Since you do not know the logo beforehand, the problem cannot be solved easy. What you could do is look for regions in the video where no difference exists between different frames of the video input. Since a logo is mostly fixed at a certain position in the video, this would give you possible interesting regions.
Then in that region you can apply multiple approaches: 1) Template matching using that logo if the situation provides it - adaptable for each new logo 2) Perform OCR on the region and maybe try to extract text - save texts to file and later select which text is interesting depending on characteristics.
Just some suggestions, but I would suggest to first define better what you actually want, since detecting any object, with no prior information is kinda impossible I guess. You will always have to give some input.
I think about searching region with no difference. But how can I do it (I use python) ? I only need detect logo region, so I need data: X coordinate of the top left corner of the logo region, Y coordinate of the top left corner of the logo region, width of the logo region, height of the logo region.
Like I suggested you need to define you're strategy better. First things is reading through documentation. However these could be steps to follow: - Create a capture stream from a video - Store first frame - Storge second frame - Subtract both frames from eachother (identical images will create a 0. - Segement regions based on this 0. - Grow smaller regions together. - Find bounding box
If yoy have tried this, and you still have questions, report back with progress.
If the logo isn't affected by transformations (projection, difference of scale, etc.) you should test template matching (samples/cpp/tutorial_code/Histograms_Matching/MatchTemplate_Demo.cpp)
The first parameter is your image, the second you template to search (ie. your logo).
The logo
The result (red rectangle).
Otherwise, it is more complex (transformations involve) you can use SIFT/SURF/etc. See Features2D + Homography in the doc.
Asked: 2013-02-07 23:46:21 -0600
Seen: 1,183 times
Last updated: Feb 08 '13
