2D or 3D camera for conveyor pick and place: choosing the system

On one belt, parts arrive mixed — boxes, tubs, small bottles — touching each other, each rotated a different way. On the second belt, the same parts must lie square to the belt at a fixed pitch, say 254 mm (10 in). Two robots move them, and the question is what to hang over the first belt: a 2D camera that reports x, y and rotation angle plus a light grid for part height, or a full 3D camera. This comparison gives the decision, the numbers that drive it, a worked example with the lens and camera arithmetic, and the checks to run before ordering. For parts that rest flat on the belt, the 2D camera plus height sensor wins, and it is not close.

The decision, up front

Parts that always rest upright, or flat on a known face, on a known belt plane need one overhead 2D camera and a height sensor. The camera gives x, y and rotation angle; the light grid, or a laser distance sensor, gives the top-of-part height so the gripper knows where to stop. That is two coordinates, one angle and one height — everything a top-down gripper consumes. Reserve 3D for the cases where height changes the grip itself: parts that lean, parts that can lie on any face, stacked product, or a gripper that must pick its landing zone from surface shape.

Once you commit to 2D, the planar side has its own guide to selecting a planar guidance system, and building a robot vision system for locating and gripping parts covers the robot handoff. If you are still asking whether the cell needs belt vision at all, the case for conveyor vision in pick and place cells applies to six-axis arms as much as to deltas.

What each system gives you

Four routes cover this cell, five if you split 3D into its two flavors:

• 2D area scan camera plus PC. Raw pixels into your software; full freedom for shape matching and touching-part separation; you write the robot handoff.
• Smart camera (vision sensor). Detection runs on the device and hands x, y and angle straight to the robot over Ethernet; no PC in the cell.
• 2D camera plus light grid. The grid adds the height of the highest point, beam by beam across the belt; beam pitch sets the height resolution.
• 3D snapshot camera. Structured light or stereo: one trigger returns a full height map per part, so lean, stack and surface shape become visible.
• 3D laser profiler. One height profile per trigger; belt motion plus an encoder builds the map line by line.

The capability column is typical for each sensor class at the standoff shown; the spread within a class is far smaller than the gap between classes, so the class decision matters more than the model number.

The numbers that drive the choice

What the height number does. If z only sets where the gripper stops — top of part minus a 10 mm approach margin — a light grid is enough: a 4 mm beam pitch separates a 120 mm box from a 30 mm bottle without ambiguity, and the worst-case error of one pitch sits inside that margin. If z selects the grip — clamp the bottle neck, land the vacuum cup on a flat patch at least 15 mm across — you need a surface, and a surface means 3D.

Touching parts. Zero gap between parts is the real vision problem in this cell, not pose accuracy. Two touching boxes look like one blob; a 2D system splits them with shape or template matching, which holds up when the shapes differ (box versus bottle) and gets fragile when identical parts touch edge to edge. Height is the cleaner separator when heights differ: a light grid shows the step between a tub and a bottle, and a 3D camera shows it precisely. Test the worst touching pair offline before buying anything.

Cycle time. At a 254 mm pitch and 0.2 m/s belt speed, a part reaches the robot every 254 / 200 = 1.27 s. Budget vision inside that: trigger, exposure, processing and handoff under about 300 ms keeps the pose fresh when the pick window opens. Every 100 ms of latency moves the part 20 mm down the belt. Stale-frame timing, not camera resolution, is what quietly ruins placement — conveyor vision scene timing covers the diagnosis.

The accuracy chain. Vision pose lands at ±0.2–0.5 mm at a 1 m working distance, robot repeatability at roughly ±0.05 mm for a mid-size six-axis arm, and belt tracking adds the rest. The camera is rarely the weakest link; the tracking chain is.

Worked example: boxes, tubs and 30 mm bottles on a 400 mm belt

Inputs, in the units the cell hands you:

• Belt width 400 mm, view length 500 mm — at least two parts in frame at the 254 mm pitch.
• Parts rest upright on the belt, heights 20–120 mm; smallest part a 30 mm diameter bottle.
• Belt speed 0.2 m/s; mounting height available about 900 mm; grip is top-down.
• Camera: 5 MP class, 1/1.8-inch sensor, 2448 × 2048 pixels, sensor width about 7.2 mm.

The rule in words: pixels across the smallest part first, then lens from magnification, then exposure from belt speed.

Resolution. Ground sample distance across the belt is 400 / 2448 = 0.16 mm per pixel. The bottle spans 30 / 0.16 ≈ 184 pixels. Rule of thumb: 100 pixels across the smallest part holds rotation to about 1° with ordinary shape fitting, so 184 px passes with margin. A sub-pixel centroid at ±0.2 px locates the center to ±0.03 mm — the robot's repeatability is the bigger term.

Lens. Magnification m = 7.2 / 400 = 0.018. Focal length f = 900 × 0.018 / 1.018 ≈ 15.9 mm, so a 16 mm C-mount lens. Checked the other way, working distance = 16 × 1.018 / 0.018 ≈ 905 mm: mount at 905 mm and the belt fills the sensor width.

Exposure. Strobe at 100 µs: the belt moves 200 mm/s × 100 µs = 0.02 mm, which is 0.02 / 0.16 ≈ 0.1 pixel of blur — sharp. Continuous light at 1 ms smears 0.2 mm ≈ 1.3 pixels across the bottle edge, and that costs angle accuracy. Strobe it. When the strobe fires but the frame is black, it is timing — fix the strobe timing instead of raising the gain.

Height. A light grid with a 4 mm beam pitch reports the 30 mm bottle and the 120 mm box equally well; the gripper descends to reported height minus 10 mm and closes. Worst case is one pitch of error, 4 mm against a 10 mm margin.

Cycle. One part per 254 / 200 = 1.27 s. At 150 ms from trigger to pose in the robot, the part has traveled 200 × 0.15 = 30 mm — inside the 500 mm view with room for a re-trigger.

Decision: every number closes with a 2D camera plus light grid. A 3D camera would add a surface map this grip never queries, paid for in cycle time and integration effort.

Choosing the hardware route

For an area scan camera into your own software, pick mono (contrast beats color for pose), global shutter if the light must run continuously, and enough pixels that the smallest part spans 100. The MV-CU020-90GC is a 2.5 MP GigE camera with a global shutter — the safe pick for continuous light. When shapes change weekly and touching parts need custom separation, take more pixels:

Six megapixels over the example's field of view give about 0.13 mm per pixel, versus 0.16 mm for the 5 MP class — every margin above tightens.

For a cell with no PC and a fixed part set, a smart camera hands x, y and angle to the robot directly, with digital I/O for the trigger and the part-ready handshake. The LS-SMC-0063 is a 1.6 MP smart camera with a 16 mm lens, built-in white light and 60 fps capture. Over the example's 400 mm belt that is roughly 0.3 mm per pixel — about 100 pixels across the 30 mm bottle, right at the limit. If the smallest part shrinks, step up a sensor size before anything else.

For the 3D route, a 3D sensor that snapshots one height map per part keeps the cycle simple: the MV-DPS200P-02 is a projected structured-light 3D camera, one trigger for one height map, no encoder. A laser profiler answers a different question — continuous height gating along the belt, classifying every part that passes — and it needs belt motion plus an encoder to turn profiles into a map. A profiler is a laser product: respect the class marked on its housing and keep the beam below eye level along the belt.

The two 3D routes sit at different points:

All 2 share: Ingress protection IP67. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

One part per route, from the three ways to build this cell:

The camera finder filters by sensor size, mount and shutter type, and the configurator walks the same chain with live parts:

The placement side: square parts on the second belt

The pick camera is only half the spec. Parts square to the belt at a fixed pitch come from hand-eye calibration and release timing. Calibrate camera-to-robot with a proper target — choosing a camera positioning grid target keeps that step honest. Rotation is the classic failure: a gripper already rotated at pick places every part crooked, and finding the full turn between pick and place is a known diagnosis. When position drifts only under rotation, work through vision pick error under rotation before touching the belt.

Then hang a small second camera over the outfeed: post-placement vision inspection catches a crooked or short-pitched part while the robot can still correct it. The case for two locates instead of one generic pick is exactly this layout — locate at the infeed, confirm at the outfeed. If the belt runs during the approach, read conveyor tracking pick hesitation before commissioning; hesitation there is timing, not vision.

Checks before you order

1. Freeze test. Strobe pulse × belt speed must stay under a third of a pixel: at 0.16 mm per pixel and 0.2 m/s, any pulse under 270 µs passes, and 100 µs passes with margin.
2. Worst touching pair. Photograph a full belt of the nastiest combination and split it offline with the algorithm you intend to ship. If identical touching parts defeat it, plan on height as the separator.
3. Height margin. The gripper approach margin (10 mm here) must exceed two beam pitches (2 × 4 mm = 8 mm). Tighter than that, use a laser distance sensor per part instead of a grid.
4. Trigger point. Place the photocell far enough upstream that the pose is computed before the pick window opens: at 0.2 m/s, 300 mm of belt buys 1.5 s.
5. Protocol. Confirm the smart camera or vision software speaks the robot's native protocol over Ethernet and that the robot accepts an external pose in millimetres and degrees.
6. Outfeed camera. Budget the verification camera now; retrofitting it after the first crooked lane is the expensive way to learn this.

When the 2D-plus-height answer is wrong

• Leaning parts. Overhead 2D reports the top surface, not the base. A 120 mm part leaning 10° puts its base about 21 mm from where the camera says it is (120 × sin 10° ≈ 21 mm) — a missed vacuum pick every time. If any part can rest more than about 5° off its normal stance, which is 10 mm of base offset at 120 mm height, go 3D.
• Stacked or overlapping product. The light grid reports one height: the top of the pile. Two parts in a column read as one. A 3D snapshot sees the step between them and plans two picks.
• Grip landing chosen from shape. Flexible pouches and clamped castings need the cup on the flattest patch, and flatness is a height map, not a silhouette.
• Anything off a plane. Parts presented in bins, racks or nested trays are a different problem: 3D bin picking for flat parts and bin picking non-planar parts with 3D guidance cover that decision.

Final checklist

• Parts rest on one known plane, upright or flat on a known face: 2D camera plus light grid — stop here.
• Smallest part spans 100+ pixels at your field of view (recheck with the calculator above).
• Strobe pulse short enough that motion stays under a third of a pixel.
• Worst touching pair splits offline with the shipping algorithm.
• Approach margin beats two beam pitches.
• Lean beyond 5°, stacking, or grip-face selection: move to a 3D snapshot camera.
• Continuous height gating instead of per-part picks: a laser profiler with an encoder.
• Outfeed verification camera in the layout and the budget.

For parts matched to these numbers, browse the vision systems topic or contact us with belt speed, part sizes and pitch — that is enough to size the whole cell.