Fast Object Detection Method Based on HaarLike Feature Classifier Cascade and AdaBoost Learning

1
Fast Object Detection Method Based on Haar-Like
Feature Classifier Cascade and AdaBoost Learning

• Paul Viola, Rainer Lienhart
• Presenter Haifeng Gong

NLPR, IA, CAS
2
Contents of This Presentation

• Object Detection Method Based on Haar-Like
  Feature Classifier Cascade and AdaBoost Learning
  originated by Paul Viola and Rainer Lienhart
  Feature Representation and Calculation, Feature
  Selection, Classifiers Training and Cascade
  Training
• Finger Mouse My experiment of this method

3
The Framework of the Method

• Haar-like Feature A feature representation
• Integral Image A new image representation and a
  fast Haar-like feature calculation method
• Perceptron Classifier made of features
• Boost Algorithm An efficient classifier
  generating algorithm
• Cascade Method for combining classifiers

4
Haar-like Features

• Example features shown relative to the enclosing
  detection window A, B, feature of 2 rectangles,
  C, D, feature of 3, 4 rectangles.
• The feature value will be calculated by the
  weighted sum of the pixels cover by the white
  rectangles subtracted by the weighted sum of the
  pixels covered by the dark rectangles

5
Haar-like Features Extension

• 45 rotated rectangle feature by Lienhart
• Feature is represented by (type, x, y, w, h)
  type 1(a), 2(b), etc. (x, y) for location in
  detection window and (w, h) for feature size
• Number of features give the window size 24x24,
  there are 117,941 different features, a large
  number

6
Integral Image, Fast Feature Calculation

• Corresponding to the origin feature and 45
  rotated feature, there are two types of Integral
  Image, called by Lienhart as Sum Area Table and
  Rotated Sum Area Table

7
Integral Image, Fast Feature Calculation

• SAT and RSAT Calculation
• Sum of Rect Calculation
• Feature value calculated as the weighted sum of
  the rectangles

8
Integral Image, Fast Feature Calculation

• Detail of the calculation of 45 rotated
  rectangle
• Use the four vertexes RSAT values
• (x-1,y-1)
• (x-h-1,yh-1)
• (xw-1, yw-1)
• (xw-1-h, yw-1h)

9
Classifier

• A classifier is a single layer perceptron
  (right).
• This classifier is a weak classifier with
  extemely high hit rate (nearly 100) and high
  false alarm rate (e.g. 3040).

10
Feature Selection

• Example images set (x1,y1),,(xn,yn) where y-1,
  1 for negative and positive examples
• Init example weights
• For t 1, , T
• Normalize the weights,
• For each feature j, train a classifier fj which
  is restricted to using a single feature. Compute
  the error as
• Choose the classifier, ht, with the lowest error
• Update the weights
• Where ei 0 if xi is classified correctly, ei 1
  otherwise
• Classifier

11
Feature Selection

• Trained Feature for Face Detection

The first and second features selected by AdaBoost
12
Cascade

• A detector is a cascade of classifiers

13
Cascade Training

• F Max acceptable false alarm rate per layer d
  Min acceptable hit rate per layer Ftarget
  overall false alarm rate.
• P positive examples set N negative examples
  set
• Fi, Di false alarm rate and hit rate of layers
  1i
• i0 F01.0 D0 1.0
• Loop while FigtFtarget
• i i1 ni0 Fi Fi-1
• While Fi gtf x Fi-1
• Use P and N to train a classifier with ni
  features
• Evaluate Fi and Di of current cascade
• Decrease threshold for the i-th classifier to
  have a hit rate of at least d x Di-1
• N Empty
• If Fi gtFtarget, N false detections of current
  cascade on negative examples

14
Implementation in OpenCV

• LoadHaarClassifierCascade
• ReleaseHaarClassifierCascade
• CreateHidHaarClassifierCascade
• ReleaseHidHaarClassifierCascade
• HaarDetectObjects
• SetImagesForHaarClassifierCascade
• RunHaarClassifierCascade
• GetHaarClassifierCascadeScale
• GetHaarClassifierCascadeWindowSize

15
Face Detection Results
MITCMU
16
Face Detection Results
Test by myself in OpenCV
17
Finger Mouse

• Purpose Use the finger movement to control mouse
  cursor
• Detect and track finger nails in video sequence
• Finger nail has not so rich features as face
• But the background on the desk is relatively
  simple

18
Finger Mouse

• Implementation Details
• Convert image to HSV color space and extract S
  component to eliminate the shadow
• Use Canny filtered S component image for
  Haar-like feature object detection
• Use ellipse fitting to refuse some false alarm
• Use Kalman filter to tracking and to refuse
  clutter

19
References

• Paul Viola, Michael J.Jones, Robust Real- time
  Object Detection, Cambridge Research Laboratory
  Technology Reports Series, CRL 2001/01, Feb, 2001
• Paul Viola, Michael J.Jones, Robust Real- time
  Object Detection, ICCV 2001
• Paul Viola and Michael Jones, Fast and Robust
  Classification using Asymmetric AdaBoost and a
  Detector Cascade, Neural Information Processing
  Systems 14, December 2001
• Rainer Lienhart, Luhong Liang, and Alexander
  Kuranov, A Detector Tree of Boosted Classifiers
  for Real-time Object detection and Tracking, ICME
  2003
• Rainer Lienhart and Jochen Maydt, An Extended Set
  of Haar-like Features for Rapid Object Detection,
  ICIP2002
• Rainer Lienhart, Alexander Kuranov, Vadim
  Pisarevsky, Empirical Analysis of Detection
  Cascades of Boosted Classifiers for Rapid Object
  Detection, Intel Labs MRL Technology Report, May
  2002
• Rainer Lienhart, Alexander Kuranov, Vadim
  Pisarevsky, Empirical Analysis of Detection
  Cascades of Boosted Classifiers for Rapid Object
  Detection, DAGM2003

20
- The End -

• Thanks

21
AdaBoost Training Method

• AdaBoost is used both to select the features and
  to train the classifiers.

22
AdaBoost Training Method

• Another AdaBoost

23
2D Haar Function