JP4772341B2 - Target identification device - Google Patents

Description

  The present invention relates to a class identification problem for determining a target type from time-series image data observed by a radar, and more particularly to a target identification apparatus that obtains a target type as a solution.

  As an example of a class identification problem, consider a class identification problem of a target such as a ship or an aircraft. 2. Description of the Related Art Conventionally, a device for classifying a target from time series image data observed by a radar is well known (see, for example, Patent Document 1). The conventional apparatus disclosed in Patent Document 1 includes a dictionary image generator, a target parameter estimator, a normalization / superimposition processor, a dictionary image selector, a type discriminator, a dictionary image file, a target specification file, and tracking data. File, radar observation specification file, observation time series image sequence file, and type discrimination result file.

  Next, the operation of the conventional apparatus disclosed in Patent Document 1 will be described. First, in the dictionary image generator, the target-specific specification information including the target three-dimensional shape data from the target-specific specification file is input to generate dictionary images in various postures of the respective candidate targets. Output to a file.

  Next, in the target parameter estimator, the tracking information of the target from the tracking data file is used as the first input, and the radar observation specification information from the radar observation specification file is used as the second input. A target parameter value which is a parameter such as a traveling direction / posture / rotational motion of each candidate target is estimated or determined and output.

  Further, in the dictionary image selector, a plurality of dictionary image data from the dictionary image file are set as the first input, and the target parameter value output from the target parameter estimator is set as the second input, and the type is determined based on these information. An image to be used at the time of selection is selected and output as dictionary image data.

  Further, in the normalization / superimposition processor, the observation time-series image sequence data from the observation time-series image file is input, and the observation time-series image sequence data is aligned with the normalization process, and then a single process is performed by the superimposition process. Generate and output observation image data.

  Next, in the type discriminator, the dictionary image data output from the dictionary image selector is used as a first input, and the observed image data output from the normalization / superimposition processor is used as a second input. Extract feature values such as contours from the observed image data, or match them using pixel intensity information of the image, select the most matched dictionary image, and output the dictionary image candidate target as the target type . This output becomes the output of the entire conventional apparatus and is input to the type discrimination result file.

  The conventional device disclosed in Patent Document 1 described above is a device that performs type determination for one image generated by normalization / superimposition processing on a time-series image sequence. There has also been proposed a conventional apparatus that performs type discrimination using a feature amount that spans between time-series images, such as a vector (see, for example, Patent Document 2). The conventional apparatus disclosed in Patent Document 2 includes an image extractor, a movement vector extractor, a type discriminator, a feature dictionary file, an observation time series image sequence file, and a type discrimination result file.

  Next, the operation of the conventional device of Patent Document 2 will be described. First, in the image-specific feature quantity extractor, the observation time-series image sequence data from the observation time-series image sequence file is input, and feature values such as brightness are extracted for each image of the observation time-series image sequence data. Output as image-specific features.

  Next, in the movement vector extractor, the observation time-series image sequence data from the observation time-series image sequence file is input, and a movement vector of a target (moving body region) such as an optical flow is extracted and output.

  Next, in the type discriminator, the feature quantity dictionary data from the feature quantity dictionary file is used as a first input, and the image feature quantity output from the image feature quantity extractor is used as a second input. 3 is used as a third input, and is used for type discrimination in combination with the image-specific feature quantity output from the image-specific feature quantity extractor and the movement vector output from the movement vector extractor. The feature value.

  Furthermore, in the type discriminator, the secondary feature values obtained by analysis such as principal component analysis for these feature values are compared with the range for each target type registered in the feature value dictionary data. Type discrimination is performed, and the result is output as a type discrimination result. This output becomes the output of the entire conventional apparatus and is input to the type discrimination result file.

JP 2002-006034 A Japanese Patent Laid-Open No. 9-249083

  However, the prior art has the following problems. In the conventional device disclosed in Patent Document 1, a plurality of time-series images are normalized and superimposed to generate one image that is a type discrimination target, so that characteristic parts that are less dependent on the observation angle are emphasized. On the other hand, there is a problem that the characteristic part appearing for each angle is not effectively used as the characteristic amount for class identification, and class identification between targets having similar shapes is difficult.

  On the other hand, in the conventional device disclosed in Patent Document 2, feature values such as brightness are extracted as independent images from an image sequence obtained in time series. Since this is not taken into consideration, the shape change information cannot be used effectively. In addition, the movement vector, which is another feature amount, is effective from the viewpoint of grasping the movement of the target, but it cannot grasp the shape change of the minute part of the target. Therefore, there is a problem that it is difficult to discriminate the type due to a slight difference between targets having similar shapes and movements.

  The present invention has been made to solve such a problem, and an object of the present invention is to obtain a target identification device that realizes class identification between targets having similar shapes and motions for image data obtained by a radar. And

A target identification device according to the present invention includes tracking data of candidate targets that have been observed by a radar and tracking data that has been generated by simulation, and is based on tracking data that is target position / velocity information and radar observation data. A target parameter estimator for estimating a target parameter estimation value relating to the motion or posture of the target, specification data for each target, which is three-dimensional shape data of the candidate target, radar observation specification data, and a target parameter estimator A dictionary time-series image sequence generator that generates dictionary time-series image sequence data for each candidate target based on the target parameter estimate, and observation time-series image sequence data obtained by observing an unknown target with a radar And the type of target of unknown type included in the observed time-series image sequence data based on the dictionary time-series image sequence data from the dictionary time-series image sequence generator Is obtained by a type discriminator do another.

  According to the present invention, similarity comparison based on observation time-series image sequence data and dictionary time-series image sequence data is performed on image data obtained by a radar, using target minute shape change information, or a dictionary Target identification that realizes class identification between targets with similar shapes and motions by performing type learning based on time-series image sequence data and determining target types of unknown type included in observation time-series image sequence data A device can be obtained.

Hereinafter, preferred embodiments of a target identification device of the present invention will be described with reference to the drawings.
The target identification device according to the present invention uses a similarity comparison based on observation time-series image sequence data and dictionary time-series image sequence data to determine the type of an unknown target included in image data obtained by a radar, or a dictionary. It is characterized by specifying by type learning based on time-series image sequence data.

Embodiment 1 FIG.
FIG. 1 is a configuration diagram of a target identification device according to Embodiment 1 of the present invention. A target identification device 10 according to Embodiment 1 shown in FIG. 1 includes a target parameter estimator 1, a dictionary time-series image sequence generator 2, and a type discriminator 3.

  In addition, as a database for input / output with the target identification device 10, a target specification file 11 in which target target specification data that is candidate target shape data is stored, and tracking data that is target position / velocity information Is stored in the tracking file 12, the radar observation specification file 13 in which the radar observation specification information is stored, and the dictionary time series image sequence data generated by the dictionary time series image sequence generator 2. Image sequence file 14, observation time-series image sequence file 15 in which observation time-series image sequence data, which is time-series image data obtained by observing a target to be subjected to type discrimination with a radar, and type discriminator The type discrimination result file 16 in which the type discrimination result determined by 3 is stored is provided.

  Next, the operation of the target identification device 10 according to the first embodiment will be described. First, the target parameter estimator 1 reads tracking information from the tracking file 12 and radar observation specification information from the radar observation specification file 13 as inputs of the entire target identification device 10. Here, the radar observation specification information includes, for example, an image sequence length, an observation time, and the like. Furthermore, the target parameter estimator 1 estimates or determines a target parameter estimated value, which is a parameter estimated value related to the target, necessary for generating the dictionary time-series image sequence data and determining the type, and outputs the estimated target parameter estimated value.

  As the target parameter estimated value, a motion parameter such as the traveling direction / rotational motion of each candidate target, or a parameter related to the posture can be considered. The target parameter estimated value is obtained from the dictionary target parameter estimated value that is the target parameter estimated value used for generating the dictionary time-series image sequence data and the observed target parameter estimated value that is the target parameter estimated value of the observation target that is the type discrimination target. Become.

  Here, the dictionary target parameter estimated value is estimated from tracking data of candidate targets that have been observed by the radar, or tracking data generated by simulation. The observation target parameter estimated value is estimated from the tracking data of the observation target that is the type discrimination target acquired by the radar. Although the observation target parameter estimated value is not shown in FIG. 1, it is output from the target parameter estimator 1 to the type discriminator 3. Thereby, the type discriminator 3 can perform the type discrimination using the observation target parameter estimated value.

  Next, the dictionary time-series image sequence generator 2 uses, as a first input, the target-specific specification information that is the shape information of the candidate target from the target-specific specification file 11 as the input of the target identification device 10 as a whole. At the time of dictionary, the radar observation specification information from the radar observation specification file 13 is used as the second input, and the dictionary target parameter estimation value output from the target parameter estimator 1 is used as the third input as the input of the entire identification device 10. Series image sequence data is generated.

  The dictionary time-series image sequence generator 2 includes a motion parameter such as the moving direction / rotational motion of the dictionary target given as the dictionary target parameter estimation value, a posture parameter, an image sequence length given as radar observation specification information, and Based on the observation time, dictionary time-series image sequence data for each candidate target is generated and output to the dictionary time-series image sequence file 14.

  The type discriminator 3 receives the observation time-series image sequence data to be type-identified from the observation time-series image sequence file 15 as a first input, and uses the dictionary time-series image sequence for each candidate target from the dictionary time-series image sequence file 14. Using the data as the second input, matching is performed between the dictionary time-series image sequence data of each candidate target and the observed time-series image sequence data, and the maximum similarity is set as the similarity of the candidate target.

  Further, after calculating the similarity of all candidate targets, the type discriminator 3 determines the candidate target having the maximum similarity as the target type, and outputs the result as the type discrimination result. This output becomes the output of the entire target identification device 10 of the present invention and is stored in the type discrimination result file 16.

  Note that the similarity of each image is calculated from the luminance difference between the pixels constituting the image and the luminance distribution, and the similarity of the entire image sequence is calculated from the sum of the similarities of each image, average / variance, and the like. A method is conceivable.

  According to the first embodiment, the observation image sequence and the dictionary image sequence are matched in units of image sequences representing the target shape change information for the radar observation image data, so that the minute shape change of the target is achieved. Class identification can be performed, and identification between targets having similar shapes and motions is possible.

Embodiment 2. FIG.
Hereinafter, a target identification device according to Embodiment 2 of the present invention will be described with reference to FIGS. FIG. 2 is a configuration diagram of the dictionary time-series image sequence generator 2 according to Embodiment 2 of the present invention. The target identification device 10 according to the second embodiment has the configuration shown in FIG. 1 and has the configuration shown in FIG. 2 as the dictionary time-series image sequence generator 2. As shown in FIG. 2, the dictionary time-series image sequence generator 2 includes a dictionary area division width determiner 21, a dictionary area divider 22, a simulation image generator 23, a divided area dictionary image determiner 24, and a dictionary time-series image. It is composed of a column determiner 25.

  In addition, as a database for inputting / outputting from / to the dictionary time-series image sequence generator 2, together with the target specification file 11, the radar observation specification file 13, and the dictionary time-series image sequence file 14 described in the first embodiment. A simulation image file 17 is further provided.

  Next, the operation of the second embodiment will be described. In the first embodiment, the target parameter estimator 1 estimates all target parameters such as motion parameters such as the target traveling direction / rotational motion and parameters related to the posture, and based on these target parameter estimates, An example is shown in which the sequence image sequence generator 2 generates a dictionary time-series image sequence.

  On the other hand, in the second embodiment, as shown in FIG. 2, in the dictionary area division width determiner 21, the dictionary area divider 22, and the simulation image generator 23, the dictionary in which the candidate target is located at the center of the area. After setting the region, simulation image data in which the candidate target is observed is generated from each divided region obtained by dividing the dictionary region. Further, in the divided region dictionary image determiner 24 and the dictionary time-series image sequence determiner 25, a target parameter estimator is generated. An example of generating a dictionary time-series image sequence based on the parameters related to the target motion estimated in 1 is shown.

  The dictionary area division width determiner 21 receives the specification data by target from the specification file 11 by target, determines the division width of the dictionary area, and outputs it as the dictionary area division width. As the specification data by target, for example, an assumed range of the rotational speed by target type can be considered. In this case, the dictionary area division width determiner 21 estimates the target average rotation angular velocity from the assumed range and sets it as the division width of the dictionary area.

  For example, when the assumed range of the rotational speed is 10 to 20 degrees / second, the average rotational angular speed is estimated to be 15 degrees / second, and the division width is set to 15 degrees. This is a case where the assumed observation time is 1 second, but the division width is adjusted so as to be proportional to the assumed observation time. For example, when the assumed observation time is 2 seconds, the division width is set to 30 degrees. Further, the assumed observation time is fixed in advance by the system, or is obtained as radar observation specification data from a radar observation specification file.

  The dictionary area divider 22 receives the dictionary area division width output from the dictionary area division width determiner 21, performs division of the dictionary area, and outputs a dictionary area division result. 3 and 4 are diagrams showing an example of dictionary area division by the dictionary area divider 22 according to Embodiment 2 of the present invention. As the dictionary area, a spherical surface as shown in FIG. 3 is conceivable, and a regular icosahedron as shown in FIG. 4 may be used. As a division method, division is performed at equiangular intervals based on the division width.

  The simulation image generator 23 receives the result of the dictionary region division output from the dictionary region divider 22 and the target-specific data from the target-specific specification file 11, and performs simulation for each candidate target observed from each divided region. Image data is generated and output to the simulation image file 17.

  The center of gravity of the candidate target is located at the center of the dictionary area, and an image obtained by observing each candidate target from the center point of each divided area is simulated based on the three-dimensional shape data of the candidate target obtained as target-specific data. Generate with Since the rotation axis direction of the candidate target is undecided, simulation image data for each rotation axis direction is generated in each divided region assuming a plurality of rotation axis directions.

  The above processing is performed as a batch process in advance, not at the time of target observation, and all simulation image data for each candidate target is stored in the simulation image file 17 as a database.

  The divided area dictionary image determiner 24 receives the simulation image data from the simulation image file 17 and the target parameter estimated value output from the target parameter estimator 1 and determines a dictionary image of each divided area for each candidate target. And output as a divided region dictionary image. As the target parameter estimated value, a rotational axis direction estimated value is given. The divided area dictionary image determiner 24 selects the corresponding simulation image data in the rotation axis direction from among the plurality of simulation image data in the rotation axis direction, and outputs it as a divided area dictionary image that is a dictionary image of each divided area. .

  The dictionary time-series image sequence determiner 25 receives the divided region dictionary image output from the divided region dictionary image determiner 24 and the radar observation specification data from the radar observation specification file 13 as input, and a plurality of divided region dictionary images. A dictionary time-series image sequence consisting of is generated and output for each candidate target. As the radar observation data, the image sequence length of the observation image sequence is given. The dictionary time-series image sequence determiner 25 determines the sequence sequence of dictionary images that satisfy the image sequence length from the divided region dictionary images that constitute the dictionary region, and sets it as the dictionary image sequence. The direction of the image sequence may be a horizontal direction, a vertical direction, a 45 degree direction, or the like.

  In the second embodiment described above, an example is shown in which the dictionary region division width determiner 21 determines one dictionary region division width and divides all dictionary regions at equiangular intervals. The width may be changed depending on the position of the dictionary area. For example, a method of dividing so that the areas of the divided dictionary regions are equal can be considered.

  In the case of division at an equal angle as shown in FIG. 3, a coordinate system with the target center of gravity as the center, the target nose direction as the X axis, the horizontal direction as the Y axis, and the height direction as the Z axis is considered. The direction of the radar with respect to the target is defined by the azimuth angle measured counterclockwise from the X axis in the XY plane in the XYZ coordinate system and the elevation angle measured in the Z axis direction from the XY plane. The area of the divided region tends to be small near an elevation angle of ± 90 degrees, and conversely tends to be large near 0 degrees, and this difference in area ratio is directly linked to a difference in sensitivity to target movement due to a difference in observation position.

  That is, there is a difference in that even if the movement is small, it extends between adjacent images at around 90 degrees, while the movement between adjacent images occurs when there is a considerably large movement at around 0 degrees. Specifically, the step width in the azimuth direction near the elevation angle of ± 90 degrees is made relatively larger than the step width near 0 degrees.

  According to the second embodiment, the target parameter is estimated only for the motion parameter such as the direction of the rotation axis, and class identification can be realized with the dependence on the parameter estimation accuracy minimized.

  Furthermore, by changing the dictionary area division width according to the dictionary area position, it is possible to eliminate the difference in sensitivity to the target movement at the target observation position and perform class identification with reduced accuracy variations depending on the target observation position.

Embodiment 3 FIG.
Hereinafter, a target identification device according to Embodiment 3 of the present invention will be described with reference to FIG. FIG. 5 is a configuration diagram of the target identification device 10 according to Embodiment 3 of the present invention. The configuration of the target identification device 10 in the third embodiment is different from the configuration of the target identification device 10 in the first embodiment in FIG. 1 in that an observation image sequence processor 4 is newly added. Further, the database for inputting / outputting from / to the target identification device 10 is the same as that shown in FIG.

  The newly added observation image sequence processor 4 is based on the input of the radar observation specification data from the radar observation specification file 13 and the observation time series image sequence data from the observation time series image sequence file 15. It has a function of processing the observation time series image sequence data and outputting the observation image sequence processing data.

  Next, the operation of the target identification device 10 according to Embodiment 3 will be described. First, the target parameter estimator 1 reads the tracking information from the tracking file 12 and the radar observation specification information from the radar observation specification file 13 as in the first embodiment. Then, the target parameter estimator 1 estimates or determines the target parameter estimated value and outputs it.

  The observation image sequence processor 4 receives the radar observation specification data from the radar observation specification file 13 and the observation time series image sequence data from the observation time series image sequence file 15 as inputs. As the radar observation data, the image sequence length and observation time of the observation time series image sequence are given. As a condition for processing the observation time series image sequence, for example, the observation time may be different from the assumed observation time, and the observation image sequence processing unit 4 uses the observation time series image sequence data according to the observation time. Is processed.

  As a first method for processing the observation time-series image sequence data, there is a method of extending the length of the observation time-series image sequence data. For example, the observation time may be longer than the expected time and the degree of target rotation between images may be large. As an expansion method, it is conceivable to insert a null image (all pixel intensities are 0) into the image sequence. The insertion position of the null image depends on the number of insertion images. For example, the insertion position is an odd or even number or the nth (n is a positive number calculated from the number of insertion images).

  As a second method of processing the observation image sequence, there is a method of reducing the length of the observation image sequence. For example, the observation time may be shorter than the expected time and the degree of target rotation between images may be small. As a reduction method, it is conceivable to delete an image at a specified position from the image sequence. Although the deletion position of the image depends on the number of deleted images, for example, the deletion position is an odd or even number or nth (n is a positive number calculated from the number of deleted images). The observed image sequence processor 4 outputs the observed image sequence processed data and the processed image sequence length processed in this way.

  The dictionary time-series image sequence generator 2 includes target specification information from the target specification file 11, observation time included in the radar observation specification information from the radar observation specification file 13, and target parameter estimator 1. A dictionary time-series image sequence data is generated by further inputting the processed image sequence length output from the observed image sequence processor 4 together with the estimated dictionary target parameter value output from, and outputs it to the dictionary time-series image sequence file 14. .

  That is, the dictionary time-series image sequence generator 2 in Embodiment 3 generates dictionary time-series image sequence data in consideration of the processed image sequence length output from the observed image sequence processor 4. Different from the dictionary time-series image sequence generator 2 in the first embodiment.

  The type discriminator 3 receives the observed image sequence processed data output from the observed image sequence processor 4 and the dictionary time-series image sequence data from the dictionary time-series image sequence file 14, and determines the type of the observed image sequence processed data. As the target data, a type discrimination process is performed and a type discrimination result is output. This output becomes the output of the entire target identification device 10 of the present invention and is stored in the type discrimination result file 16.

  That is, the type discriminator 3 according to the third embodiment takes into account, for example, a case where the observation time is different from the expected time based on the observation image sequence processing data output from the observation image sequence processing device 4. The type discrimination process is different from the type discriminator 3 in the first embodiment.

  When the dictionary time-series image sequence generator 2 according to the second embodiment is combined with the target identification device 10 according to the third embodiment, the observation time-series image processor 4 changes the dictionary time-series image sequence determiner 25 to the target time-series image sequence determiner 25. On the other hand, the processed image sequence length is output. The condition for processing the observation time-series image sequence corresponds to a case where the target rotation speed is higher than the average rotation angular speed assumed when the simulation image data is generated.

  According to the third embodiment, the observation time-series image sequence is processed and the dictionary time-series image sequence is generated in accordance with the change in the observation time and the fluctuation of the target rotational motion, so that it matches the target observation condition. Class identification.

  In addition, although the example which performs either expansion | extension and reduction | decrease of an observation time series image sequence was shown as an observation time series image sequence processing method, it is not limited to this. In addition to these, the observation time-series image sequence is expanded, reduced, or unprocessed, and the dictionary image sequence with the highest degree of similarity is found for all types to be classified. May be a candidate target evaluation value which is an evaluation result of the candidate target. By combining the images in which the lengths of the observation time-series image sequences are variously combined to be the type discrimination target, it is possible to perform class discrimination while suppressing the influence of parameter estimation errors such as target rotational motion.

Embodiment 4 FIG.
Hereinafter, a target identification device according to Embodiment 4 of the present invention will be described with reference to FIGS. FIG. 6 is a configuration diagram of the dictionary time-series image sequence generator 2 according to Embodiment 4 of the present invention. The target identification device 10 according to the fourth embodiment has the configuration shown in FIG. 1 and has the configuration shown in FIG. 6 as the dictionary time-series image sequence generator 2. As shown in FIG. 6, the dictionary time-series image sequence generator 2 includes a plurality of dictionary region division width determiners 21a, a dictionary region divider 22, a simulation image generator 23, a dictionary region division width selector 26, and a divided region dictionary image. It comprises a determiner 24 and a dictionary time series image sequence determiner 25.

  Compared with FIG. 2 showing the configuration of the dictionary time-series image sequence generator 2 in the second embodiment, the configuration of FIG. 6 is used by the plural dictionary region division width determiner 21 a instead of the dictionary region division width determiner 21. The difference is that a dictionary region division width selector 26 is newly added.

  The plural dictionary area division width determiner 21a receives the specification data by target, determines a plurality of dictionary area division widths, and outputs them as dictionary area division width candidate values. The plural dictionary region division width determiner 21a further includes a function for obtaining a plurality of dictionary region division widths as dictionary region division width candidate values with respect to the dictionary region division width determiner 21 described in the second embodiment. This corresponds to an area division width determiner.

  The dictionary area division width selector 26 also includes radar observation data, target parameter estimation values output from the target parameter estimator 1, and dictionary area division width candidates output from the plural dictionary area division width determiner 21a. It has a function of inputting a value and selecting and outputting a dictionary area division width.

  Next, the operation of the fourth embodiment will be described. In the second embodiment, an example is shown in which the dictionary region division width determiner 21 uniquely determines the dictionary region division width determiner 21, but in the fourth embodiment, it is used instead of the dictionary region division width determiner 21. A plurality of dictionary region division width candidate values are determined by a plurality of dictionary region division width determiners 21a, and an appropriate dictionary region division width is selected from a plurality of dictionary region division width candidate values by a dictionary region division width selector 26. An example of performing

  The multiple dictionary area division width determiner 21a receives the specification data by target from the specification file 11 by target, determines a plurality of division widths of the dictionary area, and outputs them as candidate dictionary area division width values. As a method for determining a plurality of division widths, a method of generating a plurality of divisions at regular intervals within an assumed rotation speed range for each target type, or a method of increasing / decreasing the division width at regular intervals centered on the average rotation angular velocity, for example, average rotation A method of determining a plurality of division widths such as 5, 10, 15, 20, 25 degrees around an angular velocity of 15 degrees is conceivable. Further, the number of candidates for division width may be fixed in advance, or may be set by the user at the time of execution.

  The dictionary area divider 22 divides the dictionary area into a plurality of dictionary area division width candidate values as in the second embodiment, and a dictionary which is a plurality of dictionary area division results based on a plurality of different dictionary area division width candidate values. Output region division result group. Furthermore, the simulation image generator 23 generates simulation image data for each candidate target for the dictionary region division result group, and outputs the simulation image data group to the simulation image file 17 as in the second embodiment.

  The dictionary region division width selector 26 outputs the radar observation specification data from the radar observation specification file 13, the target parameter estimation value output from the target parameter estimator 1, and the plural dictionary region division width determiner 21a. The dictionary area division width candidate value to be input is selected, and the dictionary area division width is selected and output. The radar observation data is given the observation time, and the target parameter estimated value is the target rotational speed. The dictionary area division width selector 26 calculates the degree of rotation per target observation time from these values, and selects the closest dictionary area division width from the dictionary area division width candidate values based on the calculated degree.

  The divided region dictionary image determiner 24 includes a simulation image data group from the simulation image file 17, a target parameter estimated value output from the target parameter estimator 1, and a dictionary region divided output from the dictionary region divided width selector 26. Using the width as input, simulation image data having a specified dictionary area division width is extracted from the simulation image data group. Further, the divided region dictionary image determiner 24 determines and outputs a divided region dictionary image for each candidate target based on the rotation axis direction given as the target parameter estimated value from the extracted simulation image data.

  Similar to the second embodiment, the dictionary time-series image sequence determiner 25 receives the divided region dictionary image output from the divided region dictionary image determiner 24 and the radar observation specification data from the radar observation specification file 13. A dictionary time-series image sequence for each candidate target composed of a plurality of divided region dictionary images is generated and output.

  According to the fourth embodiment, a single dictionary is selected from dictionaries having a plurality of dictionary area division widths in accordance with changes in observation time and fluctuations in the target rotational motion, so that the target observation conditions are met. Combined class identification can be performed.

  In addition, although the example which selects a dictionary area | region divided width uniquely from the dictionary area | region dividing width candidate value was shown, it is not limited to this. Other than that, leave multiple dictionary region division widths as they are, use dictionary image sequences with different dictionary region division widths when determining the type, and select the value with the highest similarity to the observed time-series image sequence. It is good also as a candidate target evaluation value. By using dictionary image sequences having different dictionary area division widths for type discrimination, it is possible to perform class identification while suppressing the influence of parameter estimation errors such as target rotational motion.

Embodiment 5 FIG.
Hereinafter, a target identification device according to Embodiment 5 of the present invention will be described with reference to FIGS. FIG. 7 is a configuration diagram of the type discriminator 3 according to the fifth embodiment of the present invention. The target identification device 10 according to the fifth embodiment has the configuration shown in FIG. 1, and the type discriminator 3 has the configuration shown in FIG. 7. As shown in FIG. 7, the type discriminator 3 includes a type learner 31 and a type calculator 32.

  Next, the operation of the fifth embodiment will be described. In the first embodiment, the type discriminator 3 performs the type discrimination by matching the dictionary time series image sequence and the observation time series image sequence. However, in the fifth embodiment, the type discriminator 3 is configured. An example of performing type discrimination with type learning using the type learning device 31 and the type calculation device 32 will be described. FIG. 8 is an explanatory diagram of type learning by the type learner 31 and type discrimination by the type calculator 32 in the fifth embodiment of the present invention.

  First, an example in which type learning is performed by the type learning device 31 will be described. In the fifth embodiment, a linear discriminant method that is a method for determining a boundary line is applied as a learning method. There are several linear discriminant methods. Here, a linear discriminant method based on a covariance matrix is used. The linear discriminant method based on the covariance matrix projects data to a vector w that passes through the cluster center representing the center of all data so that different types of data are separated as much as possible and the same type of data is distributed as close as possible. This is a method for discriminant analysis.

  When performing the linear discrimination based on the covariance matrix, the type learning device 31 compares the projection result y of each data with the projection result yx of each type of average vector, and the projection result y belongs to the type having the closest yx. It is determined as a thing. The boundary is defined as a straight line perpendicular to w and passing through the cluster center. FIG. 8A shows an example of type learning by the linear discriminant method. If Δ is the teacher data of the type 1 candidate target, and ■ is the type 2 candidate target teacher data, the type learning by the type learning unit 31 causes a boundary line as shown in FIG. It is determined.

  In the above example, linear discrimination based on the covariance matrix is applied to type learning, but other learning methods that have been shown to be effective in high-dimensional data such as support vector machines are applied. Also good. The support vector machine is a method to find a straight line that is a boundary based on the distance and direction from all the teacher data. The distance is maximized and the direction of the straight line seen from the teacher data is the same for each type. Determine a straight line.

  The support vector machine is considered as one of the modification methods of the linear discriminant method. Compared to general linear discriminant methods such as the linear discriminant method based on the covariance matrix described above, the support vector machine represents a function necessary for learning by a linear sum of functions called kernel functions. It is very different in that boundary definition is possible. There are a plurality of kernel functions, and the learning parameter specifies which kernel is selected.

  As described above, the appropriate type learning method differs depending on the property / distribution status of the observation time-series image sequence data. Therefore, the method may be selected based on the characteristics of each learning method.

  Next, type discrimination performed by the type calculator 32 using the type learning result output from the type learner 31 and the observation time-series image sequence data as inputs will be described. FIG. 8B shows an example of performing type discrimination. The type calculator 32 determines the type of an unknown target included in the observed time-series image sequence data based on the boundary line obtained as the type learning result.

  In FIG. 8B, the area above the boundary line to which many Δs corresponding to type 1 belong is the type 1 boundary area, and conversely, the area below the boundary line to which many ■ corresponding to type 2 belong is typed. 2 boundary region. In this case, the type calculator 32 determines the type depending on which region the observation time-series image sequence data that is the type determination target belongs to.

  In FIG. 8B, black triangles and ◇ are observation data of unknown types. The type calculator 32 determines that the black triangle is of type 1 because the distribution of black triangles is in the boundary region of type 1. Similarly, the type calculator 32 determines that ◇ is type 2 because the distribution of ◇ is in the boundary region of type 2.

  Note that type learning and type discrimination by these methods are basically methods for determining a boundary line when there are two types, so when performing type discrimination for a large number of types, for example, a tournament There is a method called a method. In the tournament method, type learning is performed with combinations of all two types, and each boundary line (discrimination plane) is determined. Next, the type discrimination of which type the discrimination target data is close to is performed with respect to a combination of a plurality of types by the remaining winning method, and the finally winning type is set as the target type.

  According to the fifth embodiment, the type determination process of the observation time-series image sequence is performed based on the type learning result of the dictionary time-series image sequence, not the matching between the dictionary time-series image sequence and the observation time-series image sequence. Can do. As a result, it is possible to perform type determination with less dependence on parameters at the time of dictionary generation, such as a dictionary image sequence generation method and dictionary area division width.

  Furthermore, by using such a type determination process, for example, even when the dictionary area division width does not match the motion of the observation target, it is possible to avoid the problem that a similar image sequence is not necessarily found, The type of target of unknown type can be determined.

Embodiment 6 FIG.
Hereinafter, a target identification device according to Embodiment 6 of the present invention will be described with reference to FIGS. FIG. 9 is a configuration diagram of the type discriminator 3 in Embodiment 6 of the present invention. The target identification device 10 according to the sixth embodiment has the configuration shown in FIG. 1, and the type discriminator 3 has the configuration shown in FIG. 9. As shown in FIG. 9, the type discriminator 3 includes a similarity calculation parameter determiner 33, a similarity upper dictionary image extractor 34, an image sequence generator 35, an image sequence evaluator 36, and a type determiner 37. The

  Next, the operation of the sixth embodiment will be described. In the first embodiment, the type discriminator 3 performs the type discrimination by matching the entire dictionary time-series image sequence and the entire observed time-series image sequence. However, in the sixth embodiment, the type discriminator 3 FIG. 3 shows an example in which the type determination of the entire image sequence is performed after matching processing is performed for each individual image constituting the observation time-series image sequence.

  The similarity calculation parameter determiner 33 receives the observation time-series image sequence data from the observation time-series image sequence file 15, and determines and outputs a similarity calculation parameter that is a parameter for calculating the similarity. Similarity calculation parameters include a similar image search target image number and a search area. Specifically, the image numbers of all images in the observation time series image sequence are designated as the similar image search target image numbers. The image sequence length of the observation time series image sequence is given from the observation time series image sequence data. As the search area, the entire divided dictionary area is designated.

  The similarity higher-order dictionary image extractor 34 includes a similarity calculation parameter output from the similarity calculation parameter determiner 33, observation time-series image sequence data from the observation time-series image sequence file 15, and dictionary time-series image sequence file. The dictionary time-series image sequence data from 14 is input, a similarity high-order dictionary image that is a dictionary image with high similarity is extracted, and is output as similarity high-order dictionary image information including position information and similarity information.

  Specifically, the similarity between all observation images constituting the observation time-series image sequence designated as the similar image search target image number and all dictionary images is calculated, and the dictionary image with the highest similarity for each observation image Are extracted as similarity high-order dictionary images, and position information on those dictionary regions and similarity information indicating the degree of similarity are output. The number of similar high-order dictionary images for each observation image may be fixed, or dynamically specified at the time of execution.

  The image sequence generator 35 receives the similarity high-order dictionary image information output from the similarity high-order dictionary image extractor 34 and generates a similarity high-level dictionary image sequence that is an image sequence composed of combinations of similarity high-level dictionary images. Then, it is output as similarity high-order dictionary image string information including position information and similarity information. Specifically, an image sequence equal to the image sequence length of the observation time-series image sequence is generated from the all similarity high-order dictionary images according to the sequence of corresponding observation images.

For example, when the images with the highest similarity L corresponding to the respective observation images are extracted and the image sequence length is K, L ^K similarity higher-order dictionary image sequences are generated. As the position information, the position number string of the similarity higher rank dictionary image constituting the image string is output, and as the similarity information, the similarity string of the similarity higher rank dictionary image constituting the image string is output.

  FIG. 10 is a view showing an example of generation of a similarity high-order image sequence according to Embodiment 6 of the present invention. FIG. 10 shows each similarity upper dictionary image extracted for a certain candidate target as a square when the similarity upper dictionary image extraction number L = 3 and the observed image sequence length K = 3. Further, a line formed by connecting three of these squares corresponds to the generated similarity high-order image sequence.

  Among the squares in FIG. 10, the squares shaded with diagonal lines indicate the similarity high-order dictionary image for the observation image with the image number 1. That is, the hatched square is obtained by extracting three dictionary images having high similarity from one candidate target dictionary image with respect to the first observation image in the observation time-series image sequence. Equivalent to. Similarly, a black square indicates a similarity high-order dictionary image with respect to an observation image with an image number 2, and a white square indicates a similarity high-level dictionary image with respect to an observation image with an image number 3.

Then, the image sequence generator 35 can generate the similarity high-order dictionary image sequence by connecting the three squares of the shaded square, the black painted square, and the white painted square in order. In FIG. 10, of the calculable similarity upper dictionary image sequence is exemplified only 4 cases, actually calculates the L ^K-number of similarities upper dictionary image sequence corresponding to all combinations Output.

  The image sequence evaluator 36 receives the similarity high-order dictionary image sequence information output from the image sequence generator 35 and the dictionary time-series image sequence data from the dictionary time-series image sequence file 14 as input. Evaluation is performed, and a candidate target evaluation value that is an evaluation result of the candidate target is determined and output. As a method of evaluating the similarity higher-order dictionary image sequence, the similarity higher-level dictionary image sequence is evaluated from the similarity evaluation value calculated from the similarity of each similarity higher-level dictionary image sequence or the position evaluation value based on the position information. A method is conceivable.

  As the position evaluation value, a distance evaluation value and an angle evaluation value can be considered. The image sequence evaluator 36 evaluates the positional relationship of the similarity high-order dictionary images constituting the similarity high-level dictionary image sequence from such position evaluation values. Specifically, for example, the distance between the dictionary images and the degree of deviation from a straight line connecting adjacent images are calculated. A distance evaluation value is obtained from the dispersion of distances between successive dictionary images, and an angle evaluation value is obtained from an angle formed by two straight lines (corresponding to line segment s1 and line segment s2 in FIG. 10) connecting three adjacent images. The evaluation value is set higher as the difference in distance between images is smaller and the angle is closer to zero.

  As the similarity evaluation value, for example, a variance of similarity between the similarity higher-level dictionary images constituting each similarity higher-level dictionary image sequence, and a similarity total value are obtained. The evaluation value is set higher as the variance of the similarity is smaller and the total value is higher.

  As an evaluation method for each similarity higher-level dictionary image sequence, a method of evaluating only from the position evaluation value, for example, among the similarity evaluation value and the position evaluation value is conceivable. Specifically, for example, with respect to the distance evaluation value and the angle evaluation value given as the position evaluation value, the similarity high-order dictionary image sequence is arranged in descending order of the angle evaluation value, and the distance in order from the top of the angle evaluation value The evaluation value is compared with a threshold value given in advance, and a similarity high-order dictionary image sequence having a distance evaluation value equal to or higher than the threshold value and the highest angle evaluation value is extracted.

  Alternatively, as a method of reflecting the similarity evaluation value in the evaluation, the highest similarity similarity dictionary image sequence having the highest similarity evaluation value is extracted from the top M angle evaluation values having the distance evaluation value equal to or greater than the threshold value. That's fine. In addition, it is also possible to obtain a total evaluation value that reflects all evaluation values, such as by setting a function having these multiple evaluation values as variables, and to extract a similarity high-order dictionary image having the highest total evaluation value.

  An example in which the evaluation of the similarity higher rank dictionary image sequence is performed based on the position evaluation value will be described with reference to FIG. Among the four similarity high-order dictionary image sequences shown in FIG. 10, the similarity is determined as the similarity high-level dictionary image sequence having the highest angle evaluation value, that is, the adjacent images are linearly positioned and the distance evaluation value is high. A similarity high-order dictionary image sequence composed of the high-order images G1, G2, and G3 is extracted. Regardless of which evaluation method is applied, the candidate target evaluation value output by the image sequence evaluator 36 is either a similarity evaluation value or a vector composed of the similarity evaluation value and the angle evaluation value.

  Note that while the observed image sequence is an image sequence arranged in a certain direction, the similarity higher-level dictionary image sequence does not necessarily have to be in a certain direction. That is, it can be turned back halfway. For example, in FIG. 10, a case is considered where the first image to the fifth image in the observed image sequence correspond to {G2, G3, G2, G1, G2} of the dictionary image.

  In this case, it is considered that folding is generated between G3 and G1, and it is possible to cope with the target vibration motion. It should be noted that a method of narrowing down a dictionary image sequence representing an unimaginable motion can be considered by incorporating a variation state of a distance between consecutive images into an evaluation index so that it can be excluded.

  The type determiner 37 receives the candidate target evaluation value for each candidate target output from the image sequence evaluator 36, and determines and outputs the type determination result of the observation target. Specifically, the type determiner 37 compares the evaluation values of the candidate targets, and outputs the candidate target name having the highest evaluation value as the type discrimination result, or selects the candidate target names having higher evaluation values in descending order of the evaluation values. And output with the ranking.

  According to the sixth embodiment, the similarity between individual observation images and dictionary images is calculated, and a dictionary image sequence is generated from images having a high similarity, so that all the assumed ones shown in the first embodiment can be obtained. Compared with the case of calculating the similarity with the dictionary image sequence, the amount of calculation at the time of calculating the similarity can be reduced.

  Furthermore, without generating the dictionary image sequence between the adjacent images in the dictionary region, after freely generating the dictionary image sequence based on the similarity of the dictionary image to each observation image constituting the observation time-series image sequence, Narrow down the distance and positional relationship between dictionary images as an index. Therefore, it is possible to generate a dictionary in accordance with changes in observation time and fluctuations in target rotational movement without being directly affected by the dictionary area division width, and class identification according to target observation conditions should be performed. Can do.

  Furthermore, by allowing a difference in distance between images, it is possible to deal with a change in acceleration of a target rotational motion. Furthermore, by evaluating the positional relationship between the dictionary images using only the angle evaluation value and the distance evaluation value, it is possible to fold back the upper similarity dictionary image sequence, and it is also possible to deal with the target vibration motion.

Embodiment 7 FIG.
Hereinafter, a target identification device according to Embodiment 7 of the present invention will be described with reference to FIGS. FIG. 11 is a configuration diagram of the type discriminator 3 in Embodiment 7 of the present invention. The target identification device 10 according to the seventh embodiment has the configuration shown in FIG. 1, and the type discriminator 3 has the configuration shown in FIG. 11. As shown in FIG. 11, the type discriminator 3 includes a neighborhood region determiner 38, a similarity higher-order dictionary image extractor 34, an image sequence evaluator 36, and a type determiner 37.

  Compared with FIG. 9 showing the configuration of the type discriminator 3 in the sixth embodiment, a neighborhood region determiner 38 is used instead of the similarity calculation parameter determiner 33, and the image sequence generator 35 is a constituent element. It has been deleted from.

  Next, the operation of the seventh embodiment will be described. In the sixth embodiment, the type discriminator 3 performs matching processing for each individual image constituting the observation time-series image sequence, and then performs type discrimination based on the entire dictionary image sequence that is a combination of the results. Indicated. On the other hand, in the seventh embodiment, the type discriminator 3 selects a plurality of candidates by performing matching processing on the images constituting the observation time-series image sequence in the order constituting the image sequence. An example is shown in which a branched image sequence is generated by connecting a plurality of candidates, and the type of the entire generated image sequence is determined.

  The neighborhood region determiner 38 receives the observation time-series image sequence data from the observation time-series image sequence file 15 and the similarity higher-level dictionary image information output from the similarity higher-level dictionary image extractor 34 as inputs, and calculates the similarity. The process of determining and outputting parameters is repeated, and the similarity higher-level dictionary image information repeatedly output from the similarity higher-level dictionary image extractor 34 is connected to generate and output similarity higher-level dictionary image string information. From the observation time-series image sequence data, the observation time-series image sequence length is given.

  Hereinafter, iterative processing procedures by the neighborhood region determiner 38 and the similarity higher-order dictionary image extractor 34 will be described. First, the neighborhood region determiner 38 first sets all dictionary regions as neighborhood regions for the first image in the observation time-series image sequence, and uses the first image as the similar image search target image number that is a similarity calculation parameter. As a search area, a neighborhood area, that is, the entire dictionary area is set and output. At this stage, there is no input of similarity higher-order dictionary image information, or null information is input.

  Next, the similarity high-order dictionary image extractor 34 calculates the similarity calculation parameter output from the neighborhood region determiner 38, the observation time-series image sequence data from the observation time-series image sequence file 15, and the dictionary time-series image sequence. The dictionary time-series image sequence data from the file 14 is input, the similarity region is extracted by using the neighborhood region specified by the similarity calculation parameter as the search region, and the similarity including position information and similarity information is included. Output as high-order dictionary image information. The number of extracted similarity higher-order dictionary images may be fixed, or may be dynamically specified at the time of execution.

  Next, the neighborhood region determiner 38 receives again the observation time-series image sequence data from the observation time-series image sequence file 15 and the similarity higher-level dictionary image information output from the similarity higher-level dictionary image extractor. Then, from the position information of the entire similarity high-order dictionary image with respect to the first observation image, which is input as the similarity high-order dictionary image information, the neighborhood region of each similarity high-order dictionary image is determined, and the second image of the observation time-series image sequence Is set as a search area for the higher-order similarity image, and a number indicating the second image is output as the search area and similarity search image number.

  Next, again, the similarity higher-order dictionary image extractor 34 receives the similarity calculation parameter output from the neighborhood region determiner 38 and the dictionary time-series image sequence data as input, and performs a search specified by the similarity calculation parameter. In the area, a higher similarity dictionary image is extracted and output as similarity higher dictionary image information including position information and similarity information.

Thereafter, the above processing is repeated between the neighborhood region determiner 38 and the similarity higher-order dictionary image extractor 34, and when the similarity higher-order images for the image sequences constituting all the observed image sequences are determined, the neighborhood region determiner 38 finally generates and outputs a similarity high-order dictionary image sequence. For example, when the images with the highest similarity L corresponding to the respective observation images are extracted and the image sequence length is K, L ^K similarity higher-order dictionary image sequences are generated.

  FIG. 12 is a view showing an example of a set search area and a generated similarity higher-order dictionary image string according to Embodiment 7 of the present invention. FIG. 12 shows each similarity upper dictionary image extracted for one candidate target as a square when the similarity upper dictionary image extraction number L = 3 and the observed image sequence length K = 3. Further, a line formed by connecting three of these squares corresponds to the generated similarity high-order image sequence.

  Among the squares in FIG. 12, the hatched squares indicate the similarity higher-order dictionary image with respect to the observation image with the image number 1. That is, the hatched square is obtained by extracting three dictionary images having high similarity from one candidate target dictionary image with respect to the first observation image in the observation time-series image sequence. Equivalent to. Similarly, a black square indicates a similarity high-order dictionary image with respect to an observation image with an image number 2, and a white square indicates a similarity high-level dictionary image with respect to an observation image with an image number 3.

  The distance from the center point of the similarity higher-order dictionary image (square with hatching) with respect to the observation image with the image number 1 is within a preset expansion / contraction coefficient that is a parameter that defines the size of the search area. A region (corresponding to a circle centered on a square painted with diagonal lines in FIG. 12) is set as a search region for the observation image of image number 2.

  The search area for the observation image with the image number 3 is, first, the distance from the center point of the similarity higher-order dictionary image (solid square) with respect to the observation image with the image number 2 as in the search area for the observation image with the image number 2. Is set to be within an expansion / contraction coefficient, and then the angle formed by the line segment connecting the images is further limited to be closer to a straight line (corresponding to the sector area centered on the black square in FIG. 12). ).

  Specifically, a search is performed in which the angle between the line connecting the hatched square and the black square and the line connecting the black square and the white square is set in advance. A region that is within an angle coefficient that is a parameter that defines the angle of the region is set as a search region for the observation image of image number 3. Similarly, the neighborhood region determiner 38 sets subsequent search regions, and the similarity higher-level dictionary image extractor 34 repeats the process of extracting the similarity higher-level dictionary image from the search region.

  Then, the neighborhood region determiner 38 determines a final similarity higher-order dictionary image sequence from the repeatedly extracted similarity higher-order dictionary images. Here, the expansion coefficient and the angle coefficient have been shown as fixed values set in advance, but may be specified at the time of execution.

  The search area after the observation image with the image number 3 is not necessarily set in one direction on a straight line, but can be set in both directions. FIG. 13 is an explanatory diagram of the search area after the observation image with the image number 3 in the seventh embodiment of the present invention. If the black square in FIG. 13 is the upper similarity dictionary area for the observation image with image number 2, the search area for the observation image with image number 3 is set like the white square in FIG.

  Thereby, the search area of the observation image of image number 3 can also include a square area painted with diagonal lines. As a result, folding of the image sequence with higher similarity is allowed, and it is possible to cope with the target vibration motion.

  The image sequence evaluator 36 receives the similarity upper dictionary image sequence information output from the neighborhood region determiner 38, evaluates the similarity upper dictionary image sequence, and determines and outputs a candidate target evaluation value. The similarity higher-order dictionary image sequence is evaluated based on the similarity evaluation value calculated from the similarity of each similarity higher-level dictionary image sequence. As the similarity evaluation value, as in the sixth embodiment, for example, the variance of similarity between the similarity higher-level dictionary images constituting each similarity higher-level dictionary image sequence, and the similarity total value are obtained. The evaluation value is set higher as the variance of the similarity is smaller and the total value is higher. The highest value among the similarity evaluation values of the similarity high-order dictionary image sequence is output as the candidate target evaluation value of the candidate target.

  The type determiner 37 receives the candidate target evaluation value for each candidate target output from the image sequence evaluator 36, and determines and outputs the type determination result of the observation target. Specifically, the type determiner 37 compares the evaluation values of the candidate targets, and outputs the candidate target name having the highest evaluation value as the type discrimination result, or selects the candidate target names having higher evaluation values in descending order of the evaluation values. And output with the ranking.

  According to the seventh embodiment, for each observed image, after calculating the neighborhood area of the corresponding higher-order dictionary image according to the order of the image sequence, the similarity with only the dictionary image in the neighborhood area is calculated. As a result, the amount of calculation at the time of calculating the similarity can be further reduced as compared to the case of calculating the similarity between all the observed images and all the dictionary images in the dictionary area shown in the sixth embodiment.

  Further, the dictionary image sequence is not generated between adjacent images in the dictionary area, but is extracted from the dictionary images in the neighboring area based on the similarity. Therefore, it is possible to generate a dictionary in accordance with changes in observation time and fluctuations in target rotational movement without being directly affected by the dictionary area division width, and class identification according to target observation conditions should be performed. Can do.

  Furthermore, by extracting within the vicinity region, a difference in distance between images is allowed, and it is possible to cope with a change in acceleration of a target rotational motion. Furthermore, by setting the neighborhood region in both directions instead of one direction, the similarity higher-order dictionary image sequence can be folded, and the target vibration motion can be dealt with.

  In the above-described embodiment, some combinations of the dictionary time-series image sequence generator 2 and the type discriminator 3 have been described. However, the present invention is not limited to this. For the target identification device 10 shown in FIG. 1 or FIG. 5, the dictionary time-series image sequence generator 2 shown in FIG. 2 or 6 and the type discriminator shown in FIG. 7, FIG. 9 or FIG. Any combination with 3 can be applied.

It is a block diagram of the target identification apparatus by Embodiment 1 of this invention. It is a block diagram of the dictionary time series image sequence generator in Embodiment 2 of this invention. It is the figure which showed the example of a dictionary area | region division by the dictionary area | region divider in Embodiment 2 of this invention. It is the figure which showed the example of a dictionary area | region division by the dictionary area | region divider in Embodiment 2 of this invention. It is a block diagram of the target identification apparatus in Embodiment 3 of this invention. It is a block diagram of the target identification apparatus in Embodiment 4 of this invention. It is a block diagram of the type discriminator in Embodiment 5 of this invention. It is explanatory drawing of the kind learning by the kind learning device in Embodiment 5 of this invention, and the kind discrimination | determination by a kind calculator. It is a block diagram of the type discriminator in Embodiment 6 of this invention. It is an illustration figure of similarity high level image sequence generation in Embodiment 6 of the present invention. It is a block diagram of the type discriminator in Embodiment 7 of this invention. It is an illustration figure of the set search area | region and the produced | generated similarity high-order dictionary image sequence in Embodiment 7 of this invention. It is explanatory drawing of the search area | region after the observation image of the image number 3 in Embodiment 7 of this invention.

Explanation of symbols

  1 target parameter estimator, 2 dictionary time-series image sequence generator, 3 type discriminator, 4 observation image sequence processor, 10 target identification device, 11 target specification file, 12 tracking file, 13 radar observation specification file, 14 dictionary time-series image sequence file, 15 observation time-series image sequence file, 16 type discrimination result file, 17 simulation image file, 21 dictionary area division width determiner, 21a multiple dictionary area division width determiner, 22 dictionary area divider, 23 simulation image generator, 24 divided region dictionary image determiner, 25 dictionary time-series image sequence determiner, 26 dictionary region divided width selector, 31 type learner, 32 type calculator, 33 similarity calculation parameter determiner, 34 Similarity upper dictionary image extractor, 35 image sequence generator, 36 image sequence evaluator, 37 type determiner, 3 8 Neighborhood determiner.

Claims (9)

Target data related to the motion or attitude of the target based on the tracking data, which is the position / velocity information of the target, and the radar observation data, including the tracking data of the candidate target observed by the radar and the tracking data generated by the simulation A target parameter estimator for estimating parameter estimates;
Dictionary time-series image sequence data for each candidate target based on the target-specific data that is the three-dimensional shape data of the candidate target, the radar observation data, and the target parameter estimation value from the target parameter estimator A dictionary time-series image sequence generator for generating
Based on the observed time-series image sequence data obtained by observing an unknown target with a radar and the dictionary time-series image sequence data from the dictionary time-series image sequence generator, the observed time-series image sequence data A type discriminator comprising: a type discriminator for discriminating the type of a target of unknown type included in. The target identification device according to claim 1,
Based on the radar observation data and the observation time-series image sequence data, the length of the observation time-series image sequence data is expanded by inserting an image with all pixel intensities being zero in the image sequence. Further comprising an observed image sequence processor for generating processed data and calculating a processed image sequence length that is the length of the observed image sequence processed data;
The dictionary time-series image sequence generator is based on the target specification data, the radar observation specification data, the target parameter estimation value, and the processed image sequence length from the observation image sequence processor. , Generate dictionary time-series image sequence data for each candidate target,
The type discriminator is included in the observed image sequence processed data based on the observed image sequence processed data from the observed image sequence processor and the dictionary time-series image sequence data from the dictionary time-series image sequence generator. A target discriminating apparatus characterized by discriminating the type of a target whose type is unknown. The target identification device according to claim 1,
Based on the radar observation data and the observation time series image sequence data, the length of the observation time series image sequence data is reduced by deleting the image at the specified position from the image sequence , and the observed image sequence processing data And an observation image sequence processor for calculating a processed image sequence length that is the length of the observed image sequence processing data,
The dictionary time-series image sequence generator is based on the target specification data, the radar observation specification data, the target parameter estimation value, and the processed image sequence length from the observation image sequence processor. , Generate dictionary time-series image sequence data for each candidate target,
The type discriminator is included in the observed image sequence processed data based on the observed image sequence processed data from the observed image sequence processor and the dictionary time-series image sequence data from the dictionary time-series image sequence generator. A target discriminating apparatus characterized by discriminating the type of a target whose type is unknown. The target identification device according to any one of claims 1 to 3,
The dictionary time series image sequence generator
A dictionary area division width determiner for determining a dictionary area division width based on the target-specific data;
Based on the dictionary region division width from the dictionary region division width determiner, the dictionary region is divided, and a dictionary region division unit that calculates a dictionary region division result;
A simulation image generator for generating simulation image data, which is image data for each candidate target observed from each divided area, based on the target-specific data and the dictionary area division result from the dictionary area divider; ,
A divided region dictionary image determiner that determines a divided region dictionary image that is a dictionary image of each divided region based on the simulation image data and the target parameter estimation value from the target parameter estimator;
Based on the radar observation data and the divided region dictionary image from the divided region dictionary image determiner, the dictionary time series image sequence data for each candidate target consisting of a plurality of divided region dictionary images is determined. A target identification apparatus comprising: a sequence image sequence determiner. The target identification device according to any one of claims 1 to 3,
The dictionary time series image sequence generator
A dictionary area division width determiner for determining a dictionary area division width as a plurality of dictionary area division width candidate values based on the target-specific data;
Based on the plurality of dictionary area division width candidate values from the dictionary area division width determiner, the dictionary area division is performed to obtain a dictionary area division result group by dividing the dictionary area;
Simulation image data which is image data for each candidate target observed from the divided areas constituting the dictionary area division result group based on the specification data by target and the dictionary area division result group from the dictionary area divider A simulation image generator for generating a simulation image data group,
A dictionary that selects a dictionary area division width from the dictionary area division width candidate value from the dictionary area division width determiner based on the radar observation data and the target parameter estimation value from the target parameter estimator An area division width selector;
Based on the simulation image data group, the target parameter estimation value from the target parameter estimator, and the dictionary area division width from the dictionary area division width selector, a simulation image is selected from the simulation image data group. A divided region dictionary image determiner that extracts data and determines a divided region dictionary image that is a dictionary image of each divided region;
Based on the radar observation data and the divided region dictionary image from the divided region dictionary image determiner, the dictionary time series image sequence data for each candidate target consisting of a plurality of divided region dictionary images is determined. A target identification apparatus comprising: a sequence image sequence determiner. In the target identification device according to claim 4 or 5,
The dictionary area division width determiner changes the dictionary area division width according to a position. In the target identification device according to any one of claims 1 to 6,
The type discriminator is:
A type learner for obtaining a type learning result based on the dictionary time-series image sequence data;
Based on the type learning result from the type learner and the observation time series image sequence data from the observation time series image sequence file, the type determination of the type unknown target included in the observation time series image sequence data A target discriminating apparatus comprising: a type calculator for performing In the target identification device according to any one of claims 1 to 6,
The type discriminator is:
A degree-of-similarity calculation parameter determiner for determining a degree-of-similarity calculation parameter that specifies each observation image and similar image search area constituting the observation time-series image sequence data;
The similarity calculation parameter from the similarity calculation parameter determiner, the observation time series image sequence data from the observation time series image sequence file, and the dictionary time series image sequence data from the dictionary time series image sequence file Based on the above, a similarity high-order dictionary image including position information and similarity information is extracted for each candidate target for each observation image constituting the observation time-series image sequence data. A similarity high-order dictionary image extractor that generates information;
Based on the similarity high-order dictionary image information, a plurality of similarity high-order dictionary image sequences that are image sequences equal to the image sequence length of the observation time-series image sequence data are generated for each candidate target, and the position information and the similarity An image sequence generator for generating similarity high-order dictionary image sequence information including degree information,
Based on the plurality of similarity higher-order dictionary image sequence information and the dictionary time-series image sequence data for each candidate target from the image sequence generator, the plurality of similarity upper-level dictionary image sequences are evaluated, and each candidate target is evaluated. An image sequence evaluator that determines candidate target evaluation values for
And a type determiner for determining the type of the unknown target based on the candidate target evaluation value from the image sequence evaluator. In the target identification device according to any one of claims 1 to 6,
The type discriminator is:
A similarity calculation parameter in which an observation image and a similar image search area are specified, the observation time series image sequence data from the observation time series image sequence file, and the dictionary time series image sequence from the dictionary time series image sequence file Based on the data, a high-similarity dictionary including position information and similarity information is extracted for each candidate target based on the observation time-series image sequence data, and a dictionary image with high similarity is extracted for each candidate target. A similarity high-order dictionary image extractor for generating image information;
A similar image search area is specified based on the similarity high-order dictionary image information from the similarity high-order dictionary image extractor, and each observation image and the similar image search area constituting the observation time-series image sequence data are included. A process for determining the similarity calculation parameter and obtaining the similarity higher dictionary image information as a response to the determined similarity calculation parameter output to the similarity higher dictionary image extractor is performed as the observation time-series image string data. A neighborhood region determiner that repeats according to the image sequence length that constitutes and generates similarity higher-order dictionary image sequence information according to the image sequence length for each candidate target;
Based on the plurality of similarity high-order dictionary image string information for each candidate target from the neighborhood region determiner, the plurality of similarity high-level dictionary image strings are evaluated, and a candidate target evaluation value for each candidate target is determined. An image sequence evaluator;
And a type determiner for determining the type of the unknown target based on the candidate target evaluation value from the image sequence evaluator.

Images