How to locate a large part when the camera field of view is too small

A part that measures 600 × 340 mm does not fit many camera snapshots, and some teach tools stop you from typing that size at all — the input field commonly ends near 500 mm. You do not need a bigger camera. You locate the part from a fiducial pair that does fit the view, stitch two snapshots when one is not enough, and drive every pallet place position from that single result with fixed, rotated offsets. This guide gives you the field-of-view arithmetic, a worked 600 × 340 mm example, numbered procedures for all three paths and the checks that keep the locate repeatable.

Choose the method before you touch the teach pendant

One snapshot has one fixed field, so start from the cause of the mismatch:

If you widen the field and still want to match the whole part, template matching a part larger than the field of view is the companion technique. A part with no flat face and nowhere to put marks is a depth problem, not a field problem: the 3D sensors category and choosing a 2D or 3D camera for conveyor pick and place take that branch.

What you need

• A camera you can mount over the part and calibrate to a world frame — the smart cameras category or the camera finder is the starting pool.
• A C-mount lens sized from the longest part dimension (16 mm in the worked example).
• Two 40 mm fiducial marks with a CAD-taught offset to the part frame.
• A fixed gantry or robot-carried mount whose camera-to-world transform is calibrated.
• The pallet slot pitch from the layout drawing, 650 × 360 mm in the example.

Path A: size the field to the part

Give the longest dimension 10% margin per side: 600 × 1.2 = 720 mm of target field. Sensor width, working distance and field are tied together:

A 1-inch format sensor is 12.8 mm wide and 9.6 mm high. So m = 12.8 / 720 = 0.0178, and at a 900 mm working distance, f = 900 × 0.0178 / 1.0178 = 15.7 mm — a 16 mm lens is the nearest stock value. Run it forward to check: m = 16 / (900 − 16) = 0.0181, so the field is 12.8 / 0.0181 = 707 mm wide and 9.6 / 0.0181 = 530 mm high. The part fills 85% of the width (600/707) and 64% of the height (340/530), with 53 mm clear past each long edge.

Resolution is the second check. An 8.8 MP sensor puts about 2,970 pixels across the width if the aspect is square-ish, and about 4,100 at 17:9, so 707 mm maps to 0.17–0.24 mm per pixel. A 40 mm fiducial then spans at least 167 pixels — comfortably past the 100-pixel floor a sub-pixel centroid wants.

Calibrate at this exact working distance, with a grid that fills the field; building a large calibration grid for a 1 m field of view shows the tiling method.

A C-mount camera with a 1-inch sensor carries this example. The LS-SMC-0017 is an 8.8 MP smart camera with a global shutter, color or mono output and a C-mount. Fit a 16 mm lens, set the working distance to 900 mm, and the arithmetic above is your field.

Path B: teach a fiducial pair, then offset to the part frame

This is the path for a capped teach tool or a part that must sit small in the image. It is the same locate-what-fits logic used for teaching long, thin parts by feature instead of the whole part, carried by a taught offset instead of a length.

1. Mark the part: two 40 mm high-contrast circles, 400 mm apart, on a line 140 mm off the part center — at (−200, +140) and (+200, +140) mm in the part frame. Print, punch or etch them.
2. Take the offset from CAD: pair midpoint to part frame is (0, −140) mm. Mark placement must be exact; a 0.5 mm marking error becomes a permanent 0.5 mm pick bias.
3. Mount and calibrate the camera at the running working distance, field sized as in Path A.
4. Locate each mark as a centroid spanning at least 100 pixels. At 0.24 mm per pixel that is a 24 mm minimum; the 40 mm marks give 167.
5. Read the angle from the pair: θ = atan2(Δy, Δx) between the two centroids.
6. Publish the part frame: part position = mark midpoint + R(θ) × (0, −140) mm.

A fiducial is a reference target placed in the view for measurement (fiducial marker); a plain high-contrast circle qualifies. Keep every mark inside the central 80% of the field — 565 × 424 mm here — because distortion and calibration error climb toward the rim. The 10% field rule for camera orientation precision makes the same argument from the accuracy side.

Rotation accuracy comes from the baseline, not the mark. At ±0.2 mm repeatability per centroid, the worst-case baseline error is 0.4 mm over 400 mm, so θ error = 0.4 / 400 = 0.001 rad ≈ 0.06°. Across the 140 mm lever from mark midpoint to part center that is 0.14 mm, and position itself holds near ±0.2 mm because the midpoint averages the two marks.

Dark marks read best on mono sensors, so this is a mono camera. The MV-SC5060XM-08M-WBN is a 6 MP mono smart camera with an 8 mm M12 lens and a GigE interface. Its rolling shutter is fine for a still part under fixed light; a part that moves during exposure needs the global-shutter sensor in the verification list below.

Path C: two snapshots, one pose

Use this when mount and lens are fixed. Say the gantry holds the camera at 500 mm with a 12 mm lens: m = 12 / (500 − 12) = 0.0246, so the field is 12.8 / 0.0246 = 520 mm wide and 9.6 / 0.0246 = 390 mm high. That holds the part's 340 mm width with 25 mm spare per side, but not its 600 mm length.

1. Teach snapshot position one centered 150 mm left of the part center and position two 150 mm right, both on the long axis.
2. Check coverage by arithmetic: each view is 520 mm wide, half-width 260 mm, so view one covers −410 to +110 mm and view two −110 to +410 mm. The part ends at ±300 mm sit 110 mm inside each view; the overlap band is 220 mm wide.
3. Put the fiducial pair inside that band, within ±110 mm of the part center, so both images see it.
4. Locate the pair in both images, convert each to world coordinates through the calibrated camera-to-gantry transform, and cross-check: the same mark must agree within 1 mm between views. Agreement says the transform is healthy; disagreement is a calibration fault, not a part shift.

The stitch output is one pose — position from the mark pair, angle from the baseline — exactly as in Path B. The tighter view also improves pixel scale to 0.13–0.18 mm per pixel on an 8.8 MP sensor. For a wrist-mounted camera the same construction scales; multiple wrist camera snapshot positions to cover a large table is the table-sized version.

Place every pallet slot from one result

A locate node returns one pose per cycle. A pallet generator is a motion tool with its own sequence and will not nest inside the locate node, so store the located frame and build each place position from it:

For 600 × 340 mm parts on a 650 × 360 mm pitch (50 mm and 20 mm gaps), the offsets from the first slot are d = (650, 0) and (0, 360) mm. At the worst skew this cell will see, θ = 2°, cos 2° = 0.9994 and sin 2° = 0.0349:

(650, 0) mm → (650 × 0.9994, 650 × 0.0349) = (649.6, 22.7) mm

(0, 360) mm → (−360 × 0.0349, 360 × 0.9994) = (−12.6, 359.8) mm

The far slot drifts 22.7 mm sideways under a 2° skew — more than gripper compliance absorbs — which is why every offset passes through R(θ). Applying raw offsets and trusting the angle to be tiny is the bug that surfaces at slot four. Structuring a vision-guided pick and place program around drop targets builds the same one-locate, many-drops pattern.

Verify the locate before you run

• Sweep the part ±5 mm and ±2° over 20 cycles: position repeatability should hold near 0.2–0.3 mm and angle near 0.06°, matching the baseline math above. Anything worse points at mark size or calibration.
• Mark zone: every mark inside the central 80% of its view — 565 × 424 mm in the Path A field, 416 × 312 mm in the Path C field.
• Mark size: at least 100 pixels across each mark at running pixel scale.
• Stitch check: shared marks agreeing within 1 mm between both snapshot views.
• Motion: a part that moves during exposure smears every centroid; use a global shutter sensor. The MV-SC3050XC-16M-ATN-H is a 5 MP color vision sensor with a global shutter and an M12 mount.

Common mistakes

• Nesting the pallet wizard inside the locate node instead of offsetting from the located frame.
• Applying slot offsets without R(θ) — the 22.7 mm drift at 2°.
• Expecting rotation from a single fiducial: one mark has no orientation; two on a known baseline do.
• Marks near the field edge, where distortion and calibration error peak — hold the central 80% rule.
• Stitching two snapshots with no shared mark, which makes a calibration drift indistinguishable from a part shift.

Work the numbers before you teach: pick the path from the table, run the field arithmetic on your longest dimension, place the marks by the 80% rule, and let one locate drive every slot. To match a camera and lens to your part size, use the camera configurator, or send the layout through the contact page for a build list. More teach and place-position methods sit on the applications topic page.