full frame keypoint extraction at >100hz computer vision, robot vision

hello,

i’m working on a framwork which extracts image keypoints (here harris-like corners) for high-speed cameras (>100hz).
the algorithmn runs that fast by an efficient approximation of the rank of the structure tensor.
i try to establish this idea at the ecv09.

besides that i try to optimize the code as much as possible for compute compatibility 1.3.
also, efficient implementation for “second stage”-algorithms are necessary, like integration, list generation and so on.
the aim is supporting real-time robot vision applications.

nevertheless, the current code extracts keypoints (752x400 image, 2048 keypoints max.) and stores them in a list for
further computation within 3ms. there is still potential to optimize the code.
matching two frames keypoints is currently in work, it seems that less than 10ms is possible for the overall correspondece calculation.
the device is a tesla c1060.

i want to ask the community if there any need for this in other research groups. the idea is to make it open source.
suggestions would also help. an opencv interface is thinkable.

greetings,

moik (munich)

Hi moik!

I’m also using CUDA for vision, including keypoint extraction and correspondence matching. In my tests I have found that it is already very quick to calculate the eigenvalues directly (even on the CPU), and this gives more accurate localization than the det & trace measure of cornerness…so when I implement this on the GPU, I will not use an approximation. This is all very fast relative to the computation that follows!

hi yahastu,

yes, a direct eigenvalue computation is achievable nowadays in rt.
with approximation of the rank 2 of the structure tensor, i dont mean that “trace,det” stuff :-)
i cant say more at this point.

got some vids?

greetings,

moik