Abstract —This paper proposes a flame detection method based on the Amount of Movement of the Flame Edge (AMFE) for detecting the fire with the characteristic of the flame edge’s constant motion. Firstly, background update has to be done to extract moving regions. And then the suspect area of similar flame color will be identified from the moving regions based on HSI color model. Finally, the paper construct curve of the AMFE of the flame suspect area, and make use of the local maxima of the wavelet coefficient of the high-frequency components to detect whether there is flame in a video sequence with the application of wavelet transform on this curve. The experiment results prove the accuracy and the effectiveness of this method.
Keywords: AMFE; HSI color model; Suspect area; Wavelet
I. I NTRODUCTION IRE is a kind of common natural disasters, which will cause a serious threat to the safety of the lives and
property [1]. Because of the uncertainty and variety of the fire, the traditional fire detection technology based on the sensor is easy to be affected by some factors of the environment, such as the area, humidity, dust, gas flow and so on, which make the traditional fire detection technology not detect a fire accurately and also have a poor real-time
performance [2].
In recent years, with the video monitor being widely utilized and the increasing improvement of digital image processing technology, video-based fire monitoring has become increasingly better. It uses the computer as the core to analyze color features and physical features of the early flame images obtained by the camera via digital image processing technology [3]. A flame detection method based on brightness and color features of the fire edge regions has been proposed in [4]. This paper detects flame using static features, but having no description of the dynamic features. Because flame edge regions and the adjacent regions have a significant difference,
Manuscript received January 23, 2013.
Hongke Xu is a professor with the School of Electronic and Control Engineering, Chang’an University, Xi’an, 710064, China. (E-mail: xuhongke@https://www.doczj.com/doc/8114519149.html,). Yanyan Qin is a graduate student with the School of Electronic and
Control Engineering, Chang’an University, Xi’an, 710064, China. (He is corresponding author, phone: 150 **** ****, E-mail: qinyanyan_happy@https://www.doczj.com/doc/8114519149.html,). Yong Pan is a graduate student with the School of Electronic and Control Engineering, Chang’an University, Xi’an, 710064, China. (E-mail:
panyong15@https://www.doczj.com/doc/8114519149.html,).
Hao Chen is a graduate student with the School of Electronic and Control
Engineering, Chang’an University, Xi’an, 710064, China. (E-mail: 317474978@https://www.doczj.com/doc/8114519149.html,). Number of funding programs: gsgzj-2011-08.
Reference [5] extracts flame suspect area with the help of brightness information, and then classify the suspect area to detect flame using support vector machine (SVM). This algorithm has high recognition rate, but with higher computational complexity. Wavelet transform is applied in the one-dimensional and two-dimensional space to analyze flame flicker frequency and the changes of pixel value in [6], which reduces false positives caused by single use of moving regions and color, but has low recognition rate on rapid shaking objects with similar flame color. Liu [7] uses color information and structural features to extract flame suspect area, and then uses Fourier transform to classify the suspect area using Fourier coefficients. This algorithm also has high recognition rate, but it is very strict to the environment.
Contour fluctuation data processing method in [8] is proposed
to detect flame using a color camera.
Based on the above researches, this paper proposes a flame detection method based on the Amount of Movement of the Flame Edge (AMFE) that is the unique feature of early flame. What we should do is to extract flame suspect area via HSI color model, and then analyze AMFE feature of the flame suspect area by using wavelet transform. II. F LAME SUSPECT AREA EXTRACTION We need to extract the suspect area of similar flame color before analyzing AMFE feature, which will make the analysis of the AMFE feature be more targeted because it is not need to analyze AMFE feature if there is no moving regions in video images or there is no suspect area of similar flame color in moving regions. What’s more, the flame suspect area relative to the whole image is just a sub-region, which will greatly reduce computation cost when analyzing the AMFE feature. The extraction of flame suspect area is divided into two parts. First of all, we need update the background image to extract moving regions using background difference [9]. And
then flame suspect area will be extracted from the moving regions based on HSI color model and morphological image processing methods.
A. Background Update and Extraction of Moving Region Moving region is determined by background estimation and update
[10]. Now, we have two adjacent frames I n , I n+1 and background image B n of n-th frame (The image B n has no target). Firstly, we get inter-frame difference image D l by using the difference between two adjacent frames and get background difference image D h by using the difference between current frame and background image, which are described by: A Flame Detection Method Based on the Amount of Movement of the
Flame Edge
Hongke Xu, Yanyan Qin*, Yong Pan, Hao Chen
F
2013 Fourth International Conference on Intelligent Control and Information Processing (ICICIP) June 9 – 11, 2013, Beijing, China
1(,)(,)(,)
(,)(,)(,).l n n h n n D x y I x y I x y D x y I x y B x y +=?=? (1) And then we can get binary images F l and F h by processing the above two difference images, that is:
1, (,)(,)0, .
l l
l D x y T F x y else >?=?? (2) 1, (,)(,)0, .h h h D x y T F x y else >?=?? (3) Finally, we can get moving region image Y n via and operation between F l and F h , that is:
1,
(,)&(,)00,
.l h n F x y F x y Y else ≠?=?? (4) Whether the background image B n+1 should be updated depends on the existence of moving regions, that is:
1(,)(1)(,),0
(,)(,), 1.
n n n n n n B x y I x y Y B x y B x y Y αα++?=?=?
=?
(5) Where, α is the background update coefficient changing in
the range 0 to 1, which is 0.9 in this paper. Equation (5) shows that background image should not be updated when there are motion regions. Otherwise, background image should be updated.
B. Extract Flame Suspect Area Based on HSI Color Model Common color model have RGB and HSI. HSI color
model is more suitable to the human eyes for the perception of scenery, where H, S, and I respectively represent the hue component, the saturation component, and the luminance component. Numerous studies show that the hue, saturation, and luminance components of flame color have its specific range in the HSI color model, based on which flame suspect
area can be extracted from the moving regions. The steps are described as follows. Step1: Flame suspect area is extracted from the moving
regions via (6).
((<((,))<)
(,) (<((,))<)(,) (<((,))<)) 0 .
min max min max
min max H H f x y H f x y S S f x y S g x y I I f x y I else ???=?
???(6) Where, g(x, y) is flame suspect area image, f(x, y) is
moving region image in the HSI color model, H max , H min , S max , S min , I max , and I min are 6 thresholds. Study [11] shows that the range of values of H, S, and I respectively are:[0,60]H ∈,[0.2,1]S ∈,[0.5,1]I ∈.
Step2: Morphological image processing methods, such as continuous opening and closing operations, are utilized to smooth the edge of the suspect area, by eliminating isolated points and filling the slits for obtaining flame suspect area, which is shown in Fig. 1.
III. F EATURE OF THE AMOUNT OF MOVEMENT OF THE FLAME
EDGE Combustible will send out a strong infrared radiation when
it is burning, which causes flames flicker that means some
pixels in video change from the fire to the non- fire or from the non-fire to the fire in a very short time. This is unique
characteristic of the flame edge’s constant motion from distracters with similar flame color. Therefore, this paper detects flame based on this characteristic after extracting
flame suspect area.
In the actual observation, we find that flame edge motion
characteristic can be determined by AMFE. This paper
defines AMFE as the number of pixels that change from the fire to the non- fire or from the non-fire to the fire in the flame
suspect area of two adjacent frames. This means the AMFE of current frame is determined by the number of pixels that have above changes between previous frame and current frame. The binary area constituted by these pixels is called edge
motion region and the corresponding binary image is called flame edge motion image, as shown in Fig. 2.
If (1,2...)i S i N = are binary images of flame suspect area based on a video sequence of N frames, then the algorithm of AMFE calculation is described in Algorithm I.
(a) (b) (c)
Fig. 1. Flame suspect area extraction: (a) Current frame RGB image, (b) (c)
Flame suspect area images via HSI color model.
(a) (b) (c)
Fig. 2. Flame edge motion binary image: (a) Current frame binary image, (b) Previous frame binary image, (c) Flame edge motion image. The white areas in images of (a) and (b) are flame suspect areas.
11Algorithm I:Calculation of AMFE
2:0
(,)(,)11i i i i i i i i i i
i i i i
for i N do K J S AND S B S OR S M B J for x y M do if f x y K K end if end for return K end for
??=====?∈==+
Where,
i J is and operation binary image between two
adjacent frames, i B is or operator binary image between two adjacent frames, i M is flame edge motion binary image. (,)x y is coordinate of the pixels in i M and the corresponding (,)f x y is pixel value (which is 0 or 1). And i K represents the value of AMFE of current frame, as shown
in (c) of Fig. 2: the number of pixels in the white area.
IV. A NALYSIS OF AMFE FEATURE VIA WAVELET
TRANSFORM Table I shows different changes of the AMFE feature of the flame suspect area for different video sequences, such as brazier flame, candle flame and car lights, of 10 consecutive frames. The value in Table I represents the value of AMFE of each frame of each video sequence.
TABLE I
CONTRAST TABLE ON AMFE CHANGES
Sequence 1 2 3 4 5 Brazier Flame 2147 135 114 568 188
Candle Flame 65 51 88 58 37
Car Lights 15 40 91 16 24
Video
Sequence
Frame 6 Frame 7 Frame 8 Frame 9 Frame 10 Brazier Flame 1147 1112 363 235 89 Candle Flame 11 10 33 31 33 Car Lights 19 25 44 50 35
changes more obvious compared to the AMFE of some distracters with similar flame color, such as candle flame and car lights, having small changes. The reason is that flame flicker is severe, causing greater edge motion, which makes the difference of AMFE between two adjacent frames be larger. Therefore, we construct curve of the AMFE of flame suspect area, taking video frames as the horizontal axis while the AMFE of current frame as the longitudinal axis. The
curve is recorded as Q, as shown in Fig. 3.
Because the difference of AMFE of fire between adjacent frames is larger, curve Q must have irregular mutation. Wavelet transform is applied to analyze the location and the degree of changes in curve Q just for its good time-frequency localization analysis capability. Because flame has obvious AMFE feature caused by edge motion, its wavelet coefficient of the high-frequency components will has local maxima, whereas the distracter’s wavelet coefficient of the high-frequency components should be near zero, as shown in Fig. 4. Based on this, we can detect flame by the number of local maxima of the wavelet coefficient of the high-frequency components within a certain range.
Input Video Frames
A M F E
Fig. 3. AMFE curve of flame and distracters.
Input Video Frames
W a v e l e t C o e f f i c i e n t o f t h e H i g h -f r e q u e n c y C o m p o n e n t o f A M F E
Fig. 4. Curve of wavelet coefficient of the high-frequency components of
AMFE.
From Fig. 3, we find that the difference of AMFE of flame (brazier flame and wastepaper flame) between two adjacent frames is large and obvious, which is resulted from the characteristic of the flame edge’s constant motion. However, the difference of AMFE of distracters (lighter flame, candle flame, and car lights) is very small and not obvious, as shown in Fig. 4, its wavelet coefficient of the high-frequency components has no severe oscillation and the absolute values of the coefficient are most in the [0, 200] range. Based on the experimental results, this paper defines the absolute values of the wavelet high-frequency coefficient of the AMFE of fire to be [250, 1500] range. The number of absolute values, in [250, 1500] range, of the high-frequency coefficient of 20 frames for brazier flame and wastepaper flame are respectively 17
and 10 whereas the number for distracters is zero in almost (in Fig.4, the high-frequency coefficient of distracters is between two red horizontal lines, whose absolute value of the longitudinal axis is 250). Therefore, we take the threshold value as 9 and make a comparison between the threshold value and the number of absolute values in [250, 1500] range of 20 frames in a video sequence to detect flame.
From what has been discussed above, we can detect fire flame via the method based on AMFE and get the flow chart of flame detection, as shown in Fig. 5.
Fig. 5. Flow chart of flame detection
V. C ONCLUSIONS
Based on the AMFE feature caused by the characteristic of the flame edge’s constant motion, this paper proposes a flame detection method. This method extracts flame suspect area based on HSI color model, and then construct curve of AMFE of the flame suspect area. And then wavelet transform is applied on this curve to analyze the location and the degree of changes in the curve of AMFE, Finally, the number of absolute values, in [250, 1500] range, of the high-frequency coefficient of 20 frames is utilized to make flame detection.
Experiment results prove that this flame detection method is accurate and effective.
We can draw the following conclusions through the experiment. The suspect area of similar flame color can be extracted accurately via HSI color model and morphological image processing methods, such as continuous opening and closing operations. The flame detection method based on AMFE can differentiate the flame and general distracters, such as lighter flame, candle flame, and car lights.
R EFERENCES
[1] Fire Department of the Ministry of Public Security. Fire Statistical
Yearbook of China 2008. Chinese People's Public Security University Press, 2008.
[2] SUN Yu, WANG Jian, YU Yang, YUAN Hong-yong, SU Guo-feng,
“Fire detection based on image and CNC fixed fire monitor
extinguishing technique,” Fire Science and Technology, vol. 22, No. 3, 2003, pp.223-225.
[3] WANG Zhen-hua, WANG Hui-qin, WANG Yan-ni, LI Hu-sheng,
“Fire detection based on video image,” Automation Technology, Information Technology, vol. 30, No. 2, 2008, pp.161-164.
[4] XU Hong-ke, FANG Jian-wu, WEN Chang-bao, “Fire detection based
on brightness and color features of the fire edge regions,” Application Research of Computers, vol. 27, No. 9, 2010, pp.3581-3584.
[5] Byoung C K, Cheong K H, Nam J Y, “Fire detection based on vision
sensor and support vector machines,” Fire Safety Journal, No. 44, 2009, pp.322-329.
[6] B.UgurToreyin, Yigithan Dedeoglu, A.EnisCetin, “ComPuter vision
based method for real time fire and flame detection,” Science direct, 2005, pp.1-5.
[7] Liu Chebin, Ahuja N, Institute B, “Vision based fire detection,”
Proceedings of the 17th International Conference on Pattern Recognition. Cambridge, UK: IEEE Press, 2004, pp.134-137.
[8] Yamagishi H, Yamaguchi J, “A contour fluctuation data processing
method for fire flame deteetion using a color camera,” IEEE 26th Annual Conference on IECON of the Industrial Electronics Society, Nagoya, Japan, 2000, pp.824-829.
[9] Haritaoglu I, Harwood D, Davis L, “W4：real-time surveillance of
people and their activities,” IEEE Trans Pattern Analysis and Machine Intelligence, vol. 22, No. 8, 2003, pp.809-830.
[10] Hang Wenbing, Peng Jianwen, Chen Chiyuan, “A new image based
real-time flame detection method using color analysis,” Proceeding of Networking, Sensing and Control, 2005, pp.100-105.
[11] LIU Jian-wei, SU Ya-xin, WAN Jun-fu, CAO Quan, XIANG
Huai-kun, “Research of the video-based highway tunnel detection algorithms,” Traffic Safety, Dec. 2007, pp.118-120.