1. Genetic Algorithms for motion estimation
Sophie Voisin
Imaging, Robotics, & Intelligent Systems Laboratory
The University of Tennessee
October 12, 2004

2. Outlines Le2i laboratory My previous work My future work at IRIS
September 21, Slide 2

3. Le2i Le2i : www.le2i.com 2 sites : Dijon and Le Creusot
Laboratoire Électronique
Informatique et Image
2 sites : Dijon and Le Creusot
September 21, Slide 3

4. Le2i Le2i : www.le2i.com 3 teams - 3 different ways :
Laboratoire Électronique
Informatique et Image
3 teams - 3 different ways :
Electronics, Architecture and Material Integration
Dijon, Le Creusot
Material
Signal Processing and Image Processing
Processing
Computer Science and Image
Dijon
Theory
September 21, Slide 4

5. My previous works DEA* research 4 months (full time) Advisor :
Doctor Albert Dipanda**
Application of Genetic Algorithms
for motion estimation based on Markov random fields :
The aim of this research was to find how we can use genetic algorithms for motion estimation, to implement this algorithm and to compare it to different existed ones.
*Master Degree * *Processing team, Dijon
September 21, Slide 5

6. DEA research Assumption Optical flow
In image sequence the color for one point of object is the same all over the time
no estimation is possible
estimation is possible
September 21, Slide 6

7. DEA research Genetic Algorithms The theory Principle
Genetic algorithms are inspired by
Darwin's theory of evolution
They were applied for the first time by
John Holland 1975
Initial population
Estimation (fitness)
Reproduction
Reproduction / Elitism
Selection
Next population / Offspring
Crossover
Estimation (fitness)
Mutation
Stop criterion no
yes
Solution
September 21, Slide 7

8. DEA research Genetic Algorithms The theory Encoding of a Chromosome
Representation of possible solution
Fitness
To estimate an individual in a population
Operators for reproduction
Selection
Crossover
Mutation
September 21, Slide 8

9. DEA research Genetic Algorithms How use them for motion estimation
Encoding of a Chromosome
Values
Vectors with amplitude equals to 0 or 1 (approximately)
2-D structure
in general 1-D structure is used
(1,-1)
(0,-1)
(-1,-1)
(1,0)
(0,0)
(-1,0)
(1,1)
(0,1)
(-1,1)
Pixel (i, j)
Pixel
(i+1, j+1)
Image Result
Image 1
Image 2
September 21, Slide 9

10. DEA research Genetic Algorithms How use them for motion estimation
Fitness
Function of energy based on optical flow and Markov random fields
*Displaced frame difference **Regularization Tikhonov or Mc Clure
September 21, Slide 10

11. DEA research Genetic Algorithms How use them for motion estimation
Operators
Selection
Tournament
September 21, Slide 11

12. DEA research Genetic Algorithms 1 How use them for motion estimation
Operators
Crossover
Chromosomes
2 points
Linear combination
Parents
Children
Children 1 1
Key 2
September 21, Slide 12

13. Mutation by small variation
DEA research
Genetic Algorithms
How use them for motion estimation
Operators
Mutations
Individual
Random mutation
Mutation by small variation
Directed mutation ?
September 21, Slide 13

14. DEA research First Tests Key Zoom x1 x4 x16 x64 Height x Width 64 x 64
Individuals per population
501
151 75
Number of generation
5,000
2,500
10,000
1,000
Explored individuals (possible solutions)
2,505,000
377,000
1,510,000
75,000
Time
2.4 hours
5.5 minutes
6.4 minutes
44.4 seconds
September 21, Slide 14

15. DEA research Improvement Results To Split images in “sub-images”
First cut in blue
Second cut in red
Results
10 minutes
64x64 images
Non homogenous background
Key
First cut Second cut Result
September 21, Slide 15

16. Results and Conclusions
Key
Results and Conclusions
Initialization
Method of Gradient
Iterative Conditional Mode
Simulated Annealing
Step
Genetic Algorithms
Random blocks of motion
Time: 1.9 s
Time: 1.5 s
Time: 5.7 s
Random motion for each point
First
Time: 5 min
Null motion for each point
Time: 2.5 s
Time: 1.8 s
Time: 6.2 s
Second
Time: 4 min
Time: 2.1 s
Time: 2 s
Time: 7.3 s
Final
Time: 9 min
Deterministic algorithms : initialization dependence
Simulated Annealing : chaotic motion for homogenous background
Genetic Algorithms : no real time
circle rotation
square translation
September 21, Slide 16

17. DEA research Results : “train sequence” 400 x 512 image
September 21, Slide 17

18. DEA research Results : “train sequence” 400 x 512 image
September 21, Slide 18

19. DEA research Results : “train sequence” 400 x 512 image
Method of Simulated
Gradient Annealing
46 s min
Genetic
ICM Algorithms
19 s h30 min
Key
September 21, Slide 19

20. DEA research The Most Important References
S . Woo - Reconstruction de Surface 3D et Estimation du Mouvement Rigide 3D à Partir d’un Système de Stéréovision Active (Application des Algorithmes Génétiques à l’ Analyse d’images). Thèse soutenue en 2003.
P. Perez - Champs de Markoviens et Analyse Multirésolution de l’ Image : Application à l’ Analyse de Mouvement. Thèse soutenue en 1993.
A. Dipanda, L. Legrand - Multiresolution Motion Estimation with Discontinuities Preservation Using MRF and Determination of the Regularization Hyperparameter. SDII, 1999, San José.
A. Dipanda, S. Woo, F. Marzani, J-M. Bilbault - 3-D Shape Reconstruction in an Active Stereo Vision System Using Genetic Algorithms. Pattern Recognition 36 (2003)
M. Zaim, A. El ouaazizi, R. Benslimane - Genetic Algorithms Based Motion Estimation
L. Li, S. J. Louis, J. N Brune - Application of Genetic Algorithms to 2-D Velocity of Seismic Refraction Data. In Proceedings of the Third Golden West International Conference on Intelligent Systems. Kluwer Academic Press. (to appear 1995).
September 21, Slide 20

21. My future work at IRIS To have my PhD Thesis French advisors :
Professor Frederic Truchetet
Doctor Sebti Foufou
September 21, Slide 21

22. 3-D Reconstruction and Segmentation
My future work at IRIS
3-D Reconstruction and Segmentation
Task 1
Topic
Applications
Task 2
Sate of the Art
Contribution
Task 3
Research
Implementation
Task 4
Results / Conclusion
Discussions / Future Research
September 21, Slide 22

23. Questions ?
Thanks
September 21, Slide 23