The megapixel myth: why 4K doesn't mean better AI CCTV

When a homeowner or business in Durban starts quoting a new CCTV system, the very first metric they usually look for on the spec sheet is the megapixel (MP) count. The assumption is entirely logical: if a 2-megapixel (1080p) camera produces a good image, then an 8-megapixel (4K) camera must produce an image four times better. And if the image is four times better, the camera’s Artificial Intelligence (AI) must be four times more accurate at detecting a person jumping a wall.

Unfortunately, in the world of computer vision and video analytics, this is a myth.

In fact, paying a premium for a higher megapixel count without understanding how AI actually processes video can result in worse detection performance, particularly under the harsh lighting conditions and frequent power interruptions we experience in KwaZulu-Natal. Higher megapixel counts do not automatically translate into better AI performance. Here is why the industry is moving away from the “megapixel race” and what you should be measuring instead.

The AI Pipeline Bottleneck

To understand why a 4K sensor is often wasted on AI, you have to look at how a neural network “sees” an image.

Cameras do not possess human cognition. When a camera looks at a driveway in Umhlanga, it doesn’t see a driveway and a gate; it sees a grid of colored pixels. To figure out if a cluster of those pixels represents a human being, the camera runs the image through an AI model (like the widely used YOLO—You Only Look Once—architecture).

Here is the catch: analyzing massive, multi-megapixel images in real-time (at 15 to 30 frames per second) requires an enormous amount of computational processing power. The processors built into the edge of standard commercial cameras simply cannot handle that load.

Many object-detection models operate on a fixed inference input that is smaller than the recorded stream. A square input around 640 pixels is common in published YOLO examples, but it is not a universal CCTV specification; manufacturers can use different model sizes, crops, secondary streams and processing pipelines.

Where a 3840×2160 frame is downscaled for inference, the detector does not receive every recorded pixel. That limits the assumption that four times the recorded pixel count must produce four times the analytics accuracy. It does not make the 4K recording pointless: the higher-resolution main stream may still preserve more detail for playback, cropping and evidence, and some analytics pipelines can use that detail differently. Check the exact camera or recorder rather than assuming one internal input size.

The Metric that Actually Matters: Pixels Per Meter (PPM)

If overall sensor resolution is largely irrelevant to the AI’s internal mechanics, what actually determines if the camera will catch an intruder? The critical factor is Pixels Per Meter (PPM) (sometimes referred to as Pixels Per Foot or spatial resolution).

PPM measures how many pixels actually cover the target object in the real physical world — pixels across the target, not across the scene. AI models need a minimum number of pixels on a target to classify it confidently.

This is not a rule of thumb someone made up. It is standardised. IEC 62676-4 defines a pixel-density ladder known as DORI, setting out what a given task requires:

Those four numbers are useful design categories, not a guarantee of recognition, identity or courtroom acceptance. A camera that gives roughly 30 px/m at the fence line is designed for detection-level detail, not an identification task. If the purpose is identification at the gate, calculate pixel density at the gate and then test the view under the expected lighting, movement and compression.

One current caveat worth knowing, because most published guidance has not caught up with it: the 2025 revision of IEC/EN 62676-4 replaces the four-level DORI model with a seven-level framework — overview, outline, discern, perceive, characterize, validate, scrutinize. DORI remains the language nearly everyone in the industry still speaks, and the underlying principle is unchanged, but a specification written today against “DORI” is being written against a superseded model. If you are drafting to a standard for a tender or an insurance requirement, work from the current standard text rather than from the familiar four letters.

This means the lens is far more important than the sensor.

Imagine a large warehouse in Pinetown. If you mount a 4K camera with an ultra-wide 2.8 mm lens, it may capture the entire yard while leaving a person at the distant fence as a very small target. Detection confidence can then fall despite the 4K label because the scene spreads the available pixels too widely.

Conversely, a 2-megapixel camera with a narrower lens aimed at the gate can put more of its available pixels on that target than the wide 4K view. It may perform the gate task better, subject to sensor quality, focus, light, motion, compression and the actual analytics pipeline. Lens and scene design decide the comparison; megapixels alone do not.

The Low-Light Penalty in a Load-Shedding Reality

Physics plays a cruel trick on high-megapixel sensors, and this is where Durban’s unique challenges come into play.

A camera sensor is a physical piece of silicon. If you have a standard 1/2.8-inch sensor, and you want to make a 2MP camera, you divide that silicon into two million light-gathering “buckets” (pixels). If you want to make an 8MP camera out of that exact same piece of silicon, you have to divide it into eight million buckets.

Because you are cramming more pixels into the same physical area, each individual pixel on the 4K camera must be significantly smaller. Smaller pixels capture less ambient light.

During the day, in the bright coastal KZN sun, this isn’t a problem. Both cameras perform flawlessly. But at night, the difference is stark. When load-shedding hits and the streetlights go dark, the smaller pixels on the 4K sensor struggle to gather enough photons. To compensate, the camera artificially boosts the signal, which introduces digital “noise” (that grainy, static-like interference you see on cheap night video).

AI algorithms struggle heavily with visual noise. The neural network cannot tell the difference between digital grain and actual movement, leading to missed detections and a massive increase in false alarms. A clean, well-lit 2MP image with larger light-gathering pixels will consistently yield better AI detection than a noisy, light-starved 8MP image in a dark environment.

When Do Megapixels Actually Matter?

It is important to state that megapixels aren’t entirely useless. We are not advocating that everyone revert to ancient analog resolutions. High resolution is highly beneficial, but only in specific, defined scenarios:

1. Forensic Digital Zoom: AI might not need 4K to detect a person, but human operators often do. If an event happens, and a site manager needs to digitally zoom in on the recorded footage after the fact to read a nametag, identify a specific tool being carried, or verify a face, the higher resolution retains clarity longer before pixelating. The AI detects the event; the 4K resolution provides the forensic evidence.
2. Massive Coverage Areas: If budget constraints mean one camera must cover a massive, open logistical yard where targets will be very far away, a 4K sensor does ensure that a distant object still occupies enough pixels (maintains a high enough PPM) to be detected by the downscaling AI.
3. Optical Character Recognition (OCR): Specialized AI tasks like Automatic Number Plate Recognition (ANPR) require immense clarity. Reading an alphanumeric plate moving at speed requires a very high PPM, which is achieved through a combination of high-megapixel sensors and tight, motorized zoom lenses.

The Installer’s Verdict

The surveillance industry’s marketing departments will always push the highest megapixel numbers because it is an easy metric to sell to consumers. But as an installer designing a system to secure a Durban property, the goal is not to buy the biggest number; the goal is to buy the most reliable detection.

Stop buying cameras based solely on the megapixel count on the box. For optimal AI accuracy, focus your budget on proper lens selection to ensure the correct Pixels Per Meter, invest in excellent supplemental lighting (or cameras with large-aperture lenses designed for low light), and ensure the camera’s processing power matches the complexity of the scene. A properly designed 4-megapixel system will outsmart a poorly deployed 8-megapixel system every single night of the week.