Vehicle Detection Using Adaboost

1
Vehicle DetectionUsing Adaboost

• Shane Brennan
• UCSC
• June 2005

2
Interface

• Input A set of features (user-defined). An
  array of positive image samples. An array of
  negative image samples. A distance function. A
  target detection rate for the cascade, OR A
  target false detection rate for the cascade.
• Output A strong classifier using each of the
  features given as input.

3
Feature Representation

• Features are not Haar-type Filters.
• Instead, find the edge map of an image (using
  Sobel), and transform this edgemap into a
  directional-edge-histogram.
• The feature will be a histogram consisting of K
  bins. Each bin will contain the relative
  frequency of edges that lie in an angle of size
  360 / K.
• This will (hopefully) work because cars have
  strong edges in only two directions, while the
  background will have edges in many directions.

4
Features and Distance Function

• Features are selected before Adaboost training.
• These features can be selected by the user to
  take advantage of the brains ability to
  discriminate between objects and non-objects.
• The distance function will return a scalar result
  based on the distance of a given image from a
  given feature.

5
The Creation of a Weak Learner

• Given A single feature, the sets of positive
  and negative images, and a positive detection
  rate or a false positive rate.
• Provide A weak learner with a threshold that
  satisfies the conditions given. Also return an
  error value which will be used later to
  rank the weak learners.

6
Distances

• The algorithmCall the distance function on each
  positive and negative image. From this, create
  two arrays negative_distances, and
  positive_distances.
• These arrays are then sorted (using any sorting
  algorithm. Quicksort is presently used).
• From the sorted arrays, the median distance for
  positive and negative samples is found by simply
  reading the midpoint in the array.
  Somedian_positive positive_distancessize/2

7
Parity

• Given the median values of the arrays, the
  parity of the weak learner can be determined.
• The parity simply indicates whether the positive
  samples contain a higher distance value than the
  lower samples, or vice versa.
• Parity 1 indicates that the median of the
  positive values is greater than the negative
  values.
• Parity -1 indicates that the median of the
  positive values is lower than the negative
  values.
• This aids in thresholding.

8
(No Transcript)
9
Thresholding

• The threshold is created using the sorted
  distance arrays (negative_distances and
  positive_distances) and either a target detection
  rate or a target false positive rate.
• For a target detection rate of 90, the threshold
  is initialy set to positive_distancessize
  .9
• This ensures that 90 of the positive training
  samples would be classified correctly.
• However, this is a rather tight threshold, and
  can be improved upon.

10
Thresholding, continued

• If the positive and negative samples are
  distributed far apart, the detection rate can be
  increased without increasing the false detection
  rate.
• The threshold is incremented (or decremented,
  depending on the parity) by small amounts until a
  negative sample is misclassified (changing the
  false detection rate).
• The final threshold is the midpoint between the
  initial threshold, and this new threshold.
• This will hopefully provide better classification.

11
(No Transcript)
12
Creating a Cascade

• From the list of features, call the Weak Learner
  function for each feature.
• Order the Weak Learners based on their error,
  with the lowest error being first in the list.
• The first item in the cascade will be the most
  discriminative (lowest error) feature. The second
  item will be the second most discriminative
  feature, etc.

13
Evaluation

• To classify an input image as containing an
  object the image must pass through each weak
  learner in the cascade.
• To pass through a weak learner, the distance of
  the input image must satisfy the following
  condition parity distance gt parity
  threshold
• As can be seen, I consider images that have a
  distance equal to the threshold to be positive.

14
Making the Detector Invariant

• The feature representation is orientation
  invariant, so we need not create separate
  cascades for each possible orientation.
• The feature representation is also invariant to
  changes in lighting, assuming that we can
  successfully detect edges at the given lighting
  level.
• The image is scanned at many different
  resolutions to detect vehicles at different
  scales. This need to not be too precise as the
  feature representation is scale invariant.

15
Future Work

• The thresholding scheme could use improvement, I
  am still thinking about ideas for this.
• Instead of finding parity (and the threshold)
  based on median distance values, I could do it
  using mean distance values and standard
  deviations. This would have the undesirable
  consequence of making the weak learners more
  vulnerable to outliers, as well as mislabled
  training images.

16
Future Work, continued

• We wish to use image-based features (parts of the
  vehicle) as well as structural-based features
  (relationships between the parts of the vehicle).
  However, it is very difficult to obtain negative
  images for the structural-based features, so we
  cannot use traditional AdaBoost for the training
  of these types of features. Instead, we must find
  another method to create classifiers for these
  features.

17

• Thank You