Airborne radar system adapted for the recognition of fixed objects on the ground; Training method and computer program products.

Abstract

Airborne radar system adapted for recognizing fixed objects on the ground; Training method and computer program output. The present invention relates to an airborne radar system (10), comprising an operational radar (12) and an operational computer (14), the operational radar being capable of producing a raw radar dataset. The radar system (10) is characterized in that the operational computer (14) is programmed to execute an artificial intelligence algorithm (40) adapted to recognize fixed objects on the ground directly from the raw radar dataset or from a normalized radar dataset derived from the raw radar dataset produced by the operational radar (12). Figure for the abstract: Figure 1

Description

Title of the invention: Airborne radar system adapted for the recognition of fixed objects on the ground; Training method and computer program products.

[0001] The invention has as its technical field that of the recognition of fixed objects on the ground, using an airborne radar system.

[0002] Fixed objects on the ground, such as bridges, water towers, lighthouses, etc., can serve as landmarks for navigation.

[0003] Object recognition from radar mapping can thus be useful for low-altitude terrain tracking without having to resort, for example, to location data from a satellite positioning device.

[0004] According to the prior art, the recognition of fixed objects on the ground consists of implementing successively:

[0005] - a first step in processing a set of raw radar data (“data set " delivered by the radar to construct an image;

[0006] - a second image analysis step to isolate shapes and classify each of these forms according to a predetermined set of object types.

[0007] However, in the case, for example, of SAR (“Synthetic Aperture Radar”) processing of a set of raw radar data, the first step monopolizes the radar, which must remain pointed at the object to be recognized while acquiring enough information to construct a SAR image.

[0008] Moreover, this construction requires significant computing resources and / or a long computation time.

[0009] The second object recognition step, when carried out automatically by expert software of the type artificial intelligence algorithm - AI resulting from supervised learning, requires populating a training database.

[0010] Such a database groups labeled SAR images, that is to say a doublet associating a SAR image and a label indicating the true class of the object present on that SAR image.

[0011] But, in addition to the nature of the object, the shape of the object on the SAR image depends on the flight parameters of the aircraft, such as its height, its speed, or on the operating parameters of the radar, such as its observation angles in elevation and bearing.

[0012] The creation of the training database then requires numerous in-flight measurement campaigns to obtain varied SAR images, followed by their analysis by expert personnel in order to label them correctly.

[0013] In-flight measurement campaigns are expensive.

[0014] Once the expert AI algorithm has been trained, it no longer evolves, for example during operational flights.

[0015] Finally, it must be used with the same type of radar as that which was used to acquire and construct the labeled SAR images of the training database.

[0016] In other words, for each type of radar, dedicated flight measurement campaigns must be carried out to obtain labeled data enabling the expert AI algorithm to be trained in a way that is adapted and specific to that type of radar.

[0017] There is therefore a need for a method to make recognition of fixed objects on the ground by means of an airborne radar in a much simpler way.

[0018] The purpose of the invention is to meet this need.

[0019] For this purpose the invention relates to a radar system comprising an operational radar and an operational computer, the operational radar being capable of producing a set of raw radar data, characterized in that the operational computer is programmed to execute a suitably parameterized artificial intelligence algorithm for recognizing fixed objects on the ground directly from the set of raw radar data or from a set of normalized radar data derived from the set of raw radar data produced by the operational radar.

[0020] According to particular embodiments, the system comprises one or more of the following characteristics, taken individually or in all technically possible combinations:

[0021] - an output of the artificial intelligence algorithm is a class of object as that estimation of the type to which a fixed object on the ground illuminated by the operational radar belongs in order to produce the raw radar data set, the object class being selected from a plurality of predetermined object classes;

[0022] - the operational radar is a frequency-modulated continuous wave radar, the raw radar dataset is then a two-dimensional dataset or in which the operational radar is a pulse radar, the raw radar dataset is then a three-dimensional dataset.

[0023] The invention also relates to a training method for obtaining a suitably parameterized artificial intelligence object recognition algorithm, adapted for execution by the operational computer of an airborne radar system conforming to the previous system. The training method is characterized in that it comprises a learning step from a training database containing a plurality of labeled raw radar datasets. A labeled raw radar dataset associates a raw radar dataset produced by a reference radar with a label equal to a class. true object for a fixed object on the ground illuminated by the reference radar to produce the raw radar dataset.

[0024] According to particular embodiments, the process comprises one or more of the following characteristics, taken individually or in all technically possible combinations:

[0025] - the label associated with a set of raw radar data produced by a radar reference results from the execution, by a reference computer, of a first program to construct a map from the set of raw radar data produced by a reference radar, then from the execution of a second program to recognize the shapes of objects in said map, the class estimated at the output of the second program being considered as the true object class constituting said label;

[0026] - the execution of the first program consists of applying a "radar" processing to aperture synthesis”, “inverse aperture synthesis radar”, or “moving ground target indication”) to the raw radar dataset produced by a reference radar to obtain said mapping;

[0027] - the reference radar being different from the operational radar, the learning step consists of: optimizing a reference encoder and a classifier by a supervised learning method from the training database whose data is produced by the reference radar; using the optimized reference encoder in inference to populate a training database with data produced by the operational radar; and, optimizing an operational encoder by an adverse method on the data from the training database, the artificial intelligence object recognition algorithm being the combination of the optimized operational encoder and the optimized classifier;

[0028] - the training database is populated with radar datasets raw data poorly or not classified by the artificial intelligence object recognition algorithm corresponding to the combination of the optimized reference encoder and the optimized classifier.

[0029] The invention also relates to a computer program product comprising software instructions which, when executed by an operational computer of an airborne radar system, implement an artificial intelligence algorithm for recognizing fixed objects on the ground from a set of raw radar data or from a set of normalized radar data derived from the set of raw radar data produced by the operational radar of the radar system.

[0030] The invention also relates to a computer program product comprising software instructions which, when executed by a computer allow the implementation of one or each step of learning of a training process that conforms to the previous process.

[0031] The invention and its advantages will be better understood upon reading the following detailed description of a particular embodiment, given solely by way of non-limiting example, this description being made with reference to the accompanying drawings in which:

[0032] [Fig-1] Fig. 1 is a schematic representation associating a radar system according to the invention, which is adapted to perform object recognition from raw data delivered by a radar, as well as a learning system;

[0033] [Fig.2] Fig.2 is a block representation of a first embodiment of the learning process according to the invention, which, using the learning system of Fig.1, allows training of the model which, once learned, is executed by the radar system of Fig.1 in the operational phase; and,

[0034] [Fig.3] The [Fig.3] is a block representation of a second embodiment of the learning process according to the invention, particularly well suited to the case where the radar for acquiring training data and the radar for acquiring new data in inference are not identical. Structure

[0035] Fig. 1 represents a preferred embodiment of a radar system 10 enabling automatic recognition of fixed objects on the ground from raw data delivered by a radar.

[0036] The system 10 is advantageously carried on board an aircraft to enable, during an operational phase, the recognition of fixed objects on the ground, such as object O.

[0037] In one example of use, the recognized objects constitute so many reference points for an aircraft navigation system (not shown in the figures) downstream of the system according to the invention.

[0038] The radar system 10 comprises a radar 12 and a computer 14.

[0039] The radar 12 includes an antenna 20 and an electronic preprocessing chain 22 for the signals delivered by the antenna 20, in order to develop a set of radar data S, as raw data delivered by the radar 12.

[0040] For example, the preprocessing chain 22 comprises successively a receiver 24, an analog / digital encoder 25, and a time-domain to frequency-domain conversion unit 26. For example, the unit 26 is suitable for applying a fast Fourier transform to the samples output from the encoder 25.

[0041] The radar 12 thus enables, during a recurrence of operation, the acquisition of a set of raw radar data S. This set corresponds to the observation by the radar 12 of a domain of space. For a ground observation, this domain includes in particular a ground surface on which possibly at least one object O rests.

[0042] Radar 12 is, for example, a pulse radar type radar.

[0043] During a recurrence, such a pulse radar emits, during an emission period, then receives, during a reception period.

[0044] This reception period is segmented into boxes by distance, of equivalent length.

[0045] On a block of N recurrences, a fast Fourier transform - FFT is performed for each cell in distance.

[0046] A decomposition of the received signal into frequency boxes is obtained.

[0047] The frequency resolution of the FFT depends on the recurrence period and the number N of recurrences.

[0048] It can therefore be considered to have frequency cells of width equal to the frequency resolution of the FFT.

[0049] The set of raw radar data S then takes the form of a 3D matrix with: in depth, the cells in distance; in abscissa, the cells in frequency; and, in ordinate, the digitized amplitude of the signals.

[0050] For a pulse radar, we therefore have one spectrum per distance cell, a spectrum being a two-dimensional matrix in amplitude and frequency.

[0051] The set of raw radar data S therefore contains as many spectra as there are boxes in distance.

[0052] For a pulse radar, it is necessary to wait for the end of the acquisition of the N recurrences necessary for the FFT, before launching the processing dedicated to the recognition of the nature of the illuminated surfaces.

[0053] The calculator 14 is a computer comprising calculation means, such as a processor, and storage means, such as a memory.

[0054] The memory stores computer program instructions, in particular those of a program 40, the execution of which allows the recognition of objects directly from the set of raw radar data S output from the radar.

[0055] Program 40 allows an object class C to be associated with an object O illuminated by radar 12 and whose raw radar data set S bears the signature. The class C is selected from a predefined set of possible object classes.

[0056] Program 40 is a suitably parameterized artificial intelligence algorithm - AI, that is to say, resulting from a learning phase, during which a generic object recognition model is progressively modified to obtain the learned model that constitutes program 40.

[0057] The learning phase corresponds to the implementation of the training process 100 of [Fig.2] using the learning system 50 also shown in [Fig.1].

[0058] The learning system 50 includes a radar learning system 60.

[0059] The radar system 60 is advantageously carried on board an aircraft to, during a plurality of flight measurement campaigns (part of the learning phase), populate a training database 80.

[0060] The radar system 60 comprises a radar 62 and a computer 64.

[0061] The radar 62 includes an antenna 70 and a preprocessing chain 72 for the signals delivered by the antenna 70 in order to develop a set of raw reference radar data Sref, as raw information delivered by the radar 72.

[0062] For example, the preprocessing chain 72 successively comprises a receiver 74, a coder 75 and a unit 76 for switching from the time domain to the frequency domain.

[0063] Calculator 74 is a computer comprising calculation means, such as a processor, and storage means, such as a memory.

[0064] The memory of calculator 74 stores, in particular, the instructions of three computer programs:

[0065] The first program 81 is a program for developing a reference image Lef (for example a SAR image) of the scene illuminated by the radar 62 from the set of raw reference radar data Sref,

[0066] The second program 82, or AI expert algorithm, is a program for recognizing objects in the Iref image obtained as output from the first program 81.

[0067] The second program 82 makes it possible to associate a reference class Cref with an object O illuminated by the radar 62, which can be distinguished on the reference image Iref.

[0068] The reference class Cref which is estimated by the second program 82 is considered to be the true class of the set of raw reference radar data Sref used to construct the reference image Iref.

[0069] The reference class Cref is therefore used as the label of the raw data represented by the reference raw radar dataset Sref.

[0070] The third program 83 is a program for labeling the raw data and saving it in the training database 80.

[0071] A labeled data is therefore a doublet associating a set of raw reference radar data Sref and its true class Cref.

[0072] The system 50 also includes a computer 94 whose memory stores a training program 92 allowing a suitably parameterized algorithm 98 to be obtained from a generic algorithm 96.

[0073] The training process consists of extracting a batch of labeled data from the database 80, and then progressively modifying the parameters of the algorithm 96 so that the class estimated by this algorithm for each of the radar data sets in the batch of labeled data gradually approaches the true class associated with each of the radar data sets in the batch of labeled data.

[0074] An algorithm 96 or 98 therefore takes as input a set of raw radar data and delivers as output a class of object.

[0075] In the embodiment of [Fig. 1], the exploitation of the content of the training database 80 is carried out after a series of flight measurement campaigns. Learning process

[0076] Fig. 2 represents a first embodiment of the learning process.

[0077] The training method 100 allows an AI object recognition algorithm to be trained from the raw data delivered by a radar.

[0078] The method 100 begins with a step 110 consisting of acquiring a set of raw reference radar data Sref of a domain observed by the radar 62.

[0079] The observed domain presents an object O, fixed on the ground, which is illuminated by the radar.

[0080] In step 120, labeled information is constructed.

[0081] More specifically, in step 130, program 81 is executed to obtain a SAR image of the observed domain.

[0082] Then, in a step 140, program 82 is executed to analyze the SAR image so as to estimate an object class from the shape of that object in the SAR image.

[0083] Program 82 performs a classification. A class is selected from a plurality of predefined object classes. The plurality of classes includes, for example, the classes "bridge", "water tower", "beacon", "lighthouse", etc.

[0084] Optionally, an additional class is provided which corresponds to the objects of a SAR image which could not be classified by program 82 into one or the other of the predefined classes.

[0085] The method 100 includes a step 150 consisting of associating the set of raw reference radar data Srefacquis in step 110 with the reference class Cref determined in step 140, as the true identifier of the fixed object on the ground illuminated by the radar 62.

[0086] This association leads to a labeled set of raw radar data.

[0087] In a step 160, the labeled set obtained in step 150 is recorded in the training database 80, as new training data.

[0088] Once the training database 80 is sufficiently populated, in a learning step 170, a generic artificial intelligence object recognition algorithm 96 is trained, validated and tested with data sets extracted from the training database 80.

[0089] Algorithm 96 takes as input a set of raw radar data and delivers as output the estimated class of the object illuminated by the radar during the acquisition of this set of raw radar data.

[0090] Since the radars 12 and 62 are of the same type, i.e. the raw radar data sets they deliver are similar, the artificial intelligence algorithm 98 that results from the learning can be loaded into the memory of the computer 14, as a program 40, and used in inference to, from a set of raw radar data newly acquired by the radar 12, automatically determine the class of an object observed by that radar.

[0091] The class to which the object belongs is for example transmitted to a navigation system of the aircraft carrying the radar system 10 to aid navigation. Second embodiment of the process

[0092] Figure 3 represents a second embodiment of the learning process.

[0093] In a second embodiment, the radars 10 and 60 are of different types.

[0094] However, the labeling of a set of raw reference data is very closely linked to characteristics of radar 62, in particular the characteristics of the analog / digital encoder 75, which enabled its acquisition.

[0095] Thus, when the reference radar and the radar used in inference are different, the labeling is difficult to share from one radar to another, due to the dispersions of their technological characteristics.

[0096] This second embodiment of the process according to the invention makes it possible to overcome this drawback by implementing, for learning, an opposing method.

[0097] Process 200 comprises the same steps 110 to 160 as process 100 of [Fig.2],

[0098] Training step 270 is similar to step 170, except that the algorithm 96 to be trained is now specific: it includes a reference encoder Eref followed by a classifier K.

[0099] The reference encoder Eref is associated with the reference radar 60.

[0100] The learning step 270 then consists of optimizing the parameters of the Eref and K models by a usual supervised approach, using the labeled data from the database 80.

[0101] This optimization can be written as:

[0102] minAref(Xref, Yref) hyefjx

[0103] where ArejQ is the cross-entropy, namely, ^ref ( ref^ Yref ) = " IL; ( ^ref ( ) )

[0105] and Xref (respectively Yref) is a set of labeled Nrefdata extracted from the database of data 80 and x'rey (respectively yre^ for each labeled data item of said lot, i integer between 1 and Nref

[0106] We thus obtain an artificial intelligence algorithm 98 consisting of an optimized reference encoder E'ref and an optimized classifier K' in series with each other.

[0107] At the end of the learning step 270, the trained model 98 is loaded into the memory of the computer 14, as a program 40.

[0108] Then, in a step 280, program 40 is used in inference.

[0109] It allows, from a set of raw radar data Xrad newly acquired by radar 12, to estimate a class Yrad, to which belongs an object observed by the radar used for inference.

[0110] As in the first embodiment, the class to which the object belongs can advantageously be transmitted to a navigation system of the aircraft carrying the radar system 10 to aid navigation.

[0111] But above all, during this use of radar 12 and program 40, a second training database 282 is populated automatically.

[0112] For example, during the inference phase, a score is calculated for each pair of data set Xrad - class Y r(td. This score is, for example, information given by the trained model 40. It then corresponds to a probability that the data set Xmd belongs to the class Yrad.

[0113] Only the Xmd datasets with a score below a predefined threshold are then recorded in the second training database 282. In other words, only the data for which the class estimation made by program 40 is not good (according to the thresholding criterion on the score) are retained.

[0114] For example, only data acquired by radar 12 and which could not be classified by program 40 are stored in database 282.

[0115] Then, when the second base 282 is sufficiently populated, in a step 290, the program 40 is optimized with a batch of data extracted from the second training base 282.

[0116] For this purpose, a discriminant network ® compares the estimates made on the one hand by the optimized reference encoder E'ref on a set of Xlvf data from the base 80, and on the other hand by an Erad encoder, which is associated with the target radar 12, on a set of Xrad data from the base 282.

[0117] The objective here is to train the target encoder Erud on radar datasets produced by radar 12, Xrwb so that the optimized classifier K' (optimized in step 270) can then be applied to the outputs of the optimized target encoder E'rad.

[0118] According to the opposing method, the two algorithms Erad and have opposing objectives: we want to optimize E^ so that rJp j] tends towards 1 (i.e. ^raWrad) / that the algorithm D cannot discriminate the output of Erad from that of E'ref, while optimizing to make this result tend towards 0 (i.e. improve D so that it is able to discriminate between the output of Erud and that of E'ref).

[0119] The function to be minimized for the Erad encoder is thus, for example:

[0120] min A ( Xrad, D )

[0121] with: 101221 A(X„* D) = )

[0123] where Xmd is a set of labeled Nraddata extracted from database 292 and xJmd for each labeled data item in said set, j an integer between 1 and Nrad

[0124] The function to be minimized for the discriminant is, for example, based on a binary cross-entropy criterion:

[0125] minA(Xre / , X^, E'^ Erad}

[0126] with

[0127] ^Xre^Er^ = ) ) )

[0128] In the following step 300, the optimized encoder E'rad, obtained at the end of step 290, is combined with the classifier K', obtained at the end of step 270, to together constitute an optimized version of program 40, called optimized program 40'.

[0129] The optimized program 40' is loaded into the memory of computer 14 in place of program 40.

[0130] In step 300, the optimized program 40' is used in inference. It is adapted to radar 12.

[0131] Advantageously, through successive iterations of steps 280 to 300, version 40 of program can thus be improved using data collected by the radar used for inference. At iteration k, E^ (optimized on X^ data) is used in inference to populate the database 282 with Xkad data. Then, Ekad is optimized using the Xk ■^rad data. Implementation variations

[0132] In a first variant, programs 81, 82 and 83, as well as training database 80, on the one hand, and learning program 92, on the other part, are implanted (or replicated) in the memory of computer 14, to be executed by the latter's processor.

[0133] In this first variant, one or more newly acquired radar data sets by radar 12 can then be labeled using expert algorithm 82 and then recorded as new labeled data in base 80.

[0134] This new data is advantageously used for training and updating the artificial intelligence recognition algorithm 40.

[0135] This first variant is particularly interesting for allowing, following an initial training of algorithm 40, an update of the parameters of this algorithm in order to broaden its recognition capabilities from operational data.

[0136] This is therefore a "real-time" training, for example during an operational flight, as opposed to a "delayed-time" training, for example following an in-flight measurement campaign to populate the training database, and to carry out the training of the recognition algorithm, for example later, on the ground.

[0137] In a second variant, independent of the previous one, it is not the raw radar data set that constitutes the applied input data for the object recognition algorithm, but a normalized radar data set.

[0138] Indeed, a set of raw radar data depends on many acquisition parameters, such as the radar beam opening, the aircraft height, the aircraft speed, the viewing angle from which the ground object is observed, etc.

[0139] Normalization then makes it possible to compare different sets of raw radar data by bringing the range and / or frequency spread of a radar data set into the same predefined domain in order to eliminate the influence of acquisition parameters as much as possible. Data normalization therefore makes it possible to significantly reduce the amount of data required for training.

[0140] Thus, in this second variant, the computer 14 (respectively the computer 64) implements, upstream of the algorithm 40 (respectively upstream of the program 83), a normalization program for the set of raw radar data delivered by the radar 12 (respectively by the radar 62).

[0141] Normalization consists of homogenizing radar datasets according to a normalization criterion

[0142] For example, one possible way to achieve normalization is as follows.

[0143] The amplitudes and frequency spread of a raw radar data set generally depend on the height H of the carrier aircraft, the site angle (also called elevation angle) a of the beam, and the aperture © of the radar beam.

[0144] Ideally, the training database should contain radar datasets for different configurations, i.e., several altitudes, elevation angles, and apertures. However, this solution would then require acquiring a large number of raw radar datasets and, consequently, increase the number of in-flight measurement campaigns needed.

[0145] It is therefore preferable that identification be carried out on normalised radar datasets rather than raw ones.

[0146] For a fixed antenna aperture and beam site, the greater the height, the more the spectrum is spread out (because of the increase in distances) and the lower the amplitudes.

[0147] Normalization then consists of relating the raw spectrum or each of a set of raw radar data to the same amplitude and frequency scales. The scales are, for example, those of a calibration spectrum.

[0148] The calibration spectrum is stored with all the shooting configuration parameters during its acquisition (height, speed, shooting angles).

[0149] To normalize a raw spectrum along the frequency axis, its spectral spread Es is compensated.

[0150] It is shown that the spectral spreading can be written, as a function of the height, the site angle and the aperture:

[0151] , / 1 i \ = ■' \ cosa+—J œslct+y) /

[0152] with Ks a coefficient equal to the ratio of the emitted frequency band to the recurrence period, H the height, a the site angle and ® the beam opening

[0153] The first step therefore consists of normalizing the raw spectrum along the frequency axis by applying a scaling factor Eso / where Es is the spectral spread of the raw spectrum and Eso is the spectral spread of the calibration spectrum.

[0154] The second step consists of adjusting the amplitudes of the raw spectrum along the amplitude axis by applying a scaling factor 42, where A is the maximum amplitude of the raw spectrum and A0 the maximum amplitude of the calibration spectrum.

[0155] In a third step, the normalized spectrum is obtained by taking a sample on R, with R - -E where F is the maximum frequency of the raw spectrum and Fo the maximum frequency of the calibration spectrum.

[0156] In a third variant, the radar is a continuous wave radar modulated in frequency - FMCW ("Frequency Modulated Continuous Wave") rather than a pulse radar.

[0157] At each recurrence, an FMCW radar emits, for a duration of emission (or illumination time), a wave taking the form of a frequency ramp, then receives, during a reception period, a received wave resulting from the reflection of the emitted wave on reflective objects present in the radar's observation range.

[0158] A Fast Fourier Transform (FFT) is performed on the received foundation samples for each recurrence.

[0159] For this transformation, the frequency interval is subdivided into N frequency cells, for example, N equal to 1024. The frequency resolution of the FFT depends on the radar recurrence period and the number of cells. Therefore, frequency cells with a width equal to the frequency resolution of the FFT can be considered.

[0160] The raw radar dataset obtained from the FFT is an amplitude-frequency spectrum. It is a two-dimensional dataset. This spectrum is characteristic of the ground object illuminated by the radar.

[0161] For an FMCW radar, the processing dedicated to recognizing the nature of illuminated surfaces can be launched as soon as a recurrence ends, since the data are available at the end of each recurrence.

[0162] In a fourth variant, if in the embodiment presented in detail above, it was a matter of using a SAR image, ISAR, GMTI, etc. images or maps can be used as support for the pattern recognition classification algorithm in an image.

[0163] In a fifth variant, instead of implementing a pattern recognition classification program, an expert analyzes the images or more generally a map resulting from SAR (“Synthetic-Aperture Radar”), ISAR (“Inverse Synthetic-Aperture Radar”), GMTI (“Ground Moving Target Indication”), etc. processing.

[0164] However, pattern recognition in SAR, ISAR, or GMTI images using artificial intelligence algorithms is now very robust. Numerous image databases are available for training such AI algorithms, enabling high-quality training. Confidence in the classification of stationary objects seen from the sky is therefore very good, even excellent. Furthermore, the use of such an algorithm allows for the automatic labeling of raw radar datasets.

[0165] It should be noted that the format of the outputs of the reference expert algorithm constrains that of the outputs of the operational recognition algorithm: the plurality of object classes is chosen in such a way that each identifier of the plurality of identifiers constituting the possible outputs of the expert algorithm corresponds to at least one class of the plurality of classes constituting the possible outputs of the recognition algorithm. Benefits

[0166] A person skilled in the art will see that the invention ultimately consists of teaching a radar system to recognize a fixed object on the ground from a set of raw radar data, based on the results given by a reference radar system.

[0167] This is therefore a supervised learning process with criteria derived from another, already trained system. There is thus a transfer of experience acquired by the reference radar system, or teacher system, to the radar system, or student system.

[0168] The normalization step makes it possible to limit the number of flight measurement campaigns required to populate a training database, so that this database is sufficiently large and complete to lead to satisfactory training of the artificial intelligence object recognition algorithm.

[0169] The normalization step of a raw radar dataset reduces the size of the radar datasets, which facilitates recognition and algorithm training. Consequently, this further reduces the number of flights required to populate the training database.

[0170] Advantageously, the generation of labeled data is carried out automatically, without human intervention. The method according to the invention eliminates the need for experts capable of recognizing the nature of objects seen from the sky directly from the data delivered by the radar. Applications

[0171] The field of application of the invention is that of airborne or satellite radars, for observation, surveillance, etc.

[0172] Particularly advantageously, it is suitable for small aircraft to aid navigation and eliminate the need to carry a satellite positioning device. The navigation system uses a map on which various reference points are marked. During flight, by recognizing specific objects or sequences of objects, the aircraft determines its position.

Claims

Demands

1. Airborne radar system (10), the radar system (10) comprising an operational radar (12) and an operational computer (14), the operational radar being capable of producing a raw radar data set, characterized in that the operational computer is programmed to execute an artificial intelligence algorithm (40) adapted to recognize fixed objects on the ground directly from the raw radar data set or from a normalized radar data set derived from the raw radar data set produced by the operational radar (12).

2. Radar system according to claim 1, wherein an output of the artificial intelligence algorithm (40) is an object class as an estimation of the type of a fixed object on the ground illuminated by the operational radar (12) to produce the raw radar dataset, the object class being selected from a plurality of predetermined object classes.

3. Radar system according to claim 1 or claim 2, wherein the operational radar (12) is a frequency-modulated continuous wave radar, the raw radar data set then being a two-dimensional data set or wherein the operational radar (12) is a pulse radar, the raw radar data set then being a three-dimensional data set.

4. A training method (100) for obtaining an artificial intelligence algorithm (40) suitable for execution by the operational computer (14) of an airborne radar system (10) according to any one of claims 1 to 3 in order to recognize fixed objects on the ground directly from a raw radar dataset or from a normalized radar dataset derived from the raw radar dataset produced by an operational radar (12), the training method being characterized in that it comprises a learning step (170) from a training database (80) containing a plurality of labeled raw radar datasets,a labeled raw radar dataset associating a raw radar dataset produced by a reference radar (62) and a label equal to a true object class for a fixed ground object illuminated by the reference radar (62) to produce the raw radar dataset.

5. A method according to claim 4, wherein the label associated with a set of raw radar data produced by a reference radar (62) results from the execution, by a reference computer (64), of a first program (81) for constructing a map from the set of raw radar data produced by a reference radar (62), and then of the execution of a second program (82) for recognizing the shapes of objects in said map, the class estimated at the output of the second program being considered as the true object class constituting said label.

6. A method according to claim 5, wherein the execution of the first program (81) consists of applying a "synthetic aperture radar", "inverse synthetic aperture radar", or "moving ground target indication" processing) to the raw radar dataset produced by a reference radar (62) to obtain said mapping.

7. A method according to claim 5 or claim 6, wherein, the reference radar (62) being different from the operational radar (12), the learning step (270) consists of: - optimizing a reference encoder (Eref) and a classifier (K) by a supervised learning method from the training database (80) whose data are produced by the reference radar; - using the optimized reference encoder (E'ref) in inference to populate a training database (282) with data produced by the operational radar (12); and, - optimizing an operational encoder (Erad) by an adverse method on the data from the training database (282), the artificial intelligence object recognition algorithm being the combination of the optimized operational encoder (E'rad) and the optimized classifier (K').

8. A method according to claim 7, wherein the training database (282) is populated with raw radar datasets that are poorly or not classified by the artificial intelligence object recognition algorithm corresponding to the combination of the optimized reference encoder (E'rdd) and the optimized classifier (K').

9. Product computer program comprising software instructions which, when executed by an operational computer (14) of an airborne radar system (10), implement an artificial intelligence algorithm (40) for recognizing fixed objects on the ground from a raw radar dataset or from a normalized radar dataset derived from the raw radar dataset produced by the operational radar (12) of the radar system (10).

10. Product computer program comprising software instructions which, when executed by a computer, enable the implementation of the learning step or each step of a training method according to any one of claims 4 to 7.

Images