EP1346330B1 - Video smoke detection system - Google Patents

Description

The invention is in the field of smoke detection based on a video image. In residential and industrial buildings, warehouses, museums, churches and the like, the smoke detection is carried out with mounted on the ceiling of each room smoke detectors, which are based for example on the principle of light scattering or light attenuation by smoke. In railway or road tunnels, on the other hand, virtually no smoke detectors are used because, because of the movement of air and stratification caused by the moving cars and trains, it is not guaranteed that the smoke produced in a fire would reach the ceiling mounted smoke detectors within a reasonable time. For this reason, so-called linear heat reporting systems such as the FibroLaser system from Siemens Building Technologies AG, Cerberus Division, are used today for tunnel monitoring in tunnels.

The WO 00/23959 A discloses a video smoke detection system including a video camera, video image comparison means, signal processing means and alerting means which is responsive to the output of the signal processing means. The signal processing means successively analyzes the images from the video camera and compares the intensity and / or color of the individual pixels or groups of pixels to decide whether there is a change characteristic of smoke.

Recently, efforts have been made to use the traffic monitoring in tunnels of existing smoke detection video systems. Since the video images for a viewer are often uninteresting and also caused by smoke only very small changes in the video image, monitoring by the staff at the screens is out of the question. If anything, the monitoring can only be done by an automatic evaluation of the video images. In a known method for automatically examining video images for the occurrence of smoke, the intensity values of the individual pixels of successive images are compared with one another. When measuring intensity values representative of a lighter image caused by the presence of smoke, the presence of smoke is signaled and an alarm is sounded.

One of the problems with this method is that smoke is not detected against a light background and even fire that produces little smoke is not detected. In addition, brightness changes, such as those caused by persons moving through the field of view of the camera, can trigger a false alarm. This problem has been solved by the fact that in addition to the actual surveillance area still investigates an outer area and interrupts the observation of the surveillance area for changes in this outer area. This method has the disadvantage that a fire may not be detected until after a certain delay and that sources of smoke are not detected in the outer area provided in addition to the monitoring area.

The present invention relates to a video smoke detection system comprising at least one means for capturing video images and having a signal processing stage in which a determination is made of the brightness of the individual pixels or groups of pixels of the video images.

The problem to be solved by the invention is to specify a video smoke detection system which enables rapid and reliable detection of smoke and is particularly suitable for use in road and rail tunnels. The smoke detection should take place in the earliest possible stage of the fire and false alarms should be virtually eliminated.

The video smoke detection system according to the invention is characterized in that the determination of the brightness of the pixels takes place by a process in which a value representative of the brightness is obtained, and in that an examination of the time characteristic of said value is characteristic of the occurrence of smoke Change takes place.

A first preferred embodiment of the video smoke detection system according to the invention is characterized in that the determination of the brightness of the pixels takes place by an edge extraction process in which an edge value is assigned to each pixel.

The smoke detection system according to the invention is based on the recognition that the occurrence of smoke leads to the fact that the contrast is reduced. In determining the brightness through an edge extraction process, the edges are smeared or disappear. This process has the advantage that the edge value is insensitive to global lighting changes.

A second preferred embodiment of the video smoke detection system according to the invention is characterized in that for each pixel a comparison of the edge value with an average value is made, and that a subsequently so-called counter image is obtained from this comparison, which indicates the temporal behavior of the edge value relative to the mean value.

Preferably, the counter image, which indicates how often the brightness of the relevant pixel has averaged above said mean value over a certain time, is updated with each comparison of the edge value with the mean value.

There is a comparison of the counter image with a threshold value and when this threshold is exceeded, a summation of one initialization value to a current value.

A third preferred embodiment of the video smoke detection system according to the invention is characterized in that, in addition to the edge extraction process, an examination of the video images, referred to below as motion detection, takes place on movements.

In road and rail tunnels, for the monitoring of which the system according to the invention is provided in the first place, the not caused by smoke coverage of edges will be almost exclusively by moving objects between the relevant edge and the camera. Since such objects do not materialize suddenly, but as a rule have moved to where they cover the edge, it can be assumed that with an edge cover that is not caused by smoke, immediately before a movement of the object covering the edge must have taken place. Thus, motion detection provides a reliable criterion for distinguishing smoke-covered edges from those covered by objects.

Both the determination of the edge values and the motion detection preferably take place on the basis of counter images, which are continuously updated with a hysteresis algorithm. For motion detection, an algorithm based on the normalized cross-correlation is preferably used.

The hysteresis algorithm preferably has a minimum and a maximum value and two threshold values lying between them, wherein the counter image jumps to the maximum value when counting up when the lower threshold value is exceeded and to the minimum value when counting down when the upper threshold value is undershot.

This hysteresis algorithm enables the use of noisy images for the detection algorithms. An edge caused by noise will not appear in the counter image, with appropriately parameterized hysteresis, and an edge will not disappear because of a single noisy image.

A fourth preferred embodiment of the smoke detection system according to the invention is characterized in that three data structures are used, a data field with information about the edges present in the respective image, a data field with a bit mask for the purpose of excreting image areas which are not to be considered for smoke detection, and the viewed image itself, with the edges and image preserved between successive iterations of the process and the bitmask re-initialized for each iteration.

A fifth preferred embodiment of the smoke detection system according to the invention is characterized in that the image and the edges are analyzed on a pixel-by-pixel basis and the analysis of the bit mask is carried out for groups of several pixels hereinafter referred to as blocks.

A sixth preferred embodiment of the smoke detection system according to the invention is characterized in that the processing of the data takes place on two paths, a first path for calculating the edges present in the image and for updating the already data present over edges, and a second path to create the bitmask, this second path comprising motion detection.

According to a seventh preferred embodiment of the smoke detection system according to the invention, the second path also comprises a check of the blocks for saturation of the device for recording the video images, in which blocks are marked with a certain number of saturated pixels and are not taken into account for the analysis of the counter image of the edges.

A further preferred embodiment of the smoke detection system according to the invention is characterized in that any image detail can be excluded from the analysis by means of a mask. Preferably, the bitmask created by motion detection and saturation checking is used to update the counter image for the elimination of image areas not to be considered for smoke detection.

A further preferred embodiment of the smoke detection system according to the invention is characterized in that prior to the decision on the presence of smoke, a check is made as to whether there is a sufficient number of edges for such a decision.

Fig. 1

ein Blockschema eines erfindungsgemässen Video-Rauchdetektionssystems,

Fig. 2-4

je ein Flussdiagramm zur Erläuterung der Funktion eines erste Ausführungsbeispiels eines erfindungsgemässen Video-Rauchdetektionssystems; und

Fig. 5

ein Flussdiagramm zur Erläuterung der Funktionsweise eines zweiten Ausführungsbeispiels des erfindungsgemässen Video-Rauchdetektionssystems.

In the following the invention will be explained in more detail with reference to embodiments and the drawings; it shows:

Fig. 1

a block diagram of a video smoke detection system according to the invention,

Fig. 2-4

a flowchart for explaining the function of a first exemplary embodiment of a video smoke detection system according to the invention; and

Fig. 5

a flowchart for explaining the operation of a second embodiment of the inventive video smoke detection system.

The inventive video smoke detection system is according to Fig. 1 essentially of a number of video cameras 1 and a common processor 2, in which the processing and evaluation of the signals of the video cameras 1 takes place. The video cameras 1 are mounted, for example, in a road tunnel and serve for traffic monitoring, for example, to monitor compliance with traffic regulations and for the detection of congestion, accidents and the like. The cameras are connected to a manned operations center, in which the traffic in the tunnel is monitored by monitors. The processors 2 are arranged in a decentralized manner, wherein a common number of, for example, 8 to 10 cameras is assigned to a common processor 2 in each case.

In the processor 2, the video images are decomposed into pixels, the individual pixels and / or groups thereof are assigned brightness values, and based on a comparison of the brightness values of the pixels with a reference value, the decision is made on the presence of Smoke. When assigning the brightness values to the individual pixels or pixel groups, it is essential that this assignment of global brightness changes, ie changes in the illumination of the entire image, is independent. This independence from the lighting can be achieved by assigning edge values to the pixels, which are indeed a derivative. The detection of smoke is based on the assumption that the edges are attenuated by smoke or disappear.

The signal processing and evaluation in the processor 2 can be divided into two in Fig. 1 be divided with pixel brightness 3 and smoke detection 4 designated function blocks. According to this division, the flowchart of FIG Fig. 2 obtaining the values representative of the brightness of the pixels (pixel brightness 3) and that of Fig. 3 their further investigation on the presence of smoke (smoke detection 4). Fig. 4 shows a flow chart of required for certain applications (smoke detection indoors, such as in aisles, foyers and the like) additional steps of the method according to Fig. 2 ,

The video images taken by each camera 1 are decomposed into pixels and digitized, whereby for each pixel with the coordinates i and j its intensity value I _{i, j is} determined, which may for example be between 0 and 255. From the intensity values I _{i, j} , the mean value M _{i, j} or the median is formed for a specific group of pixels of, for example, 3 times 3 or 5 times 5, or a value obtained by a low-pass filtering. The median has the advantage that its calculation can be done in 8-bit.

Parallel to the calculation of the mean or median, an edge value is obtained from the intensity I _{i, j} , which is done by a derivative or by a frequency analysis (high-pass filtering, for example wavelet transformation). The edge values K _{i, j of} the individual pixels can be determined, for example, by using a Roberts or a Sobel operator. Of course you can also use a more complicated operator for the edge calculation and apply it to larger areas such as 5x5 or 7x7 pixels.

It is then examined whether the edge value K _{i, j} lies above the mean or the median. If YES, a value Z _{i, j} is a number δ _whether counted to it and the previous value Z _{i, j} is replaced by the new one if NO is withdrawn _un from a value Z _{i, j} δ is a number, and the old Value Z _{i, j} is replaced by the new one. The value Z _{i, j} is a number which indicates how often the edge value and thus the brightness of the respective pixel has averaged above a certain threshold (mean or median M _{i, j} ) for a certain time. This number Z _{i, j} is referred to below as a counter image. The value range of Z _{i, j} is eg 0 to 255, the initial value of Z _{i, j} at the initialization of the system is 0. The numbers δ _un and δ _ob may be the same or different; for example, both can be equal to one.

The counter image Z _{i, j} has a particular advantage in terms of the effect of movements on the edge values. When an object moves through the image, it also moves at least one edge therethrough, and this results in the pixel having a higher edge value at the respective location of the edge, whereby the counter image Z _{i, j} increases by δ. As soon as the edge has left the relevant pixel, the counter image Z _{i, j} is reduced by δ _un , so that in total the passage of edges through the video image in the counter image Z _{i, j of} the individual pixels does not have any effect.

The finally obtained counter image Z _{i, j} thus preferably represents a value representative of the brightness of the relevant pixel. In examining the counter image Z _{i, j} , three time scales are used: the frequency of the recorded video images, for example 1/25 second, every 10th Seconds after 255 pictures and about every half an hour.

According to Fig. 3 the counter image Z _{i, j is} compared with a threshold S _z . If the counter image Z _{i, j is} below the threshold S _z , nothing happens if it is above the threshold S _z , then there is a summation, that is, a value Σ _x is increased by 1 and replaced by this new value. The initialization value Σ _x ⁰ is obtained in such a way that Σ = 0 is initialized and summed up, whereby after a certain stable phase of a few seconds, a stable value is set, which is then taken as the initialization value Σ _x ⁰ . Under normal conditions, Σ _x should be equal to Σ _x ⁰ .

If Σ _{x is} significantly larger than Σ _x ⁰ , then new edges have occurred, which can be caused by the presence of a stationary object in the image area of the video camera. Such an object may be in a tunnel, for example, a stationary car or a gear in this parked object; in both cases, the object covers a certain image area, which is in Fig. 3 With cover is designated. In the case of coverage, the initialization value Σ _x ^{0 is} redefined. Subsequently, the quotient Σ _x / Σ _x ^{0 is} formed and compared with a smoke threshold S _R. If the mentioned quotient is below the smoke threshold and thus edges are weakened or disappeared, an alarm is triggered.

The comparison of the quotient Σ _x / Σ _x ⁰ with the smoke threshold S _R is absolutely sufficient for accurate and false alarm smoke detection, as long as sharp edges in the foreground translational move, which is always the case in tunnels usually. For the smoke detection in road or rail tunnels you will therefore have a system with the in the Figures 2 and 3 use the functionality shown.

The situation is different when it comes to indoor smoke detection in which people are. It has been found that people who stand in one place and talk to each other, perform a kind of oscillating or locally reciprocating motion which, unlike a translational movement, no longer falls outside the counter image Z _{i, j} . Also problematic are movements of textures or patterns.

These movements cause new edges to emerge, which could compensate for the attenuation or reduction of edges by smoke, so that, under certain circumstances, smoke would no longer be reliably detected. In general, motion usually leads to new edges and possibly also covers edges, and that smoke does not lead to new edges, but weakens edges. An exception to this rule is smoke at a great distance, which may eventually lead to a new edge. Since the areas furthest away from the camera are in the uppermost part of the video image, this effect can be eliminated by hiding this uppermost part of the image, or it can be assumed that an edge formed by smoke will only move very slowly.

To prevent the disturbing influence of movements, the in Fig. 4 used subroutine, which serves for the elimination of movements and from the edges K _{i, j} ( Fig. 2 ). In principle, one could also assume the intensity I _{i, j} , but this would be associated with the disadvantage of the presence of interfering DC components. The difference ΔK _{i, j of} successive images is formed and compared with a motion threshold value S _B. When ΔK _{i, j is} below this threshold, there are no movements. If ΔK _{i, j} > S _B , the pixels which fulfill this condition are combined to subareas from which the motion is masked out. The latter takes place in that the counter image Z _{i, j is} not updated and the last counter image before the movement is used for the abovementioned subareas.

Signal noise is eliminated by a morphological filter (erosion). This means the following: The difference image that provides the number of changed pixels in the subareas is a binary image. You will go over this binary image with a pattern and give the pixels that match the pattern the value "1". The end of the movement is indicated by the fact that the sub-areas one after the other disappear from the picture and the edges decrease.

Fig. 5 shows a flowchart of a second embodiment of the inventive video smoke detection system, which is characterized in particular by a high robustness to interference and high reliability of smoke detection. The viewed image is in Fig. 5 denoted by the reference A.

The description of the flow chart will be preceded by some general explanations: Since the edges can be covered not only by smoke but also by located between the camera and the relevant edge objects, there is also an investigation of the observed image on movements. It is assumed that an object covering an edge did not suddenly arise at this point but has moved there.

Another point to be considered in smoke detection is that of the different time grids which, on the one hand, are to be considered and on the other hand to be distinguished from one another. There are very fast effects in the subsecond range, such as camera shake caused by a nearby truck, which can be eliminated by forming a moving average. There are mid-paced effects, such as those caused by smoke, which are in the 10 second range because the smoke takes about 10 seconds to reach the location where it is detected, and there are slow effects in about 10 minutes. Range or even slower. The latter are, for example, influences by the apparent movement of the sun. One way to distinguish these time slots and to identify effects in the right time frame are counter images with hysteresis.

A counter image is a series of values, usually the size of an image, which can be enlarged or reduced. These values are commonly used to count events, for example. Both edge detection and motion detection of the algorithm shown in the flowchart depend on counter images which are updated with a hysteresis algorithm. The hysteresis is characterized by four values, at the bottom, deep, high and uppermost, where at the bottom and at the top are counter limits that can not be fallen below or exceeded. The value deep lies above the value at the bottom and the value high lies between the values deep and at the top. If the count is between low and low, or between high and high, counting is normal, i. the counter reading is increased by one per detected event: however, if the counter reading reaches the value low and detects another event, it jumps to the top. Similarly, the counter reading jumps to the bottom when decreasing values from the top.

This hysteresis mechanism allows the use of noisy images for the detection algorithms. An edge caused by noise will not appear in the counter image, with appropriately parameterized hysteresis, and an edge will not disappear because of a single noisy image. In addition, the following relationships apply: The difference between the values low and low results in the number of consecutive frames over which a feature or event, such as an edge, must be present to be detected, and gives the difference between the upper and lower values the number of consecutive frames after which the event disappears as the counter value decreases. Since this number of individual images each corresponds to a specific period of time, these time periods represent a measure of the reaction time of the algorithm.

• Auf die Umgebung wird ein Sobel Kanten-Detektionsfilter angewendet, welches einen Wert q_Sobel liefert;
• für die Pixel der Umgebung wird ein Durchschnittswert q_sum berechnet.

The analysis presented in the flowchart begins with an edge detection 5 with a method based on a Sobel operator, for example. The algorithm analyzes the brightness of each pixel of each frame and tracks the history of the scene using one updated counter image 6 in the mentioned hysteresis mechanism. Two values are calculated for the environment of each pixel:

• To the environment, a Sobel edge detection filter is applied, which provides a value q _Sobel ;
• for the pixels of the environment an average value q _{sum is} calculated.

The two values are then compared using two scaling factors (DiffFac and SumFac): $${\mathit{f}}_{\mathit{DiffFac}}\mathit{\cdot }{\mathit{q}}_{\mathit{Sobel}}\mathit{<}\mathit{?}\mathit{>}{\mathit{f}}_{\mathit{SumFac}}\mathit{\cdot }{\mathit{q}}_{\mathit{Sum}}$$

If the left side of this inequality is greater than the right one, then the counter 6 is incremented, if not then it is decremented. In both cases the hysteresis mechanism is used.

Parallel to the edge detection 5, a motion detection 7 takes place, for which, for example, an algorithm based on the normalized cross-correlation is used, which roughly runs as follows:

The normalized cross-correlation is: $$\frac{\left|\overrightarrow{x}\cdot \overrightarrow{y}\right|}{\Vert \overrightarrow{x}\Vert \cdot \Vert \overrightarrow{y}\Vert }$$

Take small areas of the image, for example 4 by 4 pixels, at time t and consider these pixels as vectors x , The same area of the following image at time t + 1 becomes with the vector y designated. If the area has not changed at all, then x = y and the quotient according to formula 1 has the value 1. A change in the said range would change the quotient, so that the degree of this change can be used as a measure of the intensity of the change in the range.

In order to adapt to the processor used, numerator and denominator are multiplied by factors in formula 1, and the products formed thereby are written to analogously to the standardized cross-correlation according to formula 1. If the counter is smaller, then a movement has taken place and the corresponding area is marked. Sudden changes in the lighting or lighting conditions affect both sides of the inequality approximately equally, so that the described motion detection is immune to uniform changes in the image. This gives a map of the current image with 4 by 4 pixel blocks.

The next step is the calculation of the blocks that should not be taken into account in the analysis of the counter image 6 of the edges. It should all blocks of a certain number, for example, four times four, pixels are detected in the image in which events have taken place, which adversely affect the smoke detection algorithm. These blocks result in a bitmask 8 which is represented as a counter image of 1/16 the size of the complete image becomes. The size of the blocks is determined by the blocks considered in the motion detection, but can be changed.

• Die normierte Kreuzkorrelation funktioniert nur dann, wenn die Pixel des Bildes weder gesättigt noch vollständig schwarz sind;
• Grenzen eines vollständig gesättigten Bildausschnitts erscheinen als Kanten. Eine plötzliche Änderung der Beleuchtung würde den Eindruck erwecken, als ob die betreffende Kante verschoben und anschliessend verschwinden würde.
• In Bereichen von gesättigten Pixel gibt es keine Kanten. Ursprünglich erkannte Kanten verschwinden, wenn der betreffende Bereich in die Sättigung gelangt.

A next step is the correction of saturation of the video sensor. Such saturation can cause various problems:

• The normalized cross-correlation works only if the pixels of the image are neither saturated nor completely black;
• Limits of a fully saturated image section appear as edges. A sudden change in the lighting would give the impression that the edge in question had been displaced and then disappeared.
• There are no edges in areas of saturated pixels. Originally detected edges disappear when the area in question saturates.

For this reason, for each area in a saturation check 9, it is checked by comparison with a threshold value whether a certain number of pixels are saturated. If so, then the block in question is also marked. In order to avoid moving objects having "holes" because the motion detection detects only parts of a moving object, an expansion operator 10 is applied to the bit mask 8, filling any holes.

The bit mask 8 calculated in the stages of motion detection 7, saturation check 9 and expansion operator 10 is now used to update the counter image for the elimination process 11 (excretion of image areas not to be considered for smoke detection), again using the hysteresis mechanism already described ,

At this point in the algorithm, there are two counter images, the updated counter image 6 of all the pixels where edges have been detected, and the updated counter image 11 of all the blocks that are to be eliminated. The latter counter image is now used together with a parameter to modify the counter image 6 in such a way that a reliable edge estimation is possible, which is not influenced by any disturbing effects of the smoke detection. Each block in counter image 8 is compared to a threshold. If the value of the block is above this threshold, all pixels in the counter image 6 are set to a minimum value. From the counter image 6 of the edges, two quantities are now calculated, which represent the number of edges at different times. The first size is the number of pixels in the currently existing edges above a first threshold. The second size is the number of pixels above a second threshold, and this second threshold may be interpreted roughly as the number of pixels in edges present at a time.

definiert, welche die Anzahl der Pixel c_i,j im Bild der Zähler mit einem Wert,oberhalb eines Schwellwerts l zählt. Mit dieser Funktion können nun die beiden Grössen berechnet werden, indem die Anzahl der gegenwärtig an einer Kante liegenden Pixel dadurch abgeschätzt wird, dass man 1 sehr nahe an den Maximalwert W_m legt, den die Pixel c_ij erreichen können. Zur Berücksichtigung von Rauschen im Bild A wird für 1 in der Regel der Wert (W_m - k) gewählt, wobei k eine Anzahl von Frames bedeutet und beispielsweise bei einer üblichen fixen Kamera in einem Tunnel etwa 250 beträgt.To calculate these two quantities is a function $\underset{l}{\Sigma }\left(c\right)={\displaystyle \underset{i.j}{\Sigma }}\left(c_{i.j}>l\right)$

which counts the number of pixels c _{i, j} in the image of the counters having a value above a threshold value l . With this function, the two magnitudes can now be calculated by estimating the number of pixels currently at an edge by placing 1 very close to the maximum value W _m that the pixels c _ij can reach. To account for noise in image A, the value ( W _m -k ) is usually chosen for 1 , where k means a number of frames and, for example, is about 250 for a conventional fixed camera in a tunnel.

The subroutine for counting the pixels may be accompanied by an "image height" parameter which causes only the upper part, for example the upper half, of the image to be considered for smoke detection. This makes sense because smoke usually rises. In addition, any image section can be excluded from the analysis with a mask.

Before deciding whether smoke is present, it is now checked in a step 12 whether there are enough edges to make this decision. This check is necessary in the case where, for example, a large truck is directly in front of the camera and the image has no edges. Since it is impossible to detect a fire in this case, a fault signal should be triggered indicating that the algorithm can not operate under the current circumstances. To delay further actions and to be less sensitive to noise, an interrupt value is used, which may be either zero or greater than zero. In the latter case, it had been detected shortly before that there were not enough edges.

If there are fewer than the number of pixels corresponding to a minimum number of edges, then two actions can be taken: if the break value is already nonzero, it is reduced, and if it reaches one, a fault signal is triggered. On the other hand, if the break value is zero, it is increased to a value greater than zero. If there are enough edges, the break value is reset and processing continues.

If there is a sufficient number of edges for the reliable detection of smoke, in a step 13 the decision is made on the presence of smoke based on the average sum and difference of the edges. The difference is multiplied by a parameter and compared to the sum. If the sum is greater, there is no smoke; otherwise an alarm will be triggered. In both cases, the processing of the current image is finished and processing of the next one begins.

The alarm can be triggered, for example, by displaying a corresponding alarm in a manned alarm or monitoring center to which the relevant camera is connected, which causes the operator to analyze the image of the eye supplied by the relevant camera in more detail. The said center may be, for example, a police or fire station in an urban or regional base or the command center of a road tunnel.

Claims (19)

1. Video smoke detection system having at least one facility (1) for recording video images and having a signal processing stage (2), in which the brightness of individual pixels or groups of pixels of the video images is determined, wherein the brightness of the pixels is determined by a process in which a value representative of the brightness is obtained, wherein the temporal course of the cited value is examined and wherein the brightness of the pixel is determined by an edge extraction process (5), in which an edge value (K _i,j ) is assigned to each pixel, characterised in that a comparison of the edge value (K _i,j ) with an average value (M _i,j ) takes place for each pixel and that a subsequently so-called counter image (Z _i,j , 6) which specifies the temporal behaviour of the edge value (K _i,j ) relative to the average value (M _i,j ) is obtained from this comparison.
2. Smoke detection system according to claim 1, characterised in that the brightness of the pixels is determined by a frequency analysis, preferably a Wavelet analysis, in which an edge value (K _i,j ) determined by means of high pass filtering is assigned to each pixel.
3. Smoke detection system according to claim 1, characterised in that the counter image (Z _i,j , 6), which specifies how often the brightness of the relevant pixel has been set on average above the cited average value (M _i,j ) throughout a certain period of time, is updated with each comparison of the edge value (K _i,j ) with the average value (M _i,j ).
4. Smoke detection system according to claim 3, characterised in that the counter image (Z _i,j , 6) is compared with a threshold value (S _z ) and when this threshold value (S _z ) is exceeded, an initialisation value (∑ _x ⁰) is added up to form a current value.
5. Smoke detection system according to one of claims 1 to 4, characterised in that the recorded video images (A) are examined for the appearance of new edges, wherein the presence of new edges is concluded by means of a correlation calculation of temporally spaced counter images (Z _i,j , 6) and that with the presence of new edges, the initialisation value (∑ _x ⁰) is redefined.
6. Smoke detection system according to claim 4 or 5, characterised in that a quotient is formed from the current value (∑ _x ) and the initialisation value (∑_x ⁰) and this is compared with a smoke threshold value (S _R ) and that an alarm is triggered when the latter is exceeded.
7. Smoke detection system according to claim 1, characterised in that in addition to the edge extraction process (5), an examination of the video images (A) for movements, subsequently referred to as movement detection (7), takes place.
8. Smoke detection system according to claim 7, characterised in that both the determination of the edge values and also the movement detection (7) takes place with the aid of counter images (6, 11), which are continuously updated with a hysteresis algorithm.
9. Smoke detection system according to claim 8, characterised in that the hysteresis algorithm comprises a minimal and a maximum value and two threshold values lying therebetween, wherein the numerical image, when counting upwards, jumps to the maximum value when the lower threshold value is not reached and when counting downwards, jumps to the minimum value when the upper threshold value is not reached.
10. Smoke detection system according to one of claims 7 to 9, characterised in that an algorithm based on the standard cross correlation is used for the movement detection (7).
11. Smoke detection system according to one of claims 7 to 10, characterised in that three data structures are used, a data field with information about the edges available in the respective image, a data field with a bit mask (8) for the purpose of cutting out image regions which are not to be considered for the smoke detection and the observed image itself, wherein the edges and the image are retained between consecutive iterations of the process and the bit mask (8) is reinitialised for each iteration.
12. Smoke detection system according to claim 11, characterised in that the image and the edges are analysed pixel by pixel and the analysis of the bit mask (8) takes place for groups of several pixels subsequently referred to as blocks.
13. Smoke detection system according to claim 11 or 12, characterised in that the data is processed on two paths, a first path for calculating the edges available in the image and for updating the data already existing about paths, and a second path for creating the bit mask (8), wherein this second path includes the movement detection (7).
14. Smoke detection system according to claims 12 and 13, characterised in that the second path also includes an examination of the blocks for saturation of the facility for recording the video images, in which blocks are marked with a specific number of saturated pixels and are not take into consideration for the analysis of the counter image (6).
15. Smoke detection system according to claim 14, characterised in that any image sections can be excluded from the analysis by means of a mask.
16. Smoke detection system according to claim 15, characterised in that the bit mask (8) created with the aid of the movement detection (7) and the examination for saturation (9) is used here to update the counter image (11) in order to cut out image regions which are not to be considered for smoke detection (13).
17. Smoke detection system according to claim 16, characterised in that the counter image of the edges (6) is modified with the aid of the counter image (11) for cutting out the image regions which are not to be considered for smoke detection (13) such that compared with smoke detection (13), the edge estimation is largely immune to potentially interfering influences.
18. Smoke detection system according to claim 17, characterised in that prior to deciding on the presence of smoke, an examination (12) takes place to determine whether a number of edges which is adequate for such a decision are present.
19. Method of video smoke detection having at least one facility (1) for recording video images and having a signal processing stage (2) for determining the brightness of the individual pixels or groups of pixels of video images, wherein the brightness of the pixels is determined by a process, in which a value representative of the brightness is obtained, wherein the temporal course of the cited value is examined and wherein the brightness of the pixels is determined by an edge extraction process (5), in which an edge value (K _i,j ) is assigned to each pixel, characterised in that the edge value (K _i,j ) is compared with an average value (M _i,j ) for each pixel and that a subsequently so-called counter image (Z _i,j , 6) is obtained from this comparison, which specifies the temporal behaviour of the edge value (K _i,j ) relative to the average value (M _i,j ).

Images