Method and apparatus for moving object detection using principal component analysis based radial basis function network

Patent metadata

Field	Value
Publication number	US-9349193-B2
Application number	US-201414231637-A
Country	US
Kind code	B2
Filing date	Mar 31, 2014
Priority date	Mar 31, 2014
Publication date	May 24, 2016
Grant date	May 24, 2016

Abstract

A method for moving object detection based on a Principal Component Analysis-based Radial Basis Function network (PCA-based RBF network) includes the following steps. A sequence of incoming frames of a fixed location delivered over a network are received. A plurality of Eigen-patterns are generated from the sequence of incoming frames based on a Principal Component Analysis (PCA) model. A background model is constructed from the sequence of incoming frames based on a Radial Basis Function (RBF) network model. A current incoming frame is received and divided into a plurality of current incoming blocks. Each of the current incoming blocks is classified as either a background block or a moving object block according to the Eigen-patterns. Whether a current incoming pixel of the moving object blocks among the current incoming blocks is a moving object pixel or a background pixel is determined according to the background model.

First claim

What is claimed is:

1. A moving object detection method based on a Principal Component Analysis-based Radial Basis Function network (PCA-based RBF network) comprising:
receiving a sequence of incoming frames of a fixed location delivered over a network;
generating a plurality of Eigen-patterns from the sequence of incoming frames based on a Principal Component Analysis (PCA) model, wherein the PCA model comprises an optimal projection vector;
constructing a background model from the sequence of incoming frames based on a Radial Basis Function (RBF) network model, wherein the RBF network model comprises an input layer having a plurality of input layer neurons, a hidden layer having a plurality of hidden layer neurons, and an output layer having an output layer neuron, and wherein there exists a weight between each of the hidden layer neurons and the output layer neuron;
receiving a current incoming frame delivered over the network and partitioning the current incoming frame into a plurality of current incoming blocks;
classifying each of the current incoming blocks as either a background block or a moving object block according to the Eigen-patterns generated from the sequence of incoming frames based on the PCA model comprising:
calculating a projection of each of the current incoming blocks according to the optimal projection vector;
calculating a similarity level between the Eigen-pattern and the projection of each of the current incoming blocks;
determining if the similarity level exceeds a second threshold value;
if yes, classifying the current incoming block as the background block; and
if no, classifying the current incoming block as the moving object block; and

determining whether a current incoming pixel of the moving object blocks among the current incoming blocks is a moving object pixel or a background pixel according to the background model.

2. The method of claim 1 , wherein the step of generating the Eigen-patterns from the sequence of incoming frames based on the PCA model comprises:
partitioning each of the incoming frames into a plurality of sample blocks and classifying the sample blocks into a plurality of classes;
calculating a total scatter matrix according to the sample blocks;
calculating the optimal projection vector by maximizing a determinant of the total scatter matrix; and
obtaining each of the Eigen-patterns according to the optimal projection vector and the corresponding sample block.

3. The method of claim 2 , wherein the formulas for generating the Eigen-patterns from the sequence of incoming frames based on the PCA model comprise Eq.(1)-Eq.(3):

S
T

=

∑

i
=
1

M

⁢

(

b
i

-
u

)

⁢

(

b
i

-
u

)

T

Eq
.

⁢

(
1
)

wherein S T is the total scatter matrix, {b 1 , b 2 , . . . , b M } is a set of the M sample blocks in an k-dimensional block-space, u represents a mean of all the sample blocks, b i represents the i th sample block of each of the incoming frames and is classified as one of M classes {B 1 , B 2 , . . . , B M }, and M and k are positive integers,

W
opt

=

arg
⁢

max
W

⁢



W
T

⁢

S
T

⁢
W



=

[

w
1

,

w
2

,
…
⁢

,

w
m

]

Eq
.

⁢

(
2
)

wherein W opt is the optimal projection vector, and [w 1 , w 2 , . . . , w n ] represents a set of eigenvectors of S T is an empirical dimensionality value of the Eigen-patterns, m is a positive integer, and m<k, and
E ep i =W opt T b i   Eq.(3)

wherein E ep i is the Eigen-pattern corresponding to the i th sample block b i , and W opt T is the transposition of W opt .

4. The method of claim 1 , wherein the step of constructing the background model from the sequence of incoming frames based on the RBF network model comprises:
calculating a difference between an intensity value of each sample pixel of the sequence of incomi

Assignees

• Univ Nat Taipei Technology

Inventors

• Chen Bo-Hao
• Huang Shih-Chia

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/2414

  Smoothing the distance, e.g. radial basis function networks [RBFN] · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T7/204

  Physics · mapped topic

• G06T2207/20021

  Dividing image into blocks, subimages or windows · CPC title