Trajectories

[NatalyaSlepkova]

Thanks for your project! How I understand reading your code, you build trajectories of video frames use only dx and dy (i.e. H[0, 2] and H[1, 2]). I try to build trajectory of frames using his center point (x, y). So new uniform coordinates of this point in next frame are p' = H * (x, y, 1)^T. And then Euclidean coordinates are (p'[0] / p'[2], p'[1] / p'[2]). But so I get very strange plots. Could you explain me, please, where my mistake?

[AdamSpannbauer]

How I understand reading your code,

Hopefully, it wasn't too hard to understand. I hope to refactor with a focus on improving readability when I get the time.

you build trajectories of video frames use only dx and dy (i.e. H[0, 2] and H[1, 2]).

The elements of trajectory are built using the vidstab_utils.estimate_partial_transform() helper function which is just a convenience wrapper of cv2.estimateRigidTransform()¹. This helper returns [dx, dy, da]; where da = np.arctan2(H[1, 0], H[0, 0]) and dx & dy are as you described.

I try to build trajectory of frames using his center point (x, y). So new uniform coordinates of this point in next frame are p' = H * (x, y, 1)^T. And then Euclidean coordinates are (p'[0] / p'[2], p'[1] / p'[2]). But so I get very strange plots. Could you explain me, please, where my mistake?

Could you provide example code or example input-output of what you're doing? It would be very helpful in understanding the issue.

---

EDIT

¹ cv2.estimateRigidTransform() is used if using OpenCV version < 4.0.0. This function was deprecated as of OpenCV 4.0.0 in favor of cv2.estimateAffinePartial2D() (which is used in vidstab if using 4.0.0)

[NatalyaSlepkova]

You could see code https://github.com/NatalyaSlepkova/VideoTrajectories TrajectoriesCenterPoint.py is code, stable.avi is example video, centersPlot.png is plot for this video - you could see it in README