A camera that works on a bench proves very little. The same camera has to work at the end of a weld cell at three in the morning, with fume on its window, a part that arrived hot and less than a minute before the next one. This guide sets out what each automotive inspection needs, the numbers that decide the choice and the places where a technology is the wrong answer. Use it to write a specification. Then test that specification on your own parts before you buy anything.

1. Why automotive is one of the hardest vision environments

Everything on an automotive line is paced by takt time. A plant building around sixty vehicles an hour has roughly a minute per vehicle at each station. A component line stamping brackets or assembling connectors may have a second or less per part. The inspection has to trigger, expose, process and report inside whatever is left once the part has stopped moving or, more often, while it still is. A vision system that needs the line to slow down has failed its first requirement.

The line also runs around the clock, which changes what reliable means. A check that is right 99 percent of the time sounds respectable until you run the numbers. At 1,000 vehicles a day with 50 checks on each, 1 percent wrong is 500 wrong calls a day. Some are false rejects that stop the line or fill the repair bay. The rest are escapes.

Then there are the parts themselves. Clear coat and chrome reflect like mirrors. Galvanised and machined surfaces throw glare in one direction and nothing in another. Parts leave a weld cell or a press-hardening line hot enough to glow in the near infrared. Around them sit weld fume, spatter, oil mist, vibration and in some areas wash-down. None of this appears in a demonstration.

IATF 16949 turns the quality expectation into obligations. Parts have to be traceable. Measurement systems named in the control plan have to be proven capable through measurement systems analysis. Error-proofing devices have to be challenged on a schedule to show they still catch what they were installed to catch. Customers count defects in parts per million, which in practice means zero. The reason is cost. A defect caught at the station costs a rework. The same defect found at the customer's plant costs a sort, a chargeback and a difficult meeting. Found in the field it becomes a recall.

The consequence for specification is simple. Start from the defect, the tolerance, the takt and the environment. Choose the camera last.

2. Weld inspection

Judge the geometry first

A weld is a three-dimensional object. Its quality is mostly a question of shape. Is the bead tall enough and wide enough? Is the throat there? Has the seam wandered off the joint? Is there undercut at the toe or a burn-through? A 2D image can show that something is on the joint. It cannot tell you how much. For fusion welds the acceptance limits usually trace back to the ISO 5817 quality levels or the customer's own weld standard. Almost every limit in those documents is a dimension.

3D laser profiling, microns and profiles per second

A laser profiler projects a line across the weld and views it from an angle. The line bends with the surface and each image becomes one cross-section in real units. Movement along the seam stacks those cross-sections into a height map. Three numbers decide whether it can do your job. Height resolution has to be a small fraction of the tightest height tolerance. Point spacing across the seam has to put several points on the smallest feature. Profile rate has to do the same along it.

Take a profiler on a robot travelling along a seam at 200 mm per second. At 5,000 profiles per second it lays down a profile every 0.04 mm. A 0.5 mm pore is crossed by a dozen profiles, which is plenty. Drop to 500 profiles per second and the spacing is 0.4 mm. The same pore might be caught once or missed entirely. Across the seam the sum is field of view divided by points per profile. A 60 mm field over roughly 2,000 points gives 0.03 mm between points.

Zebra's AltiZ uses two cameras on a single laser line. That matters on welds because a bead shadows its own toe from one side. It delivers 1,984 points per profile at up to 11,000 profiles per second. The AltiZ 4200 raises that to 4,224 points and up to 14,000 profiles per second for small fields and fine detail. Teledyne's Z-Trak family covers the same ground with a different spread. Z-Trak Express holds 5,000 profiles per second across its full measurement range. Z-Trak2 reaches 45,000. Z-Trak LP2C gives 4,096 points per profile where width and detail are both needed. Treat every headline rate with care. It is quoted with a short exposure and usually a reduced measurement window. Dark, burnt metal needs a longer exposure. Budget below the headline figure and confirm it on a real seam.

Bright metal beside dark burn

A fresh weld puts mirror-bright spatter, matt oxide and soot-black heat tint inside the same laser line. Expose for the dark areas and the bright ones saturate, which smears the line and corrupts the height. Expose for the bright areas and the dark ones vanish. This is the high-dynamic-range problem. It is the main reason weld profiling disappoints in production after a good trial on clean samples.

The better sensors deal with it inside the head. Z-Trak2 has single-scan HDR and reflection compensation. Z-Trak Express has dual-point HDR. Laser colour helps too. A blue line gives a cleaner return from bright metal than a red one. Paired with a narrow bandpass filter it also ignores the red and infrared glow of a part that is still hot. Both the AltiZ and Z-Trak ranges come in blue and red versions.

Porosity and spatter

Porosity and spatter are surface faults that split between 3D and 2D. An open pore is a small deep hole. A profiler finds it when the point spacing is a third of its diameter or better. Spatter is a small raised bump away from the bead and is easy to see in a height map. Discolouration, soot and a missing weld are quicker to find in an intensity image. Z-Trak2 outputs reflectance alongside height, which lets one sensor serve both checks.

Thermal in-process monitoring

The checks above judge the finished bead. Thermal imaging watches the process. A long-wave infrared camera such as Teledyne's Calibir GX sees the heat left in the joint just behind the torch or in a spot weld as it cools. A cold start, a gap in the seam or a spot that never formed a full nugget shows up as a heat signature that is the wrong shape or fades too fast. Calibir GX is radiometric, comes in 640 x 480 or 320 x 240 and resolves temperature differences below 50 mK across a span of more than 600°C.

Be clear about what it is not. An uncooled microbolometer runs at tens of frames a second. It will not resolve melt pool dynamics, which need high-speed cameras with specialist filtering. Emissivity also shifts as steel oxidises, which makes absolute temperature on a weld unreliable without careful set-up. Treat thermal as a signature to compare against a known good weld. It is not a thermometer.

Where deep learning earns its place

Measure with rules. Judge appearance with deep learning. Bead height, width and position are dimensions. A rules-based tool measures them and gives a number you can put through a gauge study. What rules handle badly is appearance. Is that mark spatter, a pore or harmless scale? Is that heat tint acceptable? Zebra's Aurora deep learning tools add classification, segmentation, object detection and anomaly detection inside Aurora Design Assistant. Anomaly detection trains on good examples only, which suits a plant with thousands of good welds and very few bad ones. Deep learning is the wrong tool for the measurement itself. A network will tell you a bead looks short. It will not give you a traceable millimetre.

Spot, seam and arc welds are different problems

Resistance spot welds hide their quality. The nugget sits between the sheets where no camera sees it. Vision can confirm that every spot is present, in position, clear of the edge and free of expulsion. It can measure indentation in 3D. It cannot measure nugget diameter. Ultrasonic testing and teardown keep that job.

Laser welded and laser brazed seams are narrow, smooth and often on visible surfaces such as roof joints. The defects are small. Pores of a few tenths of a millimetre matter. This is the high-resolution end, with tight point spacing, a short field of view and a high profile rate.

Arc welds on chassis, subframes and seat structures are bigger and rougher. The checks are throat, leg length, convexity, undercut and burn-through. Resolution is easier. Access, spatter and fume are harder.

3. Surface and paint defect detection

Specular versus diffuse

A diffuse surface scatters light in all directions. Point a camera at it and you see the surface. A specular surface reflects like a mirror. Point a camera at it and you see whatever the surface is reflecting, including your own light. Primer, e-coat and unpainted castings are mostly diffuse. Clear coat, chrome, polished trim and glass are specular. Galvanised and bare pressed steel sit in between. The inspection method follows from which one you have.

Deflectometry

Deflectometry takes the mirror problem and uses it. A screen displays a known pattern, usually sinusoidal fringes. The camera looks at the reflection of that pattern in the surface, which means the surface itself becomes part of the optics. Where it is smooth the fringes return undistorted. Where there is a dent, a ding, a sink mark or a dirt inclusion under the clear coat the local slope changes and the fringes bend. Shift the pattern through several phases and the software recovers a slope map of the whole field.

The point to hold on to is that deflectometry measures slope, not height. A shallow dent tens of microns deep and several millimetres across barely exists as a height. As a change in slope it is obvious. That is why the method finds defects a trained inspector sees only by rocking the panel under strip lights.

Its limits are just as clear. It needs a specular surface and does nothing on primer. The screen has to be larger than the area inspected, more so on curved panels. Phase shifting needs several images, which means the part is stationary or the system is designed around motion from the start. A full-body paint tunnel is a specialist turnkey installation. Deflectometry on panels, trim, lamp lenses and plated parts is a components job that can be proven on a bench first.

Line scan versus area scan

Area scan suits discrete parts that stop or can be frozen with a strobe. Line scan suits anything continuous or anything large that moves at a known speed, such as coil, blanks, a body on a skid or a rotating cylinder. It builds the image one line at a time. That gives unlimited length, even lighting from a single line light and very high resolution across the width. It demands an encoder and steady motion. Teledyne's Linea range runs from 2k to 16k pixels at line rates up to 80 kHz. If the part indexes and stops, area scan is simpler. Do not force line scan onto a stop-start process.

From defect size to resolution

Start from the smallest defect you must find and allow at least three pixels across it for a rules-based tool. Deep learning likes more. Suppose the smallest defect is 0.3 mm and the panel is 1,600 mm wide. Three pixels across 0.3 mm is 0.1 mm per pixel. 1,600 mm at 0.1 mm per pixel is 16,000 pixels. That is one 16k line scan camera or two 8k cameras with an overlap.

In the direction of travel the line rate is speed divided by pixel size. A body on a skid at 100 mm per second needs 1,000 lines per second, which is trivial. Coil at 5 metres per second needs 50,000 lines per second. At that rate the camera interface and the light intensity start to decide the design. The same sum on area scan shows why large panels are hard. A 12 megapixel camera at 0.1 mm per pixel covers about 400 by 300 mm. A bonnet needs a dozen or more of them, or a camera on a robot.

Lighting geometry

Lighting geometry decides what the camera can see before any software runs. Bright field puts the light where its reflection reaches the camera. A flat surface reads bright. Anything that tilts or scatters reads dark. On diffuse surfaces it is the default. On specular surfaces it produces glare unless the source is large and even.

Dark field brings the light in at a low angle. The surface stays dark and only edges, scratches, pits and raised dirt light up. It is the best way to find scratches and inclusions on gloss. It also lifts dot-peen and laser marks out of a machined face.

A dome surrounds the part with soft light from every direction. It removes hot spots on curved and shiny parts and suppresses texture. That makes it excellent for reading and presence checks on chrome or foil. It is poor for finding shallow dents because it hides the very shading that reveals them.

Coaxial light travels down the lens axis through a beam splitter. Flat specular areas return it and read bright. Anything tilted, engraved or scratched reads dark. It works only on flat areas at short range.

Advanced Illumination makes all four geometries plus the line lights that line scan needs. Our illumination selector tool narrows the choice in six questions. The background is in our note on why machine vision lighting is so often wrong.

The false-reject problem on paint

Paint is where inspection systems get switched off. Every body carries dust that will blow off, water marks, tape residue, polish swirls and orange peel that is within specification. A rules-based system sees contrast and calls all of it. If it flags one body in ten for marks that are not defects, the repair team stops trusting it.

This is the real case for deep learning on surfaces. The optics find candidates with high sensitivity. A trained classifier then separates real defects from nuisance. Zebra's Aurora tools do this with classification and segmentation models trained on your own images. Two cautions apply. The classifier is only as good as its labelled examples, which means your quality team must agree what a defect is before training starts. No software recovers a defect that the lighting never made visible. Get the image right first.

4. Dimensional measurement

Why perspective error matters

An ordinary lens sees like an eye. Nearer things look bigger. For inspection that is harmless. For measurement it is an error source. Take a camera 300 mm from a part. If the part sits 1 mm higher on its fixture the magnification changes by one part in 300. On a 50 mm feature that is 0.17 mm of apparent size change with nothing wrong with the part. Against a tolerance of plus or minus 0.1 mm the measurement is already useless. The same effect makes the top and bottom edges of a hole in thick material appear as two circles.

A telecentric lens removes it. It accepts only rays parallel to the optical axis, which holds magnification constant through its depth of field. A feature measures the same wherever it sits in that depth. Computar's TEC series and Kowa's telecentric range cover the common sensor formats. Kowa's LM1119TC resolves 21 megapixels on sensors up to 4/3 inch. Computar's LensConnect telecentrics add motorised focus for stations nobody can reach. Pair a telecentric lens with a collimated backlight and the part edge becomes a clean step that sub-pixel tools can locate properly. Our lens calculator helps with the first pass.

The catch is physical. A telecentric lens cannot see a field wider than its own front element. C-mount telecentrics cover fields of tens of millimetres. Beyond that the lenses become large, heavy and expensive. For a 400 mm bracket a telecentric lens is the wrong answer. Use several calibrated cameras looking only at the features that matter, or move to 3D.

Calibration against a traceable standard

A camera measures in pixels. Calibration converts pixels to millimetres and corrects lens distortion by imaging a target whose feature positions are known. For the result to mean anything in an audit, that target needs a calibration certificate traceable to a national standard. The master part used for daily verification needs one too. Record temperature as well. Aluminium grows about 23 microns per metre per degree. A 500 mm casting measured 10 degrees warmer than the reference is 0.12 mm longer. Laser profilers such as Z-Trak arrive factory calibrated in real units, which removes a step. They still need verifying against a traceable artefact in the station.

2D versus 3D, gap and flush

2D measures in a plane. Hole position, pitch, diameter, outline and length are all 2D jobs. If the feature lies in one plane and can be backlit, 2D with telecentric optics is the most precise route and the least expensive. 3D measures height, form and position out of the plane.

Gap and flush is the classic automotive case. Gap is the distance between two panel edges. Flush is the height step between them. A laser profiler across the joint gives both from one profile. The difficulty is the edges. They are radiused and painted. The gap is defined between tangent points on two curves, which the software has to fit the same way every time. Dark gloss paint returns very little laser light. Clear coat can return a second reflection from beneath the surface. Test on your darkest and your lightest colour before accepting any gap and flush system.

The rule worth holding to is that the method must match the tolerance. A height tolerance cannot be proven by a 2D image.

Gauge capability and sub-pixel limits

Resolution is not accuracy. The question a quality engineer has to answer is whether the measurement system is capable, meaning its own variation is small against the tolerance. The usual measurement systems analysis threshold is that repeatability and reproducibility consume under 10 percent of the tolerance band. Between 10 and 30 percent needs justification.

Work it through. A 5 megapixel camera has about 2,450 pixels across. On a 50 mm field that is roughly 20 microns per pixel. Good edge tools on a backlit telecentric image repeat to around a tenth of a pixel. Here that is 2 microns at one sigma or 12 microns across six sigma. Against a tolerance of plus or minus 0.1 mm, a band of 200 microns, that is 6 percent. Capable. Widen the field to 200 mm with the same camera and each pixel is 82 microns. Six sigma becomes 49 microns, about 25 percent of the band. That is not capable without a higher resolution camera or a smaller field.

A tenth of a pixel is also a best case. It assumes a sharp high-contrast edge, stable lighting, a rigid mount and no thermal drift. On a front-lit machined edge with burrs, a quarter to a half pixel is nearer the truth. Claims of a fiftieth of a pixel belong in a laboratory.

5. Assembly, error-proofing and traceability

Vision sensor, smart camera or PC-based

Three tiers of hardware cover almost every assembly check. The common mistake is buying one tier too high or too low.

A vision sensor does one or two simple jobs such as presence, position or a code, with a built-in lens and light and a guided set-up. Zebra's VS20 is the example in our range. It is the right tool for the clip, the cap or the label at a single station.

A smart camera carries a fuller tool set and its own processor in one sealed body. Zebra's VS40 and VS70 run Aurora Focus. The Iris GTX runs Aurora Design Assistant on sensors from 2 to 16 megapixels in an IP67 housing with deep learning inference on the camera. Use one where there are several checks in one view or the logic is more than pass or fail.

A PC-based system puts several cameras on one controller, such as Zebra's 4Sight EV7 with Aurora Design Assistant. It is the answer when a station needs four views of a seat or a cockpit, when 2D and 3D have to be combined or when the deep learning models are large. It is overkill for one clip.

Deep learning classification

Template matching works when the correct part always looks the same. Wiring looms, foam, fabric, sealant and connectors with flying leads never do. Deep learning classification copes with that variation. It decides between seated and not seated, or right variant and wrong variant, from examples. Object detection counts and locates clips or fasteners that can sit anywhere in the view. Keep the question narrow. One model that answers one question at one station is easy to validate. One model asked to judge a whole door card is not.

Poka-yoke on the station

Poka-yoke means the station cannot pass a wrong part. The vision result has to be interlocked. The fixture does not release, the tool does not enable or the conveyor does not index until the check passes. A camera that only lights a red lamp is an advisory system. Decide as well what happens when the system cannot decide, for example with a dirty lens or a missing trigger. The safe default is fail.

Traceability and direct part marking

Traceability is usually the reason the camera is there. Every safety-relevant part carries an identity that follows it from the supplier to the vehicle record. On labels and cartons that is a printed 1D or 2D code. On the part itself it is a direct part mark, almost always a Data Matrix.

Dot-peen marks are indented dots with no colour contrast at all. They are read by shape, with low-angle light throwing a shadow in each dot. On castings the surface texture competes with the dots. Laser marks are finer and can be high contrast on anodised or coated surfaces. On bare steel and aluminium they are often an annealed or etched mark that reads well at one lighting angle and vanishes at another. Printed codes are the easy case until the label is curved, glossy or under film.

For all of them, size the optics from the module. A sound working rule is around five pixels per module for a reliable read. A 0.25 mm module then needs 0.05 mm per pixel, which a 2.3 megapixel reader holds over a field of about 95 mm. Zebra's FS range of fixed industrial scanners covers this from the compact FS10 and FS20 to the FS40, FS42 and FS70. The FS20, FS40 and FS70 can be upgraded to machine vision by licence. A station that starts as a reader can take on inspection later.

OCR and OCV

OCR reads characters it has not been told to expect. OCV checks that an expected string is present and legible. A VIN stamped into a body, a date code on a tyre and a part number cast into a housing are all candidates. Stamped and cast characters are low contrast and inconsistent, which defeated older font-trained tools. Zebra's deep learning OCR reads them without font training. Where the string is known in advance, compare it against the build record. That turns a read into a verification.

Grading to the ISO/IEC standards

Reading a code and grading it are different jobs. A read tells you this reader decoded it today. A grade tells you how much margin the mark has, which predicts whether your customer's reader will decode it after a year in service. The standards are ISO/IEC 15416 for linear codes, ISO/IEC 15415 for printed 2D codes and ISO/IEC 29158 for direct part marks. ISO/IEC 29158 modifies the printed-code method with lighting and measurements suited to marks made in the surface itself. Its current edition was published in 2025.

Formal verification needs calibrated equipment with the lighting the standard defines. An in-line reader that reports quality metrics is process monitoring. It is very useful for catching a worn stylus or a drifting laser before reads start to fail. It is not a certificate. If your customer specification demands a grade, check which of the two the wording requires.

Feeding the result to the MES

A read that is not recorded did not happen. The station should send the decoded identity, the result of each check, the measured values, a timestamp and ideally the image to the MES against that serial number. Agree the data model early. Decide who owns the image store, how long images are kept and what the station does if the MES is unreachable. Buffering locally and blocking the line are both legitimate answers. Losing the record is not.

6. Making it work on the line

PLC, robot and MES integration

Integration is where good image processing fails. The PLC handshake needs ready, trigger, busy and result valid at a minimum, plus the result itself. Tie every result to its part with an ID or a counter. A late answer must never be applied to the next part. Check that the device supports the fieldbus your plant standard requires before ordering.

Where a camera or profiler rides on a robot, three things need settling. Hand-eye calibration ties the sensor's coordinates to the robot's. Path speed must be steady through the scan or the profiler must be triggered from position. Cables in the dress pack must be rated for the motion.

Throughput and takt budgeting

Write the time budget down. Suppose a station has 1.5 seconds between part arrival and release. As an illustration, trigger to exposure takes a few milliseconds. Image transfer takes perhaps 30. Processing might take 80 for rules-based tools and several hundred for a large deep learning model on modest hardware. The PLC scan and handshake add 20 to 40. Leave half the window as margin for retries and for the day somebody adds a check.

Motion sets the exposure. A part moving at 500 mm per second imaged at 0.1 mm per pixel blurs by one pixel in 0.2 milliseconds. An exposure that short needs a strobed light with the output to match.

Environment and ingress protection

Ingress rating is the first filter. IP67 means dust tight and protected against temporary immersion. It does not mean protected against high-pressure hot wash-down, which is IP69K. AltiZ, Z-Trak and Iris GTX are IP67. Advanced Illumination's wash-down lights and Kowa's sealed JCM-WP lenses extend protection to the optics. An IP67 camera behind an unsealed lens is not an IP67 station.

Weld cells need more. Spatter pits glass. Fit a sacrificial window that maintenance can change in a minute and an air knife to keep fume off it. Monitor image brightness, which lets the system report a dirty window before it starts failing good parts. Mount cameras off the vibrating structure where possible. Check operating temperature too. Many smart cameras and profilers are rated to 45°C ambient. A housing beside a press-hardening line will exceed that without cooling.

Validation and acceptance

Agree the acceptance criteria in numbers before the purchase order. For a measurement that is a gauge repeatability and reproducibility study on production parts, with the result as a percentage of tolerance. For a defect check it is an attribute study on a defined set of good, bad and borderline parts, with the allowed escape rate and false-reject rate stated. Build that sample set early and keep it. It becomes your master set.

Then run the system at line rate for long enough to see real variation across shifts, material batches, colours and tool changes. A factory acceptance test on twenty hand-fed parts shows that the software works. It does not show that the station does.

After handover, challenge the system on a schedule with known bad parts and record the result. Log an image of every reject and a sample of passes. Put lighting intensity, focus and exposure under change control. Treat a retrained deep learning model like any other process change and revalidate it against the master set before it goes live.

7. Selection table

Each inspection is mapped to its core technology, the specification drivers that decide the choice and the matching products from our range. Treat it as a starting point. Two rows often share a station. One row sometimes needs two technologies.

8. Proving it and next steps

Everything above can be worked out on paper. None of it should be bought on paper. A weld bead, a metallic black panel and a dot-peen mark on a casting behave nothing like a test chart. The only reliable evidence is your own parts under real optics.

That is what the Clearview Insights Lab is for. Send us good parts, bad parts and the borderline ones that cause arguments. We set up the candidate technology, capture the data and show you what it can and cannot separate, with the images and numbers behind it. If the first approach does not work we will say so and try the next. If nothing works at a sensible cost we will say that too. It is a far cheaper answer before a purchase order than after one.

We are engineering-led and selective. We carry the 3D, line scan, thermal, optics, lighting and software that automotive inspection calls for, from a short list of manufacturers we know in depth, not one of everything. Our engineers are based in the UK, Spain, Germany and France. They can stand at the station with you, which is where most of the real constraints are found.

For the wider picture, see our guide to machine vision cameras and our applications by industry.