EP1776601A1

EP1776601A1 - Cfar method by statistical segmentation and normalisation

Info

Publication number: EP1776601A1
Application number: EP05763075A
Authority: EP
Inventors: Frédéric THALES Intellectual Property BARBARESCO; Jean-Pierre THALES Intellectual Property LARVOR; Bernard THALES Intellectual Property MONNIER
Original assignee: Thales SA
Current assignee: Thales SA
Priority date: 2004-07-13
Filing date: 2005-06-30
Publication date: 2007-04-25
Also published as: FR2873210B1; FR2873210A1; WO2006008227A1

Abstract

The invention relates to electromagnetic detection of targets or objects, in particular radar detection and the aim of said invention is a solution to the problem caused by the variation in the ambient signal level along the range of the radar which reduces the detection performance. In uneven regions or in the presence of large relief elements the above problem is inadequately resolved by the application to the received signal of known procedures such as maintaining a constant false alarm rate (CFAR). Said aim is achieved with a method comprising at least the following steps: a Doppler filtering (32), a determination of a map of the zones of homogeneous clutter (41), a normalisation of the received signal (42) and a detection step (43). The step of determination of a map of the zones of homogeneous clutter (41) is carried out by application of an iterative algorithm of statistical segmentation of the SEM type to the complex data from the Doppler filtering step. Said data are treated in a complex form, the algorithm being applied to the real and imaginary parts of the data. The step of determination of a map of the zones of homogeneous clutter may also comprise a complementary iterative process of fusion of the classes of environment, which permits a parametrable optimisation of the number of classes of environment used to determine the map of the zones of homogeneous clutter.

Description

TFAC process by statistical segmentation and normalization.

The invention relates to the field of surveillance and radar detection. In particular, it deals with the difficulty of detecting low surface area targets operating on rough terrain and at low altitude. It also discusses the difficulty of detecting threatening targets hidden by terrain accidents that hinder the propagation of radar waves. The disclosed invention falls within the more particular field of radar signal processing methods.

One of the main tasks of the signal processing methods is to extract from the received signal the portion of the signal corresponding to useful information. Useful information is, for example, the waves reflected by potentially dangerous or hostile targets. These reflected waves or echoes can have a variable level depending on the nature and dimensions of the target in question. As such, targets are generally classified according to their reflectivity, which is reflected in the parameter Equivalent Surface Radar or SER well known by radar equipment designers. Thus the echo reflected by a target having a small equivalent surface is a low power echo. Faced with such a target, the task assigned to radar processing consists in extracting this echo from the thermal noise or ambient clutter received by the radar receiver and accompanying this echo. By clutter is meant in an agreed manner the signal reflected by any element not strictly speaking a target in the use case considered. It can be for example the signal reflected by elements of relief, constructions, expanses of water, vegetation or atmospheric phenomena such as clouds.

To carry out this extraction, a well-known means of the prior art consists in implementing radar signal processing methods, commonly known as Constant Attack FALSE or TFAC. The concept of false alarm rate, well known to designers of radar equipment is not developed here. In general, the TFAC processes proceed for each signal element received at a time t corresponding to a given distance by relative to the radar, the estimation of the average level of signal received during a time interval Δt framing the instant considered. This estimated average level, also called "ambience", is subtracted from the received signal level. The difference is then compared to a fixed detection threshold, which threshold may be variable. Any received signal thus processed whose level exceeds the set threshold is considered as an echo representative of the presence of a target at the distance considered corresponding to the time t of reception of the signal. The conventional method of TFAC known from the prior art generally show good efficiency even in the case of low equivalent area targets for which the level of the reflected signal is low and close to the level of the atmosphere. They also make it possible to prevent nuisance tripping of alarms upon reception of a localized boiling element. On the other hand, their effectiveness is faulty in difficult operating conditions or the monitored space includes zones of different natures generating significant and abrupt variations in the level of clutter. Such an operational context is shown schematically in FIG. 1. This loss of efficiency is particularly noticeable in the estimation of the atmosphere at the places situated at the boundary between the two zones having very different levels of clutter. For these locations the application of a conventional TFAC process is reflected in a known manner, by a desensitization of the radar and an increase in false alarms. Under such circumstances the estimation of an average ambient level proves to be effective because it does not take into account sudden variations in the level of clutter occurring in the estimation time. This loss of efficiency has the serious disadvantage of rendering the radars using conventional TFAC methods incapable of detecting certain types of potentially very threatening targets that take advantage of terrain accidents to protect themselves. These threats include:

- Attack helicopters hidden behind relief elements such as hills, which take advantage of the break-up of ambient level occurring on the ridge to emerge for a brief moment from the ridge, draw a projectile in the area under surveillance and immediately recovering behind the ridge line. - Aircraft operating in mountainous terrain on low altitude terrain monitoring flights, which take advantage of a crest zone to pass from one valley to another, while remaining invisible from the radar responsible for monitoring the area. - threatening targets along a coastline in the land / sea transition zone.

In the case of a moving target, this loss of efficiency is usually attenuated by Doppler analysis. Doppler analysis of the received signal makes it possible to distinguish by its doppler frequency a threatening moving target whose signal has a spectrum centered on a Doppler frequency corresponding to its speed, the signal corresponding to the clutter whose speed is a priori low, or even zero. However, this attenuation may be insufficient, particularly in the case of a high-level clutter such as the reverberation signal of a mountainous formation, this clutter disturbing the amplitude of the spectrum of the signal received through the secondary lobes of the filter. Doppler used.

To offset this loss of efficiency in the face of threats opportunists taking advantage of sudden changes in level of clutter in the areas where they operate, the disclosed invention provides a process _^ CFAR based on a different principle. For this purpose, the invention consists of a TFAC process characterized in that it comprises at least:

a step of determining statistically homogeneous clutter areas, for which the ambient level is constant.

a step of normalizing the received signal level for each distance box,

a detection step of comparing the normalized level of received signal with a threshold. The step of determining the clutter areas calls for a statistical segmentation method known by the name of Stochastic Expectation Maximization (SEM) maximization algorithm according to the Anglo-Saxon denomination. The method effects the segmentation of the received ambient signal into ambient classes, a class being defined by its average level and the standard deviation from this average level.

In the context of the invention. The statistical segmentation method has the advantage of being applied to the complex components of the received signals and not simply to the module of the received signals.

Since this algorithm has as input parameters the number N of classes of ambiances to be determined, and being initialized by a first card arbitrarily defined, it has the advantage of being parameterizable and adaptive. Normalization is performed relative to the ambient level of the area in which the distance box is located. Each distance box is characterized by its belonging to a given area. The threshold is determined in such a way as to obtain the highest probability of detection for a chosen false alarm probability.

The method according to the invention also has the advantage of including the implementation of a complementary iterative process allowing the automatic adjustment of the number of zones by merging zones whose atmospheres are statistically close.

Other features and advantages will appear through the description made with reference to the appended figures which represent:

FIG. 1, a diagrammatic sectional illustration, along the radar pointing axis of an example of a geographical location showing abrupt changes in the nature of the clutter; FIG. 2 is a diagrammatic two-dimensional representation of the example; of Figure 1,

FIG. 3 a block diagram for implementing a conventional TFAC method,

FIG. 4 the block diagram for implementing the method according to the invention,

- Figure 5 the block diagram of the implementation of the complementary treatment of melting zones.

For the sake of clarity, the following description implicitly refers to the mode of operation of pulse radars modern and known notions of recurrence, or repetition period, burst, distance boxes and modes of burst operation. It is moreover known that the types of signal processing associated with these radars are burst processes carried out on the distance axis. It is simply recalled that the repetition period corresponds to the time interval between two times of emission of a radar pulse, during which time the radar receiver is active and that this time interval is sampled at a rate corresponding to distance division of the range of the radar in remote cells. A signal sample taken at a time t, thus corresponds to a given distance box. On the other hand, the generally employed processing methods associate by signal signals from several successive transmission pulses constituting a burst. The number of associated pulses in a burst is chosen in particular so that during the corresponding lapse of time, it is considered that the parameters associated with any detected echoes remain unchanged. The burst processing corresponds substantially to the processing of signals from a given pointing axis and is therefore a one-dimensional treatment on the distance axis.

Figure 1 illustrates in section, schematically, an example of geographical location causing difficulties in the estimation of the ambient level. This figure represents the variation of the relief as a function of the distance, along a pointing axis of the radar. This figure shows two relief elements 11 and 12 separated by a valley 13 and preventing the emission of a radar located at a point O. These two elements have faces whose surface is otherwise unequal. These faces constitute surfaces reflecting the signals emitted by the radar in the form of high-level clutter which constitutes the ambient signal in the areas of space Z ₁ and Z ₃ located above these surfaces. Conversely, between these relief elements there are lower zones Z ₅ and Z ₆ , masked by the direct emission of the radar and which reflect only thermal noise. The boundaries between zones Z ₁ and Z ₂ , Z ₂ and Z ₃ and Z ₃ and Z ₄ represented by points Pi, P ₂ and P ₃ represent limits for which the ambient signal level changes abruptly. These limits generally correspond to ridge lines, or more simply to abrupt variation of slope of the relief. For current processing methods the echoes corresponding to targets evolving in the vicinity of the ridge lines constitutes an important problem which originates from the abrupt change of the ambient signal level which has the effect of desensitizing the reception. It should be noted that the zones Z ₅ and Z ₆ which are located below the direct aim of the radar materialized by the axes 14 and 15 are areas that offers an ideal refuge to threatening targets or targets seeking to go unnoticed.

Figure 2 illustrates the same example of geographical location seen in a plan. In this figure the variations of the relief of the elements 11 and 12 are represented by the isohypse curves 21. The points Pi, P ₂ and P ₃ are located on the ridge lines 22, 23 and 24 along the axis 25 (Ox) pointed by a radar located at the point O.

FIG. 3 is a block diagram of the operating principle of a conventional TFAC method taken as an example. Conventionally, the TEFAC treatment is performed on the digitized radar signals, after characterization by doppler filtering. This is usually a processing method relating to the received signal module. In known manner, the signal samples 31 from each of the pulses composing the burst and relating to the same distance box are associated and processed by a Doppler filter bank 32. For each filter and each distance box, the distribution in this manner is thus determined. frequency of the received signal level during the duration of the burst. For example, the Doppler filter bank may be the result of an FFT operation or may result from the application of a FIR type filter or a finite impulse response filter. The levels of the spectral components obtained are then used to estimate the ambient signal level relative to each distance cell and for each frequency range corresponding to a Doppler filter. This estimation can for example be carried out in two stages as shown in box 36 of FIG. 3. The calculation of the ambient level then consists in calculating the average level of signal received over a given number of remote, eight or sixteen cells. distance boxes for example, located before and after the distance box for which it is desired to estimate the ambient level. The averages are here calculated on the amplitude of the output signals of the Doppler filters. This operation 33 for calculating the forward and backward averages is followed, as shown in box 36, by an operation of choice of the highest average. We then obtain, for each distance cell and for each Doppler filter, an average representative of the ambient environment surrounding the distance cell. Thus, in a conventional TFAC method such as that described here, the signals processed by Doppler filtering are subjected to a normalization operation which consists in calculating for each distance box the ratio between the received signal level and the level of the calculated ambient signal. which is the average before or back chosen. This operation is generally carried out, in known manner, with quantities taken in logarithmic form. The signal, thus standardized, is compared with a detection threshold, the exceeding of the threshold being the criterion for determining the presence of a target.

As previously stated, this type of TFAC process works satisfactorily for geographical areas where the ambient level remains substantially constant or varies non-abruptly. In this case the variation of the values of the front and rear averages along the range of the radar is progressive and leads for any distance box to a choice of ambience for optimal normalization and hence a satisfactory probability of detection. On the other hand, in the event of a sudden change in the ambient level, the distance boxes located in the vicinity of the transition zone are shown to be poorly normalized due to an overestimation or underestimation of the ambient level achieved by the abrupt comparison of levels of the front and rear averages. This poor signal normalization leads to a detrimental degradation of the probability of detection of a threatening target.

FIG. 4 presents in a global manner the TFAC method according to the invention. As can be seen in the figure, the method according to the invention processes the data received after doppler processing 32, as a conventional TFAC method. Likewise, it ends with the comparison 43 of the normalized signal level with a detection threshold. On the other hand, method according to the invention comprises two operations 41 and 42 which are a substitute for the calculation operations averages 33 and normalization with respect to the highest average 34. These operations are also performed not on the module but on the components in phase (I) and in quadrature (Q) of the data produced by the Doppler filtering operation 32. The function of the operation 41 is to carry out for each doppler filter a clutter zone map along the distance axis. This mapping consists of delimiting zones within which the clutter is as statistically homogeneous as possible and as different as possible from the clutter characterizing the other delimited zones. We seek here to establish areas with contrasting atmospheres. Operation 42, in turn, consists in normalizing the data with respect to the ambient level of the zones to which they belong. The detection threshold used during the operation 43 may here be an adaptive threshold determined according to the estimated parameters used to determine the zone considered.

The mapping of the space covered by the radar is carried out here by applying a statistical segmentation method implementing an iterative algorithm of Stochastic Maximization of Hope more commonly known as the SEM algorithm according to the English name (Stochastic Expectation and Maximization). This algorithm known elsewhere is not detailed here. However, it is specified that this algorithm is applied to the data from the Doppler filtering and taken in complex form. It is also specified that for reasons of convenience and robustness of calculations, the real and imaginary components of the data are considered as independent and centered Gaussian variables, with the same variance. Thus, the ambient signal processed by the radar (thermal noise and clutter) is assimilated to a complex circular Gaussian signal, ie a vector of Gaussian components centered (of zero average), of the same standard deviation and independent. Let z be the corresponding complex number, we can then write: z = x + iy = [1]

.yj

with E (z) = 0 and E (zz ^* ) = a \. E, represents here the mathematical expectation of the variable considered.

The radar ambient signal is thus modeled as a signal in which each component, in phase (x) and in quadrature (y), is a centered Gaussian variable, the two components being uncorrelated. We can write:

x and y being considered as normal distributions centered on 0 and of the respective standard deviations σ _x and σ _y .

Indeed, the radar environment can be modeled by a medium frequency voltage having for expression:

z (t) = p (t) .cos (oit + φ (t)) = x (t) .coscαt + y (t) .sinω.t [3]

with p (x) = .e and p (y) = - -. [4]

/ 2πσ ² / 2πσ "

The expressions [3] and [4] have their origin in the fact that each component is the sum of an infinity of random variables, of zero average induced by the equiprobability between the negative and positive tensions and of the same variance, the distribution voltages having no reason to be asymmetrical.

The square, w, of the ambient signal module then follows a Laplace law that can be written:

w = | z | = p ² = x ² + Ϋ [5] with: w W p (w) = 'e 2σ ² ₌ _L. _e • "* [6] 2.σ ² m _w and m _w = E (w) = 2.σ ² [7]

consequently, the modulus of the ambient signal follows a Rayleigh law having for expression:

with:

_{p (v) =} _2γ_. _e ^m wm [9]

W

As a result, the segmentation algorithm is applied to the vector consisting of the in-phase and quadrature component of the ambient signal F given by:

the use of a SEM type segmentation algorithm makes it possible to determine ambient classes, the total number N of the classes being fixed a priori.

The segmentation algorithm being an iterative process, it should be provided for initialization. For this purpose one will use a method known as a dynamic cloud method allowing a rapid determination of the different classes of environment, a class being characterized by the parameters m _× ^π , m _y ⁿ , σ _x ⁿ and σ _y ^π , with n varying from 1 to N.

The segmentation algorithm thus makes it possible to divide the geographical space into clearly defined homogeneous environment zones, each case distance constituting the range of the radar belonging exclusively to a given area.

As a result, since the different classes are defined, the normalization and detection process consists of comparing each signal sample with the parameters of the class corresponding to the zone to which it belongs. Thus, for a sample of signal belonging to a given zone belonging to the class n, the detection will consist in checking whether the following expression exceeds a fixed one or not according to the chosen probabilities of detection and false alarm.

This operation can in particular be carried out for each zone on the expression of the normalized signal as follows:

where N (0,1) represents the normal distribution of zero mean and standard deviation.

As has been said previously, an input parameter of the segmentation algorithm is constituted by the number of classes that one wishes to define. However, the choice of the number of classes necessary for a proper segmentation of any image, radar or otherwise, is a recurring problem. If the number of classes is too small, the final result does not allow to distinguish differences in ambient level between several areas, yet well marked. Conversely, in case of over-segmentation, the result is unusable because illegible. The choice depends on the number N depends in particular on the image considered, and in particular the information sought.

To help the user and thus simplify the use of the algorithm, the method according to the invention can advantageously implement a complementary iterative processing of automatic class fusion that allows to initialize the segmentation with a large number of classes and finally return, by successive iterations, to an optimum number of classes. This processing consists of starting from a relatively large number N and then estimating for each iteration of the segmentation algorithm a parameter making it possible to estimate the difference in atmosphere existing between neighboring zones belonging to different classes. It then consists of merging the areas for which the value of the estimation parameter is less than a threshold. FIG. 5 makes it possible to locate the position of this complementary treatment in the overall processing chain. The illustration of FIG. 5 presents a first step 51 which globally represents the iterative operation performed by the segmentation algorithm. In the absence of further processing, step 51 leads directly to the return of a definitive card comprising zones grouped in N classes.

Complementary processing begins with the establishment of a provisional version 53 of the map of the zones defined by step 51. This provisional map is used during a step 54 of comparison of the atmospheres associated with each zone.

The comparison parameter used to determine whether or not to merge two zones and to form only one class is a statistical distance calculation D whose definition emerges from the theory of information geometry. This statistical distance makes it possible in particular to compare independent Gaussian multivariate variables, such as those that define the prevailing atmosphere in a given area. In the problem treated by the method according to the invention, the Gaussian variable considered represents the complex data resulting from the doppler filtering step which is expressed as the sum of two independent Gaussian variables. The expression of the distance D, in the sense of the Fisher metric, between two zones marked by the indices a and b is of the following form:

with:

where mi and σi correspond respectively to m _x and σ _x , and m ₂ and 02 to m _y and σ _y .

Step 54 therefore consists in treating in pairs the different classes defined and calculating for pair the distance Dy between classes i and j. This step is followed by a step 55 which compares this distance Dy with a given threshold. In the case where the value of Dy is greater than the threshold for all the pairs of classes, the number of classes defined corresponds to the optimum number and step 55 leads to the restitution of a definitive map 52 of the homogeneous clutter zones. Otherwise, step 55 leads to a step 56 of merging the two closest classes, that is to say of the two classes for which the value of Dy is the lowest and to the decrementation of the optimal number of zones. to achieve segmentation. Step 56 leads to the establishment of an initialization card 57 which is used by the segmentation algorithm to establish a new provisional map with a number of cluttered areas 58 equal to the number of zones retained at the previous iteration minus one.

The overall operation of the method according to the invention can then be described as follows: The implementation of the method begins with an initialization iteration comprising the steps 51, 53, 54 and 55, during which the segmentation algorithm established a first set of N atmosphere class and assigns to the different areas of space one of the N given classes. For this first iteration, the number N is initialized to a given value and a first arbitrary division in N clutter zones, forming an initialization card, is provided as input to the processing algorithm. This iteration provides a first provisional map on which are calculated the interclass distances Dy which are compared to a determined threshold. If the totality of the calculated distances is greater than the set threshold, N is considered as the optimum number of classes to account for the variations of atmosphere in the considered space. It is then not necessary to proceed to a new segmentation and the provisional map of the homogeneous clutter areas becomes the definitive map 52 used for the normalization operation 42.

On the other hand, if some of the distances calculated are below the threshold, the method according to the invention continues with steps 56, 57 and 58. The classes for which the value of D is the lowest are grouped together. This regrouping or merger gives rise to the establishment of a new initialization card of the clutter zones as well as the decrementation of the number N of possible classes. The process is then extended by another iteration during which the segmentation algorithm proceeds to a new division into N-1 zones from the new initialization card. The method thus performs the number of iterations necessary to obtain a segmentation into a number N of classes for which the calculated distances Dy are all greater than the fixed threshold.

Thus, as can be seen from the description relating to FIG. 5, the automatic merge processing of the statistically close classes is integrated into the main process which implements an SEM type segmentation algorithm. This complementary treatment has the advantage of freeing the process from a precise prior determination of the optimum number of segments to restore the variations of the environment corresponding to the variations in nature of the clutter along the range of the radar. This additional processing allows starting from a large N value, to converge over time the number N to its optimum value.

In order not to hinder the efficiency of the segmentation algorithm and to limit the number of unnecessary operations, it is advantageous in practice to involve the complementary automatic merge processing only after a certain number of iterations, the first iterations being carried out without modification of the value of N nor change of the initialization card. For example, class merging can not become active only after 5 iterations as soon as the class distribution becomes sufficiently representative, the segmentation algorithm having had time to converge.

During the implementation of the method according to the invention, it is possible to consider that the number N as a parameterizable quantity by an operator for its initial value and for its minimum value. Similarly, it is possible to consider taking into account parameters relating to the minimum distance Dg between two classes, as well as the minimum number of distance cells necessary for a given area to form a particular class. This setting makes it possible both to obtain an optimal number of classes and to limit the merger. The user can, for example, access the setting of the melting threshold through a command varying between 0% and 100%, (has a value of 0% corresponding to a very low distance threshold resulting in no fusion, the value 100% corresponding to a significant distance threshold resulting in the merger of all classes into one.

Claims

1. A method of processing the radar signal implementing a Doppler filtering step (32), a step of determining a map of statistically homogeneous clutter areas (41), a step of normalizing the received signal (42) and a detection step (43), the step of determining homogeneous clutter zones implementing an iterative algorithm of the SEM type of statistical segmentation of the ambient signal received in ambient classes, a class being defined by its average level and the standard deviation from this mean level; characterized in that the statistical segmentation algorithm of the received signal is applied to the module and the phase of the data from the doppler filtering step, these data comprising a real component x and an imaginary component y.

2. Method according to claim 1, in which this iterative algorithm of statistical segmentation of the received signal is initialized by a first card (58) arbitrarily defined, this algorithm having as input parameter the number N of classes of ambiances to determine.

3. Method according to one of claims 1 or 2, wherein the step of determining a map of the homogeneous clutter areas (52) implements a complementary iterative treatment (53, 54, 55, 56) of fusion classes whose atmospheres are statistically close.

4. Method according to claim 3, in which the complementary merit repetitive treatment of the ambient classes determines the statistically close ambient classes by calculating for each pair of classes a and b the statistical distance D defined by the relation:

5. Method according to one of claims 3 or 4, wherein the complementary treatment (53, 54, 55, 56) merging classes of environments is implemented after recovery of a first exploitable map homogeneous clutter areas by convergence of the iterative segmentation algorithm.