Skip to main content

2D vs. 3D Vision: Structured Light, Laser Triangulation, Stereo, and ToF

A 2D camera tells a handling system where a part lies in a plane; a 3D sensor adds how high the part sits, how it is tilted, and what shape its surface has. UTEC Industrial designs, engineers, machines, fabricates, and installs custom material handling systems for aerospace and heavy industry from its Spokane Valley, WA facility, integrating Allen-Bradley PLC and motion control with in-house CNC machining, heat treating, and stress relief. This article compares the four common ways of measuring depth, structured light, laser triangulation, stereo, and time-of-flight (ToF), by how each one works, what limits it on a plant floor, and how its accuracy figures should be read. In the build chain, design → engineering → parts machining → fabrication → assembly → weld fatigue → stress relief → drives → controls → tuning → monitoring, a 3D sensor belongs to the controls and monitoring links, but its working range and mounting are fixed much earlier.

What does a 3D sensor measure that a 2D camera cannot?​

A 2D camera measures position and rotation in the image plane. It cannot say how high a part sits or how it is tilted, which is why stacked parts, parts in bins, and parts at varying heights push a handling system toward 3D; the category overview on what machine vision does in material handling covers that threshold and robot bin picking. This article starts where that one stops, at the choice among 3D methods.

The Hornberg handbook's 3D section has a subsection on 2.5D and 3D data, gives point clouds and registration their own subsection, and gives passive and active methods of 3D data acquisition their own subsections. Only the handbook's table of contents is used here, so the description of each group below follows the other sources cited in this article:

  • Passive methods, which use the light already on the scene. Stereo vision is one example: two calibrated views, with depth found by matching the same point in both images.
  • Active methods, which project their own light and measure how it returns. Structured light projects a coded pattern, laser triangulation projects a spot or a line, and a ToF camera times emitted light.

A related point on specifications: EMVA 1288 Release 4.0 is the European Machine Vision Association's standard for measuring and presenting the specifications of machine vision sensors and cameras. It characterizes the image sensor and camera, so it helps compare the imager inside a 3D system, but it is not a measure of that system's depth accuracy. As engineering reasoning, the failure mode to avoid is reading a camera's EMVA data as if it described the 3D result (Hornberg 2017, Ch. 10, §10.10.1–§10.10.2, pp. 754–759; Geng 2011, Advances in Optics and Photonics 3-2, §1; Scharstein and Szeliski 2002, IJCV 47, §3; EMVA 1288 Release 4.0-2021).

How should an engineer decide between 2D and 3D vision for a handling task?​

No neutral source cited here states a single decision rule, so the list below is engineering reasoning, built on the trade-offs the cited sources describe rather than taken from them:

  • Parts rest on a known surface at a known height. A flat casting on a fixture plate, a board on a transfer deck, or a panel on a nest can usually be located in 2D, provided the height variation stays inside the camera's depth of field and the pick tolerance.
  • Height, tilt, or stacking varies. Layers on a pallet, parts leaning in a returnable container, or forgings of different thickness need 3D, because a 2D system calibrated at one plane reports a shifted position when the part sits above or below that plane.
  • The measurement itself is a height. Profile, flatness, gap and flush, bead height, and volume are 3D quantities by definition.
  • The part moves or stands still. Laser triangulation builds its image from motion, so it suits a part carried past a fixed sensor. Snapshot methods, such as a ToF camera or single-shot structured light, suit a part that is stopped or moving slowly.
  • The surface cooperates or not. Plain, uniform surfaces are a hard case for passive stereo, and dark or shiny surfaces can leave gaps in laser data.

Geng lists accuracy, resolution, and speed as the three aspects often used as the primary performance indexes to evaluate 3D imaging systems, and says the three can be used to compare 3D imaging systems. As engineering reasoning, the same three are a fair way to rank any of the four methods for a given part, as long as each supplier's figures are measured the same way, a point covered later in this article (Geng 2011, Advances in Optics and Photonics 3-2, §6; Scharstein and Szeliski 2002, IJCV 47, §3 and §6; Boridy, 3D Laser Triangulation primer; LMI 2018, 15185-5.1.6.79, Filters: Gap Filling).

How does structured light build a 3D image?​

A structured-light system projects a spatially varying 2D pattern onto the part and images the deformed pattern with a camera. Depth comes from triangulation between the projector, the camera, and each surface point. Geng's tutorial sorts the patterns into two families:

  • Sequential, or multi-shot, codes. These include binary and Gray codes, phase shifting, and hybrids that combine phase shifting with Gray coding. Several patterns are projected one after another, and each pixel's code is read across the sequence.
  • Single-shot codes. These include continuously varying patterns, stripe indexing, and grid indexing, which pack enough information into one pattern to decode it from a single image.

The choice between them is the speed and motion trade-off. Geng's guidance is that multi-shot techniques can be used when the part is static and acquisition time is not tightly constrained, and may often give more reliable and accurate results, while a moving part needs a single-shot technique. He also notes that when the object moves during sequential acquisition, the 3D shape can come out distorted. On a plant floor, that is a part on a conveyor, or a part swinging on a hook, that moves between patterns. Geng also states that camera and projector calibration play a critical role in establishing the measurement accuracy of 3D imaging systems. Two points of engineering reasoning follow from the method: the projector has to be calibrated along with the camera, and, for a plain, untextured part such as a machined aluminum plate or a painted panel, the projected pattern gives the camera features that the surface itself lacks (Geng 2011, Advances in Optics and Photonics 3-2, §1–§7).

How does laser triangulation measure height?​

Laser triangulation uses the geometry of a triangle formed by the laser emitter, the camera, and the target. The distance between the emitter and the camera, called the baseline, is known. Two angles are also known, and one of them depends on where the laser light lands on the imager. With the baseline and the two angles, the height of the illuminated point follows. The key definitions come from one sensor maker's user manual for its laser point sensors:

  • Clearance distance and measurement range. The sensor measures only inside a defined window that starts at a clearance distance from the sensor face; a target outside that window returns invalid data.
  • Z resolution. Height resolution is better near the sensor than at the far end of the range.
  • Z linearity. Linearity is quoted as a percentage of the measurement range, so the same percentage means a larger error on a longer-range sensor.

A line-profile sensor extends this from a spot to a sheet of laser light. The camera views the sheet at an angle out of its plane; the stripe on the part is reduced to a height profile, and successive profiles are combined into a point cloud. One vision-system maker's guide notes that its profile data is factory-calibrated. The technology primer from another supplier states the geometric consequence: the stripe on the sensor gives the lateral (X) and depth (Z) coordinates, and the third coordinate (Y) comes from scanning, so the part or the sensor must move. An illustrative figure, assumed rather than taken from any data sheet: a linearity of 0.1% on a 200 mm measurement range is 0.1% × 200 mm = ±0.2 mm (LMI 2018, 15185-5.1.6.79, pp. 47–48; Cognex 2021, Reference Guide rev. 1.0.1.7, Theory of Operation, pp. 11–13; Boridy, 3D Laser Triangulation primer).

What limits a laser line profiler on a moving line?​

The same triangle that makes the method work also creates limits:

  • Occlusion. Because the method views the laser at an angle, the triangulation geometry creates shadows, called occlusions, where the camera cannot see the stripe. One solution the primer gives is one or two lasers with two cameras. The point-sensor manual notes that single-point sensors are often mounted with the triangulation base perpendicular to the direction of travel to avoid occlusions, and that sensors should not be installed near objects that might block the camera's view of the laser.
  • Surface. Dark and specular (mirror-like) areas can return no data and leave gaps, which the manual's gap-filling filter fills from the nearest neighboring points. A gap-filled value is therefore an estimate, not a measurement, so the software should know which points were filled.
  • Speckle and speed. Laser speckle, the grainy interference pattern of coherent light, limits resolution, and the sensor itself can limit the scan speed.
  • Mounting and travel. The line-profile guide calls for a firm mount with the direction of travel perpendicular to the laser plane, so that point clouds are repeatable.
  • Laser class. The guide's sensor uses a blue 450 nm laser rated Class 2M. IEC 60825-1:2014 applies to the safety of laser products emitting laser radiation in the wavelength range 180 nm to 1 mm, and it classifies them according to their degree of optical radiation hazard, to aid hazard evaluation and the choice of user control measures.

For a board line in a lumber mill, a weld-seam check on a fabricated frame, or a gap check on an aerospace panel, these limits decide where the sensor goes and how many sensors are needed (Boridy, 3D Laser Triangulation primer; LMI 2018, 15185-5.1.6.79, Installation: Mounting and Filters: Gap Filling; Cognex 2021, Reference Guide rev. 1.0.1.7, Mounting and Laser Specifications; IEC 60825-1:2014).

How does stereo vision recover depth, and why do plain surfaces defeat it?​

Stereo vision finds depth by matching the same scene point in two calibrated views; the shift between the two image positions is the disparity, from which depth follows. Scharstein and Szeliski's taxonomy rests on the observation that stereo algorithms generally perform some subset of four steps:

  1. Matching cost computation.
  2. Cost aggregation, over a support region around each pixel.
  3. Disparity computation and optimization, which chooses the disparity for each pixel.
  4. Disparity refinement.

Their evaluation singles out typical problem areas, including textureless and occluded regions. Textureless regions need large amounts of aggregation, and shiftable windows, like all local methods, fail in textureless areas; occluded regions are points visible in only one image, where the authors assume algorithms generally do not produce meaningful results. On a plant floor, a plain machined surface, a uniformly painted part, or a clean sheet of plate is exactly that textureless case, and as engineering reasoning the failure mode to expect is missing or wrong depth across the plain area. Geng draws the contrast with structured light: conventional stereo must extract corresponding features from a pair of images, while his example of a projected color pattern provides easy-to-identify landmarks on each surface point. Designs that add a pattern projector to a stereo pair are not evaluated by any source cited here, so this article does not rate them (Scharstein and Szeliski 2002, IJCV 47, §3, §5 and §6; Geng 2011, Advances in Optics and Photonics 3-2, §1 and §3.1).

How does a time-of-flight camera measure distance?​

A ToF camera emits light and measures how long it takes to return, for every pixel at once. Horaud and co-authors separate two families:

  • Pulsed ToF measures the round-trip time of a light pulse directly. It works outdoors and at long range.
  • Continuous-wave (CW) ToF modulates the emitted light and measures the phase difference of the return. It is usually used indoors at short range, from centimeters to several meters.

CW ToF has a built-in ambiguity. Once the round trip exceeds one modulation period, the phase wraps around and a far target reads as a near one. The authors' example is a modulation frequency of 30 MHz, which gives an unambiguous range of 0 to 5 m. As an arithmetic check, for illustration only, the paper's maximum-depth relation dmax = c ÷ (2f) gives (3 × 10⁸ m/s) ÷ (2 × 30 × 10⁶ Hz) = 5 m. A higher frequency is more accurate but has a shorter unambiguous range. As engineering reasoning from that example, the failure mode for handling is a reflective object beyond the unambiguous range, such as a far wall or a steel column, that is reported inside the working zone.

Foix, Alenyà, and Torras's survey of lock-in ToF cameras sets the expectations. These cameras offer neither higher resolution nor a larger ambiguity-free range than other range sensors. Their strengths are registered depth and intensity images at a high frame rate, in a compact, light, low-power package. Horaud and co-authors put ToF spatial resolution 10 to 100 times below that of video cameras (Horaud et al. 2016, Machine Vision and Applications 27-7, §1, §4 and §8; Foix, Alenyà and Torras 2011, IEEE Sensors Journal 11-9).

What does a ToF data sheet say about range, resolution, and error?​

One sensor supplier's data sheet for a compact snapshot ToF camera shows how the general limits turn into numbers. The values below are that one product's typical figures, not a class-wide specification, and the sheet is marked subject to change:

  • Range. The working range is up to 16 m, with lower reliability from 9 to 16 m, where individual pixels may be wrong; the result also depends on how much infrared light the target reflects (its remission).
  • Resolution and rate. 512 × 424 pixels over a 70° × 60° field of view, at up to 30 frames per second.
  • Repeatability. About 0.8 mm at 1 m and about 5 mm at 7 m, so repeatability degrades with distance.
  • Conditions behind the numbers. The accuracy and repeatability values are typical for the central 80% of the field, at room temperature, with no ambient light, at 25 frames per second, and they vary with target remission. Accuracy may degrade by up to ±10 mm (typically ±5 mm) over the operating temperature range of −10 to +50 °C.
  • Light. The emitter is an 855 nm infrared laser, Class 1 under IEC 60825-1:2014, and the sheet rates ambient light immunity at up to 50 klx, for sunlight at a measuring distance of 2.0 m.

A worked estimate, for illustration only, shows what the pixel count means; the arithmetic is this article's, not a data-sheet figure. Assume the 70° angle spans the 512-pixel axis and the target is a flat surface facing the camera. The width seen at distance d is 2 × d × tan(35°). At 1 m that is 1.40 m, which matches the 1.4 m the data sheet's own field-of-view table gives at 1 m, or 1.40 m ÷ 512 = about 2.7 mm per pixel; at 7 m it is 9.80 m, or about 19 mm per pixel. A camera that repeats to about 5 mm at 7 m can still miss a feature smaller than about 19 mm there, which suggests, as engineering reasoning, that a ToF camera suits locating a large object in a wide area better than gauging a small feature (SICK 2026, V3S145-1AAAAAA data sheet, Features, Performance, Ambient data, and Detection volume and field of view; Horaud et al. 2016, Machine Vision and Applications 27-7, §8; IEC 60825-1:2014).

Why do ToF cameras struggle with motion, ambient light, and multiple units?​

Horaud and co-authors summarize the sources of error of CW ToF cameras, and add that interference between units is an important issue when several cameras operate at the same time. The error sources in the five points below are the authors' and the data sheet's; the plant-floor applications given with them are engineering reasoning:

  • Motion blur. CW ToF needs a relatively long integration time to raise its signal-to-noise ratio, and hence its depth accuracy, and that introduces motion blur when objects move. A part on a fast conveyor or a swinging load is the case to check.
  • Background illumination. Sunlight through a bay door or strong process lighting adds light the sensor did not emit. The data sheet above rates its camera to 50 klx; the figure to request from any supplier is the ambient level at which its stated accuracy still holds.
  • Temperature. Temperature is on the authors' list, and the data sheet above says its accuracy may degrade by up to ±10 mm over its ambient operating temperature range.
  • Scattering and multipath. Light scattering and multiple-path effects are on the authors' list of error sources. As engineering reasoning, inside corners, bins, and containers, where emitted light can reach the target by more than one route, are where to expect them.
  • Interference between cameras. The authors call interference between several cameras operating at the same time an important issue. The data sheet's automatic coexistence mode is one supplier's answer; the controls design still has to plan which cameras share a field of view.

The survey by Foix, Alenyà, and Torras also reviews the limitations of these cameras and the existing calibration methods, which, as engineering reasoning, is a reminder that a ToF camera out of the box is a starting point, not a measuring instrument (Horaud et al. 2016, Machine Vision and Applications 27-7, §4; Foix, Alenyà and Torras 2011, IEEE Sensors Journal 11-9; SICK 2026, V3S145-1AAAAAA data sheet, Ambient data).

How should 3D accuracy figures be compared between suppliers?​

Geng's tutorial makes the point directly: different manufacturers may characterize accuracy in different ways, such as a mean error, a root-mean-square (RMS) error, or a ± error band, so specifications have to be compared in the same framework. A supplier quoting an RMS figure and another quoting a ± band can describe similar sensors with numbers that look very different. The sources above add three more checks:

  • What range the figure refers to. A linearity quoted as a percentage of measurement range grows with the range.
  • What conditions the figure assumes. The ToF data sheet's values hold for the central 80% of the field, at room temperature, with no ambient light.
  • What surface the figure assumes. Remission and gloss change the result.

For acceptance testing there is a standard framework. ISO 10360-13:2021 specifies acceptance tests for verifying the performance of an optical 3D coordinate measuring system when measuring lengths as stated by the manufacturer, and reverification tests that let the user periodically reverify that performance. It is applicable to verifying measuring performance if the surface characteristics of the scanned object, such as glossiness and colour, are restricted and within a cooperative range, and it does not apply to other types of coordinate measuring systems. As engineering reasoning, for a handling application that means a supplier's figure proves little about a dark forging, a wet log, or a polished aluminum part until it has been tested on those parts. The ISO page lists the 2021 edition as under systematic review, so a purchase specification should confirm on that page whether a revised edition has replaced it before citing it (Geng 2011, Advances in Optics and Photonics 3-2, §6; ISO 10360-13:2021; LMI 2018, 15185-5.1.6.79, Resolution and Accuracy; SICK 2026, V3S145-1AAAAAA data sheet, Performance).

Where does a 3D sensor sit in the build chain, and what does the mechanical design decide?​

The sensor is part of the controls and monitoring links, but most of its performance is decided upstream:

  • Design and engineering set the working distance so the part always lands inside the sensor's measurement range. A conveyor, a car deck, or a fixture that presents parts outside the clearance-distance window returns invalid data no matter how good the sensor is. As engineering reasoning, because Z resolution is better near the sensor, the design should also put the most important features toward the near end of the range.
  • Parts machining sets the repeatability of the mounts, locating pins, and nests that hold the sensor and present the part.
  • Fabrication, weld fatigue, and stress relief decide whether the frame that carries the sensor stays in position. The line-profile guide asks for a firm mount so point clouds repeat; a bracket on a frame that relaxes after welding, or one that vibrates with a nearby drive, moves the measurement.
  • Assembly fixes the orientation. The point-sensor manual notes that sensors are often mounted with the triangulation base perpendicular to travel to avoid occlusions, and the line-profile guide calls for travel perpendicular to the laser plane; as engineering reasoning, both are assembly checks, not software settings.

UTEC Industrial machines sensor mounts, nests, and fixture interfaces to tolerances as tight as ±0.001 in, the mechanical base a 3D measurement rests on. For tolerances on those interfaces, the guide to what machining tolerances to specify covers how to set them on a drawing (LMI 2018, 15185-5.1.6.79, Clearance Distance and Measurement Range, Resolution and Accuracy, and Installation: Mounting; Cognex 2021, Reference Guide rev. 1.0.1.7, Mounting).

What controls and sensing does a 3D vision station need?​

A 3D sensor is only useful when the controls tell it when to measure and put its result to use. The pieces are:

  • Encoders. One line profiler's guide says the system typically acquires images on encoder signals, so acquisition follows the speed of the moving part rather than a fixed timing, and that many applications use a rotary encoder on the conveyor. Engineering reasoning from that principle: if the encoder slips on its shaft or wheel, the profiles no longer match the distance traveled, and the point cloud is stretched or compressed along the direction of travel.
  • Triggers and part-present sensing. A photoelectric sensor or a PLC output starts a snapshot camera or a scan window when the part is in place.
  • Coexistence. Where several ToF cameras share a field of view, the controls schedule them or rely on a coexistence mode, because Horaud and co-authors call interference between several cameras operating at the same time an important issue.
  • PLC and robot link. The result moves to the controller through an event-driven, handshaked exchange. The category overview covers the PLC-side handshake, and the article on connecting a FANUC robot to an Allen-Bradley PLC over EtherNet/IP covers the robot connection. On ControlLogix 5580 and CompactLogix 5380 controllers, Rockwell Automation's controller design manual lists a consumed tag and a change in module input data among the triggers for an event task, so, as engineering reasoning, a new 3D result can start the next step without waiting for a continuous-task scan; the trigger list differs by controller family, as the category overview sets out.
  • Safety. A 3D camera that locates parts is not a safety device, a point the category overview covers. Where the station feeds a robot, the cell falls under ISO 10218-2:2025 for industrial robot applications and robot cells.

UTEC Industrial, a Rockwell Automation Recognized System Integrator, programs the Allen-Bradley ControlLogix and CompactLogix controllers, EtherNet/IP networks, and VFD and servo drives that sequence these stations (Cognex 2021, Reference Guide rev. 1.0.1.7, Encoder Specifications; Horaud et al. 2016, Machine Vision and Applications 27-7, §4; Rockwell Automation 1756-RM094N-EN-P-2025, Ch. 5 p. 43 and Ch. 4 p. 30; ISO 10218-2:2025).

What should a specification for a 3D vision station define?​

The category overview gives a general vision specification checklist; a 3D station needs these additions so that suppliers quote comparable systems:

  • Method and reason. Which of the four methods is proposed, and why it suits the part's motion and surface.
  • Working envelope. The clearance distance, the measurement range, the field of view, and the part's largest and smallest positions inside them.
  • Accuracy statement. The measure used (mean, RMS, or ±), the conditions behind it, the surface it was measured on, and whether it will be demonstrated on the customer's own parts.
  • Environment. Ambient light at the station, the temperature range, and dust or moisture, stated as values, since the ToF data sheet shows accuracy degrading over its temperature range and states its accuracy values without ambient light.
  • Laser class. The IEC 60825-1:2014 class of every emitter, so labeling and precautions are designed in.
  • Multiple sensors. How many sensors cover occluded areas, and how they avoid interfering with one another.
  • Acceptance. A test on production parts across their real range of color, finish, and position. Where surfaces are cooperative, ISO 10360-13:2021 gives the acceptance and reverification framework.

The failure mode this list prevents is a sensor that meets its data-sheet figure on a matte test block and then fails on the plant's darkest or shiniest part (Geng 2011, Advances in Optics and Photonics 3-2, §6; ISO 10360-13:2021; IEC 60825-1:2014; SICK 2026, V3S145-1AAAAAA data sheet, Performance and Ambient data; Horaud et al. 2016, Machine Vision and Applications 27-7, §4).

Related Articles

References​

  • Hornberg, A. (Ed.). Handbook of Machine and Computer Vision: The Guide for Developers and Users, 2nd ed. Wiley-VCH, 2017. ISBN 9783527413393
  • EMVA 1288 Release 4.0-2021: Standard for Measurement and Presentation of Specifications for Machine Vision Sensors and Cameras. European Machine Vision Association, 2021.
  • Geng, J. (2011). "Structured-light 3D surface imaging: a tutorial." Advances in Optics and Photonics, 3(2), 128-160. DOI 10.1364/AOP.3.000128
  • Scharstein, D., & Szeliski, R. (2002). "A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms." International Journal of Computer Vision, 47(1-3), 7-42. DOI 10.1023/A:1014573219977
  • LMI 15185-5.1.6.79: Gocator Point Profile Sensors User Manual, Gocator 1300 Series, rev. A. LMI Technologies Inc., 2018.
  • Cognex In-Sight 3D-L4000 Reference Guide, rev. 1.0.1.7: In-Sight 3D-L4000 Series Vision System Reference Guide. Cognex Corporation, 2021.
  • Boridy, R. 3D Laser Triangulation, technology primer. Teledyne DALSA (undated web documentation, accessed September 2026).
  • IEC 60825-1:2014: Safety of laser products — Part 1: Equipment classification and requirements. International Electrotechnical Commission, 2014.
  • Horaud, R., Hansard, M., Evangelidis, G., & Ménier, C. (2016). "An overview of depth cameras and range scanners based on time-of-flight technologies." Machine Vision and Applications, 27(7), 1005-1020. DOI 10.1007/s00138-016-0784-4
  • Foix, S., Alenyà, G., & Torras, C. (2011). "Lock-in Time-of-Flight (ToF) Cameras: A Survey." IEEE Sensors Journal, 11(9), 1917-1926. DOI 10.1109/JSEN.2010.2101060
  • SICK V3S145-1AAAAAA: Visionary-T Mini 3D Machine Vision Data Sheet. SICK AG, 2026.
  • ISO 10360-13:2021: Geometrical product specifications (GPS) — Acceptance and reverification tests for coordinate measuring systems (CMS) — Part 13: Optical 3D CMS. International Organization for Standardization, 2021.
  • Rockwell Automation 1756-RM094N-EN-P-2025: Logix 5000 Controllers Design Considerations. Rockwell Automation, 2025.
  • ISO 10218-2:2025: Robotics — Safety requirements — Part 2: Industrial robot applications and robot cells. ISO, 2025.

Ready to Discuss a Material Handling System?​

UTEC Industrial designs, engineers, machines, fabricates, and installs custom material handling systems for heavy industry, from the stress-relieved structure and drives to the Allen-Bradley PLC controls, tuning, and monitoring that run them, at its Spokane Valley, WA facility. Send UTEC the application, loads, and duty cycle to start a system review.

Request a Quote →

Questions? Call (509) 922-1832 or email sales@utec.co

Last updated on