29
AdaBoost & Genetic AdaBoost & Genetic algorithms: algorithms: application to application to pedestrian detection pedestrian detection Yotam Abramson Yotam Abramson Ecole des Mines de Paris Ecole des Mines de Paris 9/12/05 9/12/05 Korea-France SafeMove Workshop

AdaBoost & Genetic algorithms: application to pedestrian detection Yotam Abramson Ecole des Mines de Paris 9/12/05 Korea-France SafeMove Workshop

Embed Size (px)

Citation preview

AdaBoost & Genetic AdaBoost & Genetic algorithms: application to algorithms: application to

pedestrian detectionpedestrian detection

Yotam AbramsonYotam Abramson

Ecole des Mines de ParisEcole des Mines de Paris

9/12/059/12/05

Korea-France SafeMove Workshop

Korea-France SafeMove Workshop

Machine learning for visual object Machine learning for visual object detectiondetection

Korea-France SafeMove Workshop

ApplicationApplication

Pedestrian impact predictorPedestrian impact predictor

Calculates the probability of an impact between Calculates the probability of an impact between our car and a pedestrian.our car and a pedestrian.

If the probability is higher then a given If the probability is higher then a given threshold, an alert to the driver is issued or an threshold, an alert to the driver is issued or an action is taken (pedestrian airbag, braking…)action is taken (pedestrian airbag, braking…)

Korea-France SafeMove Workshop

Machine learning for visual object Machine learning for visual object detectiondetection

Learning algorithms for object-detection Learning algorithms for object-detection were shown to be better than any hand-were shown to be better than any hand-crafted ones.crafted ones.

Main works in the field:Main works in the field: Papageorgiou & Poggio – SVM,wavelets.Papageorgiou & Poggio – SVM,wavelets. Viola & Jones – AdaBoost and simple Viola & Jones – AdaBoost and simple

features.features.

Korea-France SafeMove Workshop

Machine learning - backgroundMachine learning - background

Support Vector Machine (SVM) – Vapnik Support Vector Machine (SVM) – Vapnik 19901990

Neural networkNeural network

Korea-France SafeMove Workshop

Machine learning Machine learning (Cont.)(Cont.)

AdaBoost (Freund & Schapire 1995):AdaBoost (Freund & Schapire 1995): A popular learning algorithm.A popular learning algorithm. Easy to understand.Easy to understand. Received a lot of attention in the machine Received a lot of attention in the machine

learning and statistics communities.learning and statistics communities.

The notion of The notion of boosting boosting (AdaBoost = (AdaBoost = adaptive boosting).adaptive boosting).

Korea-France SafeMove Workshop

AdaBoost at a GlanceAdaBoost at a Glance

Assume that we have a simple object Assume that we have a simple object classifier, that receives a rectangle in the classifier, that receives a rectangle in the image and decides if it’s the object.image and decides if it’s the object. For exampleFor example::

Korea-France SafeMove Workshop

AdaBoost at a Glance (Cont.)AdaBoost at a Glance (Cont.)

A classifier like the one shown is called a A classifier like the one shown is called a weak classifier.weak classifier. And indeed it is weak.. And indeed it is weak..

AdaBoost selects (=learns) a set of AdaBoost selects (=learns) a set of classifiers and builds a “voting system”.classifiers and builds a “voting system”.

Weak1 Weak2 Weak3 Weak4

Yes Yes No YesYes

Korea-France SafeMove Workshop

AdaBoost at a Glance (Cont.)AdaBoost at a Glance (Cont.)

Voting is not “democratic”… there is a Voting is not “democratic”… there is a weight for each weak-classifier.weight for each weak-classifier.

Weak1 Weak2 Weak3 Weak4

Yes Yes NO YesNo

0.2 0.7 0.20.2

Korea-France SafeMove Workshop

AdaBoost at a Glance (Cont.)AdaBoost at a Glance (Cont.)

The output of AdaBoost is a called a The output of AdaBoost is a called a strong classifier. strong classifier.

AdaBoost was used for face, cars and AdaBoost was used for face, cars and pedestrian detection by viola and Jones pedestrian detection by viola and Jones (2000).(2000).

Korea-France SafeMove Workshop

Weak ClassifiersWeak Classifiers

Viola & Jones

Korea-France SafeMove Workshop

We have developed new We have developed new kinds of weak classifiers.kinds of weak classifiers.

Our features are different Our features are different because because they test they test individual pixelsindividual pixels..

They deal better with the They deal better with the variation in illumination.variation in illumination.

Weak classifiersWeak classifiers

Korea-France SafeMove Workshop

Illumination independent features Illumination independent features (cont.)(cont.)

Our features are highly Our features are highly efficient (3-4 image efficient (3-4 image access operations)access operations)

2 times faster than 2 times faster than Viola&JonesViola&Jones

20% of the memory20% of the memory

BetterBetter detection rates for detection rates for pedestrianspedestrians

Korea-France SafeMove Workshop

Learning Learning processprocessusingusinggeneticgeneticalgorithmalgorithm

Korea-France SafeMove Workshop

SevilleSeville

SEmi-SEmi-automatic automatic VIsuaL VIsuaL LearningLearning

(With Dr. Yoav (With Dr. Yoav Freund, Freund, Columbia Columbia University)University)

Korea-France SafeMove Workshop

SevilleSeville

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

Korea-France SafeMove Workshop

SevilleSeville

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

Korea-France SafeMove Workshop

SevilleSeville

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

Korea-France SafeMove Workshop

SevilleSeville

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

Korea-France SafeMove Workshop

SevilleSeville

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

Korea-France SafeMove Workshop

European projectEuropean projectCAMELLIACAMELLIA

European unionEuropean union

Renault, Philips, Philips Renault, Philips, Philips semiconductor, Uni. semiconductor, Uni. Hannover, Uni. Las-palmasHannover, Uni. Las-palmas

“Smart camera”“Smart camera”

CAMELLIACAMELLIA

Korea-France SafeMove Workshop

ResultsResults

Korea-France SafeMove Workshop

ResultsResults

Korea-France SafeMove Workshop

Impact predictionImpact prediction

Korea-France SafeMove Workshop

Prediction resultsPrediction results

Korea-France SafeMove Workshop

Prediction resultsPrediction results

Korea-France SafeMove Workshop

CAMELLIA was used CAMELLIA was used also for other applicationsalso for other applications

Korea-France SafeMove Workshop

ConclusionsConclusions

We have presented a system for detection We have presented a system for detection of pedestrians.of pedestrians.

The system is based on AdaBoost and The system is based on AdaBoost and Genetic algorithms.Genetic algorithms.

The system was tested and gives good The system was tested and gives good results on real data.results on real data.

Korea-France SafeMove Workshop

Thank you for your attentionThank you for your attention