FR2814895A1 - Industrial site/public place/traffic route remote surveillance system having current image reference image compared with block transformation measuring spatial activity providing digital intrusion signal/reference comparing. - Google Patents

Abstract

The video camera surveillance system subtracts pixel by pixel the current image from a reference image. A block transform is carried out producing transformed coefficients, and obtaining a spatial activity measure over the resultant image. A digital intrusion signal is found and compared with a reference intrusion detection level (12).

Description

<Desc / Clms Page number 1>

METHOD OF DETECTING INTRUSION IN A SYSTEM OF
REMOTE MONITORING
The invention relates to a method for detecting intrusions in a remote surveillance system using at least one camera.

 Video surveillance is used in various fields of activity, particularly with regard to the surveillance of property and people: surveillance of industrial sites, surveillance of public places, surveillance of road traffic.

 The invention is particularly concerned with the surveillance of places in order to detect intrusions by means of a camera, in particular a fixed camera. In the state of the art, it is known to use fixed or mobile cameras connected to a control panel console that an operator constantly looks at. In practice, operators do not have their attention fixed on screens permanently. In addition, such a system requires having as many screens as the number of cameras.

 A known solution is to display on a small number of screens, sequentially, what films each camera. This solution reduces the number of screens and makes the monitoring of screens less monotonous to increase the vigilance of the person who monitors the screens.

 An improvement consists in doubling each camera of a presence or event detector in order to automatically display the images of a camera whose detector detects a movement on the image.

The operator then displays with a control screen what is the situation. The camera can also trigger a recording of images from cameras that have detected the event.

 In order to remove the detectors associated with each camera, it is possible to use an image processing to detect the movements on the image with respect to a reference image. Classically, this is done for fixed cameras that continuously film the same place at a single angle. A reference image corresponding to the location to be monitored is scanned and each image filmed is also scanned and compared to the reference image.

 The field of digital image coding has taken on considerable importance since the 1980s.

<Desc / Clms Page number 2>

utilisées, certaines sont dédiées au codage à réduction de débit (normes H261, JPEG, MPEG1, MPEG2, JPEG2000, etc...) et se basent toutes sur le principe de codage par transformation de blocs. Grâce à cette transformation, les algorithmes peuvent calculer des paramètres de corrélation spatiale, de corrélation temporelle, de redondance subjective ou de redondance statistique afin de comprimer fortement les images. Parmi les transformées connues, la transformée en cosinus discrète (par la suite TCD) et la transformée en odelettes sont les plus utilisées afin de déterminer la corrélation spatiale d'une image.

some are dedicated to rate reduction coding (H261, JPEG, MPEG1, MPEG2, JPEG2000, etc ...) and are all based on block transformation coding. With this transformation, algorithms can compute spatial correlation, time correlation, subjective redundancy, or statistical redundancy parameters to strongly compress images. Of the known transforms, the discrete cosine transform (subsequently TCD) and the odelette transform are the most commonly used to determine the spatial correlation of an image.

 These algorithms also use prediction techniques to further reduce the size of compressed images. For these algorithms (such as the MJPEG, MPEG1, MPEG2, JPEG2000 standards), it is possible to predict for an image at a given moment what will have changed in the image at the following moment in order to perform, for example, the motion compensation. These prediction techniques are based on the extraction of the differences between two images by subtracting said images and a motion estimation. In order to maintain an acceptable image quality, the predictor is constantly updated (by compression and decompression techniques). In this case, we speak of image 1 encoded intra or instant reference (basic predictor), non-predictive B image (not involved in the prediction) and predictive image P (temporal prediction).

 The detection systems are conventionally based on taking a reference image of the filmed scene and then on comparing the filmed image with said reference image by point-by-point differentiation (that is to say pixel by pixel ) of the filmed image and the reference image to obtain a difference image and finally triggering an alarm if the image contains moving elements, that is to say differences found on several successive images, and displaying the image on a control screen so that the operator visually controls the monitored location.

 Such a system has some defects: - a first defect comes from the fact that the monitored place must be illuminated in a homogeneous way, which is very suitable for places lit by artificial light (eg a room without window ) or indirectly (for example a corridor illuminated by reflection of light on a wall). On the other hand, if the place to be monitored is directly lit by

<Desc / Clms Page number 3>

la lumière du soleil ou s'il s'agit d'un extérieur, de nombreuses fausses alarmes peuvent se déclencher, réduisant ainsi la vigilance de l'opérateur qui ne va plus pouvoir contrôler les écrans dont les alarmes se déclenchent en permanence ; - un deuxième défaut vient également des fausses alarmes qui peuvent être déclenchées par un changement de décor d'un bureau (affiche, plante...) ; - un troisième inconvénient provient des fausses alarmes dues à un seuil de référence trop bas par rapport à la qualité de l'image ; - un quatrième défaut est que ces techniques ne s'intègrent pas facilement aux procédés de codage d'image à réduction de débit.

sunlight or if it is an outside, many false alarms can be triggered, reducing the vigilance of the operator who will no longer be able to control the screens whose alarms are triggered continuously; - a second fault also comes from false alarms that can be triggered by a change of scenery of a desk (poster, plant ...); a third disadvantage comes from false alarms due to a reference threshold which is too low compared to the quality of the image; a fourth defect is that these techniques do not easily integrate with the rate reduction image coding methods.

 The invention proposes a solution that notably makes it possible to significantly reduce false alarms by extracting from the image (or part of the image) a signature allowing detection.

 This signature is a temporal representation of the image, relatively insensitive to changes in brightness or changes of scenery but highly sensitive to intrusions, especially when they are manifested by a change in content of the image, which continues to course of time.

 On the other hand, the invention proposes to develop this signature from block transformation techniques, in particular the discrete cosine transformation TCD. It is thus a question of bringing a strong added value to its techniques without increasing in a big way the volume of computation to be carried out.

 These advantageous characteristics make this system a robust, efficient and effective process.

 According to a first aspect, the invention relates to an intrusion detection method in a remote monitoring system implementing at least one video camera providing a video image which implements for a resulting image obtained by subtracting the values, pixel by pixel a current image and a reference image, a block transform, in at least one characteristic area of the resulting image comprising N blocks, for obtaining for each block Bn, m transformed coefficients Fn, m (i, j), i and j varying from 0 to M, M denoting the maximum number of rows or columns of pixels of a block Bn, m, then calculating a mean spatial activity Az of N blocks of said zone, of the resulting image,

<Desc / Clms Page number 4>

laquelle constitue un signal numérique de détection d'intrusion qui est comparé avec un seuil prédéterminé de détection d'intrusion pour détecter une intrusion.

which constitutes a digital intrusion detection signal which is compared with a predetermined threshold of intrusion detection to detect an intrusion.

The method can implement, for the calculation of the spatial activity Az: - the calculation of the spatial activity Asu of each of the N blocks Bn, m, u varying from 0 to N-1 - the calculation of the activity mean space Az with:

L'activité spatiale Asu peut être par exemple calculée à l'aide de la formule (7) donnée plus loin dans la description.

Asu spatial activity can be calculated for example using the formula (7) given later in the description.

/ < y4 /j =-y C,. avec C ==---vec/'co'A : . /' < ?. \' < ? w/7/ vcc/ ? e'y/ë'' > ; 2 ' N l= () (1 ti

dcu représentant la luminosité moyenne du bloc de rang u. Avantageusement, le procédé met en oeuvre l'application d'une transformée temporelle unidimensionnelle à N'signaux de détection Azo,... AzM-1, calculés pour des images courantes prises à des instants to,... tN'-1, pour obtenir une matrice de coefficients transformés Do, Da... D-1, et un calcul d'un signal AT représentatif d'une évolution temporelle de l'image, avec :

The method can implement a compensation of the variation of the brightness. We then have:

/ <y4 / j = -y C ,. with C == --- vec / 'co'A:. / '<?. \ '<? w / 7 / vcc /? ey / ë ''>; 2 'N l = () (1 ti

dcu representing the average brightness of the rank block u. Advantageously, the method implements the application of a one-dimensional temporal transform to Azo detection signals, ... AzM-1, calculated for current images taken at times t0, ... tN'-1, to obtain a matrix of transformed coefficients Do, Da ... D-1, and a calculation of an AT signal representative of a temporal evolution of the image, with:

NI-J A''- ; Do est proportIOnnel à --- ; l ALu "x NI Il () 1

Le procédé peut être caractérisé en ce qu'il met en oeuvre la comparaison du signal d'évolution temporelle Ar avec un seuil et

NI-J A ''-; Do is proportIOnal to ---; l ALu "x NI Il () 1

The method can be characterized in that it implements the comparison of the time evolution signal Ar with a threshold and

<Desc / Clms Page number 5>

déclenchement d'une action de détection d'intrusion si le signal AT franchit ledit seuil.

triggering an intrusion detection action if the signal AT crosses said threshold.

 The method can implement several detection signals Az corresponding to several distinct characteristic areas of the image. At least one detection signal Az can be generated from at least one characteristic area of the image.

 In its first aspect, the subject of the invention is thus a method of processing a current video image in which a first signal associated with all or part of the image (called the characteristic area of the image) and representative of fixedness and variation characters of said zones at a given moment with respect to an instantaneous reference image. For at least one zone, it is then possible to produce at least one second signal representative of the evolution of the first signal in time.

 Several successive sub-signals can be used to elaborate the temporal variability signal of the zone of the image from which the decision will be made. Thus the invention proposes to process the block image with which a first instantaneous signal is associated: the spatial activity As representative of the instantaneous situation of a block of the resulting image, and then an image processing by zones to which is associated a second instantaneous signal: the detection signal Az representative of the variation or the instantaneous fixity of an area of the image relative to a reference image. The invention then makes it possible to treat the zones in time by elaborating the final signal of temporal variability of the zones of the image.

 In its second aspect, the invention relates to a method for detecting intrusions in a remote monitoring system, is particularly suitable for detecting an intrusion manifesting itself in the form of a misalignment and / or an obstruction of a camera.

 Video surveillance can not be effective if the field of view of the camera, or any other camera sensor used, is modified by an action on the sensor so as to make invisible the events occurring in the scene. In order to detect this type of event, the method according to the second aspect of the invention is based on the modifications induced on the content of the images. The process works

<Desc / Clms Page number 6>

ainsi sur une scène quelconque, et ne nécessite pas la présence d'un objet ou d'un marqueur spécifique dans la scène.

so on any stage, and does not require the presence of a specific object or marker in the scene.

 For this purpose, the process preferably operates in two steps.

The first (optional) is a learning phase. This one aims at extracting from the scene of the characteristic elements, in order to obtain a pertinent knowledge, representative of a normal situation. In a so-called operational second phase, the method uses its knowledge of the scene to compare the current scene with the characteristics of a normal scene, and thus perform a detection.

 The method of the invention in its second aspect is usable whenever any scene is observed by a camera pointing permanently the same field. The method itself is independent of the image coding system used, but it is very easily integrated into a standard rate reduction coding system using the Discrete Cosine Transform TCD.

 In a preferred embodiment, the TCD discrete cosine transform, recommended in JPEG and MPEG standards, is exploited. The invention takes advantage of the advantageous properties of the TCD to highlight the spatial and temporal variations in the image related to its content. This original approach thus makes it possible to rely exclusively on the processing of the video signal. The invention can implement any other linear and orthogonal block transform, for example, the Discrete Fourier transform, the Oddlet transform (DWT) whose coefficients have been reorganized to obtain the transformed blocks, or the Hadamard transform. , or Slant, or again, the discrete sinus transform.

 According to its second aspect, the invention relates to a method implementing at least one video camera providing a video image, said intrusion to be detected manifesting itself in particular by a misalignment and / or an obstruction of said camera, characterized in that it sets implemented: a) a transformation, using a block transform, of the blocks of at least one set of image control blocks comprising N blocks Bn, m of dimensions M × M pixels to obtain for each block Bn, m of the transformed coefficients Fn, m (i, j), i and j varying from 0 to M-1; b) determining, for each of the N blocks Bn, m, from the transformed coefficients Fn, m (i, j), a spatial activity bsan, m,

<Desc / Clms Page number 7>

te calcul pour au moins une image courante 1 (t) et une image antérieure) (t') prises à des instants t et t'd'un facteur d'évolution de ladite activité spatiale [bsanm (t), bsan, m (t')], de chacun des N blocs Bn, m, pondéré d'un facteur représentatif de l'évolution de la luminosité d'au moins un bloc dudit ensemble de blocs, pour constituer pour chaque bloc dudit ensemble de blocs, un paramètre caractéristique d'une intrusion qui est comparé avec un seuil prédéterminé de détection d'intrusion pour détecter une intrusion.

calculating for at least one current image 1 (t) and an earlier image) (t ') taken at times t and t of an evolution factor of said spatial activity [bsanm (t), bsan, m ( t ')], of each of the N blocks Bn, m, weighted by a factor representative of the evolution of the brightness of at least one block of said set of blocks, to constitute for each block of said set of blocks, a parameter characteristic of an intrusion that is compared with a predetermined threshold of intrusion detection to detect an intrusion.

 An intrusion detection can be declared when a predetermined number N'of blocks of the set of N control blocks reaches or exceeds at least one said predetermined threshold.

 According to a first embodiment, the method is characterized in that b implements: said determination for each of the N blocks Bn, m of the current image 1 (t) and of the previous image (t ') said spatial activity, as well as a mean spatial activity [SA (t), SA (t ')], blocks of said set of blocks; calculation for the current image 1 (t) and for the previous image 1 (t ') of a factor kn, m (t) of evolution of said spatial activity bsa (t), of each of said N blocks Bn, m, - the calculation for said successive images of a factor K (t) of evolution of the mean spatial activity {(SA (t), SA (t ')} of the current image 1 (t) relative to the previous image 1 (t ') in said set of blocks, the determination for each block Bn, m of said set of blocks of a ration parameter, m (t) characteristic of an intrusion by weighting said factor of evolution kn, m (t) of the spatial activity of the blocks by said average evolution factor K (t) of the spatial activity, which is representative of the evolution of the luminosity.

In particular, the method is characterized in that knm (t) = bsan, m (t) / bsan, m t ') and K (t) = SA (t) / SA (t') ration m (t) = kn. m (t) / K (t)
Said comparison with a predetermined threshold may implement an evaluation of the parameter rationm (t) for the current image 1 (t) and of the parameter rationm (t ') for said previous image 1 (t') and a comparison

<Desc / Clms Page number 8>

de l'évolution de ce paramètre entre ces deux images avec au moins un seuil donné de détection d'intrusion constituant un dit seuil prédéterminé de détection d'intrusion.

the evolution of this parameter between these two images with at least one given threshold of intrusion detection constituting a said predetermined threshold of intrusion detection.

 According to a second embodiment, the method is characterized in that b implements: for each block Bn, m of said zone, the computation of a spatial activity parameter bsan, m; the calculation for each block Bn, m of said set of blocks of a weighted spatial activity parameter Cn, m, representative of the contrast of said block, and the calculation for each of said blocks Bn, m of the current image 1 (t ) and of the previous image) (t ') of a factor k'n, m (t) of evolution of said weighted spatial activity, to obtain said factor of evolution of the spatial activity for each of said n blocks .

avec don, m (t) = Fn. m (O. O) coefficient transformé de rang (0,0) dudit bloc Bn, m de l'image l (t), représentant la luminosité moyenne du bloc Bn, m, et p étant une constante (par exemple p entier z 2).

et en ce que :

k'n. m (t) = Cn. m (t)/Cn. m (t')

Le procédé peut mettre en oeuvre une évaluation du paramètre k'n, m (t) pour l'image courante 1 (t) et du paramètre k'nm (t') pour l'image antérieure 1 (t') et une comparaison de l'évolution de ce paramètre entre ces deux images [I (t), I (t') ] avec au moins un deuxième seuil donné qui constitue un dit seuil prédéterminé de détection d'intrusion. The method can be characterized in that:

with gift, m (t) = Fn. m (O.O) transformed coefficient of rank (0,0) of said block Bn, m of the image l (t), representing the average luminosity of the block Bn, m, and p being a constant (for example p integer z 2).

and in that :

K'N. m (t) = Cn m (t) / C n. m (t ')

The method can implement an evaluation of the parameter k'n, m (t) for the current image 1 (t) and the parameter k'nm (t ') for the prior image 1 (t') and a comparison of the evolution of this parameter between these two images [I (t), I (t ')] with at least a second given threshold which constitutes a said predetermined threshold of intrusion detection.

The method can be characterized in that:

<Desc / Clms Page number 9>

Avec a (i, j) constantes de pondération et p étant une constante (par exemple p entier > 2)
Selon un mode de réalisation préféré de la phase de détection d'intrusion, l'image antérieure 1 (t') est l'image l (t-1) qui précède immédiatement l'image courante 1 (t), cette image antérieure étant remplacée par ladite image courante 1 (t) lorsqu'une nouvelle image courante l (t+1) apparaît et q'un dit paramètre caractéristique d'intrusion n'a pas atteint ou franchi un dit seuil prédéterminé de détection d'intrusion.

With a (i, j) weighting constants and p being a constant (for example p integer> 2)
According to a preferred embodiment of the intrusion detection phase, the prior image 1 (t ') is the image l (t-1) immediately preceding the current image 1 (t), this previous image being replaced by said current image 1 (t) when a new current image l (t + 1) appears and a said intrusion characteristic parameter has not reached or passed a predetermined threshold of intrusion detection.

 This allows refresh of the previous image that serves as a basis for comparison, as long as the situation remains normal. On the other hand, as soon as a detection threshold is crossed, the previous image is frozen, which makes it possible to improve the diagnosis.

 According to a preferred implementation, the method also implements a learning phase for selecting blocks of the image constituting said set of control blocks.

 The method can then be characterized in that the learning phase implements the selection from the set of blocks Bn, m of the image of a said set or dictionary of control blocks, for which the value of the factor d the evolution of the spatial activity remains substantially constant in a given time interval during which at least one event, chosen from among a sudden change in brightness, the extinction or the lighting of at least one lighting, and / or the moving a person in the field of the camera, as long as the camera is not subject to obstruction and / or misalignment.

 The learning phase can implement a selection of the blocks Bn, m of the image having elements oriented substantially horizontally and / or vertically.

 The method can be characterized in that it implements an obstruction detection when a decrease in the spatial activity of the image SA (t) relative to the mean spatial activity averaged over a sequence of previous images, exceeds a given threshold of activity drop and when the average spatial activity SA (t) of the blocks Bn, m of the set of control blocks of the current image has a value below a threshold of given activity.

<Desc / Clms Page number 10>

Other features and advantages of the invention will become apparent on reading the following description, given by way of non-limiting example, in connection with the drawings in which: FIGS. 1 to 3 illustrate the invention under its first aspect, Figures 1 and 2 showing the calculation modules of the activity parameters Az and AT defined below, and Figure 3 giving an example of architecture of the set; - Figures 4 to 15 illustrate the invention in its second aspect, namely: - Figures 4 and 5 are algorithms; FIGS. 6 to 8 illustrate the influence of the criteria SA, DC and CG defined below; FIG. 9 is an example of a learning algorithm; FIG. 10 is an algorithm that details the selection of oriented blocks; FIG. 11 illustrates, in the form of an algorithm, the alarm triggering logic; FIG. 12 is an exemplary architecture, and FIGS. 13 to 15 respectively illustrate an example of an evolution of the number of blocks of the dictionary during a learning process, a detection of misalignment of the camera, and a detection. obstruction of the camera.

 FIGS. 1 and 2 respectively detail for the invention according to its first aspect the first development of a signal Az representative of the variation or the instantaneous fixity of an area 6 of the image, then the elaboration of the final signal AT by MOD modules 1 for the signal Az and MOD2 for the signal AT Figure 3 illustrates the decision making according to the method in its first aspect, from the final representation.

 One can distinguish in Figure 1 a current video image l (t) provided by a C-video surveillance camera and a reference image traf. These images are derived from the aforementioned digital coding techniques. The current video image 1 (t) stored in a register 1 and the reference image Iref stored in a register 2 are identical in size, for example 256 × 256 pixels (or pixels). In the preferred example, the current video images 1 (t) and Iref are divided into 1024 blocks of M × M = 8 × 8 points.

<Desc / Clms Page number 11>

Each point corresponds to a digital datum which corresponds for example to a gray level coded on an integer, for example on 8 bits. The reference image ref is subtracted point by point from the current video image 1 (t) in order to obtain a very different result image or image that is stored in a register 3.

 The resultant image Ires which therefore represents the difference between the snapshot and the reference image; then undergoes a block-by-block two-dimensional transform to obtain a transformed IT image representative of the differences between the current video image 1 (t) and the reference image Iref. The transformed image IT is represented by a matrix MAT of coefficients Cij which is stored in a register 4.

 The transformed image IT is produced block by block for two reasons, which are on the one hand the limitation of the computing power and on the other hand for block processing of the image which will be discussed later.

 However, it is possible to make a global transformation of the entire image which will be assimilated to a single block.

 The transform made in the preferred example is a discrete cosine transform (hereinafter TCD). Other linear and orthogonal transforms can be used, among others the discrete spatial Fourier transform, the Hadamard transform, the odelette transform, etc. However, the TCD has the advantage of obtaining a number of coefficients identical to the number of points and allows the direct extraction of the average brightness of the treated block.

 The same transform can be reused according to the invention on a time stack of EN values input.

Each block of the resulting image is represented by a matrix of coefficients Cij obtained by applying the following known formula (discrete cosine transform):

<Desc / Clms Page number 12>

- dans laquelle M correspond au nombre maximum de lignes et de colonnes de points d'un bloc ;

avec f (x) =-. =- x = < 9 et f ( ) c) =7' c 0 V2

- dans laquelle i et j sont les indices des coefficients du bloc et sont compris entre zéro et M-1 ; et - x (n1, n2) correspond à la valeur numérisée du point de coordonnée n1, n2 dans le bloc avant transformation.

where M is the maximum number of rows and columns of points in a block;

with f (x) = -. = - x = <9 and f () c) = 7 'c 0 V2

in which i and j are the indexes of the coefficients of the block and are between zero and M-1; and - x (n1, n2) corresponds to the digitized value of the coordinate point n1, n2 in the block before transformation.

 For each block, a signal As representative of the variation or fixity corresponding to a spatial activity As of said block is extracted. This signal is stored in a register 5.

 The spatial activity As is a combination of at least some coefficients Cij of the matrix MAT constituting the transformed image IT and it can be for example defined as the sum of the squares of all the coefficients Cij of the matrix MAT resulting from the transform , in particular the TCD, except for the coefficient Coo which is representative of the mean luminosity of the block of the corresponding resulting image. This makes it possible to overcome to a large extent variations in the average brightness of the image.

 Thus, for the preferred example, we obtain a 32 x 32-pixel image representing the Asu spatial activity of all the blocks of the image.

Since according to this example, we carry out a sum of squares, we will be careful to code As on a sufficient number of bits (12 bits for the proposed example).

For each of the N blocks of the zone 7, which is a characteristic zone or a set of blocks of the image in which the intrusion detection is carried out, the average of the spatial activity of all the blocks of this zone is calculated. 7. Each zone is represented for the current image by the value Az obtained by the following formula:

Since the spatial activity As is coded on 12 bits, it is appropriate to encode the average spatial activity of a zone of 16 bits or more. We obtain

<Desc / Clms Page number 13>

alors un signal représentatif de la variation ou de la fixité de la zone pour l'image courante obtenue par différence entre une zone de l'image courante et la même zone de l'image de référence.

then a signal representative of the variation or the fixity of the area for the current image obtained by difference between an area of the current image and the same area of the reference image.

 If this signal is zero, it is because all the N blocks of the considered zone of the current image are identical to those, for the same zone, of the reference image. If this signal is weak, it is because the differences for said zone between the reference image and the current image are minimal.

 Generally, this is the translation into the transformed signal of the noise of the image or slight changes in brightness. On the other hand, when the signal is important, it is because the differences between the reference image and the current image for said zone are themselves important, and are likely to correspond to the detection of an intrusion. To obtain an improved intrusion detection, the last N '(for example N' = 8) values of the average spatial activity of said zone 9, obtained for the last N images captured, are stacked for each zone. These values are stored in a register 9.

représentée par une matrice 10 de coefficients D obtenus en appliquant la formule suivante.

From these values, a transform is applied, in particular one-dimensional TCD. Each intrusion zone is then

represented by a matrix 10 of coefficients D obtained by applying the following formula.

avec c = -si/= 0 et c = 7 si/ 0 V2

- dans laquelle N'correspond au nombre maximum de valeurs empilées ; - dans laquelle u est l'indice des valeurs empilées (compris entre 0 et N'-1) ; - dans laquelle f (t) est égal à la valeur de l'activité spatiale moyenne Az de la zone choisie de l'image de rang t.

with c = -si / = 0 and c = 7 if / 0 V2

in which N'corresponds to the maximum number of stacked values; in which u is the index of the stacked values (between 0 and N'-1); in which f (t) is equal to the value of the average spatial activity Az of the chosen zone of the image of rank t.

<Desc / Clms Page number 14>

On extrait alors de la matrice stockée dans un registre 10 un signal AT représentatif du niveau de variation de ladite zone dans le temps, qui correspond à une activité temporelle Ar définie comme une combinaison, par exemple la somme des carrés, de tous les coefficients Du de la matrice 10 Issue de la transformée TCD unidimensionnelle, hormis le coefficient Do qui est représentatif de la valeur moyenne des N'= 8 dernières activités spatiales moyennes. On a par exemple :

An AT signal representative of the level of variation of said zone in time, which corresponds to a temporal activity Ar defined as a combination, for example the sum of the squares, of all the coefficients Du, is then extracted from the matrix stored in a register 10. of the matrix 10 Derived from the one-dimensional TCD transform, except for the coefficient Do, which is representative of the average value of the N '= 8 last average spatial activities. For example, we have:

Thus, for each zone, an AT signal is obtained which is stored in a register 11 and which represents the variations in time of the average spatial activity of the blocks of the zone. This signal AT is representative of the differences between the blocks of the last eight current images and those of the reference image for said zone.

 If this signal is weak, it is because most of the blocks in the zone are only slightly different from those of the reference image and these differences remain constant over time.

 If this signal Ar is important, it is because one or more blocks of the area have moved from a content close to that of the reference image to a content that differs a great deal from that of the reference image. .

The numerical value of this signal will be all the more important. This signal makes it possible to quantify the movements taking place in the zone in time.

 Figure 3 illustrates the implementation of the intrusion detection. The signal Ar is generated by the modules MOD1 and MOD2 from the video images provided by the camera C.

 This signal AT is thus compared by a comparator COMP 13 to a threshold Ars whose value is stored in a register 12 so as to decide in a decision circuit DET if it is a movement of sufficient size to be signaled by the AL alarm.

 The invention according to its first aspect thus makes it possible to signal any abnormal movement taking place in a predefined zone. By judiciously positioning one or more zones in the image, the invention makes it possible to monitor only the zones of passage required and thus minimizes the number of

<Desc / Clms Page number 15>

calcul à effectuer. Elle permet également d'écarter du champ de la détection des parties de l'image comportant un éclairage (lampe ou autre) dont l'extinction, ou l'allumage, peut être dû à d'autres causes qu'une intrusion.

calculation to perform. It also makes it possible to deviate from the field of detection of the parts of the image having a lighting (lamp or other) whose extinction, or ignition, may be due to causes other than an intrusion.

Note that it is also possible to overlay several areas partially or totally to maximize detection.

 The invention also makes it possible to signal an alarm when the camera is moved (the person skilled in the art speaks of misalignment of the camera) or when the camera lens is obstructed or moved (the person skilled in the art speaks of obstruction of the camera). camera field). In such a case, the content of the image varies a lot from one image to another, but this variation does not extend over time.

 The system can be entirely realized in software or entirely realized by means of dedicated electronic components.

 The system can be implanted on a hardware device. It uses several signal processors (called DSPs of the English Digital Signal Processor) to perform image subtraction functions, TCD transformations by blocks and to calculate the spatial and temporal activities.

 The comparators, averaging and decision making are performed by a module on the DSP processor. It is important to point out that the system must process the information in real time or, failing that, at the minimum rate of images acquired and processed per second and per camera. Indeed, since one seeks to detect a movement, it is necessary that the time difference between two successive images processed is not too important to retain this movement in the image at the time of acquisition. A frame rate of one frame per second and per camera is sufficient in this respect.

 The invention presents a method and a system that performs automatic presence detection using, in particular, coding methods and modules used for low bit rate digital transmission, as is found in the field of remote monitoring.

 According to its second aspect, the invention relates to a method which is more particularly adaptable to the detection of misalignment and / or obstruction of a surveillance camera.

<Desc / Clms Page number 16>

The misalignment of a camera can be defined as an action on the camera having the effect of modifying the area of the monitored scene (rotation of the camera on its support, displacement).

 The field obstruction of a camera can be defined as an action to occupy all or part of the image by an obstacle or object placed in front of a camera lens. We speak respectively of total or partial obstruction.

 The input data of the detection method according to the second aspect of the invention is a sequence of images.

 FIG. 4 gives the block diagram of the detector of misalignment and / or obstruction of camera. A misalignment alarm, and / or partial obstruction alarm or a total obstruction alarm can be triggered when the method estimates that the detection conditions are met.

 The method is based on the principle of subdivision of the image into blocks of MxM pixels, some of which are control blocks. The control blocks are defined on the one hand by a remarkable content, including certain characteristic elements, and on the other hand by the invariance of this content.

 The control blocks are advantageously selected during a learning phase, which, when it is implemented, takes place under normal conditions, that is to say in the absence of an alarm situation. The position of the control blocks is stored, and this set is called a dictionary.

 When the learning phase is over, the operation phase starts. When a misalignment or obstruction, a number of the control blocks change content. An alarm is triggered based on the number of witness blocks that have changed content.

 The algorithm relies on the principle of invariance of the control blocks in the image in the absence of an abnormal situation. When a misalignment or obstruction occurs, a number of these feature blocks lose their stability, and a decision criterion detects a change in the content of a block.

 The algorithm is therefore organized in two distinct phases (Figure 5): learning phase and exploitation phase.

<Desc / Clms Page number 17>

The following paragraphs detail respectively the criteria of detection of a change of contents of a block of the image, the phase of learning, then the phase of exploitation.

 The criterion of change of content in the image is based on a set of parameters characterizing the content thereof. The tool used to calculate these parameters is a linear and orthogonal transform such as the Discrete Cosine Transform (TCD).

s'exprime par la formulation mathématique suivante :

In the bidirectional case and for an image subdivided in blocks of M x M pixels, the TCD makes correspond to a block M x M pixels a block of M x M coefficients. The latter express the frequency representation of the luminance amplitude of the pixels. The TCD transform

is expressed by the following mathematical formulation:

c () =V2 siu=0 c (loi) =l sluwo

Pour M. v. A'. v = 0,.., M-1. ; par exemple M = 8. et (u, v) = index de la fréquence spatiale horizontale et verticale, (x, y) = index de position horizontale et verticale bloc de pixel f (x, y) = luminance du pixel au point (x, y) dans l'image.

c () = V2 siu = 0 c (law) = l sluwo

For M. v. AT'. v = 0, .., M-1. ; for example M = 8. and (u, v) = index of the horizontal and vertical spatial frequency, (x, y) = horizontal and vertical position index pixel block f (x, y) = luminance of the pixel at the point ( x, y) in the image.

F (u, v) = coefficient transformed at the point (u, v).

The characterization parameters of the patio-temporal content use the TCD coefficients.

L'image (Ir) contient (how) blocs de pixels. La taille du bloc fn. m (x, y) est M2 pixels.

Spatial information is provided from the TCD transformed domain by computing the spatial activity of the image. This parameter accurately informs the details in the indexed (It) image at time t.

The image (Ir) contains (how) blocks of pixels. The size of the block fn. m (x, y) is M2 pixels.

1 = V (.) /. w = O. M-1 / n = 0. H-1 / m = O. W-n (3) n, nI -,

<Desc / Clms Page number 18>

Dans un premier temps la transformée TCD est appliquée à l'image entière. Le premier facteur représentatif du contenu de l'image, l'activité spatiale, est obtenu à partir de l'image transformée DCT (It).

At first, the TCD transform is applied to the entire image. The first factor representative of the content of the image, the spatial activity, is obtained from the transformed image DCT (It).

et une distribution spectrale importantes.

DCT (i) = (u, v) = DCT (fn, (v))/u, v, x, y = O.. M-1/n = 0, B.. H-1/m = 0, B.. W-l

Où (H x W) est le nombre de blocs de l'image. A very active image block spatially has a consequent representation in the transformed domain. Its coefficients have an amplitude

and a significant spectral distribution.

DCT (i) = (u, v) = DCT (fn, (v)) / u, v, x, y = 0. M-1 / n = 0, B. H-1 / m = 0, B. .. Wl

Where (H x W) is the number of blocks in the image.

SA (lut) également noté SA (t) est l'expression du contenu global de l'image l (t) à l'instant t. Il est calculé par une moyenne des (bsan. m) concernés.

Où (N x M) est le nombre des pixels de l'image (en considérant qu'au début de l'apprentissage, tous les blocs sont des blocs témoins). Two parameters are introduced, (bsa) and (SA (/ f)). They respectively represent intra-block and intra-image spatial activity. The analyzed content is that of each pixel block fn, m (x, y) positioned at n, m in the image (It). The indicator (bsanm) takes into account the alternative components of the transformed coefficient block Fin, m (u, v).

SA (lut) also denoted SA (t) is the expression of the overall content of the image l (t) at time t. It is calculated by an average of (bsan.m) concerned.

Where (N x M) is the number of pixels in the image (considering that at the beginning of the learning, all the blocks are control blocks).

 The indicator SAUt) makes it possible to qualify the richness of the image.

This parameter is affected by rate reduction coding and transmission errors.

 By way of illustration, we take the capture of images taken with artificial lighting (a), just after extinction of the light (transitional phase of rehabilitation of the camera (b), a little later compared to the extinction of

<Desc / Clms Page number 19>

 the lighting (when the camera has readjusted) (c), then after lighting (d) and after re-adaptation (e) following ignition. If one calculates the spatial activity bsan, m blocks of these images, one notes that the spatial activity varies according to the luminosity.

 Figure 6 shows the temporal evolution of the mean block spatial activity (SA) on a sequence of 290 images, affected by the events a to e above, then by a misalignment / obstruction event from the image 260. The extinction (b) and ignition (d) phases of artificial lighting, and of rehabilitation (c) and (e) are clearly identified. On the other hand, one can also note a transient fall or rise in SA just after the sudden change in luminosity. This corresponds to the re-adjustment time of camera C under the new conditions.

 The spatial activity parameters are therefore sensitive to the content of the image, but also to rapid and / or significant variations in brightness.

The definition of the intra-block spatial activity bsa chosen in the preceding paragraph can be generalized to any formula using the coefficients of the TCD. This consideration is also valid for the spatial activity As or Asu defined above in the context of the invention in its first aspect.

Where a (i, j) is a constant weighting coefficient of the component used, and p a constant (for example p integer 2)
More broadly, a spatial activity can be calculated from the results of any linear and orthogonal image transformation applied block by image block (transformed into odelettes, Haar, Hadamard, ...).

 Another use of the transformed coefficients, makes it possible to define a type of parameter called contrast local block c ,,, m, sensitive to the content of the image but little to variations in brightness.

 For this purpose, the coefficient (0, 0) of the transform, for example TCD, for example of each block, is used:

<Desc / Clms Page number 20>

. =./ (8)

dcnm représente la luminosité moyenne du bloc à la position (n, m).

. =. / (8)

dcnm represents the average brightness of the block at position (n, m).

DC (It) is the expression of the overall content of the image. It is calculated by an average of the activities (bean, m) concerned.

 FIG. 7 illustrates the variation of the parameter DC for the same sequence as in FIG. 6. The shape of the curve is substantially the same as in the case of FIG. 6, that is to say that the trace is distinguished different events that occurred during the sequence.

p étant une constante (par exemple p entier ? : 2. From an activity parameter such as bsan, m, we can define a local block contrast c ,,,:

p being a constant (for example p integer?: 2.

This results in a medium block contrast CBM.

Un contraste global de l'image CG peut être défini sur le même principe, mais à partir des paramètres globaux DC(It) et SA(It).

For the It image, this CBM (It) contrast is defined as:

An overall contrast of the CG image can be defined on the same principle, but from the global parameters DC (It) and SA (It).

<Desc / Clms Page number 21>

 The evolution of the parameter CG for the same sequence as that used previously for FIGS. 6 and 7 is represented in FIG. 8. It can be seen that only the misalignment in FIG. 260 causes a significant change in the value of CG.

 The contrast parameters are therefore sensitive to the content of the image and only to it, regardless of brightness variations.

 To define a criterion for detecting a change in content, the principle is to verify that either the spatial activity or the contrast vary little from one image to another, by avoiding luminosity variations.

 In either case, this amounts to a weighting by a factor representative of the brightness.

 In the first case, the criterion for detecting a content change based on the spatial activity block bsa (n, m) corresponding to the block (n, m) must be as free as possible from the variations in brightness, because the value spatial activity depends on the brightness.

préférence l'image immédiatement précédente l (t-1). Le rapport k n, m (t) des deux valeurs est utilisé :

k,,,,, (t) = (13)

Pour définir une signature caractéristique des éléments contenus dans l'image, et en principe indépendante de la luminosité, il faut s'efforcer de compenser les effets des variations de luminosité sur l'activité spatiale bloc. To detect a change of content, we will examine the evolution of the spatial activity bean, (t), characterizing a block of a current image, at the time f, with respect to a previous image 1 (t '), who is from

preferably the immediately preceding image l (t-1). The ratio kn, m (t) of the two values is used:

k ,,,,, (t) = (13)

To define a characteristic signature of the elements contained in the image, and in principle independent of the luminosity, one must endeavor to compensate the effects of the variations of luminosity on the block spatial activity.

 The spatial activity decreases with the illumination, in spite of the device of automatic adaptation of the camera, of a factor kn, m proper to each block, However, a variation of brightness is global, that is to say that each block in the image is normally affected in the same way. We can then estimate that, in the case of a change in illumination, the spatial activity change factor kn, m of a block, is the same as the variation of

<Desc / Clms Page number 22>

facteur K (t) de changement d'activité spatiale moyenne SA entre l'image 1 (t) et l'image 1 (t') à l'instant t avec : K (t) = SA(t)/SA(t').

factor K (t) of mean spatial activity change SA between image 1 (t) and image 1 (t ') at time t with: K (t) = SA (t) / SA (t) ').

 The mean spatial activity SA is defined above.

défini de la manière suivante : On a :

/,,, (/) =/. (t)/K (t) = [,,,, (t) /,,, ()]/ [SA (t)/SA (t')] (14)

Le critère caract est défini comme la valeur maximale

caract (t),, = max [,,, (t), 1/rat-,,, (t')]-1 (15)

Il s'agit de la valeur maximale du rapport entre ration, m (t) et ration. m (t') à laquelle on retranche le chiffre 1, c'est-à-dire : si ration, m (t) > ratiom (t'), alors caractn, m (t) = (ratio (t)/ratio (t'))-1 si ration. m (t) < ration, m (t'), alors caractn, m (t) = (ratio (t')/ratio (t)) -1 si ration, (t) = ratio. m (t'), alors caractn, m (t) = 0
De la sorte, cette valeur est toujours supérieure ou égale à 0. For blocks n, m whose factor k, m is similar to K (t), we can say that their content has changed little from one image to another, despite the fact that the effects of variation in brightness are taken into account. Conversely, for blocks whose content has changed, the values k (n.m (t) and K (t) will be very different.The criterion character, m, which is used preferentially to characterize the content change independently of the brightness is

defined as follows:

/ ,,, (/) = /. (t) / K (t) = [,,,, (t) / ,,, ()] / [SA (t) / SA (t ')] (14)

The criterion character is defined as the maximum value

Character (t) ,, = max [,,, (t), 1 / rat - ,,, (t ')] - 1 (15)

This is the maximum value of the ratio between ration, m (t) and ration. m (t ') to which we subtract the number 1, that is to say: if ration, m (t)> ratiom (t'), then character, m (t) = (ratio (t) / ratio (t ')) - 1 if ration. m (t) <ration, m (t '), then character, m (t) = (ratio (t') / ratio (t)) -1 if ration, (t) = ratio. m (t '), then character, m (t) = 0
In this way, this value is always greater than or equal to 0.

 The criterion character, m decides that a block of the image changes content between two images of type 1, one of which is a current image 1 (t) and the other an earlier image 1 (t ') when the variation block activity is very different from the variation of image activity (defined as the average spatial activity of the image). The caraco m value is always positive or zero, whether the brightness increases or decreases. This property makes it possible to use only one detection threshold instead of two It is considered that a block has changed content when this characteristic has reached a certain threshold valueVariationContent:

<Desc / Clms Page number 23>

céit'tict",, > setillvariation Contenu
On peut utiliser bien entendu un autre critère tenant compte de ratio (t) et de ratio (t'). On peut par exemple utiliser ratio (t) /ratio (t') et mettre en oeuvre deux seuils, qui constituent deux seuils d'intrusion prédéterminés.

ceit'tict ",,> setillvariation Content
One can of course use another criterion taking into account ratio (t) and ratio (t '). One can for example use ratio (t) / ratio (t ') and implement two thresholds, which constitute two predetermined thresholds of intrusion.

 An illustration of the effectiveness of this criterion is given below.

 If one calculates the parameter character, (t) of the comparison of the block spatial activity with compensation of the effects of variation of brightness, between two images, one taken in artificial lighting and the other after extinction of this lighting and rehabilitation it can be seen that the use of the compensation of the effects of variation of brightness makes it possible to limit the differences not related to a change of content. Some blocks, however, are not affected by the variation in brightness, especially those around the computer monitors turned on.

simple. En effet, le contraste est indépendant de la luminosité, donc la comparaison peut être faite directement :

) =, (') 07)

On utilise préférentiellement un critère caract'", défini comme :

rc/ = max [/., (/), 7/ (/')]-1 (18)

Il s'agit de la valeur maximale du rapport entre les deux valeur kn, m (t () et le kn, m (t') à laquelle on retranche 1, de sorte que ce rapport est toujours supérieur ou égal à 0. In the second case, the use of a contrast-based parameter to define a content variation detection criterion is more

simple. Indeed, the contrast is independent of the brightness, so the comparison can be done directly:

) =, (') 07)

We preferentially use a criterion character '', defined as:

rc / = max [/., (/), 7 / (/ ')] - 1 (18)

This is the maximum value of the ratio between the two values kn, m (t () and the kn, m (t ') to which we subtract 1, so that this ratio is always greater than or equal to 0.

(' ( {ract' C, III) > seuilV ariation Contenu (18') The criterion character, m detects a variation of contents of a block if the following condition is realized:

('({ract' C, III)> thresholdV ariation Content (18 ')

<Desc / Clms Page number 24>

En général, pour calculer le paramètre caractn, m ou caract'n m, on choisit l'image courante (t) à l'instant t et l'image précédente l (t-1) à l'instant t'= t-1. On peut également choisir une autre image antérieure 1. Lorsque le seuil défini en (18') est dépassé, on cesse de mettre à jour l'image précédente qui sert de base à la comparaison.

In general, to calculate the parameter character, m or character m, we choose the current image (t) at time t and the previous image l (t-1) at time t '= t- 1. One can also choose another previous image 1. When the threshold defined in (18 ') is exceeded, it stops updating the previous image that serves as a basis for comparison.

 The flowchart of the learning phase is shown in Figure 9. It is applied for each new image presented at the input. It implements three steps: a pre-selection of the blocks of the image, the elimination of unstable blocks, and finally the updating of the dictionary D by the elimination of the unstable blocks identified.

 After the processing of each image, a test determines whether the learning is complete, and whether the resulting dictionary D can be used by the operation phase.

 During the learning phase, the control blocks are selected and their location in the image is stored. The control blocks are those which contain characteristic elements and whose contents do not change according to one or the other criterion character, m or character, m defined above.

 The set of control blocks forms a reference dictionary D.

 For this, the changes occurring between two images separated by a duration defined according to the needs (for example 1 second) are observed block by block. Any block detected once by the criterion character m or by the criterion character n, m is eliminated from the dictionary D of the control blocks.

 The control blocks can be selected as having remarkable content. This characteristic is highlighted during the step of preselecting the control blocks. This takes place once, at the beginning of the learning phase. This step consists in selecting the blocks of structured content, in particular containing horizontal or vertical elements net. These are indeed common in the images (furniture, walls, doorway, etc ...). However, depending on the scene observed, other characteristic structures can be chosen.

 For vertical and horizontal elements, the proposed solution is the following (see the flowchart in Figure 10).

<Desc / Clms Page number 25>

The mask Sobel filters [1-1] (horizontal) and [1-1] '(vertical) are separately applied to the pixel image f. Then, the activity of the resulting pixel image whose contour was thus isolated by the Sobel filters is calculated for each block Bn, m according to formula (5) or (7). We thus obtain two images of the spatial activity oriented block, respectively horizontal asboHm) and vertical asbov m, and mean activity oriented ASoH and ASOv calculated according to (6).

 To select blocks of sufficient oriented activity, the largest oriented activity (among the horizontal and vertical oriented activity) is selected, and the latter is required to be greater than a threshold (FS) representing a ratio between the oriented activity of the considered block and the average oriented spatial activity of the image.

verticalement, et FSv pour l'activité orientée horizontalement :

Il (cl., ; bo Il (fi, fil) > asho, (fi, ni et si alors on sélectionne le bloc s mon si (asho, (n. m) > F, alors sélectionner le bloc (19)

Les blocs sélectionnés à l'issue de ce traitement forment un premier dictionnaire qui va être affiné sur une longue période et de nombreuses images par les étapes suivantes. FSH is the threshold chosen for the oriented activity

vertically, and FSv for horizontally oriented activity:

He (cl,; il (fi, wire)> asho, (fi, ni and if then we select the block s my si (asho, (n.m)> F, then select the block (19)

The blocks selected at the end of this treatment form a first dictionary which will be refined over a long period and many images by the following steps.

 The control blocks are also characterized by an invariant content between two images. A criterion for detecting a content change of a block, based on the spatial activity of the block, has been defined previously. This criterion is applied here: any location of the image where a change of content is detected in a block is eliminated from all the control blocks.

 This step is iterated over many images and for a sufficient time to place the system in a range of wide operating conditions, representative of the conditions that will normally be encountered during an operating phase. For example, the learning phase may include brightness variations (such as those of the day / night cycle), or other disturbances such as moving objects in

<Desc / Clms Page number 26>

une partie du champ de la caméra, ou une intrusion dans une zone de passage, l'allumage ou l'extinction d'une lumière ou d'un écran d'ordinateur etc...

a part of the camera's field, or an intrusion in a passage zone, the lighting or extinction of a light or a computer screen etc ...

 Another important task is carried out at this stage: it is also a question of updating the image which will be compared to the next image that will be presented. For this, one simply memorizes the current image, because no alarm is supposed to occur during the learning phase.

 The learning phase ends when the expected duration has expired. The duration of the learning phase should cover a period long enough to include the maximum of disturbing phenomena.

However, in order to shorten this delay, it is also possible to stop the learning phase IF the dictionary obtained has not varied for a long time (for example several hours).

The condition stopping the learning is therefore: Duration of learning durceMaxiPhaseLearning or D! // realDiconvanant durceMaxiSansChange (20)
Finally, it remains to verify that the learning phase has succeeded, that is to say that the dictionary is valid. For that, it suffices that it reaches a sufficient size, or in other words that there is in the image a significant number of control blocks. This condition makes it possible in particular to avoid random triggering of alarms during the operating phase: a typical example is that of a scene in which the passage zone of the objects concerns almost the entire image. This situation results in a very small number of stable blocks. During the operation or detection phase, changing the content of some blocks will be enough to trigger a false alarm.

7 (//// < ? (D/ < ;'//oM/ ? n ?) > proportionBlocsDansDico x (H x W), ( où (H x W) est le nombre de blocs de l'image. The condition validating the dictionary obtained is

7 (//// <? (D / <;'// oM /? N?)> ProportionBlocsInDico x (H x W), (where (H x W) is the number of blocks in the image.

 A module can detect the particular situations of total obstruction. It operates completely independently of

<Desc / Clms Page number 27>

 detachment detector / partial obstruction. In order to differentiate the case of a total obstruction from that of an entirely black or uniform scene, it is considered that: - a total obstruction of the field of a camera is an almost instantaneous phenomenon (the image passes from an image having content to an instant black image); - a black image from an unlit scene is obtained gradually.

 It is on this postulate that the detection of a total obstruction is based. If we abruptly pass from an image containing information to a black image, we consider that there is total obstruction. On the other hand, if the change is progressive, we do not consider the event as a total obstruction.

First, we summarize the past evolutions by calculating the average SAmoy (t) of the spatial image activity SA (t), on the MEMJMGAS images previously arrived at the current image, and on which no total obstruction has been detected.

moyenne passée sAmoy avec par exemple : :

(/) x2 < / et SA (t) < seu1ActiBlocMinimale (23) et ,,, , (/) > . SeMc/c'cM//e A decrease in brightness of the image is interpreted as a total obstruction if there is both a sharp drop in the SA (t) space activity compared to the average SAmoy spatial activity, a low value of the spatial activity. SA (t), and a significant value of the spatial activity

Average passed sAmoy with for example:

(/) x2 </ and SA (t) <only1ActiBlocMinimal (23) and ,,,, (/)>. SEMC / c'cM // e

<Desc / Clms Page number 28>

Un fonctionnement du détecteur de dépointage/obstruction peut être obtenu dans des conditions limites, de la manière suivante :
Un module a pour but d'éliminer les situations dans lesquelles l'image analysée est trop peu riche , et en particulier trop sombre, pour prendre toute décision fiable de déclenchement d'alarme. Pour cela, on utilise une condition de luminosité d'image, qui se traduit par une activité spatiale de l'image suffisante. Lorsque la luminosité baisse en dessous du seuil acceptable, par exemple à la tombée de la nuit, le détecteur suspend toute surveillance de la variation du contenu des blocs témoin, et se met en veille. Lorsque les conditions sont redevenues favorables (éclairage artificiel ou lever du jour), le détecteur se réactive automatiquement et compare le contenu de la nouvelle image à celui de la dernière image exploitable avant la mise en veille. Ainsi, tout dépointage survenu durant la période d'obscurité sera détecté.

An operation of the misalignment / obstruction detector can be obtained under boundary conditions, as follows:
A module aims to eliminate situations in which the analyzed image is too rundown, and in particular too dark, to make any reliable decision of alarm triggering. For this, we use a condition of image brightness, which results in sufficient spatial activity of the image. When the brightness drops below the acceptable threshold, for example at dusk, the detector suspends any monitoring of the variation of the contents of the control blocks, and goes to sleep. When the conditions are favorable again (artificial lighting or daybreak), the detector reactivates automatically and compares the contents of the new image to that of the last exploitable image before the standby. Thus any mishandling that occurred during the dark period will be detected.

SA> seullActlBlocM! nimal (24)
When the condition (24) is not satisfied, the misalignment / obstruction alarm is eventually reduced to FALSE, and an operation flag impossible set to TRUE. Typically, this case occurs when the observed scene is not lit at night. The detector waits for the next image to eventually reactivate the detector.

 The exploitation phase consists of monitoring the control blocks designated by the reference dictionary resulting from the learning. For that, one applies the same criterion of detection of change of contents between two successive images 1 that during the learning.

 The method counts the number of witness blocks of the dictionary, having changed content. If this number is significant, it detects a misalignment / obstruction of the camera. However, instability of reference dictionary blocks may occur when a sudden change in brightness has not allowed time for the camera to restore a nominal image. Therefore, in order to limit false alarms, a misalignment / obstruction event will only actually be reported if it is detected on at least two consecutive (or more) occasions.

 The flowchart of the operating phase is shown in Figure 11. As for the learning phase, it is applied for each image presented at the input.

<Desc / Clms Page number 29>

On retrouve également les modules identiques d'activation ou désactivation du détecteur, et de détection de l'obstruction totale.

There are also the identical modules of activation or deactivation of the detector, and detection of the total obstruction.

 The other two stages have many points in common with the learning phase.

 The first of these two steps is the identification of the blocks of the dictionary D resulting from the learning that are unstable in the observed scene. The process of triggering the misalignment / obstruction alarm is the second.

 The detection of variable content blocks is performed by calculating the characteristic parameter of the content change character, m or character m. The process is identical to that of the learning phase.

 During the learning phase, the algorithm focused on preserving only blocks Bn, m that did not change content from one image to another, according to the criterion character or character. m. These control blocks are characteristic of the scene vis-à-vis attempts of misalignment or obstruction.

During the exploitation phase, the algorithm counts the number of control blocks having changed content between two images which are either the current image and a previous image, notably the immediately preceding image, namely the current image and an image. reference that is optionally refreshed periodically. If this number is important, we detect a misalignment / obstruction of the camera: #BlocksDicoDetects> proportionBlocsAlarm x Size (Dictionary),
If this condition is fulfilled, the alarm is triggered.

 This module also updates the image that will be compared to the next image that will appear. For this purpose and according to a preferred mode, the current image 1 (t) is simply memorized, but only on the condition that a misalignment / obstruction event has not been detected, and this current image will serve as a basis of comparison. to the next current image that will be captured.

 An example of implementation of the method is given in FIG. 12. The calculation processor of the Content Change Detection Criterion (PCCC) provides the parameter character, m (t) or character m (t). This parameter is calculated as described above, based exclusively on the spatial bean activity. m (t) and on blocks of size equal to the size commonly used for coding images, ie M = 8.

<Desc / Clms Page number 30>

During the learning phase, the Partial Docking / Obstruction Detection Processor (PDDO) gradually builds the dictionary D by eliminating the blocks having changed content and therefore can not be part of the control blocks. During the operation phase the PDDO counts the blocks of this dictionary having changed content, possibly to trigger an alarm.

 The PDIU and PDOT processors operate in parallel with the PCCC and PDDO processors. The PDIU processor indicates to the PDDO processor whether the image is unusable, and in this case, the latter disables the misalignment indicator and / or partial obstruction.

Example
For an indoor scene, which can be artificially lit, and in which people can enter the field of the camera, 256 x 256 video frames were processed in blocks of 8 x 8 sizes, making a total of 1024 blocks.

 After a phase of detection of oriented blocks, the dictionary has 613 blocks (see Figure 13).

 During the learning phase during which the non-stable blocks are eliminated, the dictionary D stabilizes around 400 blocks (exactly 411 blocks).

 The vast majority of blocks in a transit zone have been eliminated.

 Outside the passage zone, other blocks have been eliminated: they are the ones that became unstable during a change of lighting during the learning process.

 The selected blocks, kept in the dictionary D, lose their stability in case of misalignment and / or obstruction, which allows a specific detection.

 Figure 14 illustrates the detector behavior.

When no event disturbs the image, the number of blocks for which a change of activity is noted is low or zero (between 0 and 2).

 In the image No. 14, a first peak (27 blocks) in 1 corresponding to the passage of a person is detected. This detection is not taken into account to trigger the alarm.

<Desc / Clms Page number 31>

 In image No. 30, an increase Il, then a stabilization 111 of the number of active blocks detected appears. It is due to a start of misalignment of the camera.

 In image No. 38, a large peak IV of the number of active blocks (100 blocks) corresponds to a frank misalignment and triggering of the misalignment alarm.

 It is therefore expected that, during a misalignment, the number of active blocks (for which the activity is detected by the evolution of the criterion character or character) becomes very important, which allows a specific detection of a misalignment, whatever the lighting variations.

 Figure 15 illustrates the case of a total obstruction of the camera. The image changes from an image with a content to an entirely black image (image # 23). A sharp drop V of the average activity AS of the image is detected as soon as the objective is obstructed (image nO 2).

 In case of progressive decrease of the luminosity of a scene filmed (by progressive passage of the light of day to the night), the fall of the average activity of the image is not brutal, but on the contrary very progressive. In such a case, there is no triggering of an alarm.

Claims (7)

1. Intrusion detection method in a remote monitoring system implementing at least one video camera providing a video image, characterized in that it implements for a resultant image obtained by subtraction between the pixel values per pixel of a current image and a reference image, a block transform, in at least one characteristic area of the image comprising N blocks, for obtaining for each block Bn, m transformed coefficients Fn, m (i, j) i and j varying from 0 to M, where M denotes the maximum number of rows or columns of pixels of a block Bnm. then calculating, from said obtained coefficients Fn, m (i, j), an average spatial activity Az of the N blocks of said zone of the resulting image, which constitutes a digital intrusion detection signal which is compared with a predetermined threshold of intrusion detection to detect an intrusion.  2. Method according to claim 1, characterized in that it implements, for the calculation of the mean spatial activity Az: the calculation of the spatial activity Asu of each of the N blocks Bn, m, with u variant from 0 to N-1.

 - The calculation of the mean spatial activity Az, with: 3. Method according to claim 1, characterized in that it implements for calculating the mean spatial activity Az: the calculation of the spatial activity As, of each of the N blocks Bn, m, u varying from 0 to N-1, - the calculation of the mean spatial activity Az, with:

<Desc / Clms Page number 33>  with p constant (for example pu 2), dcu representing the average brightness of the rank block u.  4. Method according to one of claims 1 to 3, characterized in that it implements the application of a one-dimensional temporal transform to the intrusion detection signal (Azo ... AzN'-i) catcuiés for current images taken at times t = 0 ... t = N'-1, to obtain a transformed matrix of coefficients Do, Dol ... Di and a calculation of an AT signal representative of a temporal evolution of the image with:

5. Method according to claim 4, characterized in that it implements the comparison of the time evolution signal AT with a threshold and triggering an intrusion detection action if said sign AT crosses said threshold.  6. Method according to one of the preceding claims, characterized in that it implements several detection signals (Az) corresponding to distinct characteristic areas of the image.  7. Method according to one of the preceding claims, characterized in that it implements at least one detection signal (Az) produced from at least two characteristic areas of the image.