Echo filtering method, echo filtering device and associated computer program

The method and device use a neural network to analyze radar echoes in a spatio-temporal context, improving target classification and reducing false alarms by leveraging AI for enhanced pattern recognition.

FR3135535B1Active Publication Date: 2024-11-29THALES SA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
FR2022004498
Authority / Receiving Office
FR · FR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-05-12
Publication Date
2024-11-29
Estimated Expiration
2042-05-12

AI Technical Summary

Technical Problem

Existing radar systems face challenges in distinguishing between target objects and unwanted echoes due to limited instantaneous decision-making and reliance on kinematic models, leading to high false alarm rates and resource inefficiencies.

Method used

A method and device utilizing a space-time window to construct a set of images from echo attributes, fed into a neural network for contextual pattern recognition, enabling improved target classification by placing echoes in their spatio-temporal context.

Benefits of technology

Enhances the accuracy of echo filtering by leveraging artificial intelligence for pattern recognition, reducing false alarms and optimizing computational resources by distinguishing between different types of targets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 00000017_0000
    Figure 00000017_0000
  • Figure 00000017_0001
    Figure 00000017_0001
  • Figure 00000018_0000
    Figure 00000018_0000
Patent Text Reader

Abstract

The invention relates to a method for filtering echoes associated with attribute values, comprising: determining the echoes located within a spatiotemporal window including a given echo; constructing a set of image(s), each position in an image corresponding to a respective pair of values ​​of said attributes, and inserting a specific pattern into the image at a specific position, if and only if the attribute values ​​corresponding to said position are the values ​​of said attributes for one of the determined echoes within the spatiotemporal window; estimating, based on said constructed set of image(s) (EnsIMn) for the given echo by a neural network (11), the correspondence of said given echo to at least one type of target object from a predefined set of target object types. Figure for the abstract: Fig. 1
Need to check novelty before this filing date? Find Prior Art

Description

Title of the invention: Echo filtering method, echo filtering device and associated computer program Technical field

[0001] The invention lies in the field of processing echoes received by an echo receiving system, for example of the radar type. Prior art

[0002] At various points in the received echo processing chain, for example in a radar receiver system, filters are applied at different stages of processing (beamforming filter, Doppler filter, pulse compression filter, extractor classification and filtering, Kalman filter, tracking classification and filtering, etc.), in order to promote the detection and preservation of signals reflected by targets of interest and to eliminate unwanted echoes: for example noise, interference, jamming, intentional and unintentional clutter.

[0003] This filtering is essential at each of these stages to avoid saturation (calculation and algorithmic) leading to malfunction of the different processing blocks and to ultimately result in a false alarm rate lower than the prescribed maximum rate.

[0004] Each time, a decision as to the usefulness of the presence of the echo (signal vs. noise or clutter, target of interest or not, etc.) for the continuation of the processing is made, at best based on instantaneous information (e.g., the measurements of attributes of the current plot), the instantaneous spectrum, the level of the signal relative to a local or time-smoothed environment, the position of a plot relative to a prediction of a Kalman filter of a track, etc.).

[0005] This narrow vision makes the decision very complicated. Deciding instantly whether or not to keep a radar plot in clear mode (i.e. without Doppler calculation) with an estimation of the position and the SER is not possible, to distinguish a helicopter from a wind turbine for example or a drone from a bird.

[0006] Several strategies are used in radars, mostly based on the classification of single detections and tracking filtering. A target that appears to have a temporal displacement consistent with a human object is a target of interest, so its detections are of interest.

[0007] Filtering of unit detections is based on the classification of the attributes of the unit detections. The challenge is generally to find the distribution of each attribute for the phenomenon to be filtered compared to the detections of interest. This discrimination can be done manually (by thresholding for example) or by machine learning (neural networks). A strong limitation of this approach is when the attributes detections do not allow to discriminate the targets of interest of a phenomenon to be filtered.

[0008] Tracking filtering uses kinematic models (which therefore impose their limitations) to estimate whether a detection is part of a moving target. If the motion characteristics can correspond to a target of human origin, the detection is considered of interest. This type of filtering shows limited performance against highly maneuvering targets or when it becomes necessary to define a large number of kinematic models. For each existing track, each new detection whose position corresponds to a kinematic model is evaluated. This creates a large number of hypotheses to be maintained in memory for each track. The computer responsible for tracking must therefore have a lot of dedicated memory space.

[0009] There is therefore a need to have a filtering solution which makes it possible to improve the relevance of the filtering. Summary of the invention

[0010] To this end, according to a first aspect, the present invention describes a method for filtering echoes received by an echo receiving device, each echo being associated with attribute values ​​of said echo, said method being characterized in that it comprises the following steps implemented by an electronic filtering device, for a considered echo sn: i / application of a space-time window of predetermined size comprising said echo considered and determination of the echoes located in said space-time window including said echo considered; ii / construction of a set of hnage(s) comprising at least one image, each image of the set of hnage(s) being associated with a respective predefined pair of attributes, each position in said image corresponding to a respective pair of values ​​of said attributes, distinct positions corresponding to pairs of distinct values ​​of said attributes, according to which, during the construction of said image, a determined pattern is inserted into the image at a determined position, if and only if the values ​​of the attributes corresponding to said position are the values ​​of said attributes for one of the echoes determined in the spatio-temporal window; iii / providing said set of image(s) as input to a neural network resulting from learning and adapted to estimate, as a function of said set of image(s) constructed for the echo considered, the correspondence of said echo considered to at least one type of target object from a predefined set of types of target objects; and iv / estimation by said neural network of the correspondence of said echo sn considered to at least one type of target objects among a predefined set of types of target objects.

[0011] The invention makes it possible, before each decision on the presence of an echo, to replace this presence in its spatio-temporal historical context, which greatly enriches the fields of information available for the decision.

[0012] The helicopter plot could thus appear in a continuity of plot trajectory which would immediately allow it to be identified above a wind farm.

[0013] Furthermore, artificial intelligence algorithms are very well suited to pattern recognition: the invention makes it possible to reduce the problem of filtering to a problem of image processing; it makes it possible to take advantage of the broadening of the field of information and to benefit from the contextual notion of the decision.

[0014] In embodiments, such a method will further comprise at least one of the following features: - the attributes include or are a function of one or more of the following: echo detection time, distance from the object to the origin of the echo, azimuth and / or elevation of the received echo, speed of the object, ratio between returns captured according to orthogonal polarization planes or their correlations, echo signal-to-noise ratio, Radar Equivalent Surface of the object spatial extension of the object echo, spectral spreading or presence of blade flash; - the set of hnage(s) constructed for each echo sn comprises at least two images, the pair of attributes associated with one of said images being distinct from the pair of attributes associated with another image of the set of hnage(s); - the set of hnage(s) constructed for the echo sn comprises one or more images from among an image associated with the azimuth-distance attribute pair, an image associated with the time-distance attribute pair and an image associated with the time-SER attribute pair; - the neural network is derived from a learning phase during which a learning base comprising echoes associated with their attributes and each previously associated with a type of the predefined set of target object types is used, a set of hnage(s) being constructed for each echo of the learning base in accordance with steps i and ii and according to which the set of images constructed for each echo is provided in input of the network being trained and the weights of the neurons in the network are modified according to the error between the type previously associated with the echo and the estimate provided by the neural network, in its current training state, based on said set of image(s).

[0015] According to another aspect, the invention describes a computer program intended to be stored in the memory of an electronic echo filtering device and further comprising a computer, said computer program comprising instructions which, when executed on the microcomputer, implement the steps of a method according to the invention.

[0016] According to another aspect, the invention describes a device for filtering echoes received by an echo receiving device, each echo being associated with attribute values ​​of said echo, said device being characterized in that it is adapted, for a considered echo sn, to apply a space-time window of predetermined size comprising said echo snconsidered and to determine echoes located in said space-time window including said echo snconsidered;

[0017] said device being adapted to then construct a set of hnage(s) comprising at least one image, each image of the set of hnage(s) being associated with a respective predefined pair of attributes, each position in said image corresponding to a respective pair of values ​​of said attributes, distinct positions corresponding to pairs of distinct values ​​of said attributes, so that, during the construction of said image, a determined pattern is inserted into the image at a determined position, if and only if the values ​​of the attributes corresponding to said position are the values ​​of said attributes for one of the determined echoes in said spatio-temporal window; and

[0018] said device is adapted to then provide said set of hnage(s) as input to a neural network resulting from learning and adapted to estimate, as a function of said set of hnage(s) constructed for the echo sn considered, the correspondence of said echo considered sn to at least one type of target objects from a predefined set of types of target objects.

[0019] In embodiments, such a device will further comprise at least one of the following features: - the attributes include or are a function of one or more of the following: echo detection time, distance from object to echo origin, azimuth and / or elevation of received echo, speed of object, ratio between the returns captured according to orthogonal polarization planes or their correlations, signal-to-noise ratio of the echo, Surface Equivalent Radar of the object, spatial extension of the echo of the object, spectral spreading or presence of blade flash; the set of hnage(s) constructed for each echo sn comprises at least two images, the pair of attributes associated with one of said images being distinct from the pair of attributes associated with another image of the set of hnage(s); the set of hnage(s) constructed for the echo sn comprises one or more images from among an image associated with the azimuth-distance attribute pair, an image associated with the time-distance attribute pair and an image associated with the time-SER attribute pair. Brief description of the drawings

[0020] The invention will be better understood and other characteristics, details and advantages will appear more clearly on reading the following description, given without limitation, and thanks to the appended figures, given by way of example.

[0021] [Fig.l]] [Fig.l] is a representation of a filtering device in one embodiment of the invention;

[0022] [Fig.2] [Fig.2] is a representation of a filtering method in a mode of realization of the invention;

[0023] [Fig.3] [Fig.3] represents two sets of images obtained in a mode of realization lization of the invention;

[0024] [Fig.4] [Fig.4] illustrates the application of a time window around an echo to treat in one embodiment of the invention;

[0025] [Fig.5] [Fig.5] is an example of the architecture of a neural network of a filtering device in one embodiment of the invention.

[0026] Identical references may be used in different figures when they designate identical or comparable elements. Description of the embodiments

[0027] The invention is described below, in one embodiment, in a radar echo processing system.

[0028] As is known, radar (acronym from the English radio detection and ranging) is a system that uses electromagnetic waves to detect the presence or position, as well as optionally the speed, of target objects such as, depending on the applications, flying objects (airplanes, helicopters, drones, etc.) for air traffic control, such as boats for maritime traffic monitoring, or vehicles for road traffic monitoring or rain, clouds, storms and others. weather events for weather monitoring. Waves sent by a radar transmitter are reflected by objects encountered by the incident wave, and the return signals (called radar echo or radar echo) are picked up and analyzed by the radar receiver, located for example at the same location as the transmitter.

[0029] A set of values ​​{attl_n, att2_n, ... attp_n] taken by a number p of attributes, attl, att2, ... attp, p > 2 is associated with each echo sn and defines it. The value of each attribute attl_n, att2_n, ... attp_n is determined by the receiver as a function of at least the echo sn to which the attribute relates and which characterizes the object at the origin of the echo. For example, the attributes are among the following elements, or are a function of at least one of the following elements or a combination of them: - time (timestamp) of echo detection, - distance (in particular the distance box) at which the object is located, obtained using the round trip time of the signal, - azimuth and / or elevation of the received echo, indicating the direction of the object, determined using the angular position of the antenna that picked up the return signal, - speed of the object, function of the frequency shift of the return signal generated according to the Doppler effect, - ratio between returns captured according to orthogonal polarization planes or their correlations, - echo signal-to-noise ratio, SER (Radar Equivalent Surface), o, of the object, which is expressed as follows, with P; the power density per unit area, at the level of the target illuminated by the radar and Ps the power diffused by the object per unit solid angle: spatial extension of the object echo, spectral spreading or presence of blade flash, etc.

[0030] As seen previously, it is important, in the radar processing chain, to be able to discriminate the echoes returned from target objects, i.e. which are of interest in the context of the application considered, from those returned from other objects, this on the one hand in order not to process useless echoes and to allow the necessary computational resources to be reduced and on the other hand to filter only the echoes corresponding to the targets sought and to delete the other echoes, to deduce accurate information therefrom.

[0031] To do this, in one embodiment of the invention, a filtering device radar echoes 1 is arranged in the processing chain of echoes received by the radar receiver.

[0032] According to the embodiments, the filtering device 1 is arranged at one or more of the echo filtering steps which occur along the echo processing chain, among the following: beamforming filtering, Doppler filtering, pulse compression filtering, classification and extractor filtering, Kalman filtering, tracking classification and filtering, etc.

[0033] An echo sn as considered in the remainder of the description can thus alternatively be a hit (in the example considered, approximately 10,000 hits / s), a plot, which is an aggregate of hits (in the example considered, approximately 1000 plots / second) or a track, which is a sequence of plots (in the example considered, approximately 100 tracks / s) etc.

[0034] The echoes considered come for example from reflections having taken place following an antenna rotation implemented by the radar transmitter (time necessary for an antenna rotation, for example in the range [1-10 seconds], comprising bursts of pulses, each burst having a duration Traf in the range [1 - 10 ms].

[0035] In the case considered, the set of target types sought, Enstyp_cibieS, comprises N distinct types of targets among which the following different types: helicopters, hydrometeors, biological targets, rocket, artillery, mortar, missile, aircraft, drone, ..., wind turbines, road traffic, etc. with N > 2. One of the target types of Enstyp_cibieS is for example "others" and groups together everything that is not part of the other target types of the Enstyp_cibies set.

[0036] The filtering device 1 receives the successive echoes sn (n = 1, 2, ...) detected over time and their respective set of attributes. The filtering device 1 is adapted to estimate the correspondence of the echo sn with at least one target of the set of targets Enstyp_cibieS and to deliver, for each detected echo sn received at input, an output information Outn determined as a function of this estimation.

[0037] The filtering device 1 comprises an electronic image construction block 10 and a neural network 11. In one embodiment, the image construction block 10 comprises a processor and a memory (not shown).

[0038] The image building block 10 is adapted to construct, from an echo sn and its attributes, a set of image(s), Ens1Mn, which is then provided as input to the neural network 11. The construction of the set of hnage(s) is described later, with reference to the method of [Fig.2].

[0039] The neural network 11 is a neural network resulting from prior learning.

[0040] The neural network 11 is adapted to receive, in the inference phase, the set of images Ens1Mn constructed for an echo sn and to estimate, as a function of this set of images, whether (or, in one embodiment, to what extent) the echo sn corresponds to an echo returned by a target object type that is part of the Enstyp_cibies set.

[0041] In embodiments, the neural network 11 is for example adapted to deliver, for the echo sn, the percentage of correspondence between the echo and each of the types of the set Enstyp_ciMes.

[0042] In another embodiment, the neural network 11 is adapted to compare with each other percentages obtained for the echo sn for each type of the set Ens tyP_cibies and only delivers as output, for the echo sn, the type of target having obtained the maximum percentage.

[0043] A method 100 for filtering echoes in one embodiment of the invention is shown in [Fig.2].

[0044] This method comprises a preliminary phase comprising step 101 of training the neural network 11 and step 102, which will be described later.

[0045] In the operational phase of echo processing, in a step 201, considering an echo sn defined by its attributes which are provided as input to the image construction block 10, the image construction block 10 constructs, from this input data, a set of hnage(s), Ens^. The memory of the image construction block 10 comprises, for example, software instructions, which when implemented on the processor of the construction block 10, implement step 201, detailed below.

[0046] The set of image(s) comprises K images, with K fixed > 1. For example K is greater than or equal to 2. The value of K depends on the application, the number of attributes available, the desired precision, the number of target types.

[0047] Each image of the image set Ens1Mn is associated with a predefined pair of attributes among the p attributes attl, att2, ... attp. For example, one image is associated with the pair (attl = azimuth, att2 = distance), a second image is associated with the pair (att3 = time, att2 = distance) and a third image is associated with the pair (att3 = time, att4 = SER). The attributes of a pair are distinct from each other. The pair of attributes associated with an image of Ens1Mn is distinct from the pair of attributes associated with any other image of Ens1Mn.

[0048] Each position in the image associated with (atti, attj) is thus defined by a pair of values ​​of the attribute pair (atti, attj) (in predefined ranges) and vice versa. Two distinct positions are defined by two distinct pairs of values ​​of (atti, attj).

[0049] In the case considered, an orthonormal reference frame is associated with each image with the abscissa axis representing values ​​of the first attribute of the associated pair and the ordinate axis representing values ​​of the second attribute of the associated pair. In one embodiment the attributes are reported after centering and normalization for ho mogenize the images including during the learning phase (in particular the time scale is normalized between the date of the detection to be classified (the most recent) and this same date subtracted from the time selection window).

[0050] Receiving the echo sn and its attribute values ​​{attl_n, att2_n, ... attp_n], the image building block 10 selects, from among the set of echoes received by the device 1 (and previously stored in a memory), the echoes detected in a predetermined spatio-temporal window comprising the detection time of sn and the spatial position of sn

[0051] Then the image building block 10 constructs the image associated with the pair (attl, att2), by placing a given pattern (for example a point, a cross, a pixel of a predetermined color, etc.) at the position in the image defined by the pair of attribute values ​​(attl, att2) associated with each of said echoes selected in the spatio-temporal window.

[0052] According to the embodiments, the spatio-temporal window of temporal depth L is set to the detection time of the echo sn and makes it possible to go back in time with the temporal depth L from this time.

[0053] In one embodiment, the time depth L of the space-time window is in the range of several seconds to several minutes.

[0054] In one embodiment, the spatio-temporal window surrounds (for example is centered on) the spatial position of the echo sn (i.e. the spatial position of the object at the origin of this echo).

[0055] For example, the spatial depth S of the space-time window is located in the range of 0 to k kilometers in distance (with k less than 10 km, or even with less than 5 km) and of a few 03dB in azimuth (for example a number of 0 îdH less than 10, or even less than 5), 03dB being the opening at -3dB of the transmit-receive antenna pattern of the radar.

[0056] The processing is iterated for each of the K images to be constructed. The set of images Ens1Mn determined by the construction block 10 for the echo sn is then delivered in the form of an input image vector by the block 10 to the neural network 11.

[0057] The neural network 11 receiving the set of images Ens1Mn performs, in a step 202, the convolution and classification processing as a function of the set of hnage(s) and the weights determined during the learning. It estimates whether (or, in one embodiment, to what extent) the echo sn corresponds to an echo returned by a target object forming part of the set Enstyp_cibieS. For example, the neural network 11 thus determines, for the echo sn, the percentage of correspondence between the echo and each of the types of the set Enstyp_cibieSen as a function of the set of images Ens1Mn: xi% helicopters, x2% hydrometeors, x3% biological targets, x4% rocket, x5% artillery, x6% mortar, x7% missile, x8% airplane, x9% drone, ..., xN% wind turbines: thus (100 - Xi - x2 - ... - xN) % corresponding to the “other” type.

[0058] In a step 203, the filtering device 1 determines, for each detected echo sn, an output information Outn as a function of the estimate provided by the neural network 11 on the basis of the set of images Ens1Mn and delivers it.

[0059] In one embodiment, the filtering device 1 delivers as output information Outn a vector comprising as coordinates the percentages delivered by the neural network 11 relative to the echo sn. In another embodiment, the filtering device 1 indicates as output information Outn the type of target having obtained the maximum percentage in the output data calculated by the neural network 11.

[0060] In one embodiment, the device 1 provides the echo sn and its attributes as output information Outn if and only if the target type determined as the one corresponding (with the greatest percentage if applicable) to the echo sn is part of a predetermined subset of the target types of Enstyp_cibieS, only this subset being considered relevant for the rest of the processing (strict subset, comprising a strictly lower number of types than that of the set of target types) and if not, the device 1 provides nothing as output (or provides an indication of non-relevance) relative to the echo sn: for example, thus for the rest of the processing in the processing chain, only the echoes assimilated to airplanes, helicopters, rockets, artillery, mortars, drones are kept and the echoes corresponding to wind turbines, road traffic, clutter, interference, "others" are deleted.

[0061] In [Fig.3] are represented two sets of images Ens1Mni and Ens1Mn2 each comprising 3 images (azimuth, distance), (time, distance) and (time, SER). These sets of images are delivered by the building block 10 for two echoes sni and sn2 in an embodiment of the invention, in which the set Ens1Mni was constructed for an echo sni generated by a wind turbine type object while the set Ens1Mn2 was constructed for an echo sn2 generated by an airplane type object.

[0062] In one embodiment, images for fixed attributes have a constant size.

[0063] [Fig.4] represents the echo sn (cross in bold) being processed by the building block 10 on a polar grid (distance, azimuth boxes) of the detections marked by crosses during a radar antenna rotation. The echoes selected by block 10 during step 201 are those located in the gray cells of the polar grid in one embodiment.

[0064] The proposed solution allows for visual recognition of targets across one or more views (e.g., time-distance, time-azimuth, distance-azimuth) over a temporal depth. The classification problem across context is thus reported to an image classification.

[0065] The learning phase 101 of the neural network is now described.

[0066] During the learning phase, a learning database is created, comprising a multitude of echoes (for example in the case considered 1 million) each associated with attributes and each associated with the type to which it corresponds among the types of target objects of the set Enstyp cibies (thousands of echoes per type). For each of these echoes, a set of K images (K is predefined, as is the pair of attributes associated with each image) is constructed and the weights of the neural network are defined as the learning progresses so that the classification error decreases between the classification attributed during the learning by the neural network and the known type previously associated with the echo. In the case considered, a validation phase is then carried out on a thousand examples (relating to a geographical area other than that(those) of the learning), with various radar operating modes.

[0067] In the example considered, the neural network is a Deep Learning type neural network, for example of the convolutional type (CNN, in English Convolutional Neural Network), as represented in [Fig.5]. It comprises convolution layers (convl to conv6) allowing the detection of characteristics on the image and fully connected classification layers (fully connected: fc). Each image constitutes a photograph of the context of the echo detection. The indices “1” which appear indicate the passage into one-dimensional vectors.

[0068] Each image constitutes a photograph of the context of the echo detection.

[0069] Other types of neural networks are usable in other modes of rea lisation of the invention, for example recurrent neural networks or recurrent neural networks with short-term and long-term memory, transformers etc.

[0070] The calculations implemented by the neural network are for example implemented by CNN accelerators, comprising PE (processing elements) grids implementing MAC multiplication-accumulation functions as a function of the weights and input images.

[0071] In step 102 of the preliminary phase, the neural network 11 resulting from the learning phase 101 is integrated into the (or each) filtering device 1.

[0072] Filtering according to the invention thus makes it possible to distinguish between the different types of targets and to process each of the echoes according to its type of target as determined by the filtering device. For example, echoes of the helicopter, airplane, drone type will undergo subsequent processing, in the processing chain of the radar receiver, intended to specify what type of helicopter, airplane, drone it is; echoes corresponding to types of targets considered undesirable (or not corresponding to types of targets considered useful) in the rest of the processing will be eliminated (noise, interference, jamming, intentional and unintentional clutter (ground / sea, weather, migratory birds, wind turbines, road traffic, etc.) from the rest of the processing or will be subject to specific processing: for example, cross-referencing a wind turbine type echo with the theoretical topographic survey of wind turbines, road traffic with road maps, etc.

[0073] For filtering the plots at the extractor output, in one embodiment, the device 1 replaces the role of tracking which sorts between the useful plots belonging to a trajectory having kinematic coherence and the erratic or undesirable plots.

[0074] The invention thus proposes an echo filtering solution, which: - is based on an expanded spatio-temporal vision, - places the detection in its context by extracting and analyzing its history, - broadens the scope of information available for decision-making, - is perfectly suited to the capabilities of AI solutions to detect patterns and their contexts, - provides a general and unifying framework for tracking or track before detect techniques, - significantly improves radar / lidar / sonar performance (detection / false alarm compromise).

[0075] The invention has been described above with respect to a Radar application. Of course, it is applicable to various echo analysis systems, such as, in addition to RADAR systems, SONAR or LIDAR type systems.

[0076] The invention has been described above in relation to an air traffic monitoring application. Of course, it is applicable to different applications: maritime, road, weather monitoring, etc.

Claims

Claims

1. Method for filtering echoes received by an echo receiving device, each echo being associated with attribute values ​​of said echo, said method being characterized in that it comprises the following steps implemented by an electronic filtering device (1), for a considered echo sn: - i / application of a space-time window of predetermined size comprising said considered echo sn and determination of the echoes located in said space-time window including said considered echo sn;- ii / construction of a set of hnage(s) (Ens1Mn) comprising at least one image, each image of the set of hnage(s) being associated with a respective predefined pair of attributes, each position in said image corresponding to a respective pair of values ​​of said attributes, distinct positions corresponding to pairs of distinct values ​​of said attributes, according to which, during the construction of said image, a determined pattern is inserted into the image at a determined position, if and only if the values ​​of the attributes corresponding to said position are the values ​​of said attributes for one of the echoes determined in the spatio-temporal window;- üi / providing said set of image(s) as input to a neural network (11) resulting from learning and adapted to estimate, as a function of said set of image(s) constructed (Ens iMn) for the echo sn considered, the correspondence of said echo considered to at least one type of target objects from a predefined set of types of target objects; and - iv / estimation by said neural network (11) of the correspondence of said echo sn considered to at least one type of target objects from a predefined set of types of target objects.;

2. The echo filtering method of claim 1, wherein the attributes comprise or are a function of one or more of the following: echo detection time, distance of the object at the origin of the echo, azimuth and / or elevation of the received echo, speed of the object, ratio between the returns captured according to orthogonal polarization planes or their correlations, signal-to-noise ratio of the echo, Radar Equivalent Surface of the object, spatial extension of the echo from the object, spectral spread or presence of blade flash.

3. Echo filtering method according to claim 1 or 2, according to which the set of hnage(s) (Ens^) constructed for each echo sn comprises at least two images, the pair of attributes associated with one of said images being distinct from the pair of attributes associated with another image of the set of hnage(s).

4. Echo filtering method according to any one of the preceding claims, according to which the set of hnage(s) (Ens^) constructed for the echo sn comprises one or more images from among an image associated with the azimuth-distance attribute pair, an image associated with the time-distance attribute pair and an image associated with the time-SER attribute pair.

5. Echo filtering method according to any one of the preceding claims, according to which the neural network (11) comes from a learning phase during which a learning base comprising echoes associated with their attributes and each previously associated with a type of the predefined set of target object types is used, a set of hnage(s) being constructed for each echo of the learning base in accordance with steps i and ii and according to which the set of images constructed for each echo is provided as input to the network being learned and the weights of the neurons of the network are modified as a function of the error between the type previously associated with the echo and the estimate provided by the neural network, in its current learning state, as a function of said set of hnage(s).

6. A computer program, intended to be stored in the memory of an electronic echo filtering device (1) and further comprising a computer, said computer program comprising instructions which, when executed on the microcomputer, implement the steps of a method according to one of the preceding claims.

7. Device for filtering echoes (1) received by an echo receiving device, each echo being associated with attribute values ​​of said echo, said device being characterized in that it is adapted, for a considered echo sn, to apply a spatio-temporal window of predetermined size comprising said considered echo sn and to determine echoes located in said spatio-temporal window including said considered echo sn;said device being adapted to then construct a set of image(s) (Ens1Mn) comprising at least one image, each image of the set of image(s) being associated with a respective predefined pair of attributes, each position in said image corresponding to a respective pair of values ​​of said attributes, distinct positions corresponding to pairs of distinct values ​​of said attributes, so that, during the construction of said image, a determined pattern is inserted into the image at a determined position, if and only if the values ​​of the attributes corresponding to said position are the values ​​of said attributes for one of the determined echoes in said spatio-temporal window;and said device is adapted to then provide said set of image(s) as input to a neural network (11) resulting from learning and adapted to estimate, as a function of said set of image(s) constructed (Ens1Mn) for the echo sn considered, the correspondence of said echo considered sn to at least one type of target objects from a predefined set of types of target objects.;

8. A filtering device (1) according to claim 7, wherein the attributes comprise or are a function of one or more of the following: echo detection time, distance from the object originating the echo, azimuth and / or elevation of the received echo, velocity of the object, ratio between returns captured along orthogonal polarization planes or their correlations, signal-to-noise ratio of the echo, Radar Equivalent Surface of the object, spatial extension of the echo from the object, spectral spread or presence of blade flash.

9. A filtering device (1) according to claim 7 or 8, wherein the set of image(s) (Ens^) constructed for each echo sn comprises at least two images, the pair of attributes associated with one of said images being distinct from the pair of attributes associated with another image of the set of hnage(s).

10. Filtering device (1) according to one of claims 7 to 9, in which the set of hnage(s) (Ens1Mn) constructed for the echo sn comprises one or more images from among an image associated with the azimuth-distance attribute pair, an image associated with the time-distance attribute pair and an image associated with the time-SER attribute pair.