Build vs buy: modular camera modules for UAV vision payloads

A UAV payload is a mass, bandwidth and endurance budget before it is an imaging problem, and the build-versus-buy decision for camera modules sets all three at once. This guide is for flight-system and payload engineers choosing between integrating a standard board-level module and commissioning a custom carrier. You will leave with the numbers that settle it: MIPI CSI-2 lane bandwidth for a 2 MP, 60 fps stream; ground sample distance at your mapping altitude; flight time lost per 100 g of payload; and the ISP tuning and EMVA 1288 checks that turn raw frames into field data.

Should you build a UAV camera module or buy one?

Buy a standard module unless your mission breaks one on mass, spectrum or mechanical envelope, because a global-shutter board camera streaming 2 MP at 60 fps rides a bought module with interface capacity to spare, while a custom carrier only pays back its months of bring-up at production volumes.

Three conditions break a standard part. Spectrum is the first: compact modules cluster around visible and near-infrared sensors, so a SWIR or UV mission starts outside the board-level range. Mechanical fit is the second: a narrow fuselage bay or a custom gimbal can rule out a standard board shape. Arithmetic is the third: months of carrier bring-up only amortize at production volumes.

The buy path covers most missions. Board-level cameras in this catalog connect over GigE or USB3, pair global or rolling shutters with M12, CS and C-mount lens mounts, and run from 1.3 MP to 12 MP, which spans navigation, FPV and inspection payloads. Match the sensor to the host before ordering, and walk the sensor-first checklist in the board level camera selection guide.

Two parts anchor the two ends of that range.

A global-shutter M12 module carries navigation and pilot view over GigE; the 120 fps C-mount module takes fast inspection bursts over USB3.

What does a pre-validated module take off your plate?

A bought module carries four workstreams you would otherwise run yourself: sensor-and-lens alignment with the mount set to back focus, the host driver and lane configuration, ISP tuning with lens-shading tables, and per-unit calibration records, which collapses a multi-month program into an integration job of weeks.

Alignment arrives done: mount threads and sensor height are set at the factory. M12 and M6 threads still confuse enough people that the board camera lens mount thread guide is worth ten minutes, and back focus on adjustable holders is a set-once job covered in the S-mount holder set-up guide.

Host match comes next: lane count, power tree and driver support decide whether the module streams on day one or week ten, which the CSI carrier board compatibility checks turn into a checklist.

Calibration records matter at fleet scale: shading and white-balance tables per unit, traceable to a lens serial, so a swap does not restart tuning. Revision control rides along; when a vendor revises the board, the hardware revision compatibility guide explains what changes under your feet.

Compact and bare-board modules arrive as matched sets for tight airframes:

The miniature M12 part sits in a gimbal nose; the bare-board 6 MP module disappears into your own housing stack.

How many pixels and how much ground sample distance does your mission need?

Ground sample distance decides the sensor, and GSD equals pixel size times altitude divided by focal length: 1.6 μm pixels at 50 m altitude behind a 12 mm lens resolve 6.7 mm per pixel, so a 4000-pixel sensor line covers a 27 m swath per frame.

Work the mapping case: 1.6 μm / 12 mm = 0.000133 of altitude per pixel, so at 50 m the GSD is 0.000133 × 50 m = 6.7 mm. A 12 MP 4:3 sensor divides to 4000 × 3000, so one frame covers 4000 × 6.7 mm = 26.8 m, call it 27 m of swath. A spec asking for 3 pixels across a 20 mm defect needs a GSD of 6.7 mm or better, which this stack clears at 50 m.

Tighten the spec to 15 mm defects at 3 pixels and the GSD must reach 5 mm: drop altitude to 37 m (0.000133 × 37.5 m = 5 mm) or fit a 16 mm lens at 50 m (1.6 μm / 16 mm = 0.0001; × 50 m = 5 mm). Either way the swath narrows to 4000 × 5 mm = 20 m, so the flight plan needs more lines for the same coverage.

Size the sensor before the airframe; the resolution calculator turns your smallest feature and swath into a pixel count.

The LS-BCM-0056 carries this class of payload: 12 MP of 1.6 μm pixels behind a C-mount on a GigE link.

At nadir, rolling shutter costs little at steady airspeed; near structures it shears rows. The global versus rolling shutter comparison draws that line. Swaths that outrun a 12 MP frame move to area scan cameras, which reach beyond 25 MP for wide-format mapping.

How much bandwidth does the sensor need through MIPI CSI-2?

Raw CSI-2 bandwidth is width times height times frame rate times bits per pixel: 1920 × 1200 at 60 fps and 8-bit works out to 1.11 Gbit/s of image data, about 1.27 Gbit/s on the wire with blanking, or 318 Mbit/s per lane across four lanes.

Run the navigation example: 1920 × 1200 = 2,304,000 pixels; × 60 fps = 138.24 Mpixel/s; × 8 bit = 1,105.92 Mbit/s of image data, about 1.11 Gbit/s. With 15% overhead the wire carries 1.27 Gbit/s, so four lanes run 1.27 / 4 = 318 Mbit/s per lane.

Common embedded vision processors accept up to 1.5 Gbit/s per MIPI D-PHY lane, so a four-lane link moves 6 Gbit/s and this mission idles at one-fifth of the per-lane ceiling. Growth costs little headroom: 10-bit raw lifts the wire to 1.38 Gbit/s of image data, 1.59 Gbit/s with overhead, still 398 Mbit/s per lane.

Skip the check and the failure is ugly: one team ran a sensor that asked for 1.728 Gbit/s per lane on four lanes, 6.9 Gbit/s total, past a receiver capped at 1.5 Gbit/s per lane; the fix was re-clocking the link to 1.44 Gbit/s per lane (5.76 Gbit/s total), not new silicon.

Budget lanes from the receiver ceiling downward, then buy the sensor that fits under it with the mission's bit depth. When host and camera disagree on interface, the MIPI CSI-2 to USB bridge bandwidth guide sizes the detour with the same arithmetic.

How much flight time does each added gram cost?

On a 1.5 kg quadcopter hovering 20 minutes on a 60 Wh battery, every extra 100 g of payload costs about 1.2 minutes: hover power scales with mass to the 3/2 power for the induced share, so the vision package's grams come straight off endurance.

The exponent comes from actuator disk theory: induced power scales with weight to the 1.5 power, while profile power stays roughly constant as mass changes.

Take the example platform. Hover power = 60 Wh / (1/3 h) = 180 W. Split it 60% induced (108 W) and 40% profile (72 W), a normal small-quad balance. Add 100 g: the mass ratio is 1.6 / 1.5 = 1.067, and 1.067^1.5 = 1.102, so induced power becomes 108 × 1.102 = 119 W and hover power 191 W. Endurance = 60 Wh / 191 W = 0.314 h = 18.8 minutes. That is 1.2 minutes lost per 100 g.

The induced share moves the answer: at 40% induced the same 100 g costs 4% of hover time, at 80% it costs 8%, so budget 1 to 2 minutes per 100 g on this class of platform. Bigger aircraft dilute the penalty, because the same grams are a smaller mass fraction.

Apply it to the choice: say the bought module with lens and mount lands at 60 g and the custom assembly at 200 g. The difference is 140 g × (1.2 min / 100 g) = 1.7 minutes of hover, about 8.5% of the flight, for parts that image identically. Weigh candidates with lens and mount fitted, never the bare module.

When is a custom carrier board the right call?

Commission a custom carrier when two or more of these hold: no offered module matches your spectral or shutter requirement, the airframe envelope forbids a standard board shape, your compute interface is frozen and unmatched, or annual volumes are high enough to amortize months of bring-up.

Score the four conditions honestly. Spectrum first: SWIR and UV payloads live outside the compact board range, in larger area-scan formats. Envelope second: if the module must wrap a mast, split around an antenna or flush-mount into a nose cone, custom mechanics beat adapters, and the housing design from STEP files guide shows how far that path goes.

Interface third: a frozen compute stack with no matching module forces a carrier, and the sensor compatibility checks belong in the feasibility review, not after the PCB is poured.

Volumes decide fourth. Bring-up spans power trees, lane clocking, drivers, device trees, ISP tables and regression, which is a quarter to a year of engineering. Annual counts in the hundreds amortize it; a ten-airframe prototype fleet never does. The gate is two or more broken conditions plus amortizable volume; a single broken condition usually falls to a mechanical adapter or a bridge chip, and zero means buy.

What has to be tuned before raw frames become usable images?

Raw frames are not images until the image signal processor is tuned, and the tuning order matters: bypass the ISP to capture raw, set black level, fit lens-shading gains on a uniform target, balance white under the mission light, then denoise and sharpen last.

1. Capture raw with the ISP bypassed; shading calibration needs untouched sensor response, and the raw sensor output explainer shows what a module without an ISP in line delivers.
2. Set black level and map defect pixels before anything cosmetic.
3. Fit lens-shading gains per colour channel on a uniformly lit white target; four-corner and radial models both work, and the tables belong to the lens, not the sensor.
4. Balance white under mission light: daylight sits near 6500 K and overcast runs cooler near 7500 K, so tune in the sky you fly.
5. Denoise and sharpen last, and add HDR only where propeller shadow and sun glare create extreme range.

The embedded ISP tuning guide shows why sensor and software tune as a pair rather than as separate workstations.

Clean frames feed the detection stack, and choosing pretrained vision models for UAV AI boards is the next matching problem once images flow.

How do you verify image quality before you commit the design?

Accept a design on EMVA 1288 numbers plus a hot soak, not on one sample frame: dynamic range is 20 log10 of full-well capacity over temporal dark noise, which gives 66 dB for an illustrative sensor with 10,000 e- full well and 5 e- read noise.

EMVA 1288 hands you four numbers to gate on: temporal dark noise, photo response non-uniformity, quantum efficiency and saturation capacity. Dynamic range follows from the first and last: DR = 20 × log10(10,000 e- / 5 e-) = 20 × log10(2000) = 66 dB. The EMVA 1288 comparison guide explains reading those data sheets, and every shipped unit should clear the same gate.

Soak and shake before flight trials: run the assembled module 20 minutes at worst-case enclosure temperature and re-check shading and noise, then bench-vibrate it and look for row shear on rolling shutter parts. The board camera evaluation checklist folds these into a pre-order list.

• Pin GSD from pixel size, altitude and focal length before the BOM exists; it picks the sensor.
• Budget lane rate from the receiver's per-lane ceiling downward; it picks the interface.
• Weigh the package with lens and mount, then price it in minutes of flight; it picks the mass budget.
• Capture raw before tuning anything; shading tables need untouched frames.
• Gate acceptance on EMVA 1288 numbers plus a hot soak, not one pretty frame.

The rule, restated: buy when a standard module clears the GSD, bandwidth and mass math; build only when two or more break conditions hold and volumes amortize the bring-up.

Shortlist from the camera finder or start at the camera step of the configurator; when the mission genuinely breaks a standard part, the contact page opens the custom conversation. The rest of the system set lives on the embedded vision systems topic page.