EP0818766A1 - Method for automatic detection of fires, particularly of forest fires - Google Patents

Abstract

The fire detection system has video cameras (4) surveying a fixed scene in the visible spectrum. The video output is examined in small pixel zones to find low frequency grey level variations. Each zone is then examined by measuring the decorellation time factor of the pixels in the zone. A correct pattern triggers the fire detection alarm.

Description

The invention relates to a detection method automatic fire, including forest fire, based on the recognition of smoke from at least one a video camera, in particular of the CCD type, operating in the visible spectrum. This process makes it possible to detect at faster a fire start, even of limited extent (by example less than 4 m2), that it is directly visible or masked by the relief and this day and night.

Forest fire detection is part of integral to fire fighting. Indeed, early warning of a fire allows limit the importance of the means to be used to control these fires.

• la surveillance des zones à risque par des vigies: elle demande un grand nombre d'opérateurs qui sont d'effectifs non disponibles pour la lutte sur le front;
• le survol de ces zones par des moyens aériens : l'utilisation des moyens aériens (avions, hélicoptères) est très coûteuse, de plus, son efficacité n'est pas très intéressante. En effet, le survol consiste à parcourir un itinéraire précis dans un temps établi, l'avion ou l'hélicoptère sera dans la plupart des cas éloigné du départ de feu.

The means generally used for the detection of forest fires are:

• surveillance of risk areas by lookouts: it requires a large number of operators who are not available for fight on the front;
• overflight of these areas by air means: the use of air means (planes, helicopters) is very expensive, moreover, its efficiency is not very interesting. Indeed, overflight consists of traveling a precise route in a set time, the plane or the helicopter will in most cases be far from the start of the fire.

Automatic detection systems have already been proposed, without success.

These systems include, on the one hand, methods detection using thermal cameras (see par example documents FR-A.2575572 and FR-A-2643173). Of such systems generally work in the spectrum infrared, to detect hot spots caused by flames. This technique has the disadvantages of cost high thermal imaging cameras and its ineffectiveness for fires whose foci are masked by the relief. Else apart this detection method is disrupted by the presence of "natural" hot spots, such as vehicles and solar reflections.

We also know, from document FR-A-2696939, a automatic forest fire detection process based on the detection of the smoke created by a fire, the smoke detection itself being carried out by the recognition of movements of smoke curls. Such process also has several drawbacks:

First, smoke detection based on the recognition of scroll movement has the disadvantage of require a system with very high spatial resolution, therefore great computing power. In particular the motion detection of the curls of distant smoke covered by a few pixels, for example 2 x 2 or 3 x 3, is almost impossible to differentiate from a phenomenon fixed room such as the movement of a tree, or a group of people, or any other activity.

Second, smoke recognition by the movement requires sufficiently thick scrolls. Now a smoke departure, on the contrary, consists of the emission of thin smoke, because created by very dry vegetation, then followed later by thicker fumes

Third, this kind of process often call for the storage of a smoke-free reference image: the motion detection is obtained by subtracting from the reference image with the current image to set evidence the movements of the zones. faces multiple problems and especially those generated by continual variations in illumination caused by cloudy periods.

Fourth, this process does not allow the detection of complex smoke clouds: the effects conjugates of diffusion of matter in space, of a locally maintained hearth and finally winds, generate a complex phenomenon that makes any characterization of smoke by movement extraction. For example, a relatively calm start of fire does not produce any apparent movement. However, by strong wind, the scrolls produced often have unstable, multiple movements and very difficult identifiable.

The present invention aims to avoid all of these disadvantages, and it thus proposes a process which is remarkable in particular by the fact that it makes it possible to detect a possible hearth of smoke still thin, for a early detection, and that it is able to differentiate the smoke from any other local phenomenon.

• à rechercher dans un premier temps dans les images au moins une petite zone de pixels représentant des variations temporelles de niveaux de gris particulièrement de basse fréquence,
• puis à analyser la ou chaque zone reconnue ainsi par des moyens de mesure du taux de décorrélation temporelle des pixels appartenant à cette même zone.

To this end, the method for automatically detecting fires, in particular forest fires, which is the subject of the invention, essentially consists in processing sequences of fixed images obtained by the camera (s) in the visible spectrum, in order to detect and locate a possible source of smoke, the detection of such a source consisting of:

• to firstly search the images for at least one small area of pixels representing temporal variations in gray levels, particularly of low frequency,
• then to analyze the or each zone thus recognized by means of measuring the temporal decorrelation rate of the pixels belonging to this same zone.

Preferably, to "compress" the general image of the landscape taken by the camera we select, for the search for the or each small area mentioned above, for pixels which, with regard to variations of their gray levels, frequencies between 0.3 and 0.1 Hz.

Preferably, the method consists in searching for in the images a small area of pixels with particularly chaotic variations in gray levels, by means of measuring spatiotemporal chaoticity index different pixel areas in the image.

Thus, the method of the invention differs from methods based on the detection of any movement by the fact that it is based on the analysis of complexity spatio-temporal of almost immobile objects producing local grayscale variations. In particular, any object producing variations in gray levels can be analyzed simply as a movement of material is rejected by the process as a parasite. In the principle of the process of the invention consists in recognize smoke as an inducing phenomenon in a specific area of the landscape a noise or chaos very important in space and time, unlike all other local phenomena.

Spatio-temporal chaos results in the fact that smoke introduces into the pixel area it covers temporal variations in gray levels which are almost as diverse as those produced by a random noise, but which differ from such noise because they are characteristic of a dynamic object fractal. The fractal structure of an object is generally known to be purely spatial in nature, or purely temporal, but it can also be spatio-temporal, and its highlighting allows the differentiate from that of a noise or on the contrary from that of of an organized phenomenon.

1) une grande diversité des différences de niveaux de gris entre les pixels couverts par la fumée, qui se répartissent dans le temps comme le ferait un bruit aléatoire,

2) un degré de chaoticité élevé de la zone de pixels concernée, dont la mise en évidence permet de la dissocier d'un bruit aléatoire.

3) une décorrélation temporelle élevée des signaux temporels de la plupart des couples de pixels de la zone concernée, de préférence dans les basses fréquences.

4) une nette prépondérance de basses fréquences, de l'ordre de 0,1 Hz, sur les hautes fréquences (supérieures au Hz) généralement causées par du bruit électronique.

The detection of such an object in a landscape makes it possible to affirm that there is presence of smoke, because the puffs of which it is made up which diffuse while having an unstable displacement, create a phenomenon which uniquely accumulates l of the following properties:

1) a great diversity of differences in gray levels between the pixels covered by the smoke, which are distributed over time as a random noise would do,

2) a high degree of chaoticity of the pixel area concerned, the highlighting of which makes it possible to dissociate it from random noise.

3) a high temporal decorrelation of the temporal signals of most of the pairs of pixels of the zone concerned, preferably in the low frequencies.

4) a clear preponderance of low frequencies, of the order of 0.1 Hz, over high frequencies (higher than Hz) generally caused by electronic noise.

Spatio-temporal chaos indeed generates generally a wide range of time frequencies close to those induced by a Brownian movement, of which the time spectrum is that of white noise. Unlike purely electronic noise, this noise is characterized by the fact that it contains more bass frequencies. In the case of smoke, the low frequencies of the order of 0.1 Hz are even more marked.

• courbe A : l'évolution temporelle d'un pixel "arbre en mouvement sous l'effet du vent"
• courbe B : l'évolution temporelle d'un pixel "autoroute"
• courbe "C": l'évolution temporelle d'un pixel "fumée proche"
• courbe D : l'évolution temporelle d'un pixel "fumée éloignée"

   (pour toutes ces courbes, les abscisses correspondent au temps exprimé en secondes, et les ordonnées au niveau de gris G.This is illustrated by FIG. 1 of the appended drawing, the successive curves of which represent, from top to bottom:

• curve A: the time course of a pixel "tree in motion under the effect of the wind"
• curve B: the time course of a "highway" pixel
• "C" curve: the time course of a "near smoke" pixel
• curve D: the time course of a "distant smoke" pixel

(for all these curves, the abscissas correspond to the time expressed in seconds, and the ordinates to the gray level G.

It is therefore exploited the fact that it is possible and advantageous to characterize smoke by the fact that it produces local variations in gray levels which are uncorrelated in a broad spectrum, while the others spatio-temporal phenomena induce variations in greyscale better correlated with their neighbors, or decorrelated only in the high frequencies.

• Premièrement, il exploite la propriété selon laquelle la fumée génère un large spectre de fréquences temporelles, avec un avantage pour les basses fréquences de l'ordre de 0,1 Hz, afin de compresser l'image en retenant uniquement les pixels dont les variations présentent un excès de basses fréquences par rapport aux hautes fréquences.
• Deuxièmement, il consiste à regrouper l'ensemble des pixels retenus précédemment en objets suffisamment disjoints dans l'espace et suffisamment stables dans le temps, pour exploiter l'hypothèse selon laquelle l'objet créé par la fumée est fixe ou très faiblement mobile.
• Troisièmement, il consiste à analyser chacune des zones identifiées par l'étape précédente, pour mesurer à l'aide d'un traitement numérique décorrélation et/ou de chaoticité.

In order to characterize the complexity of the behavior of different local sets of pixels, the method consists in carrying out a processing which is preferably broken down into three stages:

• First, it exploits the property that smoke generates a wide spectrum of time frequencies, with an advantage for low frequencies of the order of 0.1 Hz, in order to compress the image by retaining only the pixels whose variations exhibit an excess of low frequencies compared to high frequencies.
• Secondly, it consists in grouping together all the pixels previously selected into objects which are sufficiently disjoint in space and sufficiently stable over time, to exploit the hypothesis that the object created by the smoke is fixed or very weakly mobile.
• Thirdly, it consists in analyzing each of the zones identified by the previous step, to measure using digital decorrelation and / or chaoticity processing.

In summary, the method consists in carrying out a processing on sequences of still images of a landscape (obtained by CCD camera in the visible spectrum) for identify and locate a possible source of smoke, characterized by recognition at first a small area of pixels with variations grayscale temporal particularly low frequency, followed by analysis of this area to verify the presence of noise, decorrelation and high chaoticity.

The primary advantage of this process is that it allows both the detection of thin smoke and thick smoke since thin smoke will be initially retained not thanks to the intensity of their gray level variations but thanks to the fact that they generate low frequencies.

Other advantages of the process are as follows:

The process allows the detection of little smoke thick since they are detected by the fact that they generate particularly decorrelation indices high variations in pixel grayscale covered by smoke.

The condition of low mobility of the object to detecting makes it possible to deal with smoke as well close as in the case of distant smoke, to the extent where this mobility is quantified in relation to size of the object and not according to the information of movement contained within the area it covers.

The method makes it possible to detect sources of tiny fumes that cover only a very limited number pixels (4 pixels for example), because it analyzes basically the time structure of the levels of gray and little purely spatial information.

The process is robust under the multiple conditions of detection, such as wind speed, orientation of the sensor, meteorology, the peculiarities of the landscape, the masking of the home by different obstacles, etc., since this process involves quantifying the complexity of the object and not to seek on the contrary a more or less defined spatio-temporal structure, such as its shape, speed, color, etc.

The method makes it possible to characterize the smoke in that that it produces an effect which is not induced exclusively by movements of matter, and that is therefore unique or rare because variations in gray levels of all the pixels in an image are in the majority of cases produced by any movements of objects in the landscape at variable distance.

In cases where the pixel areas detected in the landscape at the end of the first stage are produced by variations in brightness where reflections do not not corresponding to movements of objects, and which are noise for motion detection, the process according to the invention makes it possible to reject these zones thanks to the too low result of the calculation of their indices of decorrelation.

The process makes it possible to characterize the smoke which whatever the wind speed, because the a priori effects wind disrupters which involve generating structured scroll movements are more than compensated by the accelerated diffusion of these scrolls which attenuates thus considerably the correlations induced by these scroll movements.

Regarding the implementation of the invention, the following details may be made:

In order to transform in a first step of pretreatment of the landscape image into an image containing instead of each pixel a value providing a dynamic information on this pixel, the process uses a buffer which can permanently contain at least the Last 16 images. In order to quantify this information dynamic on a scale emphasizing the bass frequencies, the system calculates a linear combination appropriate absolute values of all differences grayscale of the same pixel taken at instants different.

Advantageously, the method uses values absolute instead of signed values in order to quantify the temporal evolution of each pixel whatever the nature of the background of the landscape on which the smoke is written, so as to attenuate the influence of the brightness of the background.

In order to save calculation time, the system calculates the new transformation, for a new step of time (corresponding to a new image acquisition), using the result of the previous calculation, and in it applying the appropriate changes.

In order to form different sets of pixels disjoint in space and recognizable in time, the system uses the assumption that the envelope smoke is very weakly mobile. This allows to eliminate possible portions of the "smoke" object which would come off the main envelope.

This technique allows you to locate more precisely the focus of smoke which is defined as the lowest point of the envelope.

The last step in the process is to perform a digital analysis of each object using all the space-time information of its pixels which on the one hand is stored in the global buffer (which contains the last 16 images at least), on the other hand in a selective buffer reserved for objects so to memorize their more distant past.

This step consists of implementing more advanced algorithms relating to said information spatiotemporal. This calculation is allowed in practice, despite the complexity of these algorithms, thanks to the enormous compression of information from previous steps.

These algorithms allow to calculate different values which quantify on the one hand the degree of disorder spatio-temporal of each object, and on the other hand the degree of chaoticity of each object.

The degree of disorder is quantified using noise calculations space-time of the object, while the degree of chaoticity is quantified using calculations which provide a measure of the sensitivity rate of the evolution of the object according to small differences in the repetitive grayscale of all of its pixels at different times.

Spatio-temporal noise measurement is broken down measuring the rate of spatial decorrelation of variations grayscale, and in analysis of the distribution of differences in gray levels of pairs of pixels, which must strive for Gaussian law according to law normal.

Digital techniques of fractal analysis and characterization of chaos, are known in the community scientific and have been the subject of publications to which reference is made.

Advantageously, the application of these techniques in the context of the present invention consists in the complementary combination of noise measurement and measure of the chaos practiced there.

The detection criterion thus exploits the measurement noise and / or chaos measurement preferentially one in relation to the other according to the different detection conditions. The combination of the two measures generates a certain degree of redundancy because a pure space-time noise is at least as complex as smoke.

The combination of these two measures is made necessary due to possible confusion between noise broad spectrum and smoke. The criterion based on the measurement of chaos is better but much more difficult to be implemented under certain conditions is why it should be relayed by the criterion based on the measurement of noise and more particularly of decorrelations between pixels.

In particular, the imprecision of the measurement of degree of chaoticity of a space-time object is more large in cases where the number of pixels is insufficient, or excessive electronic noise. This difficulty is offset by the more important to the measurement of space-time noise which is easier to calculate.

Conversely, the measurement of the degree of chaoticity, when practicable in preferential cases where the number of pixels is sufficient and the electronic noise fairly weak, allows to characterize more precisely the presence of smoke, because it is then more easily discernible from a phenomenon essentially producing noise, as can happen in the case of system placed under detection conditions particularly difficult.

We can cite the example of the approach of the night, where there is a very significant increase in noise from electronic background versus information. Likewise, certain weather conditions such as mist also tend to increase background noise.

When night has completely fallen, the detection is still possible using the same algorithms, but with an automatic adaptation of the parameters of calculation. One then advantageously exploits the illumination of smoke from flames and scattering of radiation flames from smoke. So the analysis in the case night is not questioned despite the nature different from radiation, thanks to the general principle of process which makes it possible to be insensitive to variations in vision conditions.

In any case, most objects essentially producing noise are eliminated as soon as first step of the process, thanks to quantification that it performs spectral spreading of each pixel which favors low frequencies over high frequencies. Indeed, the vast majority of noises that may constitute parasites are characterized by an excess of high frequencies compared to low frequencies.

In general, the invention makes it possible to detect very quickly, in a time notably between 10 and 40 seconds, smoke produced by a start of fire with very high reliability. She thus helps valuable and essential for operators responsible for surveillance of critical areas.

Regarding the infrastructure necessary for the implementation of the process of the invention, the latter includes at least one local detection device, located on an area to be monitored or on the outskirts thereof, and comprising at least one fixed video camera connected to an autonomous processing unit, which analyzes the image of the camera according to the methods explained above. One or more detection devices are connected to a central unit, located remotely. This last, equipped with visualization means for a human operator, serves as a checkpoint and command, but an important feature is the fact that the actual smoke detection done locally, not in the unit central. Advantageously, the control and command is provided for viewing a sequence images of the part of the landscape around a hearth detected, said sequence being displayed at a rate strongly accelerated compared to its real rate acquisition, which facilitates visual interpretation gross profit obtained by the process of detection

Figure 2 représente une implantation des détecteurs, sur une étendue à surveiller;

Figure 3 est un schéma synoptique d'un dispositif de détection;

Figure 4 est un schéma synoptique de l'ensemble du système;

Figure 5 est un ornanigramme du procédé de détection;

Figure 6 est un schéma illustrant un mode avantageux de visualisation des images, pour contrôler visuellement la détection effectuée par le procédé de l'invention.

The invention will in any case be better understood with the aid of the additional description which follows, referring to the appended schematic drawing illustrating this detection method and its means of implementation:

Figure 2 shows an installation of the detectors, over an area to be monitored;

Figure 3 is a block diagram of a detection device;

Figure 4 is a block diagram of the entire system;

Figure 5 is a flow diagram of the detection process;

Figure 6 is a diagram illustrating an advantageous mode of viewing images, to visually control the detection performed by the method of the invention.

Figure 2 symbolically indicates an extent to monitor 1, supposedly rectangular, on which and at the periphery of which are installed detection 2, suitably distributed.

Each detection device 2, shown on the Figure 3, includes one to three fixed CCD video cameras 4, installed in the same observation site and connected to an autonomous processing unit 5, powered by a energy source 6 and connected to a transmission 7, generally wireless, of low speed, in order to transmit alarm information remotely, device status information, photos, etc.

Processing unit 5, capable of taking instructs the surveillance of an area without any intervention human, receives video images from cameras 4 which each scan a very large area (from 500 meters to 10 kilometers for an angle of 60 degrees per camera). Typically the observation surface of each camera 4 is 30 km2, for a total area of 90 km2 for a device 2 equipped with three cameras 4, such as mounted in Figure 3.

Advantageously, meteorological sensors 8, can detect wind direction and speed, humidity, temperature, etc. are also connected to processing unit 5, in order to have information local weather.

The preferred mode of establishment, illustrated by Figure 2, consists of placing the detection devices 2 on the area to be monitored 1 so as to cover several times each part of this range. this is desirable to solve the problems posed by "blinded" detectors (front sun, fog, ...), faulty detectors, etc.

Complementary detectors 9, oriented differently from the previous ones, allowing to calculate with high accuracy the position of the detected lights, by a principle of triangulation.

A central unit 10, which can also be designated as a control and command post, and common to a relatively high number of detection devices 2 monitoring a range 1 possibly very large, is connected to all these detection devices 2 by a network transmission 11, in particular hertzian, like the symbolizes FIG. 4. The central unit 10 comprises a control station 12 with screen, can be connected in particular to a screen managing display 13, to a printer 14, or any other peripheral equipment.

The central unit 10 can thus receive the information and photos from all devices local detection 2, each of these devices realizing smoke detection, for the partial extent it monitors, implementing the process summarized by the flowchart of Figure 5, and using for this purpose two memories 15 and 16, designated respectively as "image memory" and "object memory".

The central unit 10 is thus informed in permanence of the status of each detection device 2. An image of the landscape seen by each detector will be available and refreshed in the shortest time possible and depending on the number of detectors sharing the same communication medium.

The central unit 10 is equipped with a display suitable for managing multiple alarms the following way:

• une cartographie de la région couverte par l'ensemble des détecteurs, légèrement surdimensionnée
• l'ensemble des points et des zones de surveillance
• l'ensemble des demi-droites joignant un point de surveillance au bord de l'écran orientées dans la direction du foyer faisant l'objet d'une alarme; Cette direction est calculée simplement à partir l'abscisse et de l'ordonnée du pixel du foyer dans l'image et de l'orientation du détecteur
• les coordonnées exactes d'un foyer si plusieurs demi-droites (de plusieurs détecteurs) se coupent
• l'ensemble des petits films correspondant aux images du passé récent autour de chaque foyer , visualisés à cadence accélérée dans le prolongement des demi-droites.

When several detection devices 2 transmit one or more alarms simultaneously or successively within a short time (i.e. when at least two alarms are being interpreted), a screen composed of the following elements is presented to the operator responsible for control:

• a map of the region covered by all of the detectors, slightly oversized
• all surveillance points and areas
• the set of half-straight lines joining a surveillance point at the edge of the screen oriented in the direction of the alarmed focus; This direction is calculated simply from the abscissa and the ordinate of the pixel of the focal point in the image and the orientation of the detector.
• the exact coordinates of a focus if several half-lines (of several detectors) intersect
• all the small films corresponding to the images of the recent past around each household, viewed at an accelerated rate in the extension of the half-lines.

This visualization method is very advantageous because it provides all the information necessary for the interpretation of multiple alarms to using the same display screen. She is original at least insofar as it uses a method of original film viewing. This method is powerful because it allows to automatically view IGN or polar coordinates and the actual location of the detected focus, by intersection of the half-lines which cut this hearth.

Note that the method of the invention can, if necessary, be implemented by inhibiting certain points predefined images, which correspond to homes permanent, for example a factory chimney.

An advantageous consequence of the process, for its practical use, is to provide an operator human, placed at the control and command post, the possibility to visually control the event which has caused detection. This visual check allows a the operator to check using a sequence of images the part of the landscape, located around the detected focus, that the phenomenon detected is smoke characteristic of a fire and estimate the severity of the event. The particularity of this sequence of images is to be viewed at a very accelerated pace like illustrated by figure 6. Accelerated viewing results from the selection, from the actual film 17 presenting a sequence of images, of certain images such as the 1st, the 11th, the 21st, the 31st, etc ..., to constitute an "accelerated" film 18.

This process exploits the property that viewing at a very accelerated rate (of the order 25 to 50 frames per second) of a sequence of images smoke, improves recognition of smoke even when the spatial resolution is very low.

This property is explained by the fact that one characteristics of a fractal object, such as the smoke, is to generate an evolution of pixels that we will find whatever the time scale of observation. The signal relating to the variation of a pixel over a few seconds is similar to the signal with the same number of time samples over a period much longer (several tens or even several hundreds of seconds).

The originality of this visualization process of images is based on the fact that this way of view a film to allow recognition of a object, which results in the general case of blur the image of the object, is on the contrary in the case precise smoke particularly effective. Accelerate the viewing distant smoke comes close this smoke.

One of the benefits of accelerated viewing is to compensate for the disadvantage of viewing rate too low, such as that of acquisition by example, which makes the observer lose the notion of the smoke dynamics.

The main advantage of visualization accelerated is based on the fractal object properties of the smoke that are likely to explain why such visualization has very positive consequences on visual interpretation by a human observer. In indeed, fractal objects are known to present the same shapes or textures that appear when you change the spatial scale of visualization.

Claims (7)

Method for automatically detecting fires, in particular forest fires, based on the recognition of smoke from at least one video camera, characterized in that it consists in processing still images obtained by the camera (s) in the visible spectrum, to detect and locate a possible source of smoke, the detection of such a source consisting of: to firstly search the images for at least one small area of pixels having temporal variations in gray levels, particularly at low frequency, then to analyze the or each zone thus recognized by measuring the temporal decorrelation rate of the pixels belonging to this same zone. Detection method according to claim 1, characterized in that one selects, for the search of the or each of the aforementioned small areas, pixels which present, with regard to temporal variations of their gray levels, frequencies between 0.3 and 0.1 Hz Detection method according to claim 1 or 2, characterized in that it includes the operation of search the images for a small area of pixels with variations in gray levels particularly chaotic, by means of measurement of spatio-temporal chaocity index of the different pixel areas in the image. Detection method according to any one of Claims 1 to 3, characterized in that it comprises another measure of space-time noise. Detection method according to any one of claims 1 to 4, characterized in that it is set works by at least one local detection device (2), located on an area to be monitored (1) or at the periphery thereof, and comprising at least one camera fixed video (4) connected to an autonomous processing unit (5) analyzing the image of the camera (4). Detection method according to claim 5, characterized in that one or more detection (2) are connected to a central unit (10) located remotely, serving as a control and command post, in particular equipped with display means (12,13) for a human operator. Detection method according to claim 6, characterized in that the control and command post (10) is provided for viewing a sequence images of the part of the landscape around a hearth detected, said sequence being displayed at a rate strongly accelerated compared to its real rate acquisition.

Images