Lindley
January 13th, 2009, 03:45 PM
I have an implementation of KLT interest point tracking. Basically it picks out points and then follows them for as many frames as it can using optical flow. However, sooner or later points get lost, and must be replaced or you eventually run out.
The most obvious method of handling this is to simply start over whenever the number of tracked points gets too low (with or without retaining the last lingering successful tracks). However, for my purpose this is too slow----a scan of the entire image for new interest points takes too long.
So my solution is to divide the image up into small squares (128x128), and only re-seed a few of them each frame. This is fast enough, and keeps a reasonable number of features available at all times.
The difficulty is in choosing which squares to re-seed. I have a few ideas on this front, but I was wondering if anyone knew some research papers in this area. The only reference I've found refers to a patented proprietary algorithm, which isn't much use to me.
The most obvious method of handling this is to simply start over whenever the number of tracked points gets too low (with or without retaining the last lingering successful tracks). However, for my purpose this is too slow----a scan of the entire image for new interest points takes too long.
So my solution is to divide the image up into small squares (128x128), and only re-seed a few of them each frame. This is fast enough, and keeps a reasonable number of features available at all times.
The difficulty is in choosing which squares to re-seed. I have a few ideas on this front, but I was wondering if anyone knew some research papers in this area. The only reference I've found refers to a patented proprietary algorithm, which isn't much use to me.