Admin is hiring! Part time projects.

[Admin]

I have two jobs I need performed, and I'm willing to pay a decent amount for anyone who can perform either of them for me.


Job A: Derive inverse kinematics for a 4 DOF robotic arm. The arm will have both linear and rotating joints. If the applicant can also do trajectory planning with this arm, I'm willing to pay even more. I will have the arm design FBD finalized within the next few weeks, and will post it then. I need this job completed before the end of April.


Job B: With a camera system of the applicants choice, create a system that can calculate the location of a square in 6 DOF - x, y, z, and rotation of the square. The system must be robust in unpredictable lighting environments both indoors and outdoors, but with a plain color background. The system also must be low cost (off the shelf), lightweight, and microcontroller based (for example, CMUcam3, Blackfin, etc.). The system does not need to be real time. The deadline on this is flexible.


If you are interested, send me an IM/email. If you have any technical questions, post them to this thread.

[Fredrik Andersson]

Are there any language specifications?

What hardware are you using and how do you communicate to it? What kind of interface do you want?

[Admin]

Communication through serial/UART.

Language doesn't really matter, but preference for C. The camera should do the heavy processing, and the processed output will be sent to another microcontroller of which I'll be programming.

[Fredrik Andersson]

Communication through serial/UART.

Language doesn't really matter, but preference for C. The camera should do the heavy processing, and the processed output will be sent to another microcontroller of which I'll be programming.

Ok. What does the camera define as? Camera connected to a PC or a microcontroller based? How fast is the microcontroller/PC (if one would need to heavily optimize the code). Do you want it as an include library, an application that communicates with your other applications or more like code snippets to use?

Oh, just read Job B through (didn't do that before as that is beyond my knowledge ). That answers a lot of my questions.

Sorry if this might be a little off-topic, but I'm not so good with English, so I can't tell if "either" means that one just need to applicant on one of the jobs or have to make both of them.

[Admin]

I've been receiving a lot of resumes and requests in my email . . . I want to be clear that I am specifically looking for someone who already has specific proven expert experience with either of those jobs. This isn't a research problem or an internship!


Fred, no PC is involved. The camera will be the slave processor ("of the applicants choice"). And it must transmit processed data by serial.

[vidam]

You should be asking faculty for referrals. Contact Stefan Schaaal at USC http://www-clmc.usc.edu/~sschaal/ for names of people he can refer you to that have proven experience. Or you put an advertisement on monster. Or you could also just make it a contest and whoever comes up with the best way to solve the problem gets the job assignment. For a second, I thought this could be your biped project, but then I realized that you said robot-arm.

[rhewitt]

Can you clarify what you mean by a square that's in 3D space? do you mean a thin 3D object like a piece of paper that's square on two of its six sides?

If so, and you want full 3D, will you visually differentiate the sides (each side a different color, e.g.)?

Similarly for in-plane rotations. Rotation in-plane thru 90 degrees would result in the same visual object unless there are distinguishing marks on the square.

Will you allow for multiple cameras to handle, for example, the situation in which the square is edgewise to one camera?

Thanks,
Robin

[hgordon]

I asked Admin about this, as I wasn't clear about the task.  My understanding is that square is fixed, and it is acting as a visual reference or index.  It is the camera that is moving around, so the task is to figure out where the camera is in 3-D space relative to the position of the square by looking at the pixels defining the square.  It's not an overly complex task, but it does take a bit of math and some image processing to cleanly extract the size, distortion, and orientation of the square.  rhewitt has a good point about somehow marking the square to differentiate the sides if the range of camera motion isn't bounded.  I told Admin that if someone wanted to take on the task using the Blackfin camera board, we'd provide some support.

[rhewitt]

Thanks, that makes more sense.

Leaving aside this 4-fold ambiguity for the moment, I'd suggest extracting the four edges, then computing the homography between the four lines and a square at a specific image location, scale, and orientation. That will give you a projective transformation which in turn gives you 3D geometry.

If someone wants to collaborate, I have experience with this sort of thing (I've done forensic image analysis and currently work as a computer vision developer/researcher). I also have clean C code for a method called adaptive fuzzy thresholding that could be used at startup to determine the best thresholding for the current lighting conditions. (This is by far the best method I know of for thresholding bimodal distributions. I highly recommend using it here.)

However...

...this is only doable if the camera is fully calibrated -- including its barrel distortion.  Barrel distortion could be roughly approximated from the square itself as a one-time initialization. With manual positioning, pixel aspect ratio and image center could maybe also be determined from the square, but focal length would require additional information or an external calibration procedure.

The other feasibility criterion is the required precision. How precisely does camera pose need to be determined? Limiting factors include image resolution and range of camera distances. Also, given one image of a square in projection, some poses (i.e., some locations in 6D pose space) will have greater uncertainty than others. Before embarking on an implementation project, I'd recommend doing a feasibility study to see if a typical uc camera board will have enough resolution to provide the required precision for the worst case pose scenario. You can do this with just a PC: quantize pose space, calculate the projected image at the center of each pose voxel, then compute the rate of change along each of the six pose dimensions wrt image location.

Robin

[hgordon]

It's not my decision, but you have my vote.  It's safe to assume that the camera will have to be calibrated for the task.

[Admin]

A requirement I should have mentioned was that it should be done with minimal work, like no more than ~5 hours or so. Hence my requirement for off the shelf and experience. (Those who have done this before will not find this time limit a problem)

Sorry for being ambiguous on the project details, but being an industry project there is a lot of hush hush going on.