Electronic – Affordable stereo vision camera for industrial applications?


I am looking for the best value stereo camera for use in an agricultural vision control application. Since all the vendors do their damnest to hide the cost of their equipment, it's proving to be extraordinarily hard to get even a slight idea of what kind of money it'd cost (just as an example, the best guess based on googling for the Bumblebee X3 from Point Grey is "between $1000 to over $3500", ie. I have no clue). Thus I'd appreciate info from people with actual experience on what such equipment might cost, as well as recommendations on what the most cost-effective solution might be.

Requirements / expectations

  • Stereo vision at mid to high res (ie. 1024 x 768 is the lowest I can see working, since it's going to be plant recognition for out in the field moving equipment, but I might be wrong about the minimum needed), with full-speed capture (20fps or more). Thus Bumblebee 2 from Point Grey represents the baseline specs. Modes such as Bumblebee X3's 1280×960@16fps are also reasonable.
  • It doesn't need to be ruggedised, but needs to be ruggedisable in the end product. I have a hardware guy to work on the thing, so we can take care of putting it in a suitable compartment, dustproofing etc., but it should provide a reasonable starting point wrt. robustness of constrution and calibration. The more self-contained and durable it comes out of the box, the better, but I will look at all the options.
  • Sensibly supported SDK. Best is of course being supported by open code, but a sensibly portable vendor SDK is also acceptable. I expect the thing will run on some mix of Linux and RTOS, so a situation like "we support BOTH win7 and vista, as long as you have Visual Studio 2012" would make me a sad panda.
  • Sensibly featureful SDK. I'd rather avoid having to compute 3D information for individual pixels myself for instance, but if someone comes along and says "but mathrick, that is trivially solved in two lines of OpenCV code", I will listen.
  • Bonus points for being widely used and supported in open development community. Right now I'm looking at ROS and OpenCV as the two components that will presumably play a role, but since I'm a newcomer to the particular field of robot control systems, I don't yet have a solid idea of what exactly matters.
  • As for interfaces and interconnections used, I don't have strong preferences as long as it can be sensibly made to work on my target platform. I am similarly not particulary concerned about the footprint or power draw, since it will be mounted on a big honking tractor.
  • Target price is still up in the air a bit, but right now my guess is up to about 1300USD (ideally 1100 or less). Perhaps up to about 1500USD is not entirely out of question, I don't know yet.

Results so far

(I will update this section with everything I learn from answers or research)

  • Surveyor SVS
    Seems to be the cheapest robot-oriented stereo camera I've encountered.

    • Price: Their store is down, so the exact figure is not available, but I've seen mentions of ~$550.
    • Specs: claims to have "full-speed capture" @ 1280×1024, which I take to mean 24fps.
      I wasn't able to grasp properly what connection standards they support. WiFi is there, but no USB or FW or GigE?
    • Pros: open firmware and support code
    • Cons: very bare-bones hardware with not much in the way of ruggedness, and it seems a bit too flimsy to mount on a tractor happily driving across a dusty corn field.
  • Point Grey's Bumblebee 2
    • Price: ???
    • Specs: 1024×768@20fps
    • SDK: Proprietary. FlyCapture for image acquisition (runs on Linux), Triclops for 3D processing (Windows only)
    • Pros: self-contained, enclosed box, which I believe is fairly robust wrt calibration
    • Cons: Proprietary SDK with Windows-only 3D processing.
  • Point Grey's Bumblebee X3
    Seems to be basically Bumblebee 2 with an extra slower, higher-res mode.

    • Price: ???
    • Specs: 1280×960@16fps, 1024×768@20fps
    • SDK: Proprietary. FlyCapture for image acquisition (runs on Linux), Triclops for 3D processing (Windows only)
    • Pros: self-contained, enclosed box, which I believe is fairly robust wrt calibration
    • Cons: Proprietary SDK with Windows-only 3D processing.
  • ???

Alternative solutions

Consumer-level point-and-shoot 3D camera

Can be dirt-cheap, but these are not exactly designed to be controlled by external hardware. I don't know whether it's worth the hackiness inherent in cobbling together a solution like that, and indeed whether there are ways to make it less hacky. Feedback and suggestions appreciated.

Two individual industrial cameras together, do the 3D myself

Might be cheaper than a ready-made 3D solution while being less flimsy than Surveyor SVS if properly robust cameras are used. The downsides obviously include doing the 3D myself. It's not something I have experience with, so I don't know how hard it is to do such a thing reliably using, say, OpenCV. I also don't know how much of my own thing I'd need to provide for a ready-made stereo system, so the gap might not be that big. Feedback and suggestions appreciated.


It's low res (640×480), but well-supported and incredibly cheap. I have no idea about the fidelity in practice however, and whether it can be used for mapping multiple fine objects such as individual plants.

Edit: Turns out the PrimeSense tech used in Kinect is entirely useless outdoors. So that rules it out as a solution.

Best Answer