Choosing a camera for video analytics: criteria beyond resolution

Resolution is not enough

When choosing a camera, the first number on the spec sheet is almost always resolution. It matters, but on its own it can be misleading. For video analytics, the real question is not "how many pixels does the camera produce?" but "how many pixels does the person, vehicle or object you want to detect actually take up in the frame?"

A high-resolution camera with a very wide-angle lens, mounted far from the area of interest, can produce tiny figures that are hard to detect reliably. A more modest camera that is well placed and well framed will often do a better job. Choosing a camera for analytics is therefore a matter of geometry and real-world conditions first, and of the datasheet second.

This article walks through the criteria we look at during an audit, whether the plan is to reuse an existing camera or add a new one.

Field of view, focal length and subject size

The focal length of the lens sets the field of view: a short focal length covers a wide area, a long one shows a narrower area in more detail. The right balance depends on what you want to measure.

A useful way to think about it is pixel density on target: the number of pixels available per metre in the area of interest. There is no universal figure, because the requirement depends on the task, the AI model and the site conditions. A few principles still hold:

• counting people across an entrance line requires each person to be clearly separated from the next, which favours tight framing on the doorway;
• measuring zone occupancy (a dining room, a car park, a wash bay) tolerates smaller subjects, but the whole zone must be visible with no blind spots;
• detecting small objects (boxes, pallets, equipment) needs more detail in the relevant area;
• varifocal lenses, adjustable on installation, give useful leeway when the exact layout is not yet settled;
• very wide-angle and fisheye lenses cover a lot of space but distort the image, which not every analytics model handles equally well.

Dome, bullet, PTZ: strengths and limits for analytics

Each camera form factor was designed for a particular job, and they are not all equally suited to automated analysis.

• Dome cameras: discreet, vandal-resistant and well suited to indoor ceiling mounting. Mounted overhead, they work well for counting people at entrances. Watch out for reflections or dirt on the dome cover, which degrade the image, especially at night under infrared.
• Bullet cameras: often used outdoors, they reach further and are easy to aim. They suit car parks, loading docks and site perimeters. Their visibility acts as a deterrent, which can be a plus or a minus depending on the setting.
• PTZ (pan-tilt-zoom) cameras: they let an operator follow a scene. As the main source for analytics they are usually a poor choice: as soon as the camera moves or zooms, the configured zones and lines no longer match the image. If a PTZ has to be used, it must stay in a fixed position while analysis is running.
• 360° fisheye cameras: they cover an entire room from the ceiling. Handy for an overall view of occupancy, but they need processing that copes with the distortion.

Low light, WDR and backlighting

An AI model can only detect what the camera manages to show. Lighting conditions are therefore often more decisive than resolution.

Backlighting is the most common problem in shops and restaurants: a glass door or a bright window behind people turns them into silhouettes. Wide Dynamic Range (WDR), when it genuinely performs well, helps balance very bright and very dark areas. It is not magic, though, and a small change in camera angle is sometimes more effective.

At night, built-in infrared lets the camera see in the dark, but in black and white, with limited range and a risk of overexposing nearby subjects. Reflections on glass, rain, steam and water spray, in a car wash for instance, also disrupt the image. During an audit, we recommend looking at each camera at different times of day, not just at the time of the site visit.

RTSP/ONVIF compatibility, PoE and the network

A technically excellent camera is useless for analytics if its stream cannot be accessed. Before buying anything, check a few integration points:

• a documented RTSP stream and, ideally, ONVIF support, which makes discovery and configuration easier across different brands;
• the ability to configure multiple streams (a main stream for recording, a sub-stream for analytics), with control over resolution, frame rate and bitrate;
• Power over Ethernet (PoE), which simplifies installation to a single cable, provided the switch has enough power budget for every camera, including infrared at night;
• enough local network capacity to carry the streams to the Edge AI unit without congestion, ideally on a dedicated video network or VLAN;
• the ability to update firmware and change default credentials, which is essential for security.

Reuse an existing camera or add a dedicated one?

Starting with what you already have is often the most cost-effective route, and it is the first thing we look at. But a camera installed for security usually points at a till, a door or a stockroom, not necessarily at the area that needs measuring. Adding a dedicated camera makes sense when:

• no existing camera covers the area of interest from a usable angle;
• the installed model does not expose an accessible network stream or cannot be set up with a sub-stream;
• counting entries requires an overhead view that the security camera cannot provide;
• moving the existing camera would lose the security coverage it was installed for.

Validate the choice before you buy

The best way to avoid a poor investment is to test before rolling out: capture a few minutes of representative footage (peak hour, evening, backlit conditions), check that subjects are detected consistently, and only then finalise the equipment list. That is exactly what the feasibility study VIDEO.MA carries out before every project is for.

Our Hardware page outlines the equipment families we assess and install, and our Technology page explains our Edge AI approach. Not sure whether to reuse your cameras or add new ones? Request a quote or an audit of your installation and we will tell you, camera by camera, what can be used.