AdaBoost

1
AdaBoost Genetic algorithms application to
pedestrian detection

• Yotam Abramson
• Ecole des Mines de Paris
• 9/12/05

Korea-France SafeMove Workshop
2
Machine learning for visual object detection
3
Application

• Pedestrian impact predictor
• Calculates the probability of an impact between
  our car and a pedestrian.
• If the probability is higher then a given
  threshold, an alert to the driver is issued or an
  action is taken (pedestrian airbag, braking)

4
Machine learning for visual object detection

• Learning algorithms for object-detection were
  shown to be better than any hand-crafted ones.
• Main works in the field
• Papageorgiou Poggio SVM,wavelets.
• Viola Jones AdaBoost and simple features.

5
Machine learning - background

• Support Vector Machine (SVM) Vapnik 1990
• Neural network

6
Machine learning (Cont.)

• AdaBoost (Freund Schapire 1995)
• A popular learning algorithm.
• Easy to understand.
• Received a lot of attention in the machine
  learning and statistics communities.
• The notion of boosting (AdaBoost adaptive
  boosting).

7
AdaBoost at a Glance

• Assume that we have a simple object classifier,
  that receives a rectangle in the image and
  decides if its the object.
• For example

8
AdaBoost at a Glance (Cont.)

• A classifier like the one shown is called a weak
  classifier. And indeed it is weak..
• AdaBoost selects (learns) a set of classifiers
  and builds a voting system.

Yes
No
Yes
Yes
Yes
Weak1
Weak2
Weak3
Weak4
9
AdaBoost at a Glance (Cont.)

• Voting is not democratic there is a weight for
  each weak-classifier.

Yes
NO
Yes
No
Yes
Weak1
Weak2
Weak3
Weak4
0.2
0.7
0.2
0.2
10
AdaBoost at a Glance (Cont.)

• The output of AdaBoost is a called a strong
  classifier.
• AdaBoost was used for face, cars and pedestrian
  detection by viola and Jones (2000).

11
Weak Classifiers
Viola Jones
12
Weak classifiers

• We have developed new kinds of weak classifiers.
• Our features are different because they test
  individual pixels.
• They deal better with the variation in
  illumination.

13
Illumination independent features (cont.)

• Our features are highly efficient (3-4 image
  access operations)
• 2 times faster than ViolaJones
• 20 of the memory
• Better detection rates for pedestrians

14
Learning processusinggeneticalgorithm
15
Seville

• SEmi-automatic VIsuaL Learning
• (With Dr. Yoav Freund, Columbia University)

16
Seville

• We start by collecting 10 negative and positive
  examples. We run the learning, and classify.

17
Seville

• We now have 100 examples. We run learning, and
  the results improve.

18
Seville

• We test another sequence. We collect in the same
  way more examples. We re-run the learning and
  continue.

19
Seville

• We test another sequence. We collect in the same
  way more examples. We re-run the learning and
  continue.

20
Seville

• Throughout the phases, we use 2/3 of the set as
  training set, and 1/3 as validation set. We make
  AdaBoost rounds until the point of overfitting.

21
European projectCAMELLIA
European unionRenault, Philips, Philips
semiconductor, Uni. Hannover, Uni. Las-palmas
Smart camera
22
Results
23
Results
24
Impact prediction
25
Prediction results
26
Prediction results
27
CAMELLIA was used also for other applications
28
Conclusions

• We have presented a system for detection of
  pedestrians.
• The system is based on AdaBoost and Genetic
  algorithms.
• The system was tested and gives good results on
  real data.

29

• Thank you for your attention