Incremental learning for Robust Visual

parathyroidsanchovyΤεχνίτη Νοημοσύνη και Ρομποτική

17 Νοε 2013 (πριν από 4 χρόνια και 9 μήνες)

104 εμφανίσεις

Incremental learning for Robust Visual
Tracking

2013
-
10
-
08

Ko

Dae
-
Won


1.
PCA

2.
Face recognition for PCA

3.
Sequencial

Inference
Model

4.
Dynamical model

5.
Observation model

6.
Summary of the tracking algorithm







Contents

Incremental learning for Robust Visual Tracking








1. PCA(Principal Component Analysis)

Incremental learning for Robust Visual Tracking







X:



학습

데이터







2
. Face recognition for PCA

Incremental learning for Robust Visual Tracking


𝑖
:
얼굴

이미지
(
50



40
)










2
.
Face recognition for PCA


Incremental learning for Robust Visual Tracking


=



𝑖
𝑖



1. 2000
차원의

입력데이터를

그대로

사용하는것은

계산량




증가



메모리

증가



2. PCA(Principal Component Analysis)


이용해



차원

축소
,
특징

추출



3. Covariance matrix ( ):
계산




4.
𝑆
=
1
𝑁



(



)
𝑇

(S: 150 X 150)

2. Face
recognition for PCA


Incremental learning for Robust Visual Tracking







𝐴
:
𝑖

 










,
(



)
𝑇

:
𝑖









7.

=

𝑇





(

:



𝑇




)



2. Face
recognition for PCA


Incremental learning for Robust Visual Tracking



5.
𝑆
=
𝐴




𝐴
:
𝑖

 



𝑆
,



:
𝑖





𝑆





6.
1
𝑁




𝑇






𝑇

=
𝐴
(



)
𝑇











2. Face recognition for PCA

Incremental learning for Robust Visual Tracking










2. Face recognition for PCA


Incremental learning for Robust Visual Tracking










2. Face recognition for PCA


Incremental learning for Robust Visual Tracking










2. Face recognition for PCA


Incremental learning for Robust Visual Tracking


Visual tracking problem :

an inference task in a Markov model with hidden state
variables

3
.
Sequencial

Inference Model

Incremental learning for Robust Visual Tracking










-
Notations



𝐹


𝑖
ge

frame

at






=


,


,
𝜃

,


,
𝛼

,
𝜑

,


affine motion parameter




-
Probabilistic Formulation of Tracking


Estimate
𝑃


𝐹
0
,
1
,
2
,





=
0
,
1
,
2
,


3
.
Sequencial

Inference Model

Incremental learning for Robust Visual Tracking






Given a set of observed image
𝐼

=
𝐹
0
,
𝐹
1
,

,
𝐹

,


We aim to state the value of hidden state variable






𝑃
(


|
𝐼

)


𝑃
(
𝐼

|


)

𝑃
(


|




1
)
𝑃
(



1
|
𝐼


1
)



1

3
.
Sequencial

Inference Model

Incremental learning for Robust Visual Tracking







4
. Dynamical Model

Incremental learning for Robust Visual Tracking






=


,


,
𝜃

,


,
𝛼

,
𝜑

,



affine
motion parameter








𝑃
(


|




1
)

=



;




1
,
Ψ

Ψ
: diagonal covariance matrix whose
elements are the variances of affine
parameter (i.e
.,
𝜎

2
,
𝜎

2
,

𝜎
𝜃
2
,

𝜎

2
,

𝜎

2
,

𝜎

2
)



4
. Dynamical Model

Incremental learning for Robust Visual Tracking






We model image observations using a probabilistic
interpretation of PCA.






















































𝑃
𝐼



=

?


5. Observation Model

Incremental learning for Robust Visual Tracking





Given an image patch
𝐼


predicated by



,


We assume
𝐼


was generated from a subspace of the target

Object spanned by U and centered at
μ























































































































5. Observation Model

Incremental learning for Robust Visual Tracking












𝐼
:
𝑖𝑖

  𝑖
,
𝜇
: mean,



ε
I:
 𝑖 

𝑖

𝑖

ℎ

  𝑖











𝑃
𝑑
𝑡
𝐼



=

(
𝐼

|
𝜇
,


𝑇
+
ε
I)



𝑃
𝑑
𝑤
𝐼



=

(
𝐼

|
𝜇
,

Σ

1

𝑇
)





𝑃
𝐼



=
𝑃
𝑑
𝑡
𝐼




𝑃
𝑑
𝑤
𝐼



=

(
𝐼

|
𝜇
,


𝑇
+
ε
I)


(
𝐼

|
𝜇
,

Σ

1

𝑇
)

































































5
. Observation Model

Incremental learning for Robust Visual Tracking




Summary of the tracking algorithm

1.
Locate the target object in the first frame


and use a single particle to indicate this location

2. Initialize the
eigenbasis

U to be empty, the mean
𝜇



to be the appearance of the target in the first frame.










6. Summary of the tracking algorithm

20

Incremental learning for Robust Visual Tracking


3. Advance to next frame.



Draw particles from the particle filter,



according to the dynamical model.


4. For each particle, extract the corresponding window,



calculate its weights(
𝑃
𝐼



)













































































6. Summary
of the
tracking algorithm

Incremental learning for Robust Visual Tracking




21

21

5. Store the image window.



When the desired number of new images



have been accumulated,



perform an incremental update of
eigenbasis
, mean



and effective number of observations.



6. Go to step 3.










6. Summary
of the tracking algorithm

22

Incremental learning for Robust Visual Tracking

5. Store the image window.



When the desired number of new images



have been accumulated,



perform an incremental update of
eigenbasis
, mean



and effective number of observations.



6. Go to step 3.










6. Summary
of the tracking algorithm

23

Incremental learning for Robust Visual Tracking