| Episodes | 94 |
| Mentions | 131 |
| Cited here | 13 |
| First — last | #3 — #710 |
| Top guests | Scott Driscoll, Gerry Roston, David Kronstein |
| Related | gps · kickstarter · youtube · 3d printer · raspberry pi |
A camera is an optical imaging instrument built around an image sensor, and in modern embedded practice it is as much an electronic system as an optical one: the sensor’s clocking, power dissipation, interface bandwidth, and supply chain each impose constraints that shape the final product.[325][517] Raw image data from the sensor travels over the MIPI interface, and placing a processor directly on that interface allows a camera to function as a self-contained vision component rather than an accessory to a host computer.[517] Because concentrated light permanently destroys pixels and power limits bound achievable image quality in compact devices, camera engineering is dominated by failure modes and thermal constraints as much as by optics.[78][658]
Image sensors
Sensitivity
Sensor sensitivity is usefully expressed as required scene illumination: shooting at sixty frames per second with a 1/60-second exposure requires roughly 400 lux to fully saturate the sensor, approximately the level of ordinary office lighting.[325] Pixel size alone does not determine sensitivity. A newer sensor with pixels half the linear size of its predecessor matched the older part’s sensitivity while running considerably faster, indicating that process generation matters more than the geometric argument suggests.[325]
Failure modes
Image sensors are permanently damaged by concentrated light. Pointing a laser into a camera kills pixels outright rather than merely saturating them, a real hazard around any optical test setup involving a beam.[78]
Clocking errors are another destructive failure mode. Image sensors may contain internal phase-locked loops, so the frequency at the pin is not necessarily the frequency the silicon sees; feeding 270 megahertz into a sensor expecting 90 megahertz — having overlooked the internal PLL — permanently damaged the device, leaving dead lines in every subsequent image.[325]
Thermal and power constraints
A high-speed image sensor dissipates enough power to require real heat sinking. The practical route is a hole in the circuit board behind the sensor, allowing the package to be coupled directly to a heatsink rather than conducting heat through the laminate.[325] In small battery-powered devices, camera quality is bounded by power rather than by sensor technology, because there is a hard limit on what can be dissipated in that volume; this is why an otherwise capable sensor performs poorly in a compact product.[658]
System architecture and interfaces
Raw camera data travels on the MIPI interface, and a processor able to ingest MIPI directly can sit at the camera rather than behind a host computer, which allows a vision module to be a component instead of an accessory to a PC.[517] Placing the processor directly on the camera outputs makes it a front end: imagery is processed where it arrives, and only metadata or processed frames leave. This removes the bandwidth and latency of moving raw video to a host that will discard most of it.[517] Connecting a camera to programmable logic over the standard camera interface and then targeting the device’s signal-processing blocks is now the conventional route into vision work, rather than building a bespoke acquisition path.[535]
Separating the processing side of a camera from the sensor side turns a product into a platform: a different sensor and a different programmable logic device produce a faster camera without revisiting the system design, which is what makes a second model economically possible.[325]
Camera and display interfaces carry multiple gigabits per second and have been repurposed as general-purpose data links by making the traffic imitate a camera and a display. The approach worked but was awkward, and it was eventually replaced by a standards-based interconnect — an illustration of using the fast interface already available before designing a new one.[648]
Sensing and measurement applications
Locating something optically is long-established practice and hinges on a step that is often skipped: cameras are placed to see the working volume and are calibrated beforehand. Without that calibration, the images give relative motion and not position.[334]
Cameras can also substitute for kinematic modelling. Rather than computing inverse kinematics for a multi-axis arm, a camera is pointed at the arm, a movement is issued, the resulting position is observed, and the mapping from command to position builds up from the observations. The measurement replaces the model.[344]
Compared with radar, a camera is cheaper and richer in information, but radar returns range and angle directly, whereas an optical system requires an algorithm to infer range from the image. The radar measurement arrives already in the form the system needs, with no inference in between.[115]
Video transmission
Analog video downlinks modulate the carrier directly — for example, the video signal on a 5.8 gigahertz carrier — and degrade gracefully into noise, reflection, and ghosting. Digital links deliver high definition at high frame rates but fail differently, and the digital approach reached the point where the recording made at the receiver could exceed the quality of the one stored on the aircraft.[538]
Manufacturing and supply
Camera module vendors work at a scale where a million units is a small order. An order of ten thousand units attracts a high unit price and an eighteen-week lead time, so a new product must be priced to survive its first batches before volume brings the cost down; building in batches makes the early pricing worse than a naive cost model suggests.[517] Image sensor supply was already the tightest it had ever been before the wider component shortage, with demand exceeding manufacturing capacity rather than design capacity — a constraint that no amount of redesign around the part can relieve.[517]
Field practice
Commercial off-the-shelf vision cameras have flown on spacecraft for engineering and public imagery precisely because that footage was not mission-critical; where the requirement is documentation rather than science, custom development cannot be justified.[532]
Existing equipment can sometimes be rearranged instead of building a new mount: inverting a tripod and fitting a macro lens produced a working soldering microscope in minutes, reversible just as quickly, where the planned solution had involved printed parts and a new stand.[97]
A camera that stops working after a few uses until an application is installed and the owner registers represents a change in what buying a product means, not a firmware defect, and is worth naming as such when evaluating any connected device.[520]
References
| Episode | Title | Date |
|---|---|---|
| 78 | Alteritous Andy's Absquatulation | January 16, 2012 |
| 97 | An Interview with Eben Upton - Morbus Moilsome MakerFaire | |
| 115 | An Interview with Dr Greg Charvat - Watcher of Wraithlike Walls | September 30, 2012 |
| 325 | An Interview with David Kronstein (Tesla500) | November 30, 2016 |
| 334 | An Interview with Gerry Roston | February 1, 2017 |
| 344 | Back Into The Swing Of Things | |
| 517 | Depth and AI with Brandon Gilles and Brian Weinstein | November 15, 2020 |
| 520 | Inductance and Stuff | December 6, 2020 |
| 532 | Recalling Recalls | February 28, 2021 |
| 535 | Efinix FPGAs with Sammy Cheung | March 21, 2021 |
| 538 | Missle Man with Bruce Simson | April 12, 2021 |
| 648 | The RP1 and beyond with the Raspberry Pi Hardware team | October 22, 2023 |
| 658 | Uncle Al's Eating Garbage Again | February 12, 2024 |