Transcript Target Recognition

Target Recognition
Harmatz Isca
Supervisor: Nakhmani Arie
Semester: Winter 2007
Project goals
Create a target classification system
based on dimension reduction, using the
targets contour.
No dependence on illumination and color
Universal method works on all target types
and sizes
Fast learning for new targets
Low computational needs
The dimension reduction algorithm can be
adopted to work on all types of data.
Motivation
Tracking people
ATR– automatic target recognition
Find suspects in given areas
Look for specific characteristics of targets
Method
Change
Detection
Snakes
Dimension
Reduction
Post
processing
Result
Working Database
475 images
2176 snakes found
The snakes were divided into 3 types:
Real (339) – a snake of a person
Partial (155) – a snake were the person was
partially hidden, or a clear silhouette was not
detected
False (1682) – a snake of a random change in
the image
Change detection
Create
average
reference
image
Subtract
Detect
Get several
Find
changed
changes
the reference
background
pixels
in image
images
from
the image
= Background Image
Snakes
Level Set Evolution Without Re-initialization: A New Variational
Formulation
Chunming Li, Chenyang Xu, Changfeng Gui, and Martin D. Fox
CVPR 2005
Dimension reduction
Database
1025204119967458748574541415
2358242573320314581485681822
2536775897592432253635892125
4096825913632656149569453340
4646675364152387367421214546
7071136728572146958214343070
7124269479568128452425252573
2546693114958981676549192029
2028323782424683214667471620
1181147639735657328158761510
LLE or PCA
X
14
18
21
33
45
30
25
20
16
15
Y
15
22
25
40
46
70
73
29
20
10
• Select target snake
• Transform snake to vector
• Add snake vector to vector database
53 48748486217125827416824943
14 25255153542826391449348662
• Perform dimension reduction on vectors
• Displaying dimension reduction results in
graph
Local Linear Embedding (LLE)
For every snake in database:
Find K nearest neighbors { z1:K }
Find weight Wij for every neighbor zj
n
K
minimizing  zi   Wij zij
i=1
n
s.t.
W
j 1
ij
2
n -size of database
j 1
 1, zi ; Wij  0 if z j is not a neighbor of zi
Compute the projection to lower space where
weighted distance from neighbors is minimum
Principal components analysis
(PCA)
Calculate the covariance matrix of database
Calculate eigenvectors (ordered by eigenvalues)
Find snakes representation with eigenvectors
0.51
+
0.12
+
0.3
+
0.07
LLE vs PCA
LLE
Non-linear embedding
Local
Keeps subspace with
best local linear structure
Assumes local linearity
PCA
Linear embedding
Global
Keeps subspace with
best variance of data
Assumes global linearity
Results LLE
Grade 
d ( Snake, Database)
2
 Database
Results PCA
Grade 
d ( Snake, Database)
2
 Database
Post-processing
Steps taken to achieve better separation
between false and true snakes
Compactness: Area/Perimeter²
Adaptive Database
Target Tracking
Compactness
Grade = area/perimeter2
Dimension Reduction and Compactness
Grade = GradePCA . GradeCompactness
Adaptive Database
Unsupervised
Snakes matching a certain grade level are
added to the database. Snakes in database
with low grades are removed.
The algorithm was applied for every movie
separately
Adaptive Database
Grade 
d ( Snake, Database)
2
 Database
Tracking
Define Target of interest
For every next image:
Define search region
If “good” snake is found, then
Set target to found snake
Else
Increase search area
Move to next image
Tracking Results
Conclusions
Dimension reduction was used to find
people in images.
The method works well on clear
silhouettes.
Different post-processing methods used to
improve results, each with its own pros
and cons.
The method works with a small database
(20 snakes) and can be adopted for real
time work.
Feature Directions
Occluded target support
Improve target tracking
Multiple targets
Kalman / Particle filters
Target specific database
Adaptive grade threshold
Improved snakes