How to exclude or mask small regions of video from feature detection?

I want to try doing SLAM on some video from a helicopter. Unfortunately the video has a data overlay that interferes with the detection of features. Unfortunately, some of the overlay is in the center of the image, so selecting a rectangular ROI would be limited to a very small part of the image.

How can I tell the feature detector to ignore some portions of the image? The overlay pixel locations are very consistent, so I want to use a mask.

Answer 1

[Image Analyst]

Perhaps you can use the Color Thresholder to identify the blue-green pixels. Then use regionfill to fill them in with the surrounding "good" parts of the image.

See attached m-file.

[Nelson]

Hmmm, I can give that a try. But I'm worried that the feature detection still flag false features in those places. Also, it adds more processing at each frame.

I was hoping there's a way to short-circuit the for-loop within the feature detector so it doesn't even consider the pixels in those places -- just like it won't consider pixels outside of a ROI.

[Image Analyst]

If the graphics are in the same place each time then you can just store the binary image and use it to erase each new frame. You wouldn't have to compute it all over again for each new frame.

I don't know what for loop you're talking about. You're going to have to have a for loop to read in the frames of the video.

[Nelson]

Right, the mask can be the same. But regionfill would need to be applied to each frame, correct?

I'm assuming there is a for-loop internal to the feature detector function (e.g. detectORBFeatures).

I'm also thinking about pruning detected features based on their x,y positions, or segmenting several rectangular ROIs in the non-overlay portions, running the detector on each ROI, and concatenating the detected features.

Thanks for your ideas. I will try out your suggested approach, too.

[Image Analyst]

regionfill would need to be applied on each frame, however maybe you can just skip it an erase/blacken the pixels as long as your feature detector ignores black graphics. You can try it. Did it pick up the green/blue graphics?

[Nelson]

The cyan text certainly causes a lot of undesired detections, but the detector I tested seems to ignore pure-black (zeros). Maybe that's the convention for the detector, that black pixels are ignored. Thanks!

[Image Analyst]

Glad I could help and that it's working. Thanks getting back to me, and for Accepting the Answer. 🙂