Pouch orientation detection for random pick on a moving conveyor

A pouch on an infeed belt arrives at any angle, often wrinkled, and the robot still has to place it into a mold on a second conveyor. This guide walks the detection chain that makes that work: rotation-invariant contour matching for the silhouette, a field of view and exposure sized from belt speed, a diffused light that survives glossy film, and the encoder handshake that turns a camera result into a belt-relative pose the robot can chase. You will leave with the arithmetic for a 260 mm belt running at 300 mm/s — lens, exposure, light and controller choices — plus the checks to run before the first automatic pick. If the camera type itself is still open, start with 2D or 3D camera for conveyor pick and place: choosing the system.

How does the camera find a deformed pouch at any angle?

Rotation-invariant contour matching finds the pouch silhouette at any in-plane angle between 0° and 360°. It returns center position, angle and a match score in one pass, so a tilted, flipped or wrinkled pouch needs no second template and no orientation-specific retry.

Train the matcher on 5 to 10 good pouches covering the wrinkle range you accept, then let it search the full frame at every angle. The result is center X, Y, angle θ and a score. Accept at 0.6 to 0.7 for the first day and tighten once you see the score distribution from your own line. Because the reference is a silhouette, a change of film supplier or print layout is a retrain, not a parameter tweak.

What field of view and focal length does the belt need?

Size the field of view to belt width plus pick margin: a 260 mm belt wants roughly 290 mm of field of view. A 5 MP-class sensor meets that with a computed 11.3 mm focal length — order a standard 12 mm lens and set about 400 mm of working distance.

Example inputs: 260 mm belt, pouches up to 200 mm long, 400 mm working distance, and a 5 MP-class sensor 8.4 mm wide with 2,448 pixels across.

The computed 11.3 mm lands on a standard 12 mm lens. At 400 mm working distance that yields the 288 mm field of view — the 260 mm belt with 14 mm of margin on each side. The object-side pixel scale is 288 / 2,448 = 0.12 mm per pixel. Every other number in this guide follows from that scale.

How short must the exposure be on a moving belt?

Keep motion blur inside one pixel. At 300 mm/s belt speed and 0.12 mm per pixel, the exposure ceiling is 0.4 ms, so run 0.2 to 0.4 ms and strobe the bar light at full current during the shot to recover the lost light.

Blur is belt speed × exposure: 0.3 m/s × 0.4 ms = 0.12 mm, exactly one pixel. Shorten to 0.2 ms and the same camera holds the one-pixel rule up to 600 mm/s (0.12 / 0.0002). The price is light — a 0.2 ms exposure needs strobed output, not continuous dimming, and camera and light must fire from the same trigger. Free-running the camera adds a random phase to every pose, so trigger from the photo-eye or encoder.

Short strobed exposures also reject ambient light, which matters under a factory ceiling. The full method is in Ambient light variation in pick and place vision: control it. If the ceiling light is strongly colored, pair the strobe with a matching filter as in Bandpass filter with red LED light to block ambient light.

Which light shows a glossy film pouch cleanly?

Light the belt from one side at a low, grazing angle with a diffused white bar. Folds and edges read as bright and dark bands, and a polarizer on the lens kills most of the specular sheen that clear film throws back at the camera.

Mount 15° to 30° above the belt plane, roughly 50 to 100 mm back from the field-of-view center, and keep the camera exposure inside the strobe pulse. For the 288 mm field of view, pick a diffused bar from the bar lights whose lit length clears it with margin. The LS-BAR-1276 is a diffused white bar light, 372 × 28 mm, on 24 V.

Clear film is the exception that changes the geometry. Backlit through a translucent belt section, the pouch reads as a solid dark silhouette — the most robust input a contour matcher can get. Reserve the grazing-angle bar for belts you cannot backlight.

Should you run this on a smart camera or a vision controller?

One camera, one job, results straight to the robot: run a smart camera and skip the PC. Grow to two or more cameras, several light channels and a dedicated robot interface, and the job moves to a vision controller — check its runtime license line when ordering.

The single-camera case covers most pouch lines: image in, pose out, results onto fieldbus — the smart cameras range is built for it. The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter, a 1/1.45" sensor, an M12 mount and a Fast Ethernet interface. That covers the 12 mm lens and 0.4 ms exposure job, and it is the same architecture as Robot pick and place vision without an external PC.

If the station wants the light built in, the LS-SMC-0063 is a 1.6 MP GigE smart camera running 60 fps with a 16 mm lens and integrated polarized white light — the compact option for a tight gantry.

A second camera for fill-level or seal checks breaks the one-camera rule. A vision controller takes both streams, drives the light channels and holds the robot interface in one box. The MV-VC3301H-128G66 runs on an Intel Core i3 with 8 GB of memory, 5+ LAN ports and 4 built-in light channels. Some controllers ship with the vision runtime license included, others list it as an order line — confirm that line before you price the project.

How does the robot get a pose that is still correct at pick time?

Latch the conveyor encoder into every vision result, so each pose is stored as X, Y, θ plus the belt position at image time. The robot corrects for belt travel, which is 3 mm for every 10 ms between image and grip at that belt speed.

The handshake: the photo-eye or encoder fires the camera, the camera latches the encoder count into the result, and the robot re-reads belt position at the instant of grip. It shifts the target by (count at grip minus count at image) × mm per count. Motion platforms expose this as a conveyor-tracking function block — one implementation names it MC_TrackConveyorBelt and takes Conveyor Belt Origin, Initial Object Position and Recorded Position as inputs. Your job is to feed those values from one calibrated chain: camera frame to robot base (Eye to hand object pose transform for robot picks) and a camera fixed relative to the belt (Conveyor tracking camera placement relative to the robot).

The encoder chain deserves the same care as the camera: trigger spacing and count direction are where belt tracking silently fails — Encoder trigger distance for belt tracking: set spacing without faults. For the controls side, a random-order pick-and-place sequencing walkthrough for PLC programmers shows the handshake pattern. A simulation of two random parts arriving on one belt shows the flow before you wire anything.

What accuracy can the angle and position reading hold?

Expect about ±0.1° of angle error from the optics alone. A 120 mm pouch spans 1,000 pixels at 0.12 mm per pixel, so a one-pixel edge uncertainty at a 500-pixel radius works out to atan(1/500) ≈ 0.11°, with film deformation stacked on top.

Position follows the same logic. One pixel is 0.12 mm on the part, sub-pixel centroiding holds the reading near ±0.3 mm, and after calibration with a printed grid target at belt height the camera's share of pick error stays under ±0.5 mm. Belt bounce, gripper compliance and robot repeatability stack on top. Budget ±1 mm end to end and let the mold tolerance judge it.

Calibrate once, then prove it: hold a pouch by hand, rotate it in 15° steps, and require the reported angle to follow within ±0.3° at every step. Error that spikes near 90° means the trained contour is near-symmetric and the matcher is flipping it. Add an asymmetric feature such as the flap or a print mark to the trained contour, or resolve the flip with a second cropped match.

What do you order and check before the first pick?

Order one global-shutter camera with a 12 mm lens, one diffused bar light with a strobe-capable light controller, an encoder feed to the trigger input and the robot's tracking option — then pass four checks before the first automatic pick.

• One 5 MP global-shutter smart camera and an M12-mount 12 mm lens, sized by the field-of-view math above
• One diffused white bar light longer than the field of view, driven by a strobe-capable light controller — the lights finder filters both by size and color
• One conveyor encoder into the camera's trigger input and one photo-eye at the trigger point
• The robot's conveyor-tracking option and its function-block license
• A printed calibration grid target and a marked test pouch

The four checks: field of view confirmed at belt height rather than on the bench, strobe and exposure proven together at line speed, encoder latch proven with the marked pouch, and 20 dry-run picks inside ±1 mm before automatic running. Retrain the matcher after any film change and repeat the dry runs.

Start the light configurator from the lighting side, or send the belt drawing through the contact page and we will check the geometry with you. More belt-tracking material sits on the systems topic page.