Localize in 3D, multi camera with overlaps (stereo)

[NicksonYap]

Hi @noether

Great repo!

By configuring camerasInfo.xml i see the possibility of using multi camera, with / without overlaps in 2D

However, what about localizing in 3D, likely stereo (2 cameras with overlaps)

Regards

[noether]

Indeed,

Although I believe the performance in practice will not be so good by just adding the parameters from the cameras and their relative positions.

[NicksonYap]

If by performance you mean accuracy then yes indeed

I gave a thought and it may require both mono and stereo camera calibration (intrinsics for mono and stereo rectification for stereo)

I'm actually unsure how to perform run-time undistortion and stereo rectification together with the localization together

but given that the input image/video feed is already pre-undistorted and pre-rectified

I suppose all we need is a simple formula to convert 2D to 3D?

If so, and idea what to change? Assuming we have only 2 cameras with decent amount of overlapping, and increase from there

[NicksonYap]

I came across this multi cam calibration repo: https://github.com/ethz-asl/kalibr/wiki/multiple-camera-calibration

But I trouble figuring out the maths to actually localize :sweat_smile:

[hughhugh]

Hi. what kind of cameta do you use to implement the algorithm? Thx.

[noether]

Hej @hughhugh , I tested the algorithm with both: 2x Logitech C510 in parallel, and just the same camera alone.

[hughhugh]

@noether Thank you. How many cameras are needed in visionloc for several rover like five generally?

[noether]

@hughhugh . You might have as many cameras as you wish. The more cameras, the more area you will cover.

With a single camera Logitech C510 I was able to cover an area of around 2x2.5 meters, and the camera was at about 1.80 meters high from the table.

You can see here an example of such a setup: https://www.youtube.com/watch?v=kS_yJiaA_1Y , the robots were "E-pucks" to give you an idea of their size.

[hughhugh]

@noether Hi, I have ordered a logitech camera now from webstore. But how can I get the heading angle using this project? Environment : Ubuntu 18.04 64. Thank you.

ID (ASCII CODE of the Marker)
PosX of the Corner [pixels]
PosY of the Corner [pixels]
PosX of the Center [pixels]
PosY of the Center [pixels]
Heading [degrees], w.r.t the X axis counter clockwise [-pi, pi)

$ ./example_libvisionloc 4 I can get the position of the tag. but no heading angles. And there is a little time delay in computing the position of the tag.

[hughhugh]

@noether Hi. I have tested the ./example_libvisionloc 4 with 4 tags. It can print out the position information, but the heading angle are not given. The screen image is below:]

Thank you.

[noether]

@hughhugh

Now it should print the heading in the example (please sync with the repo).

The example is not meant to be run in "real time" but as guidance for your application. Your camerasInfo.xml describes two cameras (and you are only using one). In the example, there is a for loop that will iterate always 1000 times for each camera. Note that in the example, even if the program has detected all the makers, the for loop will continue, hence the lag.

In your application, it is up to you how to manage when you should stop looking for markers (condition for the for loop to break). If time is critical for you over detecting all the possible expected markers, then the fastest is to do not have any for loop (just one iteration).

If you have several cameras, and you found all the expected markers in the first camera, maybe you would like to add a condition so that you do not look for markers in the second camera.

It is up to you how/when to stop looking for markers depending on your application.