Image Based Localization
Experiment with Mobile robot

Josef Sivic and Tomas Pajdla
Center for Machine Perception
Czech Technical University in Prague
http://cmp.felk.cvut.cz

The mobile robot with a catadioptric panoramic camera.	The detail view of the camera.	The panoramic image shows how the camera sees itself and occludes a part of the ceiling.

The panoramic image warped onto the cylinder. Cylindrical panoramic images contain only the scene, not the camera. This is the image sequence A.

An example of the similarity matrix C. Element C(i,j) represents similarity between i-th image taken along the Route A and j-th image taken along the Route B. (Dark areas denote low similarity).	Two image sequencies were taken. Red and blue lines show the approximate path of the robot.

Description of the experiment

• See the research report: will be here

Data

• Video sequences
  Available upon request. 2 CD's, avi format.

• Images
  • Original images extracted from avi file (bmp): 14, 15
    • Scripts (mainActor for Windows) used for extracting the images from avi sequence

  • Manualy derotated images (or using conventional magnetic compass): 14, 15
  • Images derotated using ZPR on mean of rows: 14, 15
  • Images derotated using ZPR on weighted mean of rows: 14, 15

• Odometry data
  • corrected odometry data (including m-files): here,
  • azimut data for zprmnr, zprmwr and manualy derotated images: here
  • A simple map of the corridor(for matlab) is here

• Correlation in spatial domain
  Correlation matrices computed using normalized correlation of:
  • manualy derotated images: manderot
  • original images: orig
  • images derotated by ZPR mean: zprmnr
  • images derotated by ZPR weighted mean: zprmwr

• Correlation in frequency domain
  Correlation matrices of original images computed in frequency domain using search for the best correlation in:
  • whole image (in both row and column directions): here
  • row direction: here
  • sum of columns: here
  Matrices containing row index of maximum correlation found in:
  • whole image (contains both row and column indices)
  • row direction
  • sum of columns

• Proximity matrices: proxmat
  Element C(i,j) represents normalized distance between positions where i-th and j-th images were taken. If C(i,j)=1 then distance between i-th and j-th position is zero.

• All matlab m-files used for experiments: m-files

References

• Pajdla, Tomas,
  Robot Localization Using Shift Invariant Representation of Panoramic Images,
  Research Report K335-CMP-1998-170, Czech Technical University, Prague, November 1998
  PostScript

  
  
  

• Sivic, Josef and Pajdla, Tomas,
  Interactive Control of a Mobile Robot with Panoramic Camera,
  Technical Report K335-CMP-1998-165, Czech Technical University, Prague, August 1998
  PostScript

Josef Sivic
sivic@cmp.felk.cvut.cz