Ask Your Question

Best images for ORB feature point tracking

asked 2015-01-30 04:38:57 -0600

colin747 gravatar image

updated 2015-01-30 09:06:11 -0600

A have a program which detects objects in a video stream, the objects will be markers that will be attached to objects so in effect it is recognising the marker rather than the object itself.

My question is what type of marker would be best for this?

I'm aware that a complex non-repeating object is best but I'm not sure how this translates to practical application. I thought that perhaps a QR code style marker would be good but although it is easily recognised with a very high accuracy rate when I start using multiple QR codes for each object the program is finding it difficult to distinguish between them.

EDIT: As per the first comment I tried the following image but could not get the program to detect it at all (It does detect the QR codes fine).

image description

edit retag flag offensive close merge delete


I would go for a very explicit combination of simple shapes. Like a triangle with an inner circle for one, a square with an inner triangle for the second, ...

StevenPuttemans gravatar imageStevenPuttemans ( 2015-01-30 08:06:57 -0600 )edit

Hmm and if you try a more corner based feature like harris corners?

StevenPuttemans gravatar imageStevenPuttemans ( 2015-01-30 10:51:36 -0600 )edit

Is there any advantage to that method over using ORB? I only ask as the markers can be changed to whatever suits but I will try the Harris method and report back.

colin747 gravatar imagecolin747 ( 2015-02-02 03:18:25 -0600 )edit

No idea, just a wild guess :)

StevenPuttemans gravatar imageStevenPuttemans ( 2015-02-02 03:32:10 -0600 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2015-02-03 03:47:37 -0600

colin747 gravatar image

updated 2015-02-03 03:49:42 -0600

After playing about with different images I've came up with three images that are complex enough for the ORB detector to easily detect but distinct enough that it doesn't confuse them with each other. (N.B. I only stopped at three as it's all I need for the moment so I don't know if there will be a limit on the distinctiveness you can make these.)

Image one:

Image One - multiple overlapped squares and circles

Image Two:

Image Two - multiple overlapped triangles

Image Three:

Image Three - multiple overlapped stars

I have no reason to doubt that these images should not work with other feature point detection algorithms but I have not tested it. Hopefully this will be of some use for others

N.B. There was small false positives detected within the video frame which where filtered out by calculating the number of inliers within the homography and checking if this was over a pre-set value (in my case 100).

edit flag offensive delete link more

Question Tools

1 follower


Asked: 2015-01-30 04:38:57 -0600

Seen: 743 times

Last updated: Feb 03 '15