Updated by Alexander Shishkov about 13 years ago
Hello
I have implemented the algorithm described in Shi and Tomasi: "Good features to track" and more precisely in T. ZinÃer, C. GräÃl, H. Niemann: "Efficient Feature Tracking for Long Video Sequences". The algorithm is essentially the same as the existing affine tracker, but as Shi and Tomasi point out, for tracking features from frame to frame the purely translational model should be used, while "when monitoring features for dissimilarities in their appearance between the first and the current frame, on the other hand, the full affine motion system (5) should be solved."
The implemented functionality allows to cut out and store patches from the first image (which makes it easier to compare to the initial image without having to store all the images themselves). Those patches can then be tracked in the current image, given a good starting position. As described in Zinsser et al., one should first track with the purely translational model and then use this estimate and the affine distortion matrix from the previous frame to match under the affine model. The modified lkdemo.c demonstrates the functionality.
I would like to see this code upstream, as I would not have to worry as much about keeping it up to date with the recently quickly evolving OpenCV [[OpenCV]] library.
Kind regards
Markus Moll
I have implemented the algorithm described in Shi and Tomasi: "Good features to track" and more precisely in T. ZinÃer, C. GräÃl, H. Niemann: "Efficient Feature Tracking for Long Video Sequences". The algorithm is essentially the same as the existing affine tracker, but as Shi and Tomasi point out, for tracking features from frame to frame the purely translational model should be used, while "when monitoring features for dissimilarities in their appearance between the first and the current frame, on the other hand, the full affine motion system (5) should be solved."
The implemented functionality allows to cut out and store patches from the first image (which makes it easier to compare to the initial image without having to store all the images themselves). Those patches can then be tracked in the current image, given a good starting position. As described in Zinsser et al., one should first track with the purely translational model and then use this estimate and the affine distortion matrix from the previous frame to match under the affine model. The modified lkdemo.c demonstrates the functionality.
I would like to see this code upstream, as I would not have to worry as much about keeping it up to date with the recently quickly evolving OpenCV [[OpenCV]] library.
Kind regards
Markus Moll