Point Cloud Density and Scan Resolution Trade-offs

A scanner’s resolution setting is an angular step. Everything you care about — point spacing on a surface, file size, scan time, processing cost — follows from that one number and the range, and every figure below is derived from stated inputs rather than quoted from a brochure.

Resolution is an angle, density is a distance

A terrestrial scanner steps its mirror by a fixed angle between measurements. The spacing between adjacent points on a surface perpendicular to the beam is therefore the arc length:

spacing = range x angular_step_in_radians

angular step 0.036 deg = 6.283e-4 rad
  at 10 m:  10 x 6.283e-4 = 6.28 mm
  at 30 m:  30 x 6.283e-4 = 18.8 mm
  at 60 m:  60 x 6.283e-4 = 37.7 mm

angular step 0.018 deg = 3.142e-4 rad
  at 10 m:  3.14 mm      at 30 m:  9.4 mm

angular step 0.009 deg = 1.571e-4 rad
  at 10 m:  1.57 mm      at 30 m:  4.7 mm

Two consequences follow immediately and neither is optional. Density is not a property of a scan — it is a property of a scan at a range, and a single cloud from one setup has a density that varies by an order of magnitude between the near wall and the far one. And moving the scanner closer is exactly equivalent to refining the angular step, but costs nothing in points, time or storage. Halving the range halves the spacing; halving the angular step halves the spacing and quadruples everything else.

The same relation holds for airborne and mobile sensors with the terms renamed. A scanner on a drone has a fixed angular step and its range is the flying height, so spacing on the ground is set by altitude exactly as it is set by standoff distance for a tripod. What changes is that a moving platform adds a second, independent spacing: along-track spacing is set by ground speed divided by the sweep rate and has nothing to do with the angular step at all. Flying faster thins the cloud in one direction only, which is why airborne density is quoted as points per square metre rather than as a spacing — the sampling is anisotropic and a single distance would hide that.

Three settings, derived

Take a scanner covering 360° horizontally and 300° vertically — the usual coverage for a tripod instrument, which cannot see the small cone beneath itself.

directions measured = (360 / step) x (300 / step)

  0.036 deg:  10,000 x  8,333 =    83.3 million
  0.018 deg:  20,000 x 16,667 =   333.3 million
  0.009 deg:  40,000 x 33,333 = 1,333.3 million

assumption: 60 % of directions produce a return
(the rest are sky, or beyond range, or absorbed)

  0.036 deg:   50.0 million points
  0.018 deg:  200.0 million points
  0.009 deg:  800.0 million points

file size, LAS point data record format 1 at 28 bytes
  0.036 deg:   50e6 x 28 =  1.40 GB
  0.018 deg:  200e6 x 28 =  5.60 GB
  0.009 deg:  800e6 x 28 = 22.40 GB

scan time, at an assumed measurement rate of
1,000,000 directions per second
  0.036 deg:    83 s
  0.018 deg:   333 s  =  5.6 min
  0.009 deg:  1333 s  = 22.2 min

processing, at an assumed 20 microseconds per point
across the whole pipeline, on 16 cores at 70 % scaling
  0.036 deg:   50e6 x 20e-6 / (16 x 0.7) =    89 s
  0.018 deg:                                 357 s
  0.009 deg:                               1,429 s = 24 min

Every column quadruples per step. That is the shape of the whole decision: one halving of the angular step costs four times the storage, four times the scan time, four times the processing, and buys a factor of two in point spacing. Across the two steps shown, 16× the cost for 4× the spacing.

Scan time deserves separate attention because it is the column that costs the most in practice. A crew is on site for a day and the day contains a fixed number of hours; twenty-two minutes per setup against eighty-three seconds is the difference between fifteen setups and three. Since more setups is the single most effective way to reduce occlusion, incidence angle and range simultaneously, raising the resolution setting frequently produces a worse survey overall — denser data from fewer places, with more of the site never measured at all. The resolution decision and the setup-count decision are the same decision under a fixed time budget, and they are almost always made separately.

Work backwards from the smallest feature

The right setting is not the highest one you can afford. It is the one that puts enough points across the smallest feature the deliverable requires, and that is a calculation rather than a judgement.

The working rule is three points across a feature to detect it and five to model it, which is the sampling-theory intuition applied conservatively — two samples is the theoretical minimum to see something at all, and real data has noise and misses.

requirement: model a 50 mm pipe at 20 m range
  need 5 points across it -> spacing <= 10 mm
  step <= 0.010 m / 20 m = 5.0e-4 rad = 0.029 deg

  -> the 0.036 deg setting is marginal (12.6 mm at 20 m)
  -> the 0.018 deg setting gives 6.3 mm: comfortable
  -> the 0.009 deg setting gives 3.1 mm: 4x the cost of
     the setting that already met the requirement

requirement: detect a 10 mm crack at 5 m
  need 3 points across it -> spacing <= 3.3 mm
  step <= 0.0033 / 5 = 6.7e-4 rad = 0.038 deg

  -> the coarsest setting meets it, because the range
     is short. move the scanner, do not raise the setting.

The second example is the one that saves the most money in practice. Scan resolution and standoff distance are interchangeable, and standoff distance is free. Two setups at 5 m at the coarse setting cost less in every dimension than one setup at 20 m at the fine setting, and give better data because they also reduce incidence angle and occlusion.

The beam footprint caps the useful setting

There is a hard limit past which a finer angular step measures nothing new, and it is set by the beam itself. A laser beam diverges, so its spot on the target has a diameter that grows with range. Each measurement is an average over that spot, not a point.

assumption: beam divergence 0.3 mrad (3.0e-4 rad)

footprint diameter = range x divergence
  at 10 m:  10 x 3.0e-4 =  3.0 mm
  at 30 m:  30 x 3.0e-4 =  9.0 mm
  at 60 m:  60 x 3.0e-4 = 18.0 mm

compare with the point spacing at 30 m:
  0.036 deg step -> 18.8 mm spacing, 9.0 mm footprint
                    spots do not touch: undersampled
  0.018 deg step ->  9.4 mm spacing, 9.0 mm footprint
                    spots just meet: matched
  0.009 deg step ->  4.7 mm spacing, 9.0 mm footprint
                    spots overlap by half: each new point
                    is largely a re-measurement of the
                    previous one

Past the point where the spacing drops below the footprint, extra points are correlated rather than independent. They still help a little — averaging correlated measurements reduces noise somewhat — but they do not resolve a smaller feature, which is what people believe they are buying. This is why the finest setting on an instrument is rarely the right one at range, and why the same setting can be genuinely useful up close where the footprint is 3 mm.

Range noise sets a second floor underneath that one. If the instrument’s single-point range accuracy is a few millimetres, two adjacent points 1.5 mm apart differ in reported position by more noise than signal, and the local surface estimated from them is dominated by the noise rather than by the geometry. Extra points do help here, but only in the averaging sense: fitting a plane to n points reduces the fit’s uncertainty roughly as the square root of n, so quadrupling the point count halves the noise on a fitted surface. That is a real benefit and it is a different benefit from resolution. If what you need is a more accurate plane, more points help; if what you need is to see a smaller object, they do not.

Incidence angle ruins the arithmetic

Everything above assumed the surface is perpendicular to the beam. On a surface tilted by angle θ from perpendicular, the spacing along the tilt direction is stretched by 1 / cos θ and the beam footprint is stretched into an ellipse by the same factor.

spacing on a tilted surface = range x step / cos(theta)

at 30 m with a 0.018 deg step (9.4 mm perpendicular):
  theta = 0 deg   ->  9.4 mm
  theta = 45 deg  -> 13.3 mm  (x1.41)
  theta = 70 deg  -> 27.5 mm  (x2.92)
  theta = 80 deg  -> 54.1 mm  (x5.76)

A floor scanned from a tripod at eye height is at a grazing incidence almost everywhere beyond a few metres, so the floor is the sparsest and noisiest surface in a scan that nominally has millimetre spacing. Range noise also rises at grazing incidence, because the elongated footprint spans a range interval rather than a single distance. Any density budget computed from the perpendicular case is optimistic for most of the scene, and the honest version applies the cosine factor per surface.

What to actually do

1. Write down the smallest feature the deliverable must resolve and the longest range at which it must be resolved. If you cannot write it down, the resolution question has no answer and you are guessing.
2. Compute the required spacing from three to five points across that feature, then the required angular step from that spacing and the range. Add the cosine factor for the worst incidence angle you expect.
3. Check the required spacing against the beam footprint at that range. If the spacing is well below the footprint, the requirement cannot be met by resolution — move closer or use a different instrument.
4. Prefer more setups at shorter range to a higher setting from far away. It costs scan time but reduces points, occlusion and incidence angle simultaneously.
5. Estimate the resulting volume before scanning, not after. The table above scales linearly with return rate and directly with the record size, and the downstream consequences are derived on the processing cost page and the compression page.