Human Detection

1
Human Detection

• Phanindra
  Varma

2
Detection -- Overview

• Human detection in static images is based on the
  HOG (Histogram of Oriented Gradients) encoding of
  images
• Training set consists of positive windows
  (containing humans) and negative images
• For each window in the training set the HOG
  feature vector is computed and linear SVM is used
  for learning the classifier
• For any test image, the feature vector is
  computed on densely spaced windows at all scales
  and classified using the learned SVM

3
HOG encoding

• Preprocessing-
• Gamma normalize each channel using square root
  transformation in the given window
• For each channel compute gradients using -1 0 1
  and -1 0 1T and find the channel with the
  largest gradient magnitude for each pixel
• Compute gradient orientation (0 180) for each
  pixel in this dominant channel
• Descriptor computation -
• Divide the window (64x128) into dense grid of
  points with horizontal and vertical spacing equal
  to 8 pixels
• Divide the 16x16 region (block) centered on each
  point on the grid into cells of size 8x8 (i.e 4
  cells for each grid point)
• For each pixel in the current block use Trilinear
  interpolation based on gradient strength to vote
  into a 2x2x9 histogram

4
HOG encoding (Contd..)

• Different voting schemes were used for each of
  the colored regions
• Block normalization for illumination invariance
  is done on each block independently using the
  norm of the 2x2x9 vector
• The final feature vector is the collection of all
  the 2x2x9 feature vectors from all the grid
  points

Cell centers
Grid point
A Block of 16x16 pixels
5
Training

• The training set has been obtained from
  http//pascal.inriaples.fr/data/human/INRIAPerson.
  tar
• The training set consists of positive 64x128
  windows (2416) containing humans and negative
  images
• Negative windows are sampled from the negative
  images at random locations (12000)
• Initial Phase learning - Learn the SVM
  classifier on the original training set
• Generate Hard examples - Run the learned SVM on
  the negative images at all scales and window
  locations and save all the false positives
  (approx.6000)

6
Training (Contd..)

• Second Phase learning - Using the newly
  generated negative examples learn the new linear
  SVM (total positive windows 2400, negative
  windows 17000 approx)
• Following this procedure, 375 windows were
  misclassified out of the possible 19400 windows
  (using SVMLight)

7
Testing

• Given an Image - HOG feature vector is computed
  across all scales and window locations and the
  locations and scales of all positive windows are
  saved (window size 64x128)
• This procedure gives multiple detections (at many
  scales and locations)
• To fuse overlapping detections the Mean Shift
  mode detection algorithm is used
• Represent each detection in a 3D space (x y
  log(s)) and iteratively compute the mean shift
  vector at each point
• The resulting modes give the final detections and
  the bounding boxes are drawn using this final
  scale

8
Results - Detection
An example image
Detections when threshold is zero
9
Results Detection (Contd..)
Previous image
Detections when threshold is equal to one
10
Results - Detection
Detections when threshold is zero
An example image
11
Results Detection (Contd..)
Result of Mean Shift mode detection
12
Comparision
Detection Video