Vision guided picking of food products on a conveyor

Picking single food products off a moving belt — crisps, wafers, biscuits, sweets — is a timing problem before it is a robot problem. The camera has to freeze a light, fast part in well under a millisecond of exposure, hand over coordinates while the part is still inside the robot's reach, and give the gripper something it can lift without crushing. Miss any of the three and you photograph smears, chase parts you cannot reach, or scrape product off the belt. This guide works through the three decisions with the arithmetic: exposure from pixel size and belt speed, field of view and focal length from the belt width, and the trigger-to-pick window from your own cycle times. It closes with a worked cell for a 300 mm belt running 300 mm/s and a checklist for ordering.

The three decisions that set the design

A random-pick cell stands or falls on three numbers, and all three can be calculated before anything is ordered.

1. Exposure. Blur stays under one pixel when the exposure is shorter than the pixel size divided by the belt speed. At 0.135 mm per pixel and 300 mm/s that allows 0.45 ms, which means a strobed light and a global shutter — ambient light and auto exposure do not reach there.
2. Field of view. The camera must see the belt width plus margin: a 352 mm field over a 300 mm belt, which at 400 mm working distance takes a 10 mm lens on a 1/1.45"-format sensor.
3. Pick window. The part keeps moving between trigger and finished lift. An 800 ms vision-plus-robot chain at 300 mm/s consumes 240 mm of belt, so at least 350 mm of the belt must lie inside the robot's reachable band.

Singulated snack lines usually run 200-400 mm/s at the pick point, and 300 mm/s is a sound design point to carry through the rest of this guide. Each number is derived below.

Freeze the motion: exposure and the strobe

The pixel size comes from the field and the sensor, not from a datasheet guess. A 5 MP sensor in a 4:3 shape puts about 2,600 pixels across the frame, so a 352 mm field resolves 352 / 2,600 = 0.135 mm per pixel. A belt at 300 mm/s covers 0.3 mm per millisecond, so one pixel of blur allows 0.135 / 0.3 = 0.45 ms. Set the exposure to 0.4 ms: a part then smears less than a pixel even at full belt speed, which keeps centroid and edge location accurate. For half-pixel sharpness on print inspection, halve the budget to 0.22 ms.

Exposure costs light, and the bill is steep: 0.4 ms is a 125-times cut against a 50 ms exposure, about seven stops, and no continuous LED bar repays that at a sensible aperture. Pick cells therefore strobe. Drive the bar with pulse current for 0.3-0.4 ms, timed so the pulse sits entirely inside the exposure; the duty cycle stays under 1 percent (0.4 ms every 50 ms), which is what lets the LEDs run well above their continuous brightness for that instant. The result is a frame as bright as a long exposure, with the hall lights frozen out of the image. If frames still come out dark or banded, the two usual suspects are auto exposure fighting the strobe — capped in capping maximum auto exposure time — and mains flicker, which black bars in some camera images separates from exposure error.

Check the shutter type before ordering: a short strobe on a rolling-shutter sensor lights the frame row by row and splits it into bright and dark bands, while a global-shutter sensor catches the pulse in one instant. Order global shutter for any strobed, moving scene. The full blur arithmetic, including how a different blur tolerance moves the limit, is in the maximum exposure time calculation.

One eye-safety note: a bar driven in high-current pulses a few hundred millimetres from the lens is intensely bright. Mount it over the belt, shield it from eye level, and never look into it at working distance.

Field of view, focal length and depth of field

Size the optics from the belt. A 300 mm belt needs margin for overhanging edges and calibration drift: call the target 352 mm across. Put the camera 400 mm above the belt — clearance for the product and a washdown shroud, with a lens that stays ordinary.

A 1/1.45"-format sensor reads about 8.8 mm across, so the magnification is 0.025 and the focal length works out to 9.8 mm — a 10 mm lens. Check it the other way: 8.8 mm × 400 mm / 10 mm = 352 mm of field, and along the belt the same optics see 6.6 mm × 400 mm / 10 mm = 264 mm. Run your own belt and working distance through the lens focal length calculator:

Depth of field sets the aperture. Wafers lie flat but stack and tumble, so leave the focus plane room on both sides:

At f/2.8 the total depth of field is about 31 mm — roughly 15 mm either side of the set plane. That absorbs a part standing 10 mm proud of the belt, or two wafers stacked. Stopping to f/4 opens the window to about 44 mm and costs light the strobe must repay; the depth of field calculator runs your own aperture and magnification:

Resolution then follows from the field: 2,600 pixels over 352 mm give the 0.135 mm per pixel used above, so a 60 mm wafer spans 440 pixels — ample for a centre within a millimetre plus edge checks. The trade between pixels, belt speed and blur for a sorting belt is worked through in sorting conveyor camera selection.

A 5 MP colour vision sensor with a 1/1.45" global-shutter sensor, M12 mount and Fast Ethernet link matches the example optics directly. Dropping to 1.6 MP nearly doubles the pixel to about 0.24 mm and stretches the blur budget to 0.8 ms at 300 mm/s — exposure headroom in exchange for the pixels you would spend on print or small defects.

Lighting: one diffused, strobed bar over the belt

The light does two jobs at once: it supplies the photons for a 0.4 ms exposure and flattens the glare that glazed, oily or foil-wrapped snacks throw back. A diffused white bar over the belt, strobed in step with the camera, does both — diffusion spreads the LEDs into a soft field, so curved and shiny surfaces show shape instead of hotspots.

The field is 352 mm, so the lit length should exceed it. The LS-BAR-1276 is a 372 × 28 mm diffused white bar light that runs on 24 V. A 372 mm bar over a 352 mm field leaves about 10 mm of margin at each end.

Sequence the pulse from the camera's exposure output: light on, expose for 0.3-0.4 ms, light off, then transfer. Match colour to sensor — white over a colour sensor; red or infrared over mono when hall lighting must be suppressed, with a matching bandpass filter on the lens for full rejection. For a belt wider than the field, keep the same recipe on a longer bar: the 423 × 40 mm diffused red LS-BAR-1168 covers a wider window on the same 24 V drive. The rest of the bar lights range sorts by lit length, colour and diffuser.

Where the cell sits under bright skylights or hall lamps, add a simple hood around the field; ambient that survives the strobe becomes background that thresholding must fight, which the ambient light control guide treats in detail.

Trigger to pick: the timing chain

Between shutter and gripper the belt never stops, so convert every stage of the chain to millimetres of travel. At 300 mm/s the belt moves 0.3 mm per millisecond:

The stage times are planning values for a vision sensor doing one locate and a compact six-axis arm; measure your own and substitute them. The conclusion is what to design around: the robot's usable band along the belt must exceed the 240 mm the part travels during the chain, so this example needs 350 mm or more of reachable belt, and a part that enters the field with less than that remaining is skipped — it rides on and comes round again, which costs less than a rushed, misaligned pick.

Three details make the chain trustworthy:

• Trigger as the part enters the field. The field spans 264 mm along the belt; fire as the part crosses its upstream edge so the whole window remains. A trigger call and the exposure it starts are not the same instant — trigger call versus exposure time shows where the difference hides.
• Encoder truth. Track from the encoder and scale it by measurement: command a known pulse count, measure the belt travel with a ruler, and compute millimetres per pulse. Robots match conveyor speed by following that count, and the conveyor tracking set-up for robot picking page walks through the configuration side. When picks land downstream of the part, stale data is the usual cause — stale pick data timing lists the checks in order.
• Hand off position with time. Send coordinates together with the encoder count or time stamp, so the controller adds belt travel itself instead of aiming where the part used to be. A complete socket handoff for a straight-down pick is in the smart camera socket handoff guide.

Teach the pick point with a real product on a running belt, not a target in still air. The offset between where the camera sees the centre and where the cup actually lands is the cell's systematic error, and one pass with a reference target removes it — the one-target recalibration method is built for exactly that.

Belt speed: what the robot cycle can follow

Turn the chain around and read the belt limit. The window cost scales directly: 800 ms of chain consumes 240 mm at 300 mm/s, 320 mm at 400 mm/s and 400 mm at 500 mm/s. The reachable band is fixed by arm geometry — often 350-450 mm along the belt for a compact arm mounted beside it — so 300-400 mm/s fits with margin, and beyond that you lengthen the band, shorten the chain, or split the pick across two zones fed by one queue.

When picks hesitate or the arm falls behind and re-approaches, the tracking loop is asking for help: work through pick hesitation: motion and timing checks in order. With several robots sharing one belt, the queue logic decides throughput — conveyor speed control for continuous picking covers the hand-off between vision and motion.

Upstream singulation decides whether any of this can work. Random picking copes with separated parts and modest overlap; a belt of piled crisps needs a feeder that spreads them first, which is the architecture in vision guided flexible feeding. When orientation matters — print facing up, notch on a given side — add an angle output to the locate, as in pouch orientation detection for random pick.

Smart camera or camera plus PC

A vision sensor runs the locate on the device and publishes coordinates over Ethernet: one box, no PC, and the timing chain above is the whole system. That covers single-lane picking located by thresholding plus blob or pattern matching. Move to an area scan camera and a controller when the task grows — defect screening in the same image, deep-learning classification per SKU, or several cameras feeding one pick queue. The smart cameras range serves the first case, the area scan cameras range plus a vision controller the second, and both start from the same field, pixel and exposure arithmetic. If parts arrive stacked or tumbled enough that height matters, that is a 3D question; 2D or 3D camera for conveyor picking sets out where the line falls.

Handling: lift the product without crushing it

Vision places the gripper; the gripper decides whether the product survives the pick. For fragile items use vacuum with compliance: a soft bellows cup for a single wafer, or a cluster of small cups for irregular shapes, on a venturi set to the lowest vacuum that lifts reliably. Most damage happens on the descend, not the grip, so approach slowly through the last 10 mm, apply vacuum before any lateral motion, and lift vertically before the arm slews. Give each product its own approach and retreat offsets rather than one fixed path — the pattern in variable pick approach offsets keeps fragile and sturdier SKUs on one line without re-teaching. In the vision task, locate a stable feature — the centre of a round wafer, the long axis of an oval one — instead of the whole silhouette; teaching long thin parts by feature explains why the silhouette fails when the product rotates.

Checks before you order

• Global shutter on every camera in a strobed path. Rolling shutter bands under a 0.4 ms pulse.
• Exposure from your own numbers. Pixel size divided by belt speed; in this example that is 0.4 ms, driven with a controller rated for the bar's pulse current.
• Lit length over field width. A 372 mm bar covers a 352 mm field with 10 mm spare per end; wider belts take longer bars on the same 24 V drive.
• Reachable band over window. Belt speed × chain time + 100 mm of margin; decide the mount position from that, not from wherever the bracket fits.
• Encoder scale measured, not assumed. One ruler check over a known pulse count.
• One camera-to-robot target on the belt. Re-run it after any camera, bracket or belt change.
• Sealing to match the zone. Washdown and hygiene areas set the sealing the light, camera and connectors need — choose before mounting, not after.
• Skip logic in the robot program. Parts with less than the window remaining are left alone.

When the usual answer is wrong

• Piled parts. No exposure or lens fixes overlap: singulate upstream with a feeder, or move to a 3D approach that can segment by height.
• Belt beyond about 500 mm/s with a compact arm. The camera is rarely the bottleneck — two cameras and two pick zones beat one fast camera, because the window grows with speed.
• Glossy film or glaze at a shallow angle. Overhead diffusion still leaves a specular streak; drop a light to a low angle for dark-field edge reading, or add a polarizer in front of the lens.
• Product that must leave the belt re-oriented. Pick with the angle from the vision result and place rotated; do not add a second handling station for what a rotated place can do.

The example cell, assembled

Put together, the numbers describe: a 300 mm belt at 300 mm/s; a 5 MP colour global-shutter vision sensor with a 10 mm lens at 400 mm working distance; a 352 × 264 mm field at 0.135 mm per pixel; a 0.4 ms strobe from a 372 mm diffused white bar; an 800 ms trigger-to-pick chain; and 350 mm or more of reachable belt. That cell locates a 60 mm wafer within a millimetre and picks every part it can reach, part after part. Change one number and the others move — a faster belt needs a longer band or a shorter chain, a bigger field needs a longer lit length, and smaller pixels need more strobe current.

Pull the parts together in the camera finder and light finder, or work through the chain in order in the configurator:

If your belt, product or cycle differs from the example, bring the three numbers — pixel size, belt speed, chain time — to the contact page and the support team will check the window arithmetic with you before you order.