Human Detection with Simple Graphical Model

1
Human Detection with Simple Graphical Model

• Dong-Shin Kim

2
Introduction

• Key descriptor
• Distance between face and center
• Distance between left/right hand and center
• Color distribution of skin and cloth

Face
Left hand
Right hand
Cloth
3
Training stage

• 205 images
• Building color histogram for skin and cloth
• Building Mixture Gaussians for each distance
  attribute

4
Histogram model

• Skin area in the given training images are
  manually labeled and placed into the skin
  histogram
• Cloth area in the given training images are
  manually labeled and placed into the cloth
  histogram

5
Distance model
6
Testing stage

• Preprocessing
• Using watershed segmentation algorithm
• Pick 20 maximum probable candidates blobs

7
Preprocessing
8
Probability model
Face
Left hand
Right hand
Cloth
9
Constellations (1)
10
Constellations (2)
11
Conclusion Future work

• Correct Detection rate is 72 with automatic
  segmentation
• More training data needed
• Better preprocessing is important
• Need more powerful descriptor that explains human
  in the image