Lighting, Angle, Occlusion

Most disappointing deployments are disappointing for physical reasons. The model is rarely the limiting factor.

Lighting

Dynamic range is the usual killer. A doorway with bright daylight outside and a dim interior exceeds what the sensor can capture in one exposure; one of the two ends up unusable.

Backlight turns a subject into a silhouette.

Direct sun into the lens produces flare and washes out a region of the frame.

Mixed sources — daylight and sodium and fluorescent — shift colour unpredictably.

Infrared at night removes colour entirely, which breaks anything relying on it.

The fix is almost always physical: move the camera, change its angle, add or shield lighting, use a sensor with better dynamic range.

Angle and height

Too high and you see the tops of heads, which is poor for detection and useless for anything needing a face.

Too low and everything occludes everything else.

Too oblique and objects at distance are compressed into a few pixels.

Overhead is excellent for counting and poor for classification.

A camera installed for human review is usually wide, high and oblique, which is close to the worst configuration for automated detection, and this is the single most common reason existing camera estates disappoint.

Scale

Pixels on target is the number that predicts detection.

A wide-angle camera covering a large area gives very few pixels per person at the far end, and no model recovers information that was never captured.

Compute it before installing: field of view, resolution and distance give you pixels on target directly.

More cameras with narrower fields beats fewer cameras with wider ones, for detection, and costs more.

Occlusion

Partial occlusion degrades detection gracefully in good systems and abruptly in poor ones.

Crowds are the hard case: people occluding each other defeats counting more than anything else.

Static occlusion — racking, pillars, machinery — creates permanent blind zones that need mapping and stating.

Map the blind zones explicitly and report them, because a system silently not covering a third of an area is worse than one that says so.

Weather, outdoors

Rain on the lens, which is the most common outdoor failure and is fixed with a hood rather than software.

Fog and heavy rain reduce contrast at range.

Snow changes the whole scene appearance and produces motion.

Wind moves vegetation, which triggers motion-based systems continuously.

Expect materially different performance by weather, and evaluate across it rather than on a clear day.

The survey that should precede purchase

For each camera: field of view, distance to the zone of interest, pixels on target, light conditions at four times of day, occlusions, weather exposure.

Compare against what the intended detection requires.

Most sites discover that a proportion of cameras cannot support the intended application, which is much cheaper to learn before the licence is bought than after.

The four-times-of-day check

A survey step that catches most lighting failures before installation.

Photograph each camera position at dawn, midday, dusk and night, from where the camera will be.

Look for: direct sun into the lens, backlit subjects, dynamic range beyond the sensor, dark zones.

Do it in the season with the difficult sun angle, or model it.

Most sites have several hours a day of hard conditions and nobody checks until the system underperforms during them.

The fix at survey stage is moving a mount. After installation it is a project.

Mapping the blind zones

A system silently not covering an area is worse than one that says so.

Walk each camera's view and mark what it cannot see: behind racking, behind pillars, inside machinery, beyond range.

Mark it on a plan, per camera.

State it in the report, so nobody assumes coverage that does not exist.

Re-map after any change to the space, because a new rack creates a new gap.

Where a blind zone covers something that matters, that is a camera decision, not something to be resolved by tuning.