Archive for December, 2008

Dec 23 2008

paramater adaptation

Published by Martijn under PhD lab book

  • Background subtraction method should be checked. Removing behaviour switch at frame 30 seems to give much better bs result, however, shadow detection doesn’t work any longer. Maybe implement original zoran version?
  • Kalman parameters should be checked. Adaptation doesn’t seem to be working as it should.
  • Object blob detection should be changed. Localisation in top-down view should not only be dependent on red blob size, but complete voxelcarving profile at the location, centered on maxima.

Comments Off

Dec 19 2008

tracking on voxel carving

Published by Martijn under PhD lab book

  • Implemented Kalman tracker on the top-down voxelcarving result. Tracker slightly smoothes vc result and allows to keep location when no blob visible.
  • Parameters of Kalman filter should be tweaked. Thresholds of voxel carving and thresholding should be tweaked as well.
  • Top-down kalman kernel projected on all three camera views using voxel grid in the recGrid.
  • Function added to VoxelCarving to draw ellipses on fixed positions in the voxel grid, projected onto the camera images.
  • Problem exists with window updates when many open detail windows. No screen updates done.Look into window update/redraw functions.
  • Quality of tracker could be improved by taking into account the different histograms the voxel blobs map onto in the amera views, and select the best match for positioning kalman filter. This will be 2d tracking of 3d object.

Comments Off

Dec 12 2008

topdown projections cont’d

Published by Martijn under PhD lab book

  • Created VHL function for overlaying images on top of each other.A parameter can be used to merge images in such a way that black areas are fully replaced by the overlaid image.
  • Used overlayed images of top-down projection for combination with voxelcarving, showing positioning of voxel carving on actual image positions.
  • Added top down projection of Khan method

Comments Off

Dec 10 2008

top down projection in cassandra

Published by Martijn under PhD lab book

  • Integrated top down projections in Cassandra GUI.
  • Difficulty with image sizes and grid sizes. Sizes don’t seem to match.
  • Reason for non-matching sizes is OpenCV’s constraint of image sizes having to be devisable by two and computeGridProjections function always makuing the grid one voxel larger than given values, for all directions
  • Why is this done? computeGridCenterProjections¬† doesn’t do this. Makes managing projections and images kind of a hassle…

Comments Off

Dec 09 2008

voxelcarving correction and top down projection

Published by Martijn under PhD lab book

  • While Michael used a square grid for the Voxel carving space, it turns out that the actual space is more of a rectangle. Its proportions are more or less 7:12, which is far from square. To prevent distortion, the recGridXY_ parameter definig the grid size is split up into a recGridX_ and a recGridY_ parameter defining the depth and widthof the voxel grid.
  • The rest of the code was adapted to use a non-square voxel grid.
  • Top-down projection of the images finally works! In the end, only a transformation of the computed intersection points of lines from the camera, trough certain pixels towards image plane coordinates was necesarry. So no transformation of the ground plane basis or anything.
  • In the end, it turns out thet the gridProj_ variable already contains a precomputed mapping from ground plane pixels towards image coordinates. Only drawback is that the gridProj_ variable is kept very small (only about 50 pixels wide and 80 pixels deep) in order to make voxelcarving computationally efficient, which leeds to a huge loss of resolution when transforming images using this grid. A way could be deviced to create multiple sizes of the grid at the same time. A small one for the voxelcarving space and a larger one for image projections.

Comments Off