Image processing

Image processing

Benjamin Cottyn
Wouter Deryckere
Sebastiaan Dumoulein
Thomas Planckaert
Pieter-Jan Van Robays


Link to presentation:
http://beeldverwerking.vanrobays.be

download the paper

Image processing

Table of contents

  1. Situation
  2. Approach
  3. Features
  4. Segmentation
  5. SVM
  6. Markov
  7. Results & Findings
  8. Conclusion & Future Work

1. Situation

  • Helping the visually impaired
  • Robustness important
  • Different circumstances
robustness
Image processing

2. Approach

  • Look for recognizable properties
  • Fit for a feature?
  • Segmentation of the route
  • Train and use SVMs
  • Positioning via a Markov chain
Image processing

3. Features

  • Detectable characteristics of the footage
  • Based on color, lines, shapes or size

Feature Extraction

  • Inspect original footage
  • Select every recognizable property
  • Organize different properties by type
  • Convertable to possible feature?

Feature Selection

  • Bad
    • Not programmatically recognizable
    • Too short in time
    • Too general
    • Unnecessary features
  • Good
    • Unique features for a segment
    • Luminance invariant
    • Rotation invariant
    • Constant presence

Our Features

Yellow Borders

featureBoordsteen

Our Features

Grass

featureGras

Our Features

Distance

featureDistance

Our Features

Tiles: Ratio & Size

featureTegels

Our Features

Tiles: Relative Positioning

featureTegelOrientatie

Our Features

Tram Tracks

featureTram
Image processing

4. Segmentation

  • Smallest part of route, with same characteristics
  • Two-way relation with features
  • User present in only one segment

Our Segments

Image processing
Image processing

5. Support Vector Machine

  • Ability to train SVM with our features
  • Easier to add new features in the project

iFeature header

iFeatureScheme
Image processing

6. Markov Chain

  • State machine with percentage transitions
  • Markov Property: memoryless
  • Segments as states
  • Features as transitions
  • Transitions depend on travelled distance

Markov chain

Markov Chain

Transitions depend on travelled distance

7. Results & Findings

Results

  • Visual interface & Trackbar
  • Estimation of the location

Findings

  • Features depend on external parameters
  • Binary tree changed to state machine
  • Error correction with travelled distance
  • Modular implementation: easy to expand
Image processing

8. Conclusions & Future Work

  • Good result, given the circumstances
  • Not yet fully employable
  • Use multiple features for a transition
  • Needs some tweaking, but easily expandable
Image processing

Questions?

Shoot!

Image processing

Thanks!

Benjamin Cottyn
Wouter Deryckere
Sebastiaan Dumoulein
Thomas Planckaert
Pieter-Jan Van Robays