site: smoke and fire, people after hours in zones you mark, and a camera that has been covered. It also warns you when a camera goes offline. Then the photo arrives on WhatsApp, Telegram or the Foreman app while it is still happening.

Some things are set up for your site with us, not switched on by default: helmet and vest checks, forklift near-miss and idle machines. Tell us what you need to see and we will set it up with you.

A few honest limits. We do not sell cameras. We do not provide guards or a monitoring centre, so the alert comes to you and the people you choose.