Choosing CCTV Resolution and Lenses So You Actually See Faces: DORI and Pixel Density

Expensive cameras, unrecognisable faces

We see this a lot. Something goes missing from a warehouse, someone pulls up the footage, and all it shows is a figure walking past. The camera really is 4MP or 8MP, but it has a 2.8mm lens and sits 12 metres from the door. The camera is fine. The lens and mounting position don't match what you need to see.

The number that matters is pixel density: how many pixels land on the target, measured in pixels per metre (px/m). Megapixels on their own tell you very little.

DORI levels under IEC 62676-4

The formula we use on site: Pixel density = horizontal pixels ÷ scene width at the target (metres)

Worth memorising: 2MP is 1920 pixels wide, 4MP is 2560, 8MP (4K) is 3840. At 250 px/m, a face about 16 cm wide gets roughly 40 pixels. That's fewer than most people expect, and it's why a wide-angle camera never delivers faces at a distance.

Let's run some numbers

A 4 m wide entrance. A 2MP camera framed to exactly 4 m gives 1920 ÷ 4 = 480 px/m. That's nearly twice the Identify threshold. No need to pay for 8MP here.

A 40 m wide car park. A 4MP camera gives 2560 ÷ 40 = 64 px/m, which is only Observe. You'll see someone open a car door but not who it was. If you need faces, add a camera at the entry and exit where everyone has to pass.

A 2.8mm lens at 10 m. Horizontal field of view is roughly 100°. Scene width = 2 × 10 × tan(50°) ≈ 24 m. A 4MP camera delivers 2560 ÷ 24 ≈ 107 px/m. Not yet Recognize.

Need Identify at 10 m with 4MP? The scene can be at most 2560 ÷ 250 ≈ 10 m wide. Field of view ≈ 2 × atan(5/10) ≈ 53°. That's a lens of about 6mm, or a varifocal set to that point.

Lens and field of view, approximately

Actual angles depend on each model's sensor size. Check the datasheet before you order. Where we don't yet know the real distances, we like motorized varifocal lenses: the installer can set zoom and focus from the NVR without climbing the ladder again.

Mounting height and tilt

Enough pixels won't help if the camera looks down too steeply. Take a camera at 3 m looking at a 1.6 m tall person 8 m away: the downward angle is about 10°, and the face is clear. Move the same camera up to 5.5 m and the angle becomes 26°, and a cap or hard hat hides most of the face. On mine and construction sites where everyone wears a helmet, this matters a lot.

Our simple rule: face cameras go at 2.5–3 m, with the target at least twice that height away. Overview cameras can go higher, because their job is Detect and Observe.

For corridors and warehouse aisles, use corridor mode (the image rotated to a 9:16 portrait format). A 4MP camera in this mode puts 1440 pixels across the aisle and spends almost all of them on the space that matters instead of on two side walls.

Night: infrared or white light

Most cameras switch to infrared at night and give a black-and-white picture. The IR range printed on the datasheet is a maximum. The range at which you can actually see a face is often less than half of it.

Colour-at-night lines such as Hikvision ColorVu and Dahua Full-color use wide-aperture lenses (F1.0) and more sensitive sensors, and some models add a white-light LED. Seeing clothing and vehicle colours at night helps an investigation a great deal. The downside: white light can annoy neighbours and attracts insects. In Laos, insects and spider webs in front of the lens are among the top causes of false night-time alerts. Put outdoor camera cleaning in the maintenance plan.

Higher resolution has hidden costs

An 8MP camera uses roughly twice the bitrate and disk space of a 4MP camera looking at the same scene. With H.265, the planning figures we start from are about 4–6 Mbps for 4MP and 8–12 Mbps for 8MP. Across 32 cameras and 30 days of recording, that difference means several extra hard drives and possibly a bigger NVR. The full method is in our NVR storage calculation guide.

Worked example: cameras for one warehouse

The site has a 6 m vehicle gate, a 30 m wide loading dock, three racking aisles 40 m long, and an office door.

1. Vehicle gate (Identify drivers and pedestrians). A 4MP varifocal framed to 6 m: 2560 ÷ 6 ≈ 427 px/m, mounted at 3 m. Reading number plates needs a separate camera built for the job; see license plate recognition.
2. Loading dock (Observe). 30 m with 4MP gives 85 px/m, enough to Observe. For Recognize, use 8MP: 3840 ÷ 30 = 128 px/m.
3. Racking aisles (Observe/Recognize). One 4MP camera per aisle in corridor mode, around a 6mm lens.
4. Office door (Identify). A 2MP or 4MP camera framed to 3 m is plenty.

That's six cameras, and each one has a clear purpose. Ten 2.8mm cameras around the building might look safer, but none of them would give you a face.

What the datasheet won't tell you

• Backlight: a glass entrance with afternoon sun behind it turns every face into a silhouette. You need True WDR, around 120 dB, not digital WDR.
• Night: high-megapixel cameras with small sensors get noisy and smeary after dark. Look at sensor size (1/1.8", for example) and realistic IR range.
• Moving vehicles: a slow shutter blurs moving cars. Set gate cameras to a faster shutter than the default and check the night image.
• Lao conditions: heat, humidity and red dust on mine sites. Choose IP67 housings rated for high temperatures, and put surge protection on every cable that leaves the building, because rainy-season lightning is heavy.
• Form factor: turret cameras don't suffer IR glare inside a dusty dome bubble. Where vandalism is likely, specify IK10.

Mistakes we keep running into

1. 2.8mm lenses everywhere because they "see more", and then nobody can be identified.
2. Going 8MP without budgeting for the extra disk and bandwidth.
3. Cameras mounted 5–6 m up that show the tops of heads. For faces, 2.5–3 m usually works better.
4. Handing over in daylight and never coming back to check the night image.
5. Relying on digital zoom in recorded footage instead of choosing the right lens. Zooming into a recording doesn't add pixels.

What we'd do

Before choosing a brand, mark on the floor plan what each camera is for: Detect, Observe, Recognize or Identify. Then pick the lens. Most good designs are a few wide-angle cameras for the overview plus longer lenses at the points people must pass through. For general work 4MP is the sweet spot. Save 8MP for wide yards or places where you'll genuinely need to zoom into recorded footage.

If you can, put a sample camera up on site for a night or two before ordering dozens.

Once you've settled on 4MP or 8MP, put your camera count into our CCTV storage calculator to see how many terabytes you'll need. Or send us your floor plan or a few photos of the spots you want covered on WhatsApp, and we'll mark up lens choices and positions.