in Video Using Pixel Layers

naivenorthIA et Robotique

8 nov. 2013 (il y a 4 années et 2 jours)

174 vue(s)

Robust Foreground Detection
in Video Using Pixel Layers

Kedar A. Patwardhan, Student Member, IEEE,

Guillermo Sapiro, Senior Member, IEEE, and

Vassilios Morellas, Member, IEEE


IEEE TRANSACTIONS ON PATTERN ANALYSIS AND
MACHINE INTELLIGENCE, VOL. 30, NO. 4, APRIL 2008



Presented by
:曹憲中

Proposed framework

Kernel Density Estimation

Where K is some kernel and h is
a smoothing parameter called
the bandwidth.


False Alarm

= Type 1 error

= False Positives

Initial guess

Training Step


Maximum likelihood


Expectation
-
Maximization (EM)


Kernel Density Estimation (KDE)


Kullback

Leibler (KL) divergence

original 'baboon' image

initial
-
guess

the final layer after

the refinement step

Maximum likelihood


最大似然估計是一種統計方法,它用來
求一個樣本集的相關機率密度函數的參

。這個方法最早是遺傳學家以及統計
學家羅納德

費雪爵士在
1912
年至
1922

間開始使用的。

Expectation
-
Maximization (EM)


在統計計算中,
最大期望(
EM
)演算法
是在機率(
probabilistic
)模型中尋找參


Maximum likelihood
的演算法
。最大
期望經常用在機器學習和計算機視覺的
數據集聚(
Data Clustering
)領域。

Kernel Density Estimation (KDE)


核密度估計,在機率論中
用來估計未知
的密度函數
,屬於非參數檢驗方法之一,


Rosenblatt (1955)

Parsen(1962)

出,
Ruppert

Cline
基於數據集密度函
數聚類演算法提出修訂的核密度估計方
法。

Kullback

Leibler (KL) divergence


Kullback
-
Leibler Divergence
,是以它的
兩個提出者庫爾貝克和萊伯勒的名字命
名的。
KL divergence
用來衡量兩個正函
數是否相似

對於兩個完全相同的函數,
它們的

KL divergence
等於零
。在自然
語言處理中可以用

KL divergence
來衡
量兩個常用詞(在語法上和語義上)是
否同義,或者兩篇文章的內容是否相近
等等。































Proposed framework

Kernel Density Estimation

Where K is some kernel and h is
a smoothing parameter called
the bandwidth.


False Alarm

= Type 1 error

= False Positives

Online Step

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS


160x120


The algorithm was implemented using
C++
,
on a machine with
Intel
-
Pentium IV 1.8GHz

processor.


In the offline training step, we used an initial
training stack of approximately
30 frames

for
all the results, achieving a running speed of
10 frame/second

with our experimental code.


The initial layering and training steps usually
require about
5 minutes

(for layering all the
frames in the initial training stack).

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS

IMPLEMENTATION DETAILS AND
EXPERIMENTAL RESULTS

DISCUSSION AND FUTURE SCOPE


In the future, we would like to adapt the
framework described here to
multicamera

scenarios where the
different cameras may or may not
overlap and also may be of different
modalities.

DISCUSSION AND FUTURE SCOPE


The foreground models of
moving
persons

should be made more robust,
for example by adding shape
information to the global feature
-
set,
toward their use in
person identification

and tagging throughout the area of
surveillance.

REFERENCES


Kedar A. Patwardhan, Guillermo Sapiro, Vassilios
Morellas,

“A Pixel Layering Framework For Robust
Foreground Detection In Video”.



Wikipedia



Jun
-
Yi Li, “Object Extraction for Video Surveillance
System”


Thank you for your attention.