Image Subtraction for Real Time Moving Object Extraction Shahbe Mat Desa, Qussay A. Salih, CGIV’04.

Image Subtraction for Real Time Moving Object Extraction

Shahbe Mat Desa, Qussay A. Salih, CGIV’04

Outline

• Introduction

• Automatic Motion Detection

• Background Reconstruction

• Noise Reduction

• Experiment Results

• Conclusion

Introduction

• In recent years, development of motion analysis in many vision systems arise researches in conjunction with the immense attentions of employing real time application to control complex real world system such as ATM security, airport surveillance, traffic monitoring, etc.

• The basic idea of mostly automated surveillance applications is that motion detection continuously operating and the system is triggered to perform higher-level processes such as object recognition and tracking.

• Standards such as MPEG-1, MPEG-2 use block-based motion compensation and DCT techniques may be time consuming and cause complex computation.

Objectives

• Automatic motion detection

• Reference background update

• Segmentation of dynamic region from static region

• Noise reduction

• Reliable and less complex

• Real time and auto-completion

Motion mask extraction

• Motion detection issue– As a binary labeling problem:

– Automatically detecting objects

tSyxss

sls at time image),( ,

background static if ,0

objects moving if ,1

Motion mask extraction(2)

• Image subtraction– Background subtraction, dB: |fk – B|– Temporal differencing, dk-1, dk+1: |fk-1 – fk|, |fk+1 – fk|

• Calculate motion mask, motionk, with threshold Td

otherwise ,0

),(),( if ,1),(

dT|yxByx|fdyxmotion

kB

kB

otherwise ,0

),(),( if ,1 11

1

dT|yxfyx|fdmotion

kkk

kk

)()( 1 1 kkkBkkkBk motionmotionmotionmotionmotion

Background reconstruction

• Background varies from time to time as the scene changes.

%*

_

hw

pixeldynamicdynamick

MdynamicfB kk if ,

w, h: width and height of frame

dynamic_pixel: Σ (motionk)

B: background image, f0

M: motion threshold, 4%

Noise reduction

• Two operations– Erosion followed by dilation

• Erosion– With structuring element E

• Result in a value of 1 in motion mask motionk at P=(x,y) if the spatial arrangement of ones in EP fully matches that of motionk.

),( ,}|{ yxPmotionEPEmotion kPk

=> Removes isolated foreground pixels

Noise reduction (2)

• Dilation– With structuring element D

• The result is the set of all points P=(x,y) so that reflection and motionk overlap by at least one nonzero elements.

=> Adds pixels to the boundary of the object

PD̂

),( ,}|{ yxPmotionDPDmotion kPk

Result (1)

• Manual generate ground truth G

• Performance evaluator: Root Mean Square Error (RMSE) measurement

2/1

1 1

2)],(),([*

1

w

x

h

y

yxFyxGwh

RMSEG(x,y): ground truth

F(x,y): output

Results (2)

• Four different-scene test sequences:– Moving vehicles on road– People walking indoors and outdoors

Result (3)

• Compared with the common background subtraction:

SceneCommon Method

Proposed Method

A 0.0324 0.0275

B 0.0524 0.0216

C 0.1288 0.0611

D 0.1495 0.0696

Conclusion

• Proposed a less-complex method that enables high-level real time processes.

• Compared with common background extraction method, the method is more reliable.