Balluff 3D vision · how a machine sees depth

An interactive guide to 3D machine vision · methods, not products

How a machine sees depth

Every 3D sensor answers the same question: how far away is each point on the part? The methods differ only in how they ask. Some measure an angle. Some measure a time. One measures focus. This page teaches all of them by letting you operate each one.

0 points measuredmove your pointer: you are already scanning
scroll to begin ↓

The map

Three ways to ask one question

Light leaves a source, touches your part and arrives at a detector. Physics gives you exactly three things to measure about that trip: the angle it comes back at, the time it takes, and how well it focuses. Every industrial 3D method is one of these three ideas wearing different optics. Hold that map and the seven chapters below become one story.

Chapter 1 · 1D laser triangulation angle

One point of light

A laser puts a spot on the surface. A lens watches that spot from a known distance to the side, and images it onto a detector. When the surface rises, the spot slides across the detector. Height becomes position. That triangle is the foundation of the next five chapters. Grab the surface and move it.

Triangulation angle30°

The lever of the whole chapter. Steeper watching angle: more detector travel per millimeter of height.

Surface
Measured height
0.00 mm
signal strong
Drag the gray surface vertically. The red beam is the emitted laser; the blue trace is what the detector reads. Angle up, and one millimeter of height moves the spot further along the detector: finer resolution, shorter range. No free lunch. Resolution and range figures are class-level physics for a short-standoff sensor, not a product spec.

Two details separate a diagram from a real sensor. First, the detector is tilted on purpose. The surface moves along an axis the camera views obliquely, so the plane of sharp focus tilts too; tilting the detector until lens plane, detector plane and measurement axis meet in one line (the Scheimpflug condition) keeps the spot focused across the whole range. Turn the tilt off above and watch the spot blur at the ends of travel. Second, the noise floor is speckle: coherent laser light returning from a microscopically rough surface interferes with itself, and the spot's apparent center jitters by micrometers no matter how much light you add. It is the physics, not the electronics, that draws the line.

Chapter 2 · 2D laser triangulation angle

A line of triangles

Chapter 1 measured one point. Now watch that spot stretch into a line: a cylinder lens spreads the beam, the detector gains a second dimension, and the same triangle repeats in every camera column at once. One frame, one whole height profile. Hover the profile and find chapter 1 hiding in each column. The first honest cost of the angle arrives with it: whatever the camera cannot see, it cannot measure.

Laser line widtha point

It spreads on its own the first time. Then it is yours.

Part on the table
Camera angle30°

Steeper angle: taller line deflection per millimeter, and longer shadows.

This profile
0 points
0 lost to shadow
Drag the part left and right under the line. Two different shadows steal data: the laser's own shadow (a wall blocks the sheet of light) and the camera's shadow (the surface is lit, but the camera cannot see past an edge). They fall on different sides. Real profile sensors run this at hundreds of profiles per second up to the tens of kilohertz. One method, several names: 2D laser triangulation, laser line profiling, sheet of light, even the simplest structured light. The geometry answers to all of them; it is chapter 1's triangle, in parallel.

Look closely at the camera image: the line is a few pixels wide and roughly Gaussian across its width. The sensor does not read the brightest pixel; it computes the center of gravity of the intensity in each column and lands between pixels. That sub-pixel trick is where micrometer resolution comes from, and it is exactly what a glinting weld bead or a saturated highlight corrupts first.

Chapter 3 · 3D laser scanning angle

Motion makes it 3D

One profile is a slice. Move the part under the line and stack the slices: now you have a surface. The entire trick is knowing exactly where the part was when each slice was taken. That is the encoder's job, and you are about to see what happens without it.

Profile trigger

One profile every fixed millimeter of travel, no matter the speed.

Trigger pitch1.0 mm

Or grab the belt and pull it yourself. The belt drive is deliberately uneven, like a real line.

Scan
0 profiles
y pitch set by the encoder
The belt speed wobbles on purpose. Encoder-triggered, the wobble does not matter: profiles land every millimeter of travel and the part comes out true. Timer-triggered, every speed change stretches or compresses the reconstruction. Same sensor, same part, different geometry.

This is the workhorse of inline 3D inspection, and its resolution budget splits three ways: X and Z come from the optics of chapter 2, but Y is a choice you make at the encoder. Halve the trigger pitch and you double the rows, until the sensor's profile rate caps how fast the belt may run. Speed and point density trade against each other through one number.

Chapter 4 · structured light angle

A pattern instead of a line

Chapter 2's laser line was already structured light at its simplest: one known stripe. Why sweep it when a projector can paint the whole part at once? Stripes flood the surface; a camera at a baseline watches them bend over the shape. Each pixel sees the stripes shift as the pattern steps, and the phase of that shift encodes height. No motion, full field, one catch: phase repeats. Which stripe are you in?

Fringe period8 mm

Finer stripes: finer height, and more identical-looking stripes to confuse.

Pixel under your pointer
φ = —
hover the camera image to probe a pixel
Four exposures of the same stripes, each shifted a quarter period. From one pixel's four brightness values the sensor fits a sine and reads its phase: that is the height, modulo one stripe. The ghost surfaces are all the heights that phase could mean. A single coarse pattern settles it. Nudge the part between exposures and watch the whole reconstruction tear: the multi-shot bargain.

This is the snapshot family: dense clouds of an entire stationary surface in tens to a few hundred milliseconds, with micrometer-class precision on cooperative parts. Its enemies are motion during the sequence, ambient light washing out fringe contrast, and surfaces that fight the pattern itself: mirrors glint, and translucent materials scatter light beneath the surface and blur the stripes from within. Single-shot pattern codes exist for moving scenes; they buy speed with coarser, sparser data.

Chapter 5 · stereo vision angle

What two cameras agree on

Hold a thumb up at arm's length, look past it, and close one eye, then the other. The thumb jumps; the far wall barely moves. That jump is disparity, and it is the only thing stereo measures. Two cameras a known baseline apart play the role of your two eyes. Blink between their views below and watch the near block jump farther than the far plate. Then face the question the whole method rests on: which pixel over there matches this one here?

Scene surface

Smooth metal gives the matcher nothing to hold.

Baseline120 mm
Your match
no lock
drag in the right image to search
Blinking the views is the whole idea of stereo in one gesture: the shift you see is the depth, and widening the baseline makes every jump bigger. The matching game is the price of it. The marked patch lives somewhere on the drawn line in the right image: the epipolar line, the one honest shortcut in stereo. Your correlation score plots below as you search. When the surface is blank the peak is a plateau and every position lies equally well. Project a speckle pattern and the same surface locks instantly: that is active stereo, and it is why industrial stereo heads carry a projector.

The depth arithmetic is one line: Z = f·B/d. Its consequence rules the method: differentiate and the error grows as Z squared. Double the distance and the same sub-pixel matching error costs four times the depth uncertainty. Widen the baseline to fight back and the two views grow so different that matching itself suffers, and the occlusion gap between them widens. Stereo is generous up close, honest at range about what it cannot promise.

Chapter 6 · time of flight time

Ask the time, not the angle

Every triangle on this page dies slowly with distance: the angles flatten and the geometry runs out. Time of flight does not care. Send light out, time the round trip, and distance is time: 300 millimeters every nanosecond, at two meters or two hundred. The catch is what a millimeter costs: 6.7 picoseconds. Fire the pulse and try to catch it.

Flavor

Drag the target. Light is fast; the clock has to be faster.

Round trip
a 3 GHz CPU ticks every 333 ps: fifty times too coarse for a millimeter
Pulsed sensors time single flashes with picosecond-class statistics over many shots. Phase sensors modulate the light continuously and read the returning phase; past c divided by 2f the dial wraps and a far target reads near, until a second frequency breaks the tie. The same wrap you met in chapter 4, wearing different units.

Time of flight buys range and speed with precision: millimeter to centimeter class where triangulation reaches micrometers. Its systematic errors are physical, not electronic. In a concave corner light arrives by two paths and the pixel reports a blend, so corners bow inward. At a depth edge one pixel straddles near and far and reports a point floating between them, the flying pixel every ToF cloud carries. It is the method you pick when the scene is big, the standoff is long, and a millimeter is good enough.

Chapter 7 · chromatic confocal focus

When the surface fights back

A mirror finish defeats every triangle above: the light obeys the reflection law, leaves at the mirror angle and never reaches the lens. Chromatic confocal gives up on angles altogether. It sends white light straight down through a lens built to focus each color at a different depth, and asks which color comes back in focus. Color becomes the ruler.

Surface

Drag the surface up and down.

Spectrometer reads
Everything happens on one axis: no baseline, no shadow, no angle for a mirror to exploit. A pinhole rejects every wavelength except the one focused exactly on the surface, and a spectrometer names it. On transparent parts both faces reflect and two peaks return: the gap between them is the glass thickness, measured through the glass.

The price is written in the bars above. The rainbow of focal points spans micrometers to a few millimeters of range, the spot is micrometers wide, and covering an area means scanning point by point, which sends you back to the motion of chapter 3. Chromatic confocal is the specialist you call when the surface is polished, transparent or steep, and the tolerance has a µ in it.

The gauntlet

The same part, seven ways

Here is one deliberately hostile part: a machined step, a polished dome, a black rubber gasket, a glass window and a deep bore. Pick a zone. Every method reports honestly, and the reasons all come from the chapters you just operated.

The reference

The whole map, one table

Everything above, condensed to the figures an engineer compares. Every number is an order-of-magnitude class for the method, never a product spec.

MethodReturnsHeight precisionWorking scaleMotion neededFirst thing to break it
1D laser triangulationone distancesub-µm to tens of µmmm-class ranges, short standoffnone for a point; scan for morespecular or dark surfaces, sharp edges
2D laser profileheight profile Z(x)µm classlines of tens to hundreds of mmnone per profileocclusion shadows, gloss
3D laser scanningordered point cloudµm class in X and Z; Y is the trigger pitchparts of any length that move pastpart or sensor must move, encodedlosing track of the motion
structured lightdense cloud, full fieldµm to tens of µmfields of tens of mm to about a meternone; part still during the patternsmotion mid-sequence, gloss, translucency
stereo visiondepth map per shotmm class, worsening with distance squared0.3 m to tens of mnonesurfaces with nothing to match
time of flightdepth map per framemm to cm0.5 m to hundreds of mnonecorners (multipath), depth edges (flying pixels)
chromatic confocalpoint height, or thicknessnm to sub-µmwindow of µm to a few mmscan for arealeaving the tiny window

Two cousins earn a mention without a chapter. Photometric stereo lights the part from several directions with one camera and recovers surface slope, not metric height: unbeatable for making scratches and embossing visible, wrong tool for a dimension. Scanning LiDAR is chapter 6's physics swept over big scenes: stockpiles, vehicles and warehouses rather than parts.

The decision

Which method fits your part

Describe your measurement problem in five answers. The ranking below is the physics of this page applied to your part, with every judgment traced back to its chapter. Nothing you choose leaves this browser.

This ranking weighs method physics only: class-level resolution, range, speed, surface behavior and motion requirements. A real selection also weighs price, integration and the specific sensor's datasheet. Treat this as the shortlist you walk into that conversation with.