An Eigendecomposition Method for Detecting Occluded Objects

1
An Eigendecomposition Method for Detecting
Occluded Objects

• Rodney Roberts, Faculty
• Hyun Geun Yu, Ph.D. Student
• Edmond Dupont, MSEE Student

2
Eigenspace Methods

• also known as
• Principal Component Analysis (PCA)
• Singular Value Decomposition (SVD)
• Karhunen-Loeve transformation
• applications
• image compression
• human face recognition
• object/pose recognition

3
Eigenspace Methods

• advantages
• appearance based (no model)
• relatively fast on-line calculation
• disadvantages
• slow off-line computation
• difficult to deal with occlusion and background
  clutter

4
Example

• Training Images

5
Subspace Representation

• Eigenimages
• Reconstructed Images

6
SVD-Based Object/Pose Detection - No Occlusion
Original Images Normalize Size Normalize
Intensity Average Image
7
SVD-Based Object/Pose Detection - No Occlusion
Average subtracted Eigenimages
8
SVD-Based Object/Pose Detection - No Occlusion
Projection onto Eigenspace
9
Problems when occlusion is present

• Segmentation becomes difficult
• Scale normalization cannot be done
• Intensity normalization will not work
• Projection onto the eigenspace will be affected

10
Proposed method

• Approach
• Localization generate candidate locations
• Eigenspace image comparison evaluate candidate
  locations to simultaneously register the object
  and identify its pose
• Differences
• based purely on appearance
• uses a quad-tree data structure
• Scenarios
• occlusion
• background clutter

11
Quad-tree structure
12
Quad-tree eigenspace image comparison -Level 1
13
Quad-tree eigenspace image comparison - Level 2
14
Quad-tree eigenspace image comparison -Level 3
15
Conclusions

• algorithm using localization and quad-tree
  eigen-decomposition is effective for dealing with
  occlusion
• accuracy is proportional to the difficulty of the
  problem
• computational expense is proportional to the
  difficulty of the problem
• much more computationally expensive than
  unoccluded objects in controlled environments

16
Vegetation Classification
classified image
u1 vs u2
17
20-meter Accuracy
u1
classified image
18
80-meter Accuracy
classified image
u1
19
Singular Vector Comparison
80 meters
20 meters
20
Singular Vector Comparison
refined 80 meters
20 meters