CCTV video analytics and AI: what works and what to promise customers

Almost every camera data sheet now promises "AI". Customers have seen impressive demonstrations and they expect the system to spot any intruder, flag every suspicious person and never raise a false alarm. Real-world results tend to be more modest, and they vary according to the setup.

Analytics can add real value, and it's also an opportunity to earn higher margins. But overpromising is a good way to create an unhappy customer. This guide looks at what's commonly useful, what to be careful of, and how to set expectations.

What analytics really means

Analytics software analyses video automatically and creates events or searchable data. It may run on the camera, on a recorder or server, or in the cloud. The place it runs determines cost, scale and what happens if the network fails.

Older "motion detection" simply looks for pixel changes, so shadows, rain, insects and headlights trigger it. Modern systems try to recognise what the object is, such as a person or a vehicle. That's where most of the real gain has come from.

Useful, proven features

Person and vehicle detection

This is the most useful upgrade for most customers. It greatly reduces nuisance alerts compared with basic motion detection, and allows faster searches through recordings. It's not perfect. Very small objects at a distance, poor light and unusual angles still cause errors.

Line crossing and intrusion zones

An alert is raised when a person or vehicle crosses a line or enters an area, usually outside of business hours. This works well for perimeters, yards and restricted spaces, especially in combination with person detection.

Smart search

Finding a person in a day's recordings manually takes ages. Filtering by motion in a region, by person or vehicle type, or by attributes such as clothing colour can cut this down dramatically. Customers often value this more than alerts, and it's easy to demonstrate.

People counting

Useful in retail and in some buildings, for footfall analysis and occupancy. Accuracy depends a lot on the camera position, typically overhead. Test it against a manual count.

Number plate recognition

This can work very well in the right setup, with a dedicated camera, a controlled angle, fast shutter speed and proper lighting. It rarely works from a general overview camera. Plate formats differ by region, so check that the product supports your local ones.

Tamper detection

An alert when a camera is covered, moved or blurred. It's simple, and can be useful for important positions.

System health monitoring

Alerts for camera failure, disk errors or lost connections aren't "AI", but they could be the most valuable alert of all, because they show a problem before footage goes missing.

Features that need more caution

Face recognition

This raises serious legal, ethical and accuracy concerns. Rules differ widely and are changing, with some places imposing strict controls on biometric data. Accuracy can also vary across demographic groups and conditions, and a false match could have severe consequences for an innocent person.

Before you offer it, recommend that the customer take legal advice on their use case, and be sure they have a lawful basis and a clear policy. Many companies choose not to offer it at all, except for specific use cases, like staff access with consent. Don't use it just because the camera offers it. Note that many cameras also offer "face detection", which only finds that there's a face, and that's quite different from identifying who it is.

Behaviour and threat detection

Claims about loitering, fights, falls or "suspicious behaviour" should be tested on real scenes. Some can work in constrained settings, but false alarms can quickly erode trust. Set expectations carefully, and don't sell these as a safety guarantee.

Weapons and fire or smoke detection

Video-based detection can complement other systems, but it should never replace a certified fire alarm or the customer's own security procedures. Be very careful about what you claim. If a customer relies on it for life safety, the requirements are entirely different, so involve specialists.

Thermal analytics

Thermal cameras can detect people in total darkness and across long perimeters. They can be very effective for perimeter protection, but they're more expensive and have different image characteristics. Elevated-temperature screening has its own accuracy and regulatory questions, so be careful about such claims.

Why analytics fail in the real world

Almost all analytics disappointment can be traced back to the installation.

• Target too small. Each analytic needs a minimum object size. A person at the far end of a yard might be just a few pixels high.
• Poor angle. Algorithms are trained on certain views. Check the manufacturer's recommended mounting height and angle.
• Bad light. Low light, backlight and headlights all affect accuracy.
• Weather and plants. Rain, snow, swaying trees and spider webs all generate events.
• Wrong settings. Default sensitivity is rarely ideal.
• Too many alerts. Staff quickly learn to ignore them.
• Low bitrate. Heavy compression can harm detection on some systems.

Good design is the foundation. Our camera placement guide applies here too, along with each manufacturer's specific guidance.

How to set it up well

1. Start with the outcome. "Alert the guard when a person enters the yard after 8 pm."
2. Choose the camera and position to suit it. Check the target size at the far end of the zone.
3. Draw zones carefully. Exclude roads, trees and neighbours' property.
4. Set schedules. Most alerts only matter during certain hours.
5. Use object filters. Limit alerts to people or vehicles.
6. Test in real conditions. Walk through the zones in day and at night, in the types of clothing people really wear.
7. Tune for a week. Review the false alerts, then adjust settings. Plan and quote for this time.
8. Define who responds. An alert nobody acts on has no value. Who gets it, and what do they do next?

How to measure performance

Before going live on anything important, run a simple test.

• Detection rate: Walk the area 20 times, and count how many were caught.
• False alarms: Count alerts with no real event over several days and nights.
• Conditions: Test in daylight, at night, and in poor weather if possible.
• Response time: How long does the alert take to reach the right person?

Record your results, and share them with the customer. Don't quote a manufacturer's accuracy claims as a promise. These usually come from ideal conditions.

Hardware and capacity

• Check how many channels of analytics a recorder or server can handle, and which features it supports on each one.
• Consider same-brand combinations. Analytics events may not transfer between brands, even using ONVIF. See our ONVIF guide.
• Consider data storage for metadata and thumbnails.
• If cloud analytics are involved, compare the ongoing subscription costs and the upload needs.
• Keep firmware up to date, as analytics often improve through updates. Test after each update.

Privacy and legal questions

Analytics can create new data about people. Counting visitors may raise few concerns, but tracking individuals, identifying faces or inferring behaviour can raise many.

Encourage your customer to consider:

• whether the feature is necessary for their purpose
• whether staff, visitors and the public have been told about it
• what data is kept, and for how long
• who can access it
• whether the law imposes extra duties, particularly for biometric data

You aren't their lawyer, but you should raise these questions. Disable features the customer doesn't need.

Selling analytics honestly

Promise outcomes you've tested, and describe limits plainly.

Don't say: "Our AI will catch all intruders."

Do say: "This camera can alert you when a person enters the yard after hours. We'll tune it during the first week. It won't catch every case, particularly in very poor weather, so we recommend keeping your existing alarm too."

Good first offers

• Person and vehicle alerts for after-hours perimeters
• Smart search for faster investigations
• Health alerts for system failures
• People counting for a retail customer with a specific need

Price the tuning

Analytics take time to configure and tune. Include this in your quote as a line item rather than doing it for free. Consider offering periodic reviews under a maintenance contract, as scenes change over time, with new plants, moved stock or new lighting.

Show a demo

Demonstrate on a similar site or on the customer's own premises using a sample camera. Real false alarms and real detections are far more convincing than marketing videos.

Common mistakes

• Enabling alerts on every camera. Customers soon turn them all off.
• Using a wide overview for analytics. Targets are too small.
• Promising face recognition without considering the law. It can cause legal trouble for the customer and for you.
• Ignoring the response. Alerts with no plan achieve nothing.
• Not testing at night. Performance can change dramatically.
• Treating analytics as a substitute for fire or life safety systems. They aren't.

Used well, analytics make a system far more useful, not by replacing people but by directing their attention where it's needed. The installers who do well here are the ones who test before they promise.