The Camera Decides the Rest of the Architecture

Adding a camera is rarely a component decision. The sensor sets the data rate, the data rate sets the processor class, the processor sets the power budget, the power budget sets the thermal design, and the optics set the enclosure geometry. Pick the sensor casually and you have implicitly chosen a board, a battery, and a housing.

Work the other direction. Start from what the image has to accomplish, in measurable terms: read a barcode at 12 in (300 mm), detect a person at 30 ft (9 m), classify a defect 0.02 in (0.5 mm) across, or produce video a human will watch. Each implies a different sensor, and only one of them cares about megapixels.

Resolution: More Pixels Is Often Worse

Marketing sells megapixels; engineering buys pixels on target. The useful calculation is how many pixels land across the smallest feature you need to resolve. Barcode reading typically wants two to three pixels per narrowest bar. Face detection wants roughly 40 pixels across the face. A machine-vision defect check often wants three to five pixels across the defect. Compute the required pixel count from the field of view and the feature size, then stop.

Pushing past that number costs three ways. Pixel size shrinks, so each pixel collects fewer photons and low-light performance degrades sharply. The data rate rises, forcing a faster interface and a bigger processor. And the lens has to resolve what the sensor samples, which means a more expensive optic. A 2 MP sensor with 3 micron pixels frequently outperforms an 8 MP sensor with 1.1 micron pixels in the conditions where products actually operate: indoors, at night, under a plastic window. Compare pixel size and optical format, not headline resolution. A 1/2.8 in sensor gathers roughly four times the light of a 1/5 in sensor at the same pixel count.

Shutter, Frame Rate, and Dynamic Range

Rolling shutter exposes rows sequentially. It is cheaper, more common, and perfectly fine for static scenes and slow motion. Anything moving quickly across the frame skews, and anything under pulsed lighting bands. Global shutter exposes every pixel simultaneously, costs roughly two to four times as much for equivalent resolution, and is mandatory for scanning parts on a conveyor, reading codes on moving objects, or any camera on a vibrating platform.

Frame rate should be derived, not chosen. If an object crosses the field of view in 200 ms and you need three clean looks at it, you need at least 15 fps with an exposure short enough to freeze the motion. Higher frame rates raise the data rate proportionally and shorten the exposure window, sending you back to pixel size and lens aperture.

Dynamic range decides whether the product works at a window or a doorway. A standard sensor delivers roughly 65 to 70 dB, enough for controlled lighting. Scenes with a bright background and a dark subject need wide dynamic range, typically 100 dB or more, achieved through multi-exposure fusion or a split-pixel design. Wide dynamic range introduces its own artifacts on moving subjects, so test it on your actual scene rather than trusting a datasheet figure.

Interface: Where the Processor Choice Gets Locked In

The output interface is the most consequential line on the datasheet because it determines what can receive the data.

  • MIPI CSI-2 is the standard for embedded cameras. It is fast, low power, and short reach, typically under 8 in (200 mm) on a controlled-impedance board. It requires a host with a dedicated camera port and an image signal processor, which effectively means an application processor running Linux rather than a microcontroller. This is usually the fork described in microcontroller vs embedded Linux.
  • Parallel DVP survives on low-resolution sensors and small microcontrollers. It burns a dozen or more I/O pins and tops out well below HD at usable frame rates, but it lets a capable microcontroller handle simple imaging without a Linux stack.
  • USB with the UVC class is the fastest path to a working system. The module handles its own processing and enumerates as a standard webcam on any host. It costs more per unit, adds latency, and gives you limited control over exposure behavior, but it removes months of driver work.
  • Serializer links such as GMSL or FPD-Link carry camera data many feet over coax, which is how automotive and industrial systems place a sensor far from the compute board.

Whatever you pick, plan the physical routing. Camera modules almost always sit on a separate small board connected by a flex, which makes flex and rigid-flex construction part of the camera decision rather than an afterthought.

Bare Sensor or Finished Module

A bare sensor is a die on a package. Choosing one means also choosing a lens, a lens holder, an IR cut filter, an active alignment process, and a supplier who can do all of it repeatably. It gives the lowest unit cost at volume and full control over optics, and it adds optical design, focus tolerance analysis, and a production alignment step.

A camera module ships as a tested assembly with a known field of view and focus. Below roughly 20,000 units per year the module almost always wins on total cost once engineering is included.

Either way, do not underestimate tuning. Getting good images out of a raw sensor requires ISP work: auto exposure, auto white balance, lens shading correction, noise reduction, and color calibration for your specific lens and window. A professional tuning engagement runs 15,000 to 60,000 dollars and takes six to twelve weeks. Modules from a vendor with an existing tuning package for your host processor let you skip most of it.

Supply Lifetime, Heat, and What Comes After the Image

Image sensors follow the consumer phone market, which means aggressive end-of-life cycles. Ask specifically for an industrial or automotive grade part with a published longevity commitment, often seven to ten years, and confirm the exact module part number rather than the sensor die. A module vendor swapping lens suppliers silently changes your field of view and breaks your tuning. Treat that risk the same way you would any other single-source component, using the tactics in component shortage planning.

Sensors also self-heat, and heat raises dark current and noise. A sensor streaming at full rate inside a sealed housing can run 20 to 30 F above ambient, visibly degrading night images, so the camera belongs in the thermal design from the start. Finally, decide where the pixels are processed, since running inference on-device changes the processor requirement entirely, as covered in computer vision at the edge. If the application is heat rather than light, the tradeoffs in choosing a thermal imaging payload apply instead.

Get the Camera Chain Specified Before You Order Parts

Projects House specs camera subsystems from the requirement backward: pixels on target, sensor and optics selection, interface and processor pairing, module versus custom stack, tuning plan, and a supply strategy that survives the sensor market. Send us your scene, working distance, and lighting conditions through our contact form.