Favourable characteristics of robotics research camera?

Question

I use Imaging Source CCD firewire digital cameras to capture video streams for use in robotics research. These cameras are reasonably priced and performance is (from my perspective) decent. Unfortunately, there are a few aspects to the video produced by these cameras which is unsatisfactory such as susceptibility to motion blur and glare.

I do not have much experience with photography or cameras and I'm not sure what performance I should expect from such devices. Are motion blur and glare unavoidable problems?, what characteristics should I be looking for in a camera which can address these problems?

I work exclusively in Linux so any suggested solution would have to be standards compliant.

Thanks.

Answer 1

Controlling the vision quality is the same as for photography - your lens will impact the FOV and the image quality in how focused it will be in F/1.4, glares (coating on the lens). The cheap computar and fujinon lenses have a nice price, but they have a lot of glares and cannot really focus at F1.4. Schneider, Azure and Kowa are much better in terms of resolution, coatings, and build quality.

The shutter time compared to your target speed affects the motion blur. For example, when I made agricultural engineering with a case where vision cameras sample a row of trees at 10km/h with a certain FOV and a certain resolution I had to keep the shutter under 10ms. I used 5MP Basler Pilot GigE. You basically have to translate the speed in real world to "pixels in the image per millisecond" and keep your shutter short enough that it doesn't move within an exposure. This conversion depends on resolution, FOV, and distance to the target. In machine vision the lighting (probably like in photography) is 80% of the job done.

As for the best camera to get, that also depends on your calculations, what resolution do you need; go for the lowest possible as it makes the bandwidth and processing lighter - even consider if you can use subpixel accuracy instead of going up in resolution. I once tried to go up in resolution and I didn't get better accuracy anyway, just 4-16x processing time.

What's the lowest acceptable FPS? I needed 90FPS for a project, so I got Basler Ace 640gc and 640gm (colour and monochrome). Do you need colour or can you use monochrome with a filter on it and thus get better images? A real pixel value is better than an interpolated one. Be aware that built-in SDK bayer conversion is usually crappy NN.

Then there's the sensor CCD/CMOS/CMOSIS IR (2x quantum efficiency in NIR light)/InGaas etc..

Features that I do look for in a camera apart from these dependent ones is the transfer protocol, shutter type (global / rolling), DR, and pixel element size.

I like the solid transfer protocols like GigE and CameraLink. GigE is most convenient through a POE switch or POE injector, if the bandwidth is enough (it usually is), just make sure to put it through Intel Pro ethernet. A standard laptop connection is fine for testing but you get a lot of exceptions you have to handle explicitly.

I weigh large pixel sizes high, and I consider rolling shutters a deal-breaker, and DR can be handy in uncontrolled lighting.

I have good experience with Basler (great value for the money), JAI (albeit expensive) and FLIR A-series for thermographic vision. PtGrey FireFly looked like a webcam, and PtGrey Digiclops was so bad we sent it back for another one and the 2nd was bad, too, and they said they couldn't fix it.

Answer 2

Point Grey are the go-to people for robotic & industrial vision cameras.

Basically you're looking for a camera with a high sensitivity, so that it makes the most of available light. A large sensor size, a CMOS sensor (instead of CCD), and a lens with a low f-number (bigger aperture) are all indicators of better sensitivity.

These factors are, however, not directly comparable: the f-number is dependent on the sensor size (a smaller aperture is, in effect, bigger if the sensor is smaller), and amongst sensors of the same size the sensitivity varies considerably between manufacturers, and the exact technology used.