EP1496483B1 - Method and apparatus for the detection of flames - Google Patents

Abstract

A flame detection procedure (2) analyses a video (1) image (3) for bright spots and moving areas using a weighted accumulation matrix (4) constructed from the differences between successive images with low intensity and so dark moving objects filtered (5) out to leave small areas of interest for integration (6) and flame presence probability (7) estimation (8). Includes INDEPENDENT CLAIMs for equipment using the procedure.

Description

The present invention relates to a method and a device for detecting flames in a monitored area referred to below as a monitoring space by analyzing at least one parameter of a radiation occurring in the interstitial space.

WO 02/093525 A1 discloses an apparatus and a method for the simultaneous processing of first and second image processing, especially for the detection of flames on the images. The apparatus includes an image sensor for creating video images, a frame grabber for receiving a first and a second image (frame), a processor for processing the captured images, and an output unit.

Previously known devices, as described for example in US Pat. Nos. 4,866,420 and 4,280,058, contain at least one sensor which evaluates the flicker frequency spectrum of the radiation, whereby signals located outside a specific frequency band are evaluated as interference signals. Thus one uses the typical flickering of the flames in a very low-frequency oscillation range as a feature for distinguishing between the radiation emitted by a flame and interfering radiation. The determination of the frequency band takes place in the simplest case by the sensor upstream filter or by this downstream frequency-selective amplifier, in both cases, a certain passband of, for example, 5 to 25 Hz is obtained.

Apart from that, even with optimal tuning of the frequency band to the flickering of flames malfunctions and false positives can not be ruled out because it happens again and again that accidental Intensity changes of the ambient radiation are in the passband. Such changes in intensity may be caused, for example, by shadows or reflections from vibrating or slowly moving objects, reflections of the sunlight on water surfaces, or flickering or fluctuating light sources.

The invention will now be given a method for the detection of flames, which is characterized by a high noise immunity at low cost and.

This object is achieved according to claims 1 and 10.

According to the method according to the invention, the search for regions of high light intensity and local flickering movement takes place with the aid of an accumulation matrix which is obtained from the weighting factor-weighted difference images of successive intensity images, wherein the weighting factor indicates how much the accumulation matrix is accumulated on the difference images.

A first preferred embodiment of the method according to the invention is characterized in that the video images are generated at a specific frequency and intensity images are obtained therefrom.

A second preferred embodiment of the method according to the invention, characterized in that the coordinates of the brightest pixels are searched with the aid of the accumulation matrix.

A third preferred embodiment is characterized in that an interesting image area containing the brightest or the brightest pixels and reduced in relation to the original image area is defined and analyzed for the presence of a flame.

Further preferred embodiments of the method according to the invention emerge from the dependent claims 7 to 9.

Preferred embodiments of the device according to the invention are claimed in the dependent claims 11 to 15.

With the proliferation of CCTV systems and systems, it can be assumed that in many cases a video camera will be present in a surveillance room, so that a separate sensor does not have to be installed for flame detection, which undoubtedly means a cost reduction. A further cost reduction results from the restriction of the evaluation to the possibly containing a flame image sections, which allows a significant reduction of computer power. It can also be assumed that the evaluation of these image sections is sufficiently robust to interference.

In the following, the invention is explained in more detail by way of example with reference to a block diagram showing a device according to the invention for detecting flames. The reference numeral 1 denotes a video camera which has video output via an output to an evaluation stage 2 supplies, wherein the evaluation stage 2 can be integrated into the camera 1 or connected to this. The evaluation level 2 can be provided at the site of the camera 1 or in the immediate vicinity of this or it can also be spatially distant from the camera 1, wherein in the latter case between camera 1 and evaluation level 2 is a communication link.

The evaluation stage 2 contains a processor (not shown), which has an algorithm for the localization of flames found in the images of the camera 1 and the subsequent analysis of the corresponding image detail. According to the representation, the extraction of intensity and / or chrominance images X _ij (t) (hereinafter referred to as intensity images) from the video sequences supplied by the camera 1 takes place in a first process of the algorithm designated image acquisition; i and j are the coordinates of each pixel. The frequency of these pictures is at least 15 frames per second, the picture size for example 352 times 288 pixels. Intensity images are obtained because it can be assumed that a flame is a place of high light intensity and also has a characteristic hue.

• In einem ersten Schritt erfolgt die Bestimmung des Maximalwerts max [X_ij(t)] und des Mittelwerts mean [X_ij(t)] der Intensität und daraus wird eine Helligkeitschwelle q(t) bestimmt, wobei gilt: $$q\left(\mathit{t}\mathit{+}1\right)={\mathrm{\lambda }}_1\max \left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right],\mathrm{wenn\; mean}\mathit{\left[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{\right]}<{\mathrm{\lambda }}_1\max \left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right],$$
  
  und $$q\left(\mathit{t}\mathit{+}1\right)={\mathrm{\lambda }}_2\left\{\max \left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right]-\mathrm{mean}\mathit{\left[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{\right]}\right\}+\mathrm{mean}\left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right],$$
  
  in allen anderen Fällen. λ₁ und λ₂ sind Konstante, die zwischen 0 und 1 liegen, wobei beispielsweise λ₁ gleich 0.68 und λ₂ gleich 0.05 ist.
• Mit Hilfe dieser beiden Bedingungen wird ein die Flammeneigenschaften berücksichtigender Gewichtungsfaktor w_ij(t) bestimmt: $$w_{\mathit{ij}}\left(\mathit{t}\right)=X_{\mathit{ij}}\left(\mathit{t}\right),\mathrm{wenn}{X}_{\mathit{ij}}\left(\mathit{t}\right)>\max \mathit{[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{-}\mathit{q}\left(\mathit{t}\right),$$
  
  und $$w_{\mathit{ij}}\left(\mathit{t}\right)=0$$
  
  in allen anderen Fällen.

In a next process designated preprocessing 4, flames in the intensity images X _ij (t) are searched for and the localized flames are localized in corresponding image sections. This localization takes place by means of a so-called accumulation matrix, which is formed in the following way:

• In a first step, the determination of the maximum value max [X _ij (t)] and the mean value mean [X _ij (t)] of the intensity and from this a brightness threshold q (t) is determined, where: $$q\left(\mathit{t}\mathit{+}1\right)={\mathrm{<figure-callout\; id="1"\; label="\lambda "\; filenames="imgf0001.png"\; state="\{\{state\}\}">\lambda </figure-callout>}}_1\mathrm{Max}\left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right].\mathrm{if\; mean}\mathit{\left[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{\right]}<{\mathrm{<figure-callout\; id="1"\; label="\lambda "\; filenames="imgf0001.png"\; state="\{\{state\}\}">\lambda </figure-callout>}}_1\mathrm{Max}\left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right].$$
  
  and $$q\left(\mathit{t}\mathit{+}1\right)={\mathrm{\lambda }}_2\left\{\mathrm{Max}\left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right]-\mathrm{mean}\mathit{\left[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{\right]}\right\}+\mathrm{mean}\left[{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\right].$$
  
  in all other cases. λ ₁ and λ ₂ are constants that lie between 0 and 1, where, for example, λ ₁ is 0.68 and λ ₂ is 0.05.
• With the aid of these two conditions, a weighting factor w _ij (t) which takes into account the flame properties is determined: $$w_{\mathit{ij}}\left(\mathit{t}\right)=X_{\mathit{ij}}\left(\mathit{t}\right).\mathrm{if}{X}_{\mathit{ij}}\left(\mathit{t}\right)>\mathrm{Max}\mathit{[}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{-}\mathit{q}\left(\mathit{t}\right).$$
  
  and $$w_{\mathit{ij}}\left(\mathit{t}\right)=0$$
  
  in all other cases.

This means that all pixels with an intensity below the value max [X _ij (t)] - q (t) , ie dark objects, are filtered out and ignored. As will be seen shortly, they are dark, moving objects that are filtered out.

Since one can assume that a flame is recognizable as a movement of high light intensity, a difference image is formed by comparing successive images to find out such a movement; since dark objects fall out due to the definition of the weighting factor w _ij (t), moving, dark objects that can not be flames are filtered out when forming the difference image. So you are looking for areas of high light intensity and local flickering and that means that, for example, a stationary light source that does not flicker would not be interpreted as a flame and also not a transversely across the interstitial space moving lamp.

For the difference image Q _ij (t) : $$Q_{\mathit{ij}}\left(\mathit{t}\right)=\left|{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\right)\mathit{-}{\mathit{X}}_{\mathit{ij}}\left(\mathit{t}\mathit{-}1\right)\right|w_{\mathit{ij}}\left(\mathit{t}\right)$$

Then the accumulation matrix A _ij (t) is determined from the difference image Q _ij (t) : $$A_{\mathit{ij}}\left(\mathit{t}\right)=\mathrm{\alpha }{A}_{\mathit{ij}}\left(\mathit{t}\mathit{-}1\right)+\left(1-\mathrm{\alpha }\right)Q_{\mathit{ij}}\left(\mathit{t}\right)$$

α is a constant between 0 and 1, which indicates how much the difference image Q _ij (t) flows into the accumulation matrix A _ij (t) . For α = 0, the accumulation matrix becomes equal to the difference image, and for α = 1, the difference image no longer has any influence because A _ij (t) equals A _ij (t-1) . The accumulation matrix is formed primarily to obtain a smoothed image with no noise and short-term changes.

As the last step of preprocessing 4, the pixel or pixels [i _m , j _m ] (t) having the highest value are searched by means of the accumulation matrix and a so-called interesting image area ROI containing this or these pixels is defined, in which there could be a flame: $$\left[{\mathit{i}}_{\mathit{m}},,{\mathit{j}}_{\mathit{m}}\right]\left(\mathit{t}\right)=\left\{\left(\mathrm{i},,\mathrm{j}\right)|\mathrm{Max}\left[{\mathit{A}}_{\mathit{ij}}\left(\mathit{t}\right)\right]\right\}$$

This determination thus provides the coordinates of the pixels with local flickering and maximum brightness. It will usually be a single pixel, but of course also several brightest pixels can be determined, what you can use a multi-channel selection and preferably specifies minimum distances between the individual pixels.

In the next process designated by analysis 5, a significant data reduction first takes place in that the analysis does not take place in the entire original image of 352 by 288 pixels but in the image region ROI of interest having a reduced size of, for example, 32 by 32 pixels. This results in a reduction to one hundredth. Of course, this reduction can also be lower, for example, to a fiftieth, or even much stronger.

• Mittlere Helligkeit L(t) des interessierenden Bildbereichs X _ROI (t): $$L\left(\mathit{t}\right)=\left[\mathrm{mean\; von}{X}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}\right]$$
  
• Chrominanz C(t) des interessierenden Bildbereichs X _ROI (t):
  • C(t) = [Anzahl der C_ij(t)|_ROI] ∈ "Feuer-Chroma-Sektor"/R(t), wobei C_ij(t) das Chrominanzpaar (V_ij, U_ij) des Bildes X_ij(t) zur Zeit t bezeichnet. YUV ist eine bekannte Darstellung des Farbraums, mit zwei Farbkomponenten U und V auf der x- bzw. der y-Achse und der Intensität Y auf der z-Achse, wobei die Länge des Vektors vom Nullpunkt zu einem Pixel in der UV-Ebene die Farbsättigung dieses Pixels angibt. Der Feuer-Chroma-Sektor R(t) ist ein Sektor des Farbraums in der UV-Ebene, in dem für eine Flamme typischen Farbbereich, der insbesondere die Farbe rot enthält.
• Anzahl der aktiven Pixel R(t) der Akkumulationsmatrix A _ROI(t): $$\begin{array}{c}\begin{array}{cc}\begin{array}{c}R\left(\mathit{t}\right)=\left[\mathrm{Anzahl\; der}{A}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}>{\mathrm{\eta }}_1\right];\end{array}& 1\le {\mathrm{\eta }}_1<\mathrm{Z}\left(\mathrm{Z}=\mathrm{Gesamtanzahl\; Pixel\; des\; interessierenden\; Bildbereichs\; ROI}\right),\mathrm{beispielsweise\; ist}{\mathrm{\eta }}_1=30\end{array}\end{array}$$
  
• Sättigungsgrad S(t) des interessierenden Bildbereichs X _ROI (t): $$\begin{array}{ccc}S\left(\mathit{t}\right)=\left[\mathrm{Anzahl\; der}{X}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}>{\mathrm{\eta }}_2\right]\mathrm{.}& 1\le {\mathrm{\eta }}_2<\mathrm{Z}& \mathrm{beispielsweise\; ist}{\mathrm{\eta }}_2=5\end{array}$$
  

Then, for each image area of interest, the following image information is determined:

• Average brightness L (t) of the image area of interest X _ROI (t) : $$L\left(\mathit{t}\right)=\left[\mathrm{mean\; of}{X}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}\right]$$
  
• Chrominance C (t) of the image area of interest X _ROI (t) :
  • C (t) = [number of C _ij (t) | _ROI ] ∈ "Fire Chroma Sector" / R (t) , where C _ij (t) denotes the chrominance pair (V _ij , U _ij ) of the image X _ij (t) at time t. YUV is a well-known representation of the color space, with two color components U and V on the x and the y-axis and the intensity Y on the z-axis, where the length of the vector from the zero point to a pixel in the UV plane Indicates the color saturation of this pixel. The fire chroma sector R (t) is a sector of the color space in the UV plane, in the color region typical for a flame, which in particular contains the color red.
• Number of active pixels R (t) of the accumulation matrix A _ROI ( t ): $$\begin{array}{c}\begin{array}{cc}\begin{array}{c}R\left(\mathit{t}\right)=\left[\mathrm{number\; of}{A}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}>{\mathrm{<figure-callout\; id="1"\; label="\eta "\; filenames="imgf0001.png"\; state="\{\{state\}\}">\eta </figure-callout>}}_1\right];\end{array}& 1\le {\mathrm{\eta }}_1<\mathrm{Z}\left(\mathrm{Z}=\mathrm{Total\; number\; of\; pixels\; of\; the\; image\; area\; ROI\; of\; interest}\right).\mathrm{for\; <figure-callout\; id="1"\; label="example\; \eta "\; filenames="imgf0001.png"\; state="\{\{state\}\}">example\; </figure-callout>}{\mathrm{\eta }}_{}{\mathrm{}}_1=30\end{array}\end{array}$$
  
• Saturation degree S (t) of the image area of interest X _ROI (t) : $$\begin{array}{ccc}S\left(\mathit{t}\right)=\left[\mathrm{number\; of}{X}_{\mathit{ij}}\left(\mathit{t}\right){|}_{\mathrm{ROI}}>{\mathrm{<figure-callout\; id="1"\; label="\eta "\; filenames="imgf0001.png"\; state="\{\{state\}\}">\eta </figure-callout>}}_2\right]\mathrm{,}& 1\le {\mathrm{\eta }}_2<\mathrm{Z}& \mathrm{for\; example}{\mathrm{\eta }}_2=5\end{array}$$
  

• Mittelwert Helligkeit:     F_L = L
• Mittelwert Frequenz:    F_F = F
• Mittelwert Amplitude:     F_M = M
• Mittelwert Feuer-Chroma-Pixel:     F_C = C
• Mittelwert aktive Pixel:     F_R = R
• Mittelwert Sättigung:     F_S = S

In order that the result remains stable, a time integration of the image information determined in the analysis 5 is then carried out in a process designated by extraction 6. For example, if the integration is done over 1 second, it will span 25 frames in PAL format. Thus, the average brightness, the chrominance, the active pixels and the degree of saturation are integrated over the time t from t ₀ to t _n and are given the following properties:

• Average brightness: F _L = L
• Mean frequency: F _F = F
• Mean value amplitude: F _M = M
• Average Fire Chroma Pixel: F _C = C
• Mean active pixels: F _R = R
• Mean saturation: F _S = S

The mean value of the frequency is obtained, for example, by counting the pixels of the average brightness L (t) . The frequency caused by the characteristic flickering of a flame is an important quantity for the detection of a flame because it is within a defined narrow range of usually between 1 Hz and 10 Hz.

$${\mathit{\Psi }}_L=\mathit{\Gamma }\left({\mathit{F}}_{\mathit{L}}\right)=0,\mathrm{wenn}{F}_L<{\mathrm{\delta }}_{\mathrm{L}}$$

und so weiter für die anderen Eigenschaften.In the subsequent process, designated by pattern recognition 7, the probability that a flame is present is calculated from the properties obtained during extraction 6. In this case, for example, it is examined for each of the above properties whether the mean value is above or below a threshold value and the probability is set equal to one or equal to zero. Then, from the probabilities of all n properties, a total probability is formed. $${\mathit{\Psi }}_L=\mathit{\Gamma }\left({\mathit{F}}_{\mathit{L}}\right)=1.\mathrm{if}{F}_L>{\mathrm{\delta }}_{\mathrm{L}}$$

$${\mathit{\Psi }}_L=\mathit{\Gamma }\left({\mathit{F}}_{\mathit{L}}\right)=0.\mathrm{if}{F}_L<{\mathrm{\delta }}_{\mathrm{L}}$$

and so on for the other properties.

$$\mathit{\Pi }\left(t\right):=\frac{1}{N_F}{\displaystyle \sum _n}{\mathit{\Psi }}_n=\left({\mathit{\Psi }}_{\mathit{L}}{\mathit{w}}_{\mathit{L}}+{\mathit{\Psi }}_{\mathit{F}}{\mathit{w}}_{\mathit{F}}+{\mathit{\Psi }}_{\mathit{M}}{\mathit{w}}_{\mathit{M}}+{\mathit{\Psi }}_{\mathit{C}}{\mathit{w}}_{\mathit{C}}+{\mathit{\Psi }}_{\mathit{R}}{\mathit{w}}_{\mathit{R}}+{\mathit{\Psi }}_{\mathit{S}}{\mathit{w}}_{\mathit{S}}\right)/N_F$$

für w_i gilt 0 ≤ w_i ≤ 1, wobei die Werte w_i empirisch bestimmt werden. N_F ist die Summe der w_i über alle i. Total probability Π (t): $$\mathit{\Pi }\left(t\right):=1/N_F\mathrm{\Sigma }{\mathit{\Psi }}_n=\left({\mathit{\Psi }}_L\mathrm{,}{w}_L+{\mathit{\Psi }}_F\mathrm{,}{w}_F+{\mathit{\Psi }}_M\mathrm{,}{w}_M+{\mathit{\Psi }}_C\mathrm{,}{w}_C+{\mathit{\Psi }}_R\mathrm{,}{w}_R+{\mathit{\Psi }}_S\mathrm{,}{w}_S\right)/N_F$$

$$\mathit{\Pi }\left(t\right):=\frac{1}{N_F}{\displaystyle \underset{n}{\Sigma }}{\mathit{\Psi }}_n=\left({\mathit{\Psi }}_{\mathit{L}}{\mathit{w}}_{\mathit{L}}+{\mathit{\Psi }}_{\mathit{F}}{\mathit{w}}_{\mathit{F}}+{\mathit{\Psi }}_{\mathit{M}}{\mathit{w}}_{\mathit{M}}+{\mathit{\Psi }}_{\mathit{C}}{\mathit{w}}_{\mathit{C}}+{\mathit{\Psi }}_{\mathit{R}}{\mathit{w}}_{\mathit{R}}+{\mathit{\Psi }}_{\mathit{S}}{\mathit{w}}_{\mathit{S}}\right)/N_F$$

for w _i , 0 ≦ w _i ≦ 1, where the values w _{i are} determined empirically. N _F is the sum of w _i over all i.

wenn Π(t) > κ

dann I(t) = I(t-1) + σ₊ (gesättigt zu S₊, wenn I(t) > S₊ in allen anderen Fällen gilt I(t) = I(t-1) - σ_- (gesättigt zu S_- (üblicherweise 0), wenn I(t) < S.

σ₊ und σ_- sind in der Regel gleich +1.In the process designated by decision 8, the decision is then made as to whether an alarm is triggered. This process includes an integration in which the total probability Π (t) is integrated over successive images. The integration starts at zero and counts an increment for each Π (t) > κ (κ is a threshold) and subtracts an increment for each Π (t) <κ. If I (t) denotes the value of the integral, then: $$\mathrm{I}\left(\mathrm{t}=0\right)=0$$

if Π (t) > κ

then I (t) = I (t-1) + σ ₊ (saturated to S ₊ , if I (t) > S ₊ in all other cases I (t) = I (t-1) - σ _- (saturated to S _- (usually 0) if I (t) <S.

σ ₊ and σ _- are usually equal to +1.

• Wenn I(t) > β (β ist eine Schwelle), wird Alarm ausgelöst, in allen anderen Fällen nicht.

With the help of the integral I (t) , the decision is now made as to whether an alarm is triggered:

• If I (t) > β (β is a threshold), the alarm is triggered, but not in all other cases.

The described device has the advantage that can be used in many applications on already installed video cameras and installation of special flame sensors is not required, which undoubtedly means a cost reduction. A further cost reduction results from the restriction of the evaluation to the possibly containing a flame image sections, which allows a significant reduction of computer power. It can also be assumed that the evaluation of these image sections is sufficiently robust to interference.

Claims (15)

1. Method for detecting flames in a monitored zone, referred to hereinafter as monitoring space, by analysis of at least one parameter of a radiation that occurs in the monitoring space, in which case a video image of the monitoring space is generated and zones having high light intensity and local flicker motion are sought in said video image, in which case, in a first step, said zones are localized and the relevant image excerpts are subsequently analyzed with regard to the presence of a flame, characterized in that the search for zones having high light intensity and local flicker motion is effected with the aid of an accumulation matrix [A_ij(t)], which is obtained from the difference images of successive intensity images [X_ij(t)], said difference images being weighted with a weighting factor, the weighting factor specifying the extent to which the difference images influence the accumulation matrix [A_ij(t)].
2. Method according to claim 1, characterized in that the video images are generated with a specific frequency and intensity images [X_ij(t)] are obtained therefrom.
3. Method according to claim 1, characterized in that all pixels having a brightness that lies below a predetermined threshold and thus all moving, dark objects are filtered out during the formation of the difference images and thus of the accumulation matrix [A_ij(t)].
4. Method according to claim 3, characterized in that the coordinates of the brightest pixels are sought with the aid of the accumulation matrix [A_ij(t)].
5. Method according to claim 4, characterized in that an image region of interest [ROI] which contains the brightest pixel or pixels and is reduced with respect to the original image is defined and analyzed with regard to the presence of a flame.
6. Method according to claim 5, characterized in that the size of the image region of interest (ROI) amounts at most to one fiftieth of the size of the original image.
7. Method according to claim 6, characterized in that the image information items of brightness [L(t)], chrominance [C(t)], number of active pixels [R(t)] above a specific intensity threshold, and the saturation [S(t)], are determined in the image region of interest [ROI].
8. Method according to claim 7, characterized in that said image information items are integrated over a specific time and thus over a plurality of images and their mean value is determined in that the mean value of the frequency [F] and the mean value of the amplitude [M] are determined as additional parameters during the integration, and in that the probability for the presence of a flame is calculated for each of said mean values.
9. Method according to claim 8, characterized in that the probabilities of the mean values are used to calculate an overall probability for the presence of a flame in the reduced image region [ROI], in that said overall probability is integrated over a plurality of images, and in that an alarm is triggered when a threshold is exceeded by the integrated value.
10. Device for detecting flames in a monitored zone, referred to hereinafter as monitoring space, by analysis of at least one parameter of a radiation that occurs in the monitoring space, characterized by a video camera [1] with an evaluation stage [2] for the images supplied by the camera [1], the evaluation stage [2] having a processor with an algorithm for the localization of regions having high light intensity and local flicker motion in the images of the camera [1] and the subsequent analysis of the corresponding image excerpts with regard to the presence of a flame, and the algorithm containing a process, referred to below as preprocessing [4], during which an accumulation matrix [A_ij(t)] is determined for the search for zones having high light intensity and local flicker motion, said accumulation matrix being obtained from the difference images of successive intensity images [X_ij(t)], said difference images being weighted with a weighting factor, the weighting factor specifying the extent to which the difference images influence the accumulation matrix [A_ij(t)].
11. Device according to claim 10, characterized in that the algorithm contains a process, referred to below as image obtaining [3], during which intensity images [X_ij(t)] are obtained from the video images generated with a specific frequency.
12. Device according to claim 10, characterized in that during the preprocessing [4], with the aid of the accumulation matrix [A_ij(t)], the coordinates of the brightest pixels are determined and an image region of interest (ROI) which contains the brightest pixel or pixels and is reduced with respect to the original image is defined.
13. Device according to claim 12, characterized in that, the algorithm contains a process, referred to below as analysis [5], for the analysis of the image region of interest [ROI] during which analysis the image information items of brightness [L(t)], chrominance [C(t)], number of active pixels [R(t)] above a specific intensity threshold, and the saturation [S(t)] are determined.
14. Device according to claim 13, characterized in that the algorithm contains a process, referred to below as extraction [6], during which said image information items are integrated over a specific time and thus over a plurality of images and the mean values of the image information items are determined, in that the mean value of the frequency [F] and the mean value of the amplitude [M] are determined as additional parameters during the integration, and in that the probability for the presence of a flame is calculated for each of said mean values.
15. Device according to claim 14, characterized in that the algorithm contains a process of pattern recognition [7] and a process of decision [8], during which the probabilities of the mean values are used to calculate an overall probability for the presence of a flame in the reduced image region [ROI] and said overall probability is integrated over a plurality of images and an alarm is triggered in the event of a threshold being exceeded by the integrated value.

Images