It’s long time no post now, so I thought why not post something :) So, here goes this little augmented reality experiment. Let me warn you first, it is not impressive at all, nothing fancy on 3D side, and tracking – to be fair – sucks. This is in part because it does not use any solid technology such as FLARToolKit, just something hastily hacked in haXe, for a change. In other part, it uses very simple fixed threshold filter, that can (and does) harm tracking. To save me some typing, I have made this short video explaining what this experiment is about:

If you have seen how bad it is in the video, but still want to play with it, here you go. The code uses slightly modified blob detection code by Tomek and sandy for 3D part. Make sure you have a webcam, piece of paper, and red, green and blue marken pens.

Advertisements

Like this:

LikeLoading...

Related

9 Responses to “Augmented reality based on blob detection”

I actually saw your reply on your homography post and was like: what do you mean how? Homography is a transformation between two planes, and camera focal plane is nothing different. So, you have two base vectors for your plane, and the third is cross product; three vectors together form a matrix that you want. For another bit of discussion see here or, in short, this is linear problem with 4 points and non-linear with just 3, and I am very tempted to try some gradient descent on this (the above could go as 1st approximation of solution).

Hey Makc3d! I am glad you’ve found the blob detection useful!

Amazing!

hi,

nice take :)

wanting to do the same, I remained stuck on the invert projection matrix (2D > 3D coordinates system)…

how did you do that ?

I actually saw your reply on your homography post and was like: what do you mean how? Homography is a transformation between two planes, and camera focal plane is nothing different. So, you have two base vectors for your plane, and the third is cross product; three vectors together form a matrix that you want. For another bit of discussion see here or, in short, this is linear problem with 4 points and non-linear with just 3, and I am very tempted to try some gradient descent on this (the above could go as 1st approximation of solution).

There two well known tracker which use blob detection (corner detection actually) –

Studierstub tracker

http://studierstube.icg.tu-graz.ac.at/handheld_ar/stbtracker.php

and PTAM tracker (source available non-commercial)

http://www.robots.ox.ac.uk/~gk/PTAM/

Both use FAST at initial detection stage

Is the haXe source code for this available? I would love to see how it works.

sorry I have never got to finish it.

Pixelblender version of HSV Threshold:

http://gasubasu.com/2009/11/21/hsv-threshold-using-pixelblender/

спасибо :)