Stereo camera in a radioactive environment: what actually degrades

You are mounting a stereo depth camera in a radioactive environment — a hot cell, a beamline enclosure, a contaminated workshop — and the point cloud has to stay trustworthy for the life of the job. The field itself will not bend the cloud, and there is no calibration you can take against a dose reading. What dose does is damage the camera slowly: dark current rises in the CMOS sensors and the texture projector ages. Here are the dose numbers that matter, a worked budget, a decision for each dose band, and the checks that catch degradation before your data does.

What a radioactive field does to a stereo point cloud

A stereo depth camera measures depth from the disparity between two images of the same scene. Both sensors are silicon, so they respond to light from roughly 400 to 1000 nm — the visible band plus the near-infrared band a texture projector uses, the same window discussed in SWIR vs NIR vs visible imaging: which band to inspect with. Gamma rays and neutrons carry no light in that window, so the field itself adds no signal to the images and no shift to the calibration. Frame by frame, the cloud is untouched.

Two exceptions exist. Cherenkov light and radioluminescence are real light inside the window, and both are scene problems you handle in software, covered below. Cumulative dose changes the sensor and the projector themselves, which is a hardware-lifetime problem, covered next. Neither is a geometric error, so keep all geometry work — intrinsics, extrinsics, hand-eye — exactly as you would run it outside the field.

The number that decides: total ionizing dose

CMOS imagers degrade with accumulated dose, not with instantaneous rate: ionization builds interface states in the pixel, and dark current rises with the total. Published total-ionizing-dose tests of CMOS image sensors report dark current climbing through the dose range with sharp steps at 75–100 krad(Si) and again at 175–210 krad(Si).

That gives three working bands, and the decision that goes with each:

• Under 1 krad(Si) accumulated over the service life: run a standard industrial stereo camera, no special measures.
• 1 to 75 krad(Si): dark current grows gradually. Refresh dark frames, cap exposure times, keep a qualified spare.
• Above 75 krad(Si): the sharp steps begin. Shield the camera, move it back, or replace it on a schedule and treat it as a consumable.

Worked example: a two-year inspection cell

Take the dose rate at the camera position from the radiation survey — the number at the mount, not at the source — then multiply by the hours the camera will actually run.

Case A, inspection cell: the survey reads 0.5 mGy/h at the mount, and duty is 8 h/day for 220 days/year over 2 years, so 3,520 h.

D = 0.5 mGy/h × 3,520 h = 1,760 mGy = 1.76 Gy = 0.176 krad(Si)

The first sharp step, 75 krad(Si), is 750 Gy — more than 400 times this budget. At 0.5 mGy/h the camera would need about 1.5 million hours to reach it. Decision: standard stereo camera, no shielding, quarterly accuracy checks.

Case B, beamline zone at 10 Gy/h: a 40-hour campaign (4 h/day for two weeks) delivers 10 × 40 = 400 Gy = 40 krad(Si), inside the gradual band, with the first sharp step 75 hours away at full rate. Decision: shield or shorten the exposure, refresh the dark frame mid-campaign, and replace the camera when the campaign ends.

How dose noise reaches the point cloud

Dark current from dose is non-uniform across the array, so dark-signal non-uniformity and readout noise rise with it. In a stereo camera that noise lands directly on the disparity search: less texture contrast means noisier disparity, and depth error follows the parallax law

with Z the range, sigma_d the disparity noise in pixels, f_px the focal length in pixels and B the baseline.

Run the numbers for a typical small stereo rig. A 2.5 mm lens on 3 µm pixels gives f_px = 2.5 / 0.003 ≈ 833 px; with B = 95 mm, a disparity noise of 0.2 px — typical subpixel matching on good texture — gives sigma_Z = 2² × 0.2 / (833 × 0.095) ≈ 10 mm at Z = 2 m. Double the disparity noise to 0.4 px, a level dark-current growth can produce, and depth noise doubles to 20 mm; step back to 4 m and the same 0.2 px gives 40 mm. Dose therefore shows up first at long range and on low-contrast texture, and as frame-to-frame jitter rather than a constant offset — the levers in stereo depth jitter and IMU drift troubleshooting apply unchanged, because the matcher cannot tell noise sources apart.

Heat stacks on dose: silicon dark current roughly doubles every 6 to 8 °C, so a camera warmed by shielding or cell ambient ages faster, an interaction covered in depth camera performance vs operational temperature range. After capture, temporal and spatial filters buy some of the loss back; start from noisy stereo point cloud fixes: thresholds, sensing mode, depth range and point cloud filtering before surface reconstruction: the pipeline.

Light the camera can see: Cherenkov and radioluminescence

A radioactive scene can make light a stereo camera sees. Cherenkov radiation peaks in the ultraviolet and blue, and its visible part lands inside the response window above, so water-shielded sources and very active surfaces add a blue glow to the left and right images. That glow is false texture: the matcher locks onto it and depth in that region goes wrong. Radioluminescence from scintillating deposits or a contaminated viewport does the same at lower intensity.

Handle both as scene hygiene, not calibration. Watch the live feed, and when a glow appears, mask those pixels or exclude that region from matching before disparity. A thermal camera watching the same cell sits in the long-wave infrared, where the glow is invisible — the infrared cameras category covers them, with an LWIR unit such as the LS-IRC-0044 as one example — but budget its dose life the same way.

Choosing the strategy for your dose band

Within 3D sensors, structured-light stereo units fit the low and middle bands well, because the whole camera is swapped as a unit when dose retires it:

All 2 share: Data interface Gigabit Ethernet (1000 Mbit/s); Humidity 20% to 85% RH (non-condensing); Certification CE, RoHS. Values as printed in each product's data sheet; open a part for its drawing and full specifications.

The MV-DLS300P-04 sweeps its texture with a moving laser beam and the MV-DPS200P-02 projects a fixed pattern, so either can sit in the replace-as-a-unit plan.

The texture projector is usually the first sub-system to fade, because it is an active laser source run at duty. A passive stereo pair, or running with the projector off under fixed scene lighting, removes that path — the trade-offs are laid out in depth camera has no IR projector: passive stereo and laser class.

Move it back, shield it, or consume it

Distance is the first shield to reach for: for a localized source the dose rate falls with the square of distance, so doubling the standoff quarters the rate. In the 10 Gy/h zone of case B, moving the camera from 1 m to 2 m drops the rate to 2.5 Gy/h and stretches time-to-first-step from 75 h to 300 h. Lead and leaded-glass ports handle gamma; polyethylene handles neutrons; both cost light and working distance.

Standoff costs depth accuracy, because depth noise grows with the square of range. Earn it back with a wider baseline — the reasoning and the limits are in long-range stereo depth accuracy with a wide baseline.

A watching camera behind lead glass is then ordinary lens selection. To cover a 2.4 m cell face from 3 m with a 1/1.8-inch sensor 7.2 mm wide: magnification m = 7.2 / 2400 = 0.003, focal length f = 3000 × 0.003 / 1.003 ≈ 9 mm; a 12 mm lens narrows the view to 7.2 × 3000 / 12 = 1.8 m. The lens focal length calculator runs this arithmetic, the camera finder filters by mount and interface, and the configurator walks the chain.

For parts you can present to a port, a line profiler keeps its sensor behind shielding and scans through the gap:

The same DP3000 series also offers the MV-DP3062-01D V2.0 binocular single-line variant. A time-of-flight unit such as the MV-DT01WMU swaps disparity matching for a modulated flood source, which changes which sub-system ages first but not the budget arithmetic above.

Verify with a target, never with a dose reading

The question "can the point cloud be calibrated against a radioactivity value" has a firm answer: no. Dose produces noise and drift, not a repeatable geometric offset, so there is no curve that maps a dose reading to a displacement you could subtract. Do this instead:

1. Calibrate geometry the ordinary way, with targets sized to the measurement volume — the rule is in stereo calibration accuracy: the measurement volume decides.
2. At commissioning, fit a plane to a machined flat at working range and record the scatter; the method is in 3D flatness inspection: fitting a plane to a point cloud.
3. Re-measure the same flat quarterly, monthly once accumulated dose passes 10 krad(Si). When the scatter doubles against commissioning, retire the camera — that trigger needs no dosimetry.

When the usual answer is wrong

Neutrons: the krad(Si) steps above come from gamma and X-ray tests. Neutrons add displacement damage in the silicon lattice on top of ionization, so when the neutron component is comparable to the gamma component, halve the bands you plan against and put polyethylene between source and camera.

Pulsed machines: prompt radiation during beam-on can fog whole frames at once. Capture in beam-off windows, and let a watchdog restart the interface after single-event upsets rather than debugging them mid-shift.

Very hot zones: above a few Gy/h, keep silicon out of the volume entirely — a periscope mirror with the camera outside the shield wall is the classical hot-cell layout, and it converts the problem into lens and mirror selection.

Checks before you commit

• Survey the dose rate at the camera position, gamma and neutron separately, not at the source.
• Multiply by planned run hours and divide Gy by 10 to read krad(Si).
• Read the band off the table and pick the row: standard, maintained, or shielded.
• Decide projector duty now: texture on only during capture extends projector life.
• Fix the retirement trigger in writing: plane-fit scatter doubling against commissioning.
• Box a spare with brackets pre-fitted; swap time in a radiation area is what costs dose to people.

Send the survey numbers and the standoff you can live with through the contact page and the applications team will size the camera, lens and shielding trade, or work from the 3D topic page for the capture and filtering steps around it.