CN101393603A - Method for recognizing and detecting tunnel fire disaster flame - Google Patents

Abstract

The invention relates to a method for identifying and detecting flame of a tunnel fire hazard. The method comprises the following steps: performing a pretreatment of eliminating illumination on an input video stream; performing motion detection on the video stream to obtain a moving pixel; performing color detection on the video stream to obtain a color pixel with flame characteristics; searching communication areas which are formed by all mutually connected pixels according with same characteristics; calculating out the perimeters and the areas of the obtained communication areas to perform shape analysis; and performing area variation analysis on each communication area to finally judge whether a fire hazard happens. The method, based on a fire hazard detection system with image processing, can correctly identify flame information from an image sequence with background noise according to the characteristics of flame images that the color, the shape and glittery characteristic, the position, the area and the brightness of the flame images change over time, thus the method can separate the real flame from a glittery vehicle lighting in a tunnel, greatly reduce the false alarm rate, and achieve the aim of fire hazard monitoring.

Description

The method of a kind of identification and detection tunnel fire disaster flame

Technical field

The present invention relates to the method for a kind of identification and detection tunnel fire disaster flame, the car light of real flame and flicker in the tunnel can be made a distinction, thereby reduced rate of false alarm greatly.Disturb very big fire hazard monitoring place (as tunnel, night fire hazard monitoring etc.) to be applied to the various light that are subjected to.

Background technology

Based on the fire monitoring method of Flame Image Process utilized in the short time that fire just taken place flame image color, shape, blinking characteristics with and position, area, the time dependent characteristic of brightness, from the image sequence that contains background noise, correctly discern flame information, reach the purpose of fire monitoring.Compare with traditional temperature sense, cigarette sense fire detection alarm system, following advantage arranged based on the fire detecting system of Flame Image Process:

1. because the large scale of CCD camera, a kind of feasible means of large scene, large space being carried out detection are provided based on the fire detecting system of Flame Image Process.

2. applied widely.Because CCD camera and external environment are isolated, and therefore, can carry out the quick detection at fire initial stage based on the fire detecting system of Flame Image Process under the environment that conventional fire sniffers such as high temperature, Gao Chen can't normally be brought into play.

3. accuracy rate height, reaction velocity are fast.Owing to can from image, find out whether have fire to take place intuitively, in case therefore system produces warning, only need switch to the zone of warning, just can from monitor, directly confirm.Its reaction velocity and accuracy rate are all good than traditional fire hazard monitoring system.

4. be convenient to the cause of fire investigation.The CCD camera can be taken the overall process that fire begins, and these image informations will directly be kept in the memory device of pulpit.Therefore, can consult these image informations very easily afterwards, analyze culprit.

5. be convenient to utilize other functions of computer development.

Because its large scale, be specially adapted to large scales such as megastore, cinema, tunnel, large-sized workshop workshop, hangar, large ship, cargo hold, large space indoor and outdoor building place based on the fire detecting system of Flame Image Process.Consider in the present tunnel or the closed-circuit TV monitoring system of high-rise and high, therefore can in a cover hardware system, realize above two kinds of monitoring functions fully based on the hardware degree of coupling of the fire hazard monitoring detection system of Flame Image Process.

Fire monitoring method based on Flame Image Process mainly carries out discriminance analysis from aspects such as color characteristic, textural characteristics, shape facility and motion features at present.

Color is the notable attribute of image, compares with other features, and color characteristic calculates simple, stable in properties, and is all insensitive for rotation, translation, dimensional variation, shows very strong robustness.Color characteristic comprises color histogram, main color, mean flow rate etc.

Texture analysis is an important research direction of computer vision always, and its method can roughly be divided into statistical method and structural approach.Statistical method is that the space distribution information of the color intensity of image is added up, can be further divided into again traditional based on model statistical method and based on the method for spectrum analysis, as Markov random field model, Fourier spectral characteristic etc.Structural approach at first suppose texture pattern by texture cell according to certain regularly arranged composition, so texture analysis just becomes and determines these unit, their spatial disposition of quantitative test.

Shape analysis at first will split object from background, re-use the similarity comparison that the whole bag of tricks such as circularity, rectangle degree, square carry out shape.Shape facility has the unchangeability to translation, rotation, convergent-divergent, and the expression of shape can be divided into based on the border with based on regional 2 classes usually.Shape facility based on the border can comprise complicated border with less parameter, as the Fourier descriptor.Shape facility square invariant commonly used based on the zone is described.Because the similarity of shape relatively is still a very difficult problem, thereby makes in field of video processing at present and be used less.

Motion feature is the key character of video lens, has reflected that the time domain of video changes, and the behavioral characteristics main contents that also user can provide during video frequency searching often.The method of motion analysis has spatiotemporal mode, two-dimensional parameter motion model, pixel-recursive method and the bayes method etc. of the method based on optical flow equation, block-based method, MPEG motion vector, section.

Summary of the invention

The object of the present invention is to provide the method for a kind of identification and detection tunnel fire disaster flame, this method can make a distinction the car light of real flame and flicker in the tunnel, thereby has reduced rate of false alarm greatly.

Above-mentioned purpose of the present invention is achieved through the following technical solutions:

1. the method discerning and detect tunnel fire disaster flame is characterized in that comprising the steps:

1) video flowing of input is rejected the pre-service of illumination: for black and white that is filmed by the video camera that is installed in tunnel top under the various situations in the tunnel or color video picture, at first coloured image is converted into gray level image, use the method for gamma transformation to reject unnecessary illumination then, wherein the threshold value of gamma transformation is dynamically determined by the maximal value of pixel grey scale in the computed image;

2) video flowing is carried out motion detection, obtain the motion pixel: to the image after the illumination pretreatment that is obtained in the step 1, use has the time-domain difference method of fixed threshold and carries out motion detection, at first the initialization background image utilizes the relevance between frame and the frame to come background image updating and foreground image according to present frame then;

3) video flowing is carried out color detection, acquisition has the flame characteristic color pixel: have the pixel of flame color by extracting in training video and picture, analyze its intensity level or RGB component value, if the color value of current pixel is positioned at the pixel range inside that meets the flame color feature, then this pixel is judged as the pixel with flame color, enters the detection of next stage;

4) search meets the connected region of same characteristic features and interconnective pixel composition to all: for the image after motion detection and the color detection, carry out the search of connected region; Connected region search comprises zone marker and two steps of range searching: at first use mask method respectively motion pixel region, flame color pixel region and the pixel region that belongs to flame fringe to be carried out mark, use the BFS (Breadth First Search) algorithm to carry out the search of connected region then;

5) connected region of gained is calculated its girth and area, carry out shape analysis: shape analysis comprises: the border that the method for using depth-first search algorithm combining form to learn is extracted each connected region; Calculate the girth on each connected region border respectively; Calculate the area of each connected region; Calculate the circularity of each connected region;

6) to each connected region, carry out area and change component analysis, judge at last whether fire takes place: comprise in this step that mark belongs to the pixel in flame fringe zone; The connected region of using the above-mentioned pixel of BFS (Breadth First Search) algorithm search to be formed; Set up data structure and store the connected region that finds; Use arrives first the connected region of the order coupling front and back frame correspondence of handling earlier; Calculate the area change amount of corresponding connected region, judge whether fire takes place.

The method of identification of the present invention and detection tunnel fire disaster flame, be based on the fire detecting system of Flame Image Process, can according to flame image color, shape, blinking characteristics with and position, area, the time dependent characteristic of brightness, correct identification flame information from the image sequence that contains background noise, the car light of real flame and flicker in the tunnel is made a distinction, reduce rate of false alarm greatly, can reach the purpose of fire monitoring.

Description of drawings

Fig. 1 is the process flow diagram of the inventive method.

Embodiment

Be distributed in the tunnel with the tunnel outside each highway section camera taken screen transition become simulating signal by Optical Fiber Transmission to the CCTV controller of pulpit.On the CCTV controller, simulating signal is converted into digital signal, a part is sent on the computer monitor screen that is positioned at each Control Room, and another part is sent in the digital monitoring main frame, and some be encoded (being generally mpeg encoded) stored in the DVR.

As shown in Figure 1, the present invention's invention comprises that input video stream changes processes such as component analysis by pre-service, motion detection, color detection, connected region search, shape analysis and area, produce a comprehensive judged result the most at last.

Below each step is elaborated:

1. pre-service

Because influence and interference ratio that relative confined space such as tunnel is subjected to illumination are bigger, directly original image is used for motion detection and will produces unacceptable effect, thereby the operation after the influence therefore need be by reducing the influence of illumination someway as far as possible.Adopt the method for gamma transformation (being also referred to as power time conversion) to handle.

The citation form of power time conversion is:

s＝cr ^γ (1)

Wherein c and γ are positive constant.Sometimes consider side-play amount (i.e. measured output when being input as 0), formula (2) is also write and is s=c (r+ ε) ^γIn any case, side-play amount normally shows deriving of demarcation, and generally neglects in formula (3).The part value of γ is mapped to the broadband output valve to the dark value in input arrowband in the power time conversion.Also set up when on the contrary, importing the high value.

Method among the present invention at first changes into the gray level image (if there is) to coloured image, travels through entire image then, finds gray-scale value g the highest in all pixels.Promptly

g＝max{g ₁，g ₂，...，g _M×N} (2)

Use the deformation formula of power time conversion then:

s＝c(r-g) ^γ (3)

And

s＝c(r+L-1-g) ^γ (4)

Wherein L is the number of greyscale levels (being generally 256) of grayscale image.

2. motion detection

Through after the pre-service, just can carry out motion detection to the frame of video of output.The purpose of motion detection is preliminary car light and the flame of distinguishing motion.

Use the method for time-domain difference to judge motion pixel and moving region.(x, y) gray-scale value of the pixel on is designated as g being positioned at coordinate in the i+1 frame _i(x, y), (x, y) background pixel value on the coordinate is designated as B in first frame ₀(x, y).

Initial situation B ₀(x, y)=g ₀(x, y); For each frame, the next frame background pixel value of being predicted is upgraded according to present frame background pixel value and current actual pixel value afterwards:

$B_{i+1}(x,y)=\left\{\begin{array}{cc}{\mathrm{\&alpha;B}}_i(x,y)+(1-\mathrm{\&alpha;})g_i(x,y)& \mathrm{if}|g_i(x,y)-g_{i-1}(x,y)|<T\\ B_i(x,y)& \mathrm{else}\end{array}\right.---\left(5\right)$

Wherein α is a scale-up factor, the speed of expression context update, and generally its value is near 1.

At last, if satisfy following inequality, then think coordinate position (x, y) pixel on is the motion pixel:

|g _i(x，y)-B _i(x，y)|>T (6)

3. color detection

Have the pixel of flame color by in training video and picture, extracting, analyze its intensity level (black and white picture or video) or RGB component value (colour picture or video).The color value of note current pixel is I _R, I _G, I _B(black and white then is intensity values of pixels I _g), if then satisfy following condition, then this pixel is judged as the pixel with flame color, enters the detection of next stage:

L _R1＜I _R＜L _R2, L _G1＜I _G＜L _G2, L _B1＜I _B＜L _B2Or L _G1＜I _g＜L _G2

4. connected region search

After motion pixel and element marking with flame color come out, will use a kind of method that is called mask (mask) that mark is carried out in the zone.

Three masks (all correspondence positions all are changed to 0) that have identical size with primitive frame of initialization at first, these three masks are respectively applied for motion detection, color detection and the area that will mention afterwards changes component analysis.After the motion detection step, all motion pixels value on the correspondence position in the motion mask in the present frame is changed to 1; Equally, through after the color detection step, all pixel values on the correspondence position in the color mask that meet the flame color feature in the present frame are changed to 1.

For the motion mask, use the breadth First algorithm to search for connected component then.

Use the necessary and sufficient condition of breadth First algorithm to have three:

1. one group of concrete state is arranged, and state is each situation that problem may occur; The state space that all states are constituted is limited; Problem scale is less.

2. in the answer process of problem, can from a state according to problem given condition, change another one or several state into.

3. can judge the legitimacy of a state, and clear and definite one or more dbjective states are arranged.

4. problem to be solved is: find out dbjective state according to given original state, or according to given original state and done state, find out a path from the original state to the done state.

At first, construct a queue data structure, a search starting point is specified in the position that is marked as the motion pixel in mask arbitrarily, writes down its coordinate, and makes it enter formation; Then with current as basic point, search for its 8 adjacent pixels, if there is the pixel that is marked as motion in its 8 adjacent pixels, then join the team according to the order of sequence (noting its coordinate equally), in mask, the value on the correspondence position is labeled as simultaneously and handled according to the precedence of search.When can not find the point that meets described condition, then search stops.

Remember that all collection of pixels are V in the frame, all collection of pixels that are marked as motion are V in the same frame _m, and $V_M\&SubsetEqual;V.$ Being used for the formation of storing moving pixel is designated as Q.

Original state:

$V_m=\{v_m^1,v_m^2,...,v_m^k\},$ 0<k≤M×N

The first step: get i ∈ 1,2 ..., k} and

$V_m=V_m-\left\{v_m^i\right\},$ $Q=Q+\left\{v_m^i\right\};$

Second step: if $\&Exists;v_m^j,v_m^j\&Element;V_m$ And $v_m^j\&Element;N_8\left(v_m^i\right),$ Then $V_m=V_m-\left\{v_m^j\right\},$ $Q=Q+\left\{v_m^j\right\};$

The 3rd step: if $\&ForAll;v_m^j\&Element;N_8\left(v_m^i\right),$ $v_m^j\&NotElement;V_m$ And $v_m^j\&Element;Q,$ Then $Q=Q-\left\{v_m^i\right\}.$

Each order (first in first out) of pressing FIFO takes out a pixel from Q Repeat three steps of the first step to the, up to satisfying following end condition:

And

In like manner can carry out connected component search for the color mask.

5. shape analysis

(1) extract the border of each connected region:

Use the depth-first search method to extract the border of connected region.

The frontier point set of a connected region of note is E, $E\&Subset;V;$ The set of all boundary pixels is designated as V _e

Original state: A pixel v who belongs to the connected region border _i

The first step: E=E+{v _i, V _e=V _e-{ v _i;

Second step: right $\&ForAll;v_j\&Element;N_8\left(v_i\right),$ If $\&Exists;v_k\&Element;N_4\left(v_j\right)$ And $v_k\&NotElement;V_m,$ E=E+{v then _j, V _e=V _e-{ v _j;

Repeat this two steps afterwards at every turn, up to satisfying following end condition:

(2) calculate the girth on each connected region border respectively:

In (1) step during the depth-first search connected region, use be recursive algorithm, every increase one deck of recursive tree so, the variable that is used for depositing the connected region perimeter value is also from increasing 1, when recurrence finished, what obtain was exactly the perimeter value of this connected region naturally.

(3) calculate the area of each connected region:

When the BFS (Breadth First Search) connected region, use formation to store pending pixel, so every next pixel is entered team, is used for depositing the variable of connected region area value also from increasing 1, when satisfying end condition, what obtain is exactly the area value of this connected region naturally.

(4) calculate the circularity of each connected region:

The connected region girth that note is calculated is C, and area is S, and then circularity can be calculated as:

D _circle＝C ²/4πS (7)

The circularity that is calculated illustrates then that more near 1 the shape of connected region is regular more, and then it is that the probability of flame is low more.

6. area changes component analysis

(1) mark belongs to the pixel in flame fringe zone:

Still use the method for mask to come mark to belong to the pixel in flame fringe zone.If an intensity values of pixels is labeled as 1 in the mask corresponding position so, otherwise is labeled as 0 less than certain preassigned intensity level P (intensity level of expression flame kernel belongs to the edge of flame than little this pixel that just illustrates of this value).

(2) search for the connected region that above-mentioned pixel is formed:

Use the BFS (Breadth First Search) algorithm to search plain each connected region, process can no longer describe in detail with reference to the 4th joint.

(3) set up data structure and store the connected region that finds:

Utilization structure body array is stored the connected region that finds, and is used for storing length and wide, the lower left corner apex coordinate of connected region boundary rectangle, the area of connected region self and the number of times that current connected region is judged as flame region of sequence number, the connected region boundary rectangle of connected region respectively.

Last is done explanation at this: if current connected region is judged as flame region in a frame, this method does not think at once that current region is a flame region; When having only the number of times that is judged as flame region continuously when this same connected region to surpass a threshold value, think that just this zone is a flame region.

(4) connected region of frame correspondence before and after the coupling:

All connected region set are C on the note current video frame, and then initial situation is

Afterwards whenever finding a connected region R _iThe time, it is added among the set C by the order of FIFO:

C＝C+{R _i}

When handling next frame, at first check first connected region of current connected region set C, if this zone has exceeded indication range, promptly

Then its deletion from the connected region set; Otherwise do not do any operation.

Then for each region R _i, itself and the connected region when pre-treatment are carried out the comparison of area change amount; If detect new connected region R _k, k〉and max{i}, then add it among connected region set C to: C=C+{R _k.

(5) the area change amount of the corresponding connected region of calculating:

$\mathrm{\&Delta;A}=\frac{\mathrm{dA}}{\mathrm{dt}}=\frac{A_{R_i}-A_R}{t_{i+1}-t_i}$

Wherein

The area of pairing connected region in the expression previous frame, and A _RThe area of the current connected region that compares of expression.

At last, if T _Low＜Δ A＜T _High, this connected region may be flame region so.

Those of ordinary skill in the art will be appreciated that, above embodiment is used for illustrating the present invention, and be not as limitation of the invention, as long as in essential scope of the present invention, all will drop in the scope of claims of the present invention variation, the modification of the above embodiment.

Claims (2)

1. the method discerning and detect tunnel fire disaster flame is characterized in that comprising the steps:

1) video flowing of input is rejected the pre-service of illumination: for black and white that is filmed by the video camera that is installed in tunnel top under the various situations in the tunnel or color video picture, at first coloured image is converted into gray level image, use the method for gamma transformation to reject unnecessary illumination then, wherein the threshold value of gamma transformation is dynamically determined by the maximal value of pixel grey scale in the computed image;

2) video flowing is carried out motion detection, obtain the motion pixel: to the image after the illumination pretreatment that is obtained in the step 1, use has the time-domain difference method of fixed threshold and carries out motion detection, at first the initialization background image utilizes the relevance between frame and the frame to come background image updating and foreground image according to present frame then;

3) video flowing is carried out color detection, acquisition has the flame characteristic color pixel: have the pixel of flame color by extracting in training video and picture, analyze its intensity level or RGB component value, if the color value of current pixel is positioned at the pixel range inside that meets the flame color feature, then this pixel is judged as the pixel with flame color, enters the detection of next stage;

4) search meets the connected region of same characteristic features and interconnective pixel composition to all: for the image after motion detection and the color detection, carry out the search of connected region; Connected region search comprises zone marker and two steps of range searching: at first use mask method respectively motion pixel region, flame color pixel region and the pixel region that belongs to flame fringe to be carried out mark, use the BFS (Breadth First Search) algorithm to carry out the search of connected region then;

5) connected region of gained is calculated its girth and area, carry out shape analysis: shape analysis comprises: the border that the method for using depth-first search algorithm combining form to learn is extracted each connected region; Calculate the girth on each connected region border respectively; Calculate the area of each connected region; Calculate the circularity of each connected region;

6) to each connected region, carry out area and change component analysis, judge at last whether fire takes place: comprise in this step that mark belongs to the pixel in flame fringe zone; The connected region of using the above-mentioned pixel of BFS (Breadth First Search) algorithm search to be formed; Set up data structure and store the connected region that finds; Use arrives first the connected region of the order coupling front and back frame correspondence of handling earlier; Calculate the area change amount of corresponding connected region, judge whether fire takes place.

Images