Sunteți pe pagina 1din 27

AuthorsChuang-Jan Chang, Shu-Lin Hwang*

Ming Chi University of Technology


New Taipei City, Taiwan
MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Introduction
IP-CAM :a major role in the digital surveillance

system
The cascaded AdaBoost-based face detection
system proposed by Viola
Support real-time detection

Implemented by image library-OpenCV: Open Source

Computer Vision Library, provided by Intel, 2005

MCUTEE Omnidirectional Imaging & Surveillance Lab

Introduction (cont.)
The performance of the Alt2 cascade (OpenCV) in

video is worse than that in static images


Training data set in the Alt2 cascade cannot take the

localized character from the special IP-CAM video

Localized sampling optimum (LSO) for local IP-

CAM video
An efficient motion detection model is adopted
for practical applications

MCUTEE Omnidirectional Imaging & Surveillance Lab

Introduction (cont.)
Train our classifier
Used the method provided by Open CV
Training samples

FERET face database


Some local samples were taken from a DM355 IP-CAM Video

Overall performance using 30% local samples


97.9% detection rate improvement and 54.5%

detection time reduction than the Alt2 cascade

MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Haar-like Rectangular Features


The extension technology from Haar wavelet

transformation
Viola applied it to image retrieval
Lienhart added the 45-degree angular features

MCUTEE Omnidirectional Imaging & Surveillance Lab

AdaBoost Face Detection


AdaBoost: binary classification method

Feature-based AdaBoost approach proposed by Viola


integral images: calculate rectangular feature value fastly
learning algorithm: AdaBoost based training method

the cascade decision-making classifier

Alt2 cascade: implemented with OpenCV Library

MCUTEE Omnidirectional Imaging & Surveillance Lab

Cascade decision-making classifier


Strong classifier is the one stage of the cascade

classifier
Train the eigenvalue and determine the thresholds
used in decision-making of the feature form one stage
Combine all trained stages

MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Our System Structure


Our detection framework follows Viola & Jones

classical work
Two differences
added localized sampling features
the motion detection preprocess was added

The self-training classifier with the LSO method


Implemented by Open CV classifier training tools
Face detection process with motion detection
Based on OpenCV cvHaarDetectObjects ()
MCUTEE Omnidirectional Imaging & Surveillance Lab

Face Detection Process


Face detection with motion detection flowchart

MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Experimental Environment
Hardware and Software
Hardware

Software

CPU: Intel Core 2 Duo E4600 O.S.: Microsoft Windows XP


2.4GHz
Service Pack 3
RAM: 2 Giga Byte
VGA: ASUS 7300GT 256MB
DDR2

Develop Software:
Microsoft Visual Studio 2008

Open CV 2.0

IP Camera: DM355IPNC-MT5

MCUTEE Omnidirectional Imaging & Surveillance Lab

Testing Data
DM355 IP-CAM Testing video information

MCUTEE Omnidirectional Imaging & Surveillance Lab

Training samples and Testing Video


Face database
FERET face database (Information Technology Laboratory

of the National Institute of Standards and Technology )


All samples are resized to 20*20

DM355 IP-CAM testing video

MCUTEE Omnidirectional Imaging & Surveillance Lab

The Samples Creation for LSO


The LSO sample creation is the same as FERETs
include the whole face, the hair and retain sufficient edges

All samples are resized to 20*20

Train new classifiers for face detection


performed by createsample.exe & haartraining.exe (OpenCV)

Add LSO samples by 10%, 30% and 50% to FERET

positive samples (total 1000) samples)

MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Face Detection Measurement Criteria


Discuss the accurate detection frames and the error
detection frames separately

Ideal face detection system should have high detection


rate and a low error rate
(1) Detection rate = number of correctly detected frames /total number of faces.
(2) Rate of mistakenly detected frames = number of mistakenly detected frames
/total number of detected frames.
(3) Total number of detected frames = number of correctly detected frames
+number of mistakenly detected frames.

(4) Actual detection rate = detection rate -rate of mistakenly detected frames.

MCUTEE Omnidirectional Imaging & Surveillance Lab

Experimental Results before Added LSO


Adding motion detection could actually improve the

detection rate and reduce the rate of false positives


Average actual detection rate is poor (only 35.12%)

MCUTEE Omnidirectional Imaging & Surveillance Lab

Experiments after Adding LSO


Five types of casecade classifiers were used

Set parameter scale_factor to 1.1 & min_size to


20*20 (in cvHaarDetectObjects())
Classifier
Alt2

Introduction
OpenCV built-in Classifier with large training samples

nonLSO Self-training Classifier with FERET subset (1000 training samples)


LSO-10% Self-training Classifier with FERET subset and 10% Local samples (100 training
samples)
LSO-30% Self-training Classifier with FERET subset and 30% Local samples (300 training
samples)
LSO-50% Self-training Classifier with FERET subset and 50% Local samples (500 training
samples)
MCUTEE Omnidirectional Imaging & Surveillance Lab

The performance comparison without


motion detection after LSO
The classifier with the local samples, the actual detection
rate is better than Alt2 and nonLSO
Add local samples more than 30%, the increasing detection rate

gradually slows down

Add local samples, detecting time increases only slightly

MCUTEE Omnidirectional Imaging & Surveillance Lab

The performance comparison with


motion detection after LSO
Those classifiers with LSO still elicit better performance
The highest average detection rate reaches a detection
efficiency of 70% (LSO-30%)
The detection times are clearly short: LSO classifier
average only 35ms

MCUTEE Omnidirectional Imaging & Surveillance Lab

OUTLINE
Introduction
Related Work

System Structure and Experimental Process


Experimental Environment and Testing Data

Experimental Results
Conclusions

MCUTEE Omnidirectional Imaging & Surveillance Lab

Conclusions
Have developed face detection test bench based

on the cascaded AdaBoost approach


Use and modify the Open-CV library functions

Combine motion detection preprocess and self-

training face classifiers with LSO from a local IP-CAM


video

The overall system with 30% local samples can

get better performance


detection rate improvement and detection time

reduction than Alt2 cascade


MCUTEE Omnidirectional Imaging & Surveillance Lab

Face Detection Test Bench


VS 2008 & OpenCV 2.0

MCUTEE Omnidirectional Imaging & Surveillance Lab

MCUTEE Omnidirectional Imaging & Surveillance Lab

S-ar putea să vă placă și