下載/瀏覽

Download Report

Transcript 下載/瀏覽

Early fire detection algorithm based on
irregular patterns of flames and
hierarchical Bayesian Networks
日期:2010/12/06
作者:ByoungChul Ko , Kwang-Ho
Cheong and Jae-Yeal Nam
報告者:楊智雁
Fire Safety Journal 45 (2010) 262-270
南台科技大學
資訊工程系
大綱
2
1
Introduction
2
Color information
3
Bayesian Networks
4
Experimental results
5
Conclusion
1. Introduction
 Most current fire alarm systems are based on infrared
sensors
 They are usually unable to provide any additional
information
 Location and size of the fire and the degree of burning
3
1. Introduction (c.)
 The use of cameras for fire detection and have applied
a variety of visual features
 Color, motion, edge and the shape of a fire region
 The axonX company produces SigniFire
 Fastcom Technology produced the SFA (fire and
smoke alert) system
4
1. Introduction (c.)
 The present study first detects candidate fire regions
by detecting moving regions and fire-colored pixels
 Applied to nodes of hierarchical Bayesian Networks
5
1. Introduction (c.)
6
2. Color information
x is moving if
I n ( x)  Bn ( x)  Tn ( x)
X is non-moving
otherwise
7
2. Color information (c.)
8
3. Bayesian Networks
 The structure of the proposed hierarchical Bayesian
Networks is composed of three layers
 One final goal node
 One sub-goal node
 Four observation nodes
9
3. Bayesian Networks (c.)
10
3. Bayesian Networks (c.)
P(W , O1 F ) P( F )
P( F W , O1 ) 
P(W , O1 F ) P( F )  P(W , O1  F ) P(  F )
4
P(O2,O3,O4,W )  P(W )   P(Oi W )
i 2
11
3. Bayesian Networks (c.)
12
3. Bayesian Networks (c.)
13
3. Bayesian Networks (c.)
14
4. Experimental results
15
4. Experimental results (c.)
16
4. Experimental results (c.)
17
4. Experimental results (c.)
 [Töreyin] TP rate of 72%, an FP rate of
0.9%, and an M rate of 27.1%
 [Ko]
TP rate of 88.6%, an FP rate of
1.2%, and an M rate of 10.2%
 [Our]
TP rate of 95.3%, an FP rate of
0.4%, and an M rate of 4.3%
18
5. Conclusion
 Relatively inexpensive equipment, rapid response
time, and fast confirmation through surveillance
monitors
 The experimental results show that the proposed
approach had a more robust identification of noise
19
5. Conclusion (c.)
 Small regions were considered noises and were thus
removed from both Töreyin’s proposed algorithm and
ours
 Reducing false alarms and missed fire regions remains
an ongoing challenge for successful fire detection in
real-world environments
20
南台科技大學
資訊工程系