Head Detection

1
Head Detection Tracking Final
Report 03/09/2004

• Yisheng Chen

2
Goals

• Given a clip of video with a single person
• Detect the head region
• Track the head region
• Find the orientation of the head

3
Head Detection

• Assumptions
• Head is located at an extreme position of the
  body image and is approximately indexed by the
  highest projection histogram
• Two concave corners define the bottom line of the
  head region

4
Footage
5
Head Geometric Features

• There is an increase and decrease downward the
  scanlines in the silhouette
• According to the y-axis projection histogram to
  find the bounding box
• Use a partial ellipse to match the head contour
• Use the concave corners to segment head region

6
Head Bounding Box
7
Head Segmentation (1)

• Use a partial ellipse to represent the head
  region
• Equation

8
Partial Ellipse

• How to get x0, y0, a and b?

9
Partial Ellipse Results
10
Kalman Filter

• Use Kalman Filter to estimate x0, y0, a and b
• Assume x0 and y0 take 1-degree motion
• a and b take 0-degree motion

11
Kalman Filter Results
12
Head Segmentation (2)

• Use convex and concave corners to identify the
  head region

13
Head Orientation

• Training samples
• 36 frames, a complete 180-degree head rotation

14
Training Samples

• Extract the head regions by the second method
  (concave corners) and resize them to 1616 images

15
Feature Vector

• Use a feature vector X to represent the template
• Divide the template into 44 grids
• X (W1, W2, , W16)

16
K-Nearest Neighbor Classifier

• A test sample is classified into class I if the
  distance function (pixel-wise or motion vector
  comparison) has the largest number of nearest
  neighbors belonging to class I among the 7
  nearest neighbors.
• I choose 7 because there are 6 classes
• If the sample has the same number of nearest
  neighbors to class i and j, then choose the
  nearest one.

17
(No Transcript)
18
Methods Comparisons
19
Partial Ellipse Match

• Due to the incorrect assumption, the bounding box
  is incorrect in cases, and we get incorrect
  partial ellipse match.

20
Methods to find the partial ellipse

• Initial bounding box
• Extended bounding box (according to the face
  orientation)
• Rotated partial ellipse
• Using PCA to find the major axis of the partial
  ellipse (PCA ellipse)
• PCA ellipse with dynamic phase
• Using Kalman Filter to estimate the center

21
Expended BBox Rotated Ellipse

• Using the Head Orientation Detector, we modify
  the bounding box according to head orientation.
• In the previous example, we know the head has 90
  degree rotation, so expand the bbox to the left
  and down.
• Search the tilting angle of the partial ellipse
  according to the head orientation

22
Major Axis from PCA
23
Dynamic Phase
24
Illustrations for Partial Ellipses
25
(No Transcript)
26
Metric

• Use the per pixel distance from the head contour
  to the partial ellipse as the matching function
• PCA 0.2232 (1.6250 s)
• PCA phase 0.2175 (1.7180 s)
• Brute Search 0.1519 (1.5930 s)
• Brute phase 0.1480 (2.7350 s)

27
Final Result
28
Further Work

• The head region is really small now (1616), what
  if apply my methods to a larger image?
• Retrieve 3D head position according to the head
  position, orientation and others.
• Besides head, can we find and track arms, legs
  and torso?
• Combine head orientation and other information,
  to find the body orientation.

29
References

• Head Segmentation and Head Orientation in 3D
  Space for Pose Estimation of Multiple People by
  Sangho Park, J.K. Aggarwal, 2000
• Human Face Segmentation and Identification by
  Saad Ahmed Sirohey, 1993
• Hydra Multiple People Detection and Tracking
  Using Silhouettes by I. Haritaoglu, D. Harwood,
  L. Davis, 1999
• An Introduction to the Kalman Filter by Greg
  Welch and Gary Bishop, 2003
• Pattern Recognition and Image Preprocessing by
  Sing-Tze Bow, 1992