Whoever turns the laser measure already knows where it pointed.
Picture someone standing in the middle of a dark room, slowly turning with a laser distance meter in hand. At each step of the turn they write down one number: the distance to the wall. There is no need to write “now I pointed at 37 degrees”, because the next line of the notebook is always the next step of the turn.
After a full turn the notebook holds one column of distances, and that is enough to draw the room. A car LiDAR does the same with 64 or 128 stacked lasers, ten turns a second. Each laser draws a ring around the vehicle.
The usual way to store the result is to convert it all into loose points in space, with x, y and z. That is three numbers where there was one, and the order in which the shots happened is lost. Our proposal is to keep the notebook, not the drawing.
Three steps, and the compressor is the usual one.
1. The ring becomes a channel. The sequence of distances from one laser, turn after turn, is a time series like that of any sensor. It is the format TUBE already compresses: it predicts the next value from the previous ones and spends bits only when the measurement strays from the prediction by more than the agreed tolerance.
2. A flat wall costs almost nothing. Along a façade or the asphalt the distance changes slowly and the prediction holds. The cost concentrates where the beam jumps from one object to another: the corner of the truck, the edge of the pole.
3. The angles come from the sensor grid. The compressor knows how far the head turns between two shots and the tilt of each laser. Only the small deviation from that grid travels, measured with a step that follows the distance: fine up close, wider far away, always the same error in centimetres.
Illustration with 3 mm of range noise and a simple predictor. With a tolerance below the noise, almost every shot is a surprise: noise does not compress. The bit estimate is a teaching model, not the codec. The real bytes are in the measurements section.
The laser’s error is a stroke, not a box.
The laser errs along its own ray: it reads 20.03 m where there were 20.00 m. It barely errs in direction. When the point is stored as x, y and z with a tolerance per axis, the uncertainty becomes a box around the point, with slack in directions where the sensor never erred.
Storing the range, the uncertainty goes back to being a stroke along the beam. The question “how far is that object?” gets an answer with the least slack possible.
The guarantee we give is the easiest one to audit: every reconstructed shot lies within an agreed distance of the original, measured in a straight line in space. It is the same yardstick as a LAZ file at the equivalent scale, and that is how we compared.
Two sensors, three precisions, two outside yardsticks.
We ran it on two sensors from different makers: a Velodyne HDL-64E (the one in the public KITTI set, 64 beams) and an Ouster OS2-128 (the maker’s public sample, 128 beams, raw packets). That is 22 turns of each, about 2.6 and 2.9 million points.
The yardsticks are LAZ, the standard compressed point-cloud format, and G-PCC, MPEG’s point-cloud codec, in the reference software and in three configurations (the best one counts). All with the same maximum error per point, checked point by point.
Against LAZ the per-beam mode won all six cases, by 16% to 44%. Against G-PCC it won at fine precision and lost at coarse precision. The same compressor applied to x, y and z loses to LAZ every time: the gain comes from the geometry, not from the codec.
Shorter bar is better. G-PCC: MPEG TMC13 reference software, lossless geometry on the same grid, best of three configurations.
| Sensor | Tolerance | LAZ | G-PCC (best) | TUBE on x, y, z | Per beam | ÷ LAZ · ÷ G-PCC |
|---|---|---|---|---|---|---|
| Velodyne HDL-64E (KITTI) | 1 mm | 1.804 | 1.503 | 2.240 | 1.253 | 0.69 · 0.83 |
| 1 cm | 0.801 | 0.852 | 1.102 | 0.622 | 0.78 · 0.73 | |
| 5 cm | 0.430 | 0.297 | 0.574 | 0.352 | 0.82 · 1.19 | |
| Ouster OS2-128 | 1 mm | 1.989 | 1.231 | 2.335 | 1.112 | 0.56 · 0.90 |
| 1 cm | 1.029 | 0.531 | 1.246 | 0.701 | 0.68 · 1.32 | |
| 5 cm | 0.565 | 0.140 | 0.722 | 0.474 | 0.84 · 3.38 |
On a third source, the LOLA laser altimeter orbiting the Moon, the result repeats: one orbit in the order the shots happened takes 1.497 bytes per point, against 2.399 for LAZ (38% smaller, 10 cm tolerance). The same data as a reordered cloud, without the time axis, had lost to LAZ by 2% to 5%.
On a coarse grid, the octree wins.
At a 5 cm tolerance, G-PCC in octree mode came out ahead: by 16% on the Velodyne sensor and by 3.4 times on the Ouster. On the Ouster it was already winning at 1 cm.
The reason is easy to see alongside. When the grid is coarse and the sensor is dense, several shots land in the same cell. The octree stores which cells are occupied and how many points each holds. The per-beam mode keeps storing one range for every shot, in order.
These are different promises. The octree gives back a cloud. The per-beam mode gives back the sensor’s record, with every shot in its place in the scan. For a map or a detector working in 10 cm cells, the octree is the right tool, and that is what we recommend in that case.
– of cells hold a single point, in the real data ·
“No points” is not “no obstacle”.
The most common question asked of a point cloud is whether something is in a safety zone. The usual answer checks whether any point fell inside. But the cloud cannot tell two opposite situations apart: the sensor looked and nothing was there, or the sensor could not look, because a truck was in the way.
In the sensor frame the difference shows. Every beam that crossed the zone and came back from behind it is proof of empty space along that ray. Every beam that stopped short, or never came back, proves nothing.
We measured this on the 22 KITTI turns, over 1,760 zones of 3 × 3 m. Where the usual query answered “no obstacle”, we placed an object the size of a child and asked whether any beam would touch it. 40% of the time, none would: the scan would come out identical. At 6 m it almost never happens. At 45 m it happens in two thirds of the cases.
So the query returns, along with the verdict, the coverage: the fraction of the beams aimed at the zone that actually went through it.
Plan-view sketch, with 900 beams in a single plane. On the real sensor the count uses every ring and the tolerance of the compressed file.
| Distance to the zone | “No obstacle” | Object would fit unseen |
|---|---|---|
| 6 m | 147 | 0.1% |
| 12 m | 106 | 8.7% |
| 20 m | 193 | 37.1% |
| 30 m | 318 | 46.7% |
| 45 m | 297 | 65.8% |
| Coverage | Zones | Object would fit unseen |
|---|---|---|
| < 50% | 756 | 56.2% |
| 50–90% | 155 | 0.1% |
| ≥ 90% | 150 | 0.0% |
On the left, the result of the protocol. On the right, an observation made afterwards on the same sample: with half or more of the beams going through, the hidden object showed up once in 3,050 tries. It is a single stretch, and the coverage threshold is the user’s decision. A strict “proven empty”, with every beam going through, never occurred in the real data: 13% of the shots have no echo.
When to use the per-beam mode.
The file is the sensor’s record
Forensics, audits, reprocessing: when every shot must come back in its place, with a known maximum error.
The precision is fine
Up to about 1 cm is where it beat both yardsticks. On a very dense sensor, only at 1 mm.
For a map in large cells, use an octree
If the destination is a 10 cm grid and the order of the shots does not matter, G-PCC takes less space.
Ask for the coverage with the verdict
“No return, coverage 93%” and “no return, coverage 4%” are different answers to the same question.
The circle one of the sensor’s lasers draws around the vehicle on each turn. Here, a telemetry channel.
The maximum straight-line distance between each stored point and the original. It holds for all of them, with no averaging.
Of all the beams aimed at a zone, the share that went through it and came back from behind.
Where this may be wrong.
Two short samples
22 turns of each sensor, a little over two seconds of street. Rain, dust and other sensors may give different numbers.
G-PCC has a mode we did not measure
We compared against lossless geometry on a grid, the only way it can guarantee a maximum error per point. Its native lossy mode merges points and targets mean error; under such a criterion it may win.
The technique is not ours
Storing spinning LiDAR in the sensor frame is known, and G-PCC itself has an angular mode. What we measured is that a telemetry compressor, unchanged, does this job with error per shot.
On one sensor the grid was rebuilt
KITTI delivers points already converted. We recovered the rings from them; on a raw sensor file the angles would cost even less.
Coverage is a single-sample result
The table by coverage band was observed after the experiment, on one urban stretch. It is a hypothesis to confirm, not a threshold to adopt.
No echo does not say why
Open sky, a dark surface and glass give the same silence. Coverage treats them all as “not seen”, which is the safe side of the error.
Does this apply to your sensor?
Tell us in two lines which sensor it is, how much data it produces and what has to be proven later. The first conversation is to see whether this mode solves your case, and if it does not, we say so.
← Telemetry certified · stickybit.com.br
- Velodyne HDL-64E: KITTI raw,
2011_09_26_drive_0048, 22 turns (Geiger, Lenz, Stiller and Urtasun, 2013). Only the measurements appear here; the scene in the animations is synthetic. - Ouster OS2-128: the maker’s public sample,
OS2_128_bridge_sample, 22 turns, read withouster-sdk. - LAZ:
laspywithlazrs, point format 0, scale equal to twice the tolerance per axis. - G-PCC: MPEG TMC13 reference software (
MPEGGroup/mpeg-pcc-tmc13), octree, octree with angular mode and predictive with angular mode; lossless geometry on the grid. - LOLA: record
LOLARDR_092590139from PDS Geosciences (NASA), one orbit, 289,080 valid shots. - Common yardstick: straight-line error per point of at most √3 times the tolerance per axis, checked on every point.