Camera Placement

The single best predictor of whether a vision system works is where the cameras are. It is a physical problem and it is solved with a ladder rather than a model.

Pixels on target

The number that predicts detection, and it is computable before anything is installed.

Resolution, field of view and distance give it directly.

A person at the far end of a wide-angle view is a handful of pixels, and no model recovers information the sensor never captured.

Different tasks need different densities: presence detection needs few, classification needs more, reading text or plates needs many more again.

Compute it per camera against the intended task, and expect a proportion of an existing estate to fail the test.

Angle

Overhead: excellent for counting, minimal occlusion, consistent scale, and it sees the tops of heads rather than faces — which is a privacy advantage as well as a technical one.

Oblique at moderate height: general purpose, more occlusion, variable scale across the frame.

High and wide: what most existing estates have, installed for human review, and close to the worst configuration for automated detection.

Low: everything occludes everything.

Match the angle to the task rather than to what is already mounted, which is the decision that costs money and saves the project.

Lighting

Avoid pointing at bright sources. A camera facing a window or a doorway spends part of every day with an unusable frame.

Watch the sun path across the year, not the day you surveyed.

Add lighting rather than raising sensitivity, which is cheaper than a better camera and more effective.

Consistent illumination beats bright illumination.

Infrared removes colour, which breaks anything relying on it and is worth knowing before specifying a night application.

Coverage and blind zones

Map them explicitly, per camera, on a plan.

Static occlusion from racking, pillars and machinery creates permanent gaps.

State the gaps rather than letting the system silently not cover a third of the area.

Overlap at boundaries if tracking across cameras matters, and accept that tracking across non-overlapping views is a much harder problem.

The survey

Before purchase, per camera: field of view, distance to the zone of interest, computed pixels on target, light conditions at four times of day, occlusions, weather exposure, mounting stability.

Compare against the requirement for the intended detection.

A day of survey work regularly saves a failed deployment, and it produces the evidence for the camera budget that the proposal omitted.

Retrofit or install

Existing cameras were installed for a different purpose and mostly do not suit this one.

Reusing them is the default assumption and it is usually wrong, which is why deployments underperform their demonstrations.

Budget for camera work from the start, as a line item, or the project will be judged on hardware that was never suitable.

Where the budget will not stretch, reduce the scope rather than accepting cameras that cannot support the task. A system working properly on four zones is worth more than one failing across twenty.

Computing pixels on target

Four numbers and a calculation, done before anything is bought.

Sensor horizontal resolution.

Horizontal field of view at the distance of interest, from the lens specification.

Distance to the zone.

The size of the thing being detected.

Divide to get pixels across the target.

Compare against the requirement: presence detection needs few, classification more, reading detail many more.

Do it per camera, and expect part of an existing estate to fail.

Reducing scope rather than accepting bad cameras

The decision that saves a project when the camera budget will not stretch.

A system working properly on four zones is worth more than one failing across twenty.

Accepting unsuitable cameras produces a deployment judged on hardware that was never adequate, and the judgement lands on the technology rather than the mounting.

Say which zones the current cameras can support, with the pixel arithmetic.

Deploy there, measure, and use the result to fund the rest.

This also produces a working reference inside your own organisation, which is more persuasive than any vendor's.