1 / 34

CAP 5415 Computer Vision Fall 2004

CAP 5415 Computer Vision Fall 2004. Dr. Alper Yilmaz Univ. of Central Florida www.cs.ucf.edu/courses/cap5415/fall2004 Office: CSB 250. Recap Circle Fitting. Circle equation Change x 0 , and y 0 compute r Change only r and compute ( x 0 , y 0 ) from image gradient angle 

lotus
Download Presentation

CAP 5415 Computer Vision Fall 2004

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CAP 5415 Computer VisionFall 2004 Dr. Alper Yilmaz Univ. of Central Florida www.cs.ucf.edu/courses/cap5415/fall2004 Office: CSB 250 Alper Yilmaz, Fall 2004 UCF

  2. RecapCircle Fitting • Circle equation • Change x0, and y0 compute r • Change only r and compute (x0, y0) from image gradient angle  • Increment (x0, y0 , r) in accumulator array • Find the local maxima Alper Yilmaz, Fall 2004 UCF

  3. RecapGeneralized Hough Transform • For shapes with no analytical expression • Requires learning of r-table • Distance vector and its angle for each boundary pixel to object centroid • Finding a shape in an image • For input edge-map compute angle (can be from image gradient) • For each distance vector corresponding to edge angle increment a 2D accumulator array corresponding to (x0, y0) Alper Yilmaz, Fall 2004 UCF

  4. RecapMedial Axis Transform • Represents object shape: skeleton of an object • Computed using an iterative algorithm, • Inverse transform exist Alper Yilmaz, Fall 2004 UCF

  5. RecapInterest Points • High texture variation in neighborhood of a pixel • Movarec’s operator • Compute directional intensity variations in 4x4 neighborhood • Pick points with high intensity variations in a neighborhood • Harris corner detector • Compute a moment matrix M from gradients in a neighborhood • Min eigenvalue of M higher than a threshold indicates corner Alper Yilmaz, Fall 2004 UCF

  6. Object Motion Alper Yilmaz, Fall 2004 UCF

  7. Motion • Projection of object motion in real-world (3D)results in motion in image plane (2D) • 2D motion is defined over a sequence of frames • Computed by using brightness constancy constraint • Intensity of a moving pixel does not change over-time Alper Yilmaz, Fall 2004 UCF

  8. What is its use? • Lots of uses • Motion Detection • Track object behavior • Correct for camera jitter (stabilization) • Align images (mosaics) • 3D shape reconstruction • Video Compression Alper Yilmaz, Fall 2004 UCF

  9. Measurement of motion at every pixel Alper Yilmaz, Fall 2004 UCF

  10. Measurement of motion at every pixel Alper Yilmaz, Fall 2004 UCF

  11. Visual Mosaics Alper Yilmaz, Fall 2004 UCF

  12. Visual Mosaics Alper Yilmaz, Fall 2004 UCF

  13. Geo Registration Alper Yilmaz, Fall 2004 UCF

  14. Video Segmentation Alper Yilmaz, Fall 2004 UCF

  15. Structure From Motion Alper Yilmaz, Fall 2004 UCF

  16. Optical Flow Alper Yilmaz, Fall 2004 UCF

  17. Optical Flow • Flow vector in image space (2D) • Taylor series expansion of right side around t Alper Yilmaz, Fall 2004 UCF

  18. Optical Flow Brightness constancy equation Equation of a line in (u,v) space Alper Yilmaz, Fall 2004 UCF

  19. Optical Flow • We know Ix, Iy, and It from images • For every point these values provide one equation • We have 2 unknowns: u, v • Solution lie any where on the line v u Alper Yilmaz, Fall 2004 UCF

  20. d . Optical Flow v p • Let (u’, v’) be true flow • True flow has two components • Normal flow: d • Parallel flow: p • Normal flow can be computed • Parallel flow cannot u Alper Yilmaz, Fall 2004 UCF

  21. Computing True Flow • Horn & Schunck • Schunk • Lukas & Kanade Alper Yilmaz, Fall 2004 UCF

  22. Horn & Schunck • Define an energy function and minimize • Differentiate w.r.t. unknowns u and v laplacian of u laplacian of v Alper Yilmaz, Fall 2004 UCF

  23. Horn & Schunck • Laplacian controls smoothness of optical flow • A particular choice can be 2u=u-uavg, 2v=v-vavg. • Rearranging equations • 2 equations 2 unknowns • Write v in terms of u • Plug it in the other equation Alper Yilmaz, Fall 2004 UCF

  24. Horn & Schunck • Iteratively compute u and v • Assume initially u and v are 0 • Compute uavg and vavg in a neighborhood Alper Yilmaz, Fall 2004 UCF

  25. Schunck • If two neighboring pixels move with same velocity • Corresponding flow equations intersect at a point in (u,v) space • Find the intersection point of lines • If more than 1 intersection points find clusters • Biggest cluster is true flow v u Alper Yilmaz, Fall 2004 UCF

  26. Lucas & Kanade • Similar to line fitting we have seen • Define an energy functional • Take derivatives equate it to 0 • Rearrange and construct an observation matrix Alper Yilmaz, Fall 2004 UCF

  27. Lucas & Kanade 1 Alper Yilmaz, Fall 2004 UCF

  28. Lucas & Kanade 2 Alper Yilmaz, Fall 2004 UCF

  29. Discussion • Horn-Schunck and Lucas-Kanade optical method works only for small motion. • If object moves faster, the brightness changes rapidly, derivative masks fail to estimate spatiotemporal derivatives. • Pyramids can be used to compute large optical flow vectors. Alper Yilmaz, Fall 2004 UCF

  30. u=1.25 pixels u=2.5 pixels u=5 pixels u=10 pixels image It image It+1 Gaussian pyramid of image It Gaussian pyramid of image It+1 Coarse to Fine Optical Flow Estimation Alper Yilmaz, Fall 2004 UCF

  31. Alper Yilmaz, Fall 2004 UCF

  32. Alper Yilmaz, Fall 2004 UCF

  33. Alper Yilmaz, Fall 2004 UCF

  34. Alper Yilmaz, Fall 2004 UCF

More Related