Synchronizing a physical object's position and rotation to its digital twin at 8-10 m looks like a depth-sensor job, until you put numbers on the error. A compact stereo depth camera drifts by tens of centimetres over that range, against a tolerance of 1-5 cm. More cameras do not fix that; more pixels on a known target do. A single 20 MP area scan camera watching a printed board holds a few millimetres at the same distance. This guide gives you the two formulas that decide the choice, the worked arithmetic for an 8 m cell, camera and lens picks for the job, and the checks to run before ordering parts.
Can a camera hold 1-5 cm of tracking accuracy at 8-10 m?
Yes — with a coded marker in the scene: a 20 MP camera viewing a 1 m planar board at the working distance resolves 0.77 mm per pixel and tracks it to 1-3 mm, while compact stereo depth misses by 12-19 cm at the same range.
The deciding quantity is the target's size in pixels, not the number of cameras. Position error on the wall scales with the pixel footprint. Range error from a marker scales with how precisely the board's span is measured. Stereo range error scales with distance squared. When both the object and the wall are accessible — the normal case when a real object must drive a model — a printed board wins that comparison outright.
Two architectures fit an 8-10 m cell:
- Marker tracking — one camera, fixed or riding on the object, watches a printed ChArUco-style board. Pose comes from board corners, at millimetre class.
- Depth sensing — a stereo or time-of-flight camera measures the object against a natural background; compact modules land in the 10-20 cm class at this range.
The first meets 1-5 cm with an order of magnitude of margin. The second does not, unless you build a wide baseline that no integrated module offers.
Why does stereo depth error grow with the square of distance?
Stereo range comes from a small triangulation angle, so the z = f·b/d relation turns 0.2 px of disparity noise into about 12 cm of depth error at the working distance and 19 cm at 10 m for a 120 mm baseline and 2.7 mm lens.
The relation and the arithmetic:
f is focal length, b the baseline, d the disparity, and Δd the disparity noise in sensor units. Take a typical compact depth module: b = 0.12 m, f = 2.7 mm, 3 µm pixels, and a working Δd = 0.2 px = 0.6 µm.
- At z = 8 m: Δz = (64 m² × 6×10⁻⁷ m) ÷ (2.7×10⁻³ m × 0.12 m) = 3.84×10⁻⁵ ÷ 3.24×10⁻⁴ m ≈ 0.12 m
- At z = 10 m: Δz = (100 m² × 6×10⁻⁷ m) ÷ (3.24×10⁻⁴ m²) ≈ 0.19 m
Hitting 5 cm at that range needs Δd ≤ (0.05 m × 3.24×10⁻⁴ m²) ÷ 64 m² ≈ 2.5×10⁻⁷ m, or 0.08 px — sub-tenth-pixel disparity on unprepared textures, which compact modules do not reach. Extra cameras add redundancy, not baseline or focal length; that is why two cameras beat five for a depth map. The same physics drives why depth camera noise grows with distance, and how a wide baseline restores long-range stereo accuracy.
Static arithmetic even understates the problem. Frame-to-frame jitter and onboard IMU drift add wander on top of the 0.2 px figure; a stereo depth jitter and IMU drift troubleshooting guide lists the usual sources.
How many millimetres is one pixel at the working distance?
Footprint is field width divided by pixel count, field width is sensor width times distance divided by focal length, and a 5472-pixel, 13.2 mm sensor behind a 25 mm lens at the working distance resolves 0.77 mm per pixel over a 4.22 m field.
W is sensor width, z the camera-to-board distance, f the focal length, N_px the horizontal pixel count. The example assumes a 1-inch-class sensor of 5472 × 3648 pixels. Every result moves linearly with W and z, inversely with f and N_px:
- 50 mm at the working distance: 0.39 mm per pixel, a 2.11 m field.
- 25 mm at 10 m: 0.96 mm per pixel, a 5.28 m field.
- A 4000-pixel, 7.4 mm sensor (1/1.7-inch) with a 16 mm lens: 0.93 mm per pixel at the working distance.
Design to this number. Whatever error budget you set in centimetres, the sensor delivers it in fractions of a pixel, and that is the entire case for resolution over triangulation at long range.
What can a 20 MP camera and a marker board actually deliver?
About 1,300 pixels fall across a 1 m board at the working distance, turning 0.1-0.3 px of corner repeatability into 0.1-0.3 mm of raw noise and a calibrated system in the 1-3 mm band, ten to fifty times inside the 1-5 cm target.
The worked example, end to end:
- Field width: 13.2 mm × 8000 mm ÷ 25 mm = 4224 mm.
- Footprint: 4224 mm ÷ 5472 px = 0.772 mm per pixel.
- Board span: 1000 mm ÷ 0.772 mm ≈ 1295 px across the board.
- Lateral noise: 0.2 px × 0.772 mm ≈ 0.15 mm at the board plane.
- Range from scale: 0.2 px in 1295 px is a relative precision of 1.5×10⁻⁴, so range follows to 8000 mm × 1.5×10⁻⁴ ≈ 1.2 mm.
Those are sensor-level numbers. A system budget adds lens distortion residuals, board flatness, print accuracy and thermal drift; a calibrated 25 mm lens on a rigid mount lands the total pose error at 1-3 mm. Print the board at true size and measure it: marker size errors surface directly as distance errors, and board dimension mistakes bias the pose the same way. The target-size rule from pose calibration transfers unchanged — more pixels across the board, tighter every angle.
Which camera and lens fit an 8-10 m tracking cell?
A 20 MP mono sensor on USB 3.0 with a 25 mm C-mount lens covers the example cell: mono buys sensitivity and skips Bayer interpolation, USB 3.0 handles short cable runs, and the 25 mm lens holds the worked-example field of view.
The MV-CS200-10UM is the 20 MP, mono, USB 3.0, C-mount camera on a 1-inch sensor that this example assumes.
- Sensor format
- 1"
- Lens mount
- C-mount
- Interface
- USB3
- Resolution
- 16-25 MP
Its rolling shutter costs nothing here while the camera stands still. When the camera sits more than a few metres from the PC, the MV-CS200-10GC is the same 20 MP IMX183 imager on GigE. GigE runs up to 100 m over one cable.
- Sensor format
- 1"
- Lens mount
- C-mount
- Sensor format
- 1/1.7"
- Interface
- USB3
When 12 MP is enough, the LS-CMR-0166 reads 12 MP at 30 fps over USB 3.0 on a 1/1.7-inch rolling-shutter sensor.
Lens choice is a footprint choice. Stay at 25 mm for the 4.2 m field of the worked example, or step to 35 mm for 0.55 mm per pixel when the wall only fits a 3 m field. Check the image circle covers the sensor format, because vignetting wastes the corner pixels where board corners land. Filter the camera finder by resolution, interface and shutter to compare the field.
Do four corner cameras with IR LEDs work better?
They can reach millimetre class at 2-4 m, but only with mono or NIR-sensitive cameras, synchronised 850 nm illuminators and a calibration for every camera; their reason to exist is occlusion-free coverage, not accuracy at the far wall.
The geometry is generous at close range: an LED centroid localizes to 0.05-0.1 px on that footprint, so each camera places one LED to about 0.1 mm. Two or more calibrated cameras with a multi-metre baseline triangulate the object to low millimetres — the same band as the marker, at four times the hardware.
Make it work:
- Colour sensors carry IR-cut filters, so the corner cameras must be mono units, illuminated at 850 nm.
- Spread 6-8 LEDs over the object so every camera sees several; solve pose whenever two cameras have line of sight, and coast through short dropouts by prediction.
- Calibrate every camera's extrinsics, then test the calibration before you trust it.
Build the single-camera marker rig first. Add the LED rig only when the board's line of sight is genuinely blocked for long stretches.
When should you still choose stereo depth?
When no target can live in the scene: a marker needs half a metre to a metre of surface and a fixed, calibrated viewpoint, and if the pattern cannot be mounted or the viewpoint moves, depth sensing with a wide baseline stays the tool.
The arithmetic sets the bar. With b = 0.5 m and f = 8 mm, the same 0.2 px disparity noise at the working distance gives Δz = (64 m² × 6×10⁻⁷ m) ÷ (8×10⁻³ m × 0.5 m) ≈ 9.6 mm — a centimetre-class instrument, at the cost of a rig you calibrate and maintain. Purpose-built hardware sits in the 3D sensors category: laser profilers, stereo and structured-light cameras, and time-of-flight units, each specced for its own range band. When a depth readout sticks at one value during integration, the range window and subpixel settings are the usual cause.
What do you check before ordering?
Four numbers decide the build: the board's span in pixels, the footprint at the farthest working distance, the exposure that keeps motion blur under half a pixel, and the cable run that picks GigE or USB 3.0.
- Board span. Keep 1,000 px or more across the board at the farthest distance; pose precision scales as one over that span. At 10 m with the example optics, a 1 m board spans 1,036 px.
- Print quality. Measure the printed board. A 0.5% scale error becomes 0.005 × 8000 mm = 40 mm of range bias — most of the whole budget, from the paper alone.
- Image circle. The lens must cover the sensor format, or vignetting eats the corner pixels the pose solve depends on.
- Rigid mounting. At 0.77 mm per pixel, 1 mrad of camera sway shifts the image by 8 m × 0.001 ≈ 8 mm, about 10 pixels; stillness during capture decides image quality for any camera whose pose matters.
- Exposure. Blur = speed × exposure ≤ 0.5 px. At 0.5 m/s and 0.77 mm per pixel, stay at or below 0.386 mm ÷ 500 mm/s ≈ 0.77 ms.
- Cable. USB 3.0 for bench distances, GigE beyond a few metres.
- End-to-end proof. Move the board a tape-measured 100 mm and require the pose to read 100 mm within a couple of millimetres before the rig goes live.
For the cell described here the decision is one mono 20 MP camera, a 25 mm C-mount lens and a measured 1 m board: millimetre-class pose from a single viewpoint, no occlusion logic, no multi-camera calibration. Size the combination in the camera configurator, browse the 3D imaging topic page for the depth-side detail, or contact the applications team to review your cell before ordering.
Questions
How large should the marker board be at 10 m?
Scale the board to its span in pixels. At 10 m the example optics resolve 0.96 mm per pixel, so a 1 m board covers about 1,036 pixels and a 0.5 m board only 518. Keep 1,000 pixels or more across the board. If the wall cannot take a metre-wide board, a 35 mm lens cuts the footprint to 0.69 mm per pixel and restores about 1,450 pixels on the same 1 m board.
Can one camera measure rotation as well as position?
Yes, a planar board gives all six degrees of freedom from one image. With about 1,300 pixels across the board, in-plane rotation repeats to roughly 0.01-0.05 degrees, from 0.2 px of corner repeatability spread over the board span. Tilt is coarser because it leans on perspective distortion; keep the board normal within about 20 degrees of the viewing axis, or add a second camera, when tilt accuracy matters.
Can the projected screen image act as the marker?
It can, but a printed board tracks better. Screen content, brightness and refresh rate change between frames, and projector gamma shifts apparent edge positions, all of which feed noise into corner detection. Mount a printed 1 m-class board beside or around the screen and keep the display for content. The camera then sees a static, high-contrast pattern at a constant exposure, which is what sub-pixel repeatability needs.
How fast can the object move before the pose blurs?
Cap motion blur at half a pixel: exposure in seconds ≤ 0.5 × footprint ÷ speed. At 0.77 mm per pixel that allows 0.77 ms at 0.5 m/s and 7.7 ms at 0.05 m/s. Rolling-shutter sensors add shear during fast pans, so at speeds approaching 1 m/s either shorten the exposure further, move the board slowly, or step up to a global-shutter camera.
How accurate are the IR LED corner cameras in practice?
Low millimetres at 2-4 m. A 6 MP camera with a 60 degree field resolves 1.1 mm per pixel at 3 m; an LED centroid localizes to 0.05-0.1 px, so each camera places one LED to about 0.1 mm. Triangulation from two or more calibrated cameras with a multi-metre baseline lands the combined pose in the 1-3 mm band. Occlusion, not optics, decides how often you get a full solution.


