Docstoc

1 Real-Time Stereo-Matching for Micro Air Vehicles Pascal Dufour

Document Sample
1 Real-Time Stereo-Matching for Micro Air Vehicles Pascal Dufour Powered By Docstoc
					  Midterm Master Thesis Presentation


Real-Time Stereo-Matching
  for Micro Air Vehicles

              Pascal Dufour
               11.12.2009




                    1
                     Outline

•   Introduction

•   Stereo-Matching in Practice

•   Future work

•   Demo


                           2
Requirements
•   Object avoidance on a MAV
    •   Stereo matching with two
        cameras

•   Resources are limited
    •   Same hardware as iPhone

    •   Onboard DSP

•   Robustness

                                   3
                               Stereo Vision

                QuickTime™ an d a
                GIF decompressor
          are need ed to see this p icture.




•   Similar to human stereoscopy

•   Closer objects are shifted more if the eye is
    moved
                                              4
• The objects shift
 along the horizontal
 axis determines the
 depth of the object




                        5
     In Practice
•   A complete system is
    needed!

1.Image acquisition

2.Disparity computation

3.Post-processing

4.Publishing of results
               6
                Image Acquisition
•   Baseline
    •   tradeoff between runtime and distance of last disparity before
        background

•   Camera shutter synchronization

•   Camera exposure synchronization



                                     7
                 Undistortion and
                  Rectification
•   Images are distorted

•   Undistort

•   Rectify to get epipolar lines

•   Floating point is slow on
    Overo!
    •   own fixed-point routine for DSP
        would be nice
                                    8
Matching Scanlines




        9
From scanlines to Disparity

•   Find corresponding
    pixels

•   The shift defines
    the disparity



                         10
                Block Matching
•   Instead of finding
    corresponding
    pixels, find blocks

•   Difficulties:

    •   Low-texture
        regions

    •   No vertical
        variance

    •   Repeating textures   11
•   Efficient with sliding
    windows

•   Compute only changes


•   Can be applied
    horizontally and
    vertically
•   Runtime is
    independent from
    window-size

                             12
• Sum    of absolute differences (SAD)

  •   Similarity is the difference in pixel intensities

• Sum    of squared differences (SSD)

  •   Similarity is the squared difference of pixel
      intensities

  •   More robust to differently exposed images


                             13
14
       Dynamic Programming

•   Optimization problem

•   Best possible match

•   ‘Smoothes’ over low
    texture regions


                           15
                  on the DSP
•   DP is about as fast as block matching, but has a lot
    of conditionals, not suitable for DSP

•   Block matching algorithms can be vectorized on
    the DSP

•   If exposure is the same in both pictures, SAD and
    SSD give equivalent results

•   Tradeoff of window sizes
                            16
             Post Processing




•   Robustness can be increased with post processing

                           17
             Confidence Map




•   Remove areas with low textures

•   Variance in image defines confidence
                           18
             Median Filtering



•   Remove out of place pixels

•   3x3 Median Filtering provided by DSP library

•   5x5 or 7x7 would be nice for robustness
                           19
                   Difficulties

•   Horizontal lines can give wrong matches

•   Occlusions

•   Gradients, reflections

•   Differences in images because of self mounted
    lenses

                             20
                   Future Work
•   Implementation DSP

    •   Undistortion and rectification

    •   SAD

    •   Confidence map

•   System integration

•   Realistic testing
                               21
Demo



 22

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:15
posted:8/12/2011
language:English
pages:22