JP2019086464A - Radar device and radar signal processing method thereof - Google Patents

Abstract

To heighten the capacity of detecting a small target having a small RCS and a low reception SN.SOLUTION: A radar device of an embodiment of the present invention converts a received signal in a pulse repetition period in an N (N≥1) hit transmitted pulse to a Doppler axis frequency domain and generates range-Doppler data, divides the range-Doppler data by a range-Doppler axis into a plurality of blocks, generates an amplitude intensity histogram with regard to each of the plurality of blocks, discriminates the presence of a target for each of the divided blocks by a multilayer neural network from the histogram of each of the plurality of blocks and extracts a target detection candidate block from the result of discrimination, converts range-Doppler data in block range of the target detection candidate block to a one-dimensional vector, specifies a target detection cell on the range-Doppler axis from the one-dimensional vector by a multilayer neural network, and calculates at least one of a range and a speed from the target detection cell.SELECTED DRAWING: Figure 1

Description

  The present embodiment relates to a radar device and a radar signal processing method thereof.

  In conventional radar devices, CFAR (Constant False Alarm Rate, Non-Patent Document 2) processing is often used after pulse compression (Non-Patent Document 1) and FFT (Fast Fourier Transform) in general. However, in the CFAR processing, when detecting a small target having a small RCS (Radar Cross Section) and a low received SN, the received amplitude may not be detected because it does not exceed the CFAR threshold.

  Although there is a method (patent document 1) which detects a target from an amplitude histogram of a block unit of a range-Doppler axis as a measure against this, in order to extract a feature representing a histogram shape etc. from an amplitude histogram shape further, If the extraction method did not fit the data, it could become undetectable.

JP, 2015-176665, A

Ouchi, "Basics of synthetic aperture radar for remote sensing", Tokyo Denki University Press, pp. 131-149 (2003) CFAR (Constant False Alarm Rate) processing, Yoshida, 'Revision radar technology', The Institute of Electronics, Information and Communication Engineers, pp. 87-89 (1996) DBSCAN (Density-based Spatial Clustering of Applications with Noise), Sebastian Raschka, Python, 'machine learning programming', Impress, pp. 319-323 (2016) Neural network, Okaya, Yuka deep learning ', Kodansha, pp.7-26 (2014) Activation function, Okaya, Yuka deep learning ', Kodansha, pp. 10-12 (2014) Loss function, Okaya, Yuka deep learning ', Kodansha, pp. 14-16 (2014) Convolutional NN, Okaya, Yuka Deep Learning ', Kodansha, pp. 79-91 (2014) Convolutional NN (up sampling), Fujita, Sakai implementation deep learning ', Ohmsha, pp. 40-45 (2016)

  As described above, in the conventional radar apparatus that performs CFAR processing after performing pulse compression and FFT processing, when detecting a small target having a small RCS (Radar Cross Section) and low reception SN, the reception amplitude is CFAR. There was a problem that did not go beyond the threshold of

  This embodiment is made in view of the above-mentioned subject, and an object of the present invention is to provide a radar installation which can raise detection capability of a small target with small RCS and low receiving SN, and its radar signal processing method.

  In order to solve the above problems, according to the present embodiment, in the coherent radar apparatus that transmits pulses of N (N ≧ 1) hits, FFT processing is performed on the PRI axis, and pulses are generated as necessary. A block is divided into (Nh, Mh) (Nh 1, 1, Mh 1 1) on the range-doppler axis using compressed range-doppler data (RD data), a histogram of amplitude intensity is created, and a one-dimensional vector of the histogram Is input to the multilayer NN for block detection to extract a target detection candidate block, and RD data of the block range is converted to a one-dimensional vector and input to the multilayer NN for cell specification to detect the range-Doppler axis Identify the cell

  That is, in the radar apparatus according to the present embodiment, a feature quantity such as an amplitude histogram is extracted from the block-divided signal using the transmission / reception signal of the Doppler range axis to extract a target presence / absence block by multilayer NN. For the data of the extracted block, it is possible to specify the cell of the range-doppler axis in the block by the multi-layer NN and output the distance and the speed.

FIG. 1 is a block diagram showing a schematic configuration of a radar device according to a first embodiment. The figure for demonstrating block utilization of range-Doppler data, vector preparation of an amplitude histogram sequence, and utilization of a multilayer neural network in 1st Embodiment. FIG. 4 is a diagram showing a configuration in which units are combined using weighting and a bias coefficient in each layer of the multi-layered neural network in the first embodiment. The figure for demonstrating the method to specify the detection cell of a range-Doppler axis | shaft in the detected target block in 1st Embodiment. The block diagram which shows schematic structure of the radar apparatus which concerns on 2nd Embodiment. The figure which shows a mode that the pseudo | simulation high-resolution-ized range-Doppler data is obtained by a range-Doppler axis | shaft in 2nd Embodiment. The block diagram which shows schematic structure of the radar apparatus which concerns on 3rd Embodiment. The block diagram which shows the flow of a process of a convolutional neural network in 3rd Embodiment. The figure which shows a mode that the distance and speed of a target are acquired by a convolution neural network in 3rd Embodiment. The block diagram which shows schematic structure of the radar apparatus which concerns on 4th Embodiment. FIG. 14 is a conceptual diagram for explaining processing of the DBSCAN method applied to the fourth embodiment. FIG. 14 is a conceptual diagram showing how a target and misdetection are discriminated by cluster analysis in the fourth embodiment.

  Hereinafter, embodiments will be described with reference to the drawings. In the description of each embodiment, the same parts will be denoted by the same reference numerals, and overlapping descriptions will be omitted.

First Embodiment
The radar apparatus according to the first embodiment will be described with reference to FIGS. 1 to 4.

  FIG. 1 is a block diagram showing a schematic configuration of the radar device according to the first embodiment. This radar device is roughly divided into a transmission system, a reception system, and a signal processing system.

  In the transmission system, a transmission signal such as a chirp signal generated by the transmission signal generation unit 1 is converted into a high frequency signal by the frequency conversion unit 2 and pulse modulation is performed by the pulse modulation unit 3 to transmit / receive separation unit 4 such as a circulator. It transmits as a radar wave from the antenna part 5 via.

  In the reception system, the frequency conversion unit 11 converts the reflected signal of the radar wave received by the antenna unit 5 through the transmission / reception separation unit 4 into an intermediate frequency, and the AD conversion unit 12 converts it into a digital signal.

  In the signal processing system, the received signal converted into a digital signal is subjected to pulse compression (Non-Patent Document 1) in the pulse compression unit 13 as necessary, and in FFT (Fast Fourier Transform) processing unit 14 PRI (Pulse Repetition Interval) FFT processing of the Doppler axis (slow-time axis) between the two is performed to obtain range-doppler (RD) data. Then, the RD data is divided into predetermined blocks by the block division unit 16, and an amplitude histogram (specifically, a vector of amplitude histogram sequence) is generated by the histogram generation unit 17 for each division unit, and a multilayer NN (Neural Network, Non-Patent Document 4) The block extraction unit 18 identifies the presence or absence of a target for each divided block by multilayer NN, and extracts a block of a target detection candidate. Subsequently, the block vectorization processing unit 19 converts the extracted RD data of the block range of the target detection candidate into a one-dimensional vector, and the multilayer NN cell identification unit 20 multilayers the detection cells of the range-Doppler axis from the one-dimensional vector. The range and speed are detected from the detection cell by the range / speed detection unit 22.

  In the above configuration, radar signal processing of the present embodiment will be described with reference to FIGS. 2 to 4. Fig. 2 is a diagram for describing RD data divided into blocks, histogram creation, and utilization of multilayer NN, and Fig. 3 is a diagram showing a configuration in which units are combined using weighting and a bias coefficient in each layer of multilayer NN. FIG. 4 is a diagram for describing a method of specifying a detection cell of the range-Doppler axis in a target block detected by the multilayer NN.

  If the digital reception signal obtained by the above reception system is subjected to pulse compression as needed by the pulse compression unit 13, and FFT processing on the slow-time axis between PRI is performed by the FFT 14 to obtain RD data The RD data is input to the RD data block division unit 16.

  As shown in FIG. 2A, the RD data block division unit 16 divides RD data RD (r, fd) (r: range, fd: Doppler frequency) into regions (1 to Nh, 1 to Mh) (Nh). Divide into blocks of 1,1, Mh ≧ 1). Although FIG. 2A shows the case where the blocks do not overlap, the blocks may be slid and overlapped as a countermeasure in the case where there is a target near the boundary between the blocks. . The RD data of each divided block is input to the histogram generation unit 17. The histogram generation unit 17 generates an amplitude histogram in units of divided blocks as shown in FIG. 2 (b). The amplitude histogram created here is input to the multilayer NN block extraction unit 18 arranged in multiple layers as shown in FIG. 2 (c).

  In the multilayer NN block extraction unit 18, each unit is combined using weighting and a bias coefficient, as shown in FIG. Expressed by a mathematical expression, the following equation is obtained.

  An activation function (non-patent document 5) which is a non-linear function is applied to the output Zm (m = 1 to M) to obtain an output ym (m = 1 to M).

  As described above, the RD data RD (r, fd) is divided into blocks of the regions (1 to Nh, 1 to Mh), and the amplitude histogram generated for each block has the amplitude strength on the horizontal axis set by a predetermined fixed value. In the case of normalization, if there is a signal of a different amplitude with a larger intensity than that of the other blocks in the block, its shape changes. Therefore, in the multilayer NN block extraction unit 18, the presence or absence of the target is identified by the multilayer NN by using the change of the amplitude histogram shape. The multi-layer NN has a function of extracting feature quantities that are easy to determine the presence or absence of a target from the amplitude histogram shape, so the presence or absence of a target is more than the method of extracting a fixed feature quantity from an amplitude histogram (Patent Document 1) It can be expected that the ability to distinguish

  In order to obtain optimal parameters (weights and biases) at this time, the presence or absence of a target is learned as a teacher signal using RD data in the case of target presence and absence. That is, the amplitude histogram generated from the RD data for learning is a column vector, and the loss function (non-patent document 6) due to the difference between the final output and the non-targeted / targeted identification output is minimized. Determine the parameters. Then, using this parameter, an inference process (inference process by multilayer NN based on learned parameters) is executed on unknown input data for each block to extract a block having a target. As a result, since the range of RD data in which a target exists can be limited, the certainty of cell identification shown below can be improved.

  With reference to FIG. 4, a method of specifying a detection cell of the range-Doppler axis in the detected target block will be described.

  First, the block vectorization processing unit 19 converts the amplitude data of the target detection block extracted by the multilayer NN block extraction unit 18 into a column vector. The number of dimensions of the column vector is: all Doppler cells × all range cells. In addition, in order to specify a cell among all the cells, an output having the same number of units as the number of input units (the number of all cells) is required. In order to learn the parameters (weights and biases) in this case, using the RD data in the presence and absence of the target, create a teacher signal with 1 for the cell in which the target is present and 0 for the nonexistent cell. Learn based on the teacher signal. Therefore, the multilayer NN cell identification unit 20 inputs all cells of RD data for learning as a one-dimensional column vector, and a loss function due to the difference between the final output of the presence / absence of all cells and the learning output (Non-Patent Document 6) Determine the parameters of each layer so as to minimize Using this parameter, inference processing (inference processing by multilayer NN based on learned parameters) is executed on RD data of a block in which a target exists to specify a target cell. The range / speed detection unit 22 calculates and outputs a target range and speed based on the range (distance) of the identified cell and the Doppler frequency. The velocity can be calculated from the Doppler frequency by the following equation.

  As described above, in the radar apparatus according to the first embodiment, in the coherent method in which N (N ≧ 1) hit pulses are transmitted, FFT processing is performed on the PRI axis, and pulse compression is performed on RD as necessary. The data is used to divide the block into (Nh, Mh) (Nh 1, 1, Mh) 1) in the range-Doppler axis, generate a histogram of amplitude intensity, and generate a one-dimensional vector of the amplitude histogram into multilayer NN for block detection A target detection candidate block is extracted and RD data of the block range is converted into a one-dimensional vector and input to the multi-layer NN for cell specification to specify a detection cell of the range-Doppler axis. That is, in the above radar apparatus, the feature quantity of the amplitude histogram is extracted from the block-divided signal using the transmission / reception signal of the range-Doppler axis, and the target presence / absence block is extracted by the multilayer NN. Furthermore, for RD data of the extracted block, the distance and speed are calculated and output by specifying cells of the range-doppler axis in the block by the multilayer NN.

  Here, in the case of a radar apparatus that detects a target using CFAR with conventional pulse compression or Doppler filter, it is common to use a target amplitude intensity to raise detection when the threshold of CFAR is exceeded. The In this case, even if there are multiple reflection points for each target, it is a method to detect by comparing the point with the maximum amplitude intensity and the threshold, so detection is not possible when the amplitude intensity at the maximum point is low There was a problem called. On the other hand, in the present embodiment, the fact that there are a plurality of reflection points of each target is used, so that the detection probability when the amplitude intensity at the maximum point is low can be improved.

Second Embodiment
The first embodiment is a method using RD data as it is. In this case, when the block is subdivided, it is assumed that the number of cells decreases and the features of the amplitude histogram can not be extracted correctly. Therefore, this measure will be described in the present embodiment.

  FIG. 5 is a block diagram showing the configuration of a radar system according to the second embodiment. The radar apparatus of this embodiment has substantially the same configuration as that of the first embodiment, but the RD data subjected to signal processing such as pulse compression (13) and FFT (14) is transmitted to the pseudo high resolution processing unit 15 Input and perform pseudo-resolution on at least one of the range axis (fast-time axis) and the Doppler axis (slow-time axis). For this purpose, the pseudo high resolution processing unit 15 performs inverse Fourier transform (IFFT) shown in the following equation on RD data.

Here, although both the range axis and the Doppler axis are described as a method for increasing the resolution in a pseudo manner, in the case of the resolution enhancement in any one of the axes, it is sufficient to process only that axis.

  In the RD data RDf subjected to the two-dimensional inverse conversion as described above, zeroing is performed on the range axis and the Doppler axis, and the number of dimensions is Nz × Mz (Nz> N, Mz> M) as RDfz. . A two-dimensional FFT is performed on the RD data RDfz.

As a result, as shown in FIG. 6A, RD data including erroneous detection is subjected to pseudo-high resolution on the range-Doppler axis, and the reflection point of the foot of the peak of the amplitude due to the target is also observed near the target. It can be expected that the number of reflection points will increase as shown in FIG. 6 (b). Detecting the target by the same processing as the first embodiment using this RD data makes it easy to detect the feature value of the amplitude histogram, and the target cell is specified also in the cell specification in the block where the target exists. It can be expected to be easy to do.

  As described above, in the radar apparatus according to the second embodiment, at least one of the range axis (fast-time axis) and the Doppler axis (slow-time axis) is subjected to IFFT processing on RD data, and each axis is processed. By applying the processing of the first embodiment to RD data that has been subjected to zero-filling and pseudo-high resolution processing by FFT processing, pseudo high resolution of RD data is made to increase the number of cells, and an amplitude histogram is obtained. Since it is easy to extract feature quantities such as, etc., it is possible to extract a target candidate block and to specify a target cell.

Third Embodiment
In the second embodiment, the method of achieving high resolution of RD data by pseudo high resolution and inputting to multilayer NN to perform block extraction and cell specification has been described. In this case, since the input signal is column vectorized at the time of cell specification, the information on the cell position of the range-doppler axis is lost, and the feature of the position on the range-doppler axis is not included. In this embodiment, a method of specifying a cell using position information of the range-doppler axis will be described.

  FIG. 7 is a block diagram showing a schematic configuration of a radar system according to the third embodiment. The radar apparatus of this embodiment includes a CNN (Convolutional Neural Network, non-patent documents 7 and 8) cell identification unit 21 in place of the block vectorization processing unit 19 and the multilayer NN cell identification unit 20 shown in FIG. 1. That is, the same processing as in the second embodiment is performed until the block with the target is extracted by the multilayer NN, but after the block is extracted, the CNN cell identification unit 20 performs the convolution NN in the cell identification. Use.

  As shown in FIG. 8, the CNN includes an input layer 31, a convolution layer 32, a pooling layer 33, a convolution layer 34, a pooling layer 35, a plurality of stages of upsampling layers 36, and an all coupling layer 37. A filter that holds the position information is applied, and the pooling layer 33 outputs the maximum value and the average value of the predetermined range of cells after the convolution filter. From this, it is possible to absorb an error such as slight displacement of input data. This is repeated as necessary to make it easy to extract feature quantities. Since CNN reduces the number of cells of the input two-dimensional data, the output includes the number of convolutions and the pooling layer followed by the upsampling layer 36 (non-patented) so as to output more than the number of cells before pseudo high resolution. A multi-layer NN is constructed by adding the reference 8) and the like. The learning method is the same as that of the second embodiment.

  As described above, in the radar device according to the present embodiment, as shown in FIG. 9, the target cell is obtained by using the feature quantity including the positional relationship of the multiple reflection points of each target on the range-doppler axis by CNN. Since it is possible to specify the target, it is possible to extract the target more efficiently than when inputting to the multilayer NN by column vectorization of RD data in the first embodiment. Thus, it is possible to output the target range and speed as in the second embodiment. Also, the RD axis data is subjected to IFFT processing on the range axis (fast-time axis) and the Doppler axis (slow-time axis), zero-padded to each axis, and subjected to FFT processing to realize pseudo high resolution RD The data is divided into (Nh, Mh) (Nh 1, 1, Mh 1 1) blocks on the range-Doppler axis, a histogram of amplitude intensity is generated, and a one-dimensional vector of the histogram is made into multilayer NN for block detection. It inputs and extracts the block of a target detection candidate, RD data of the block range is input into CNN, and the detection cell of a range-Doppler axis | shaft is specified. That is, by increasing the number of cells by pseudo-resolutioning RD data, feature quantities such as amplitude histograms can be easily extracted, and target candidate blocks can be easily extracted. Furthermore, the feature quantity considering up to the reflection point arrangement of the range-Doppler axis can be used by the convolutional NN (CNN), and the target cell can be specified with high accuracy.

Fourth Embodiment
In the first to third embodiments, the method of identifying the target detection cell in the extracted block after the block extraction by the multilayer NN has been described. In this case, erroneous detection may occur depending on the degree of learning of the multi-layer NN. In this embodiment, an example of the countermeasure will be described.

  FIG. 10 is a block diagram showing a schematic configuration of a radar system according to the fourth embodiment. In this embodiment, the block detection by the multilayer NN and the cell identification by the multilayer NN (first and second embodiments) and the CNN (third embodiment) are the same as in the first to third embodiments. However, the present invention is characterized in that after the cell is specified by the multilayer NN cell specifying unit 20, discrimination is performed by the cluster analysis unit 23 using the density difference between the target and false detection in the range-Doppler axis.

  As an example of the above cluster analysis, there is a DBSCAN (Density-based Spatial Clustering of Applications with Noise, Non-Patent Document 3) method shown in FIG. This specifies a gate radius (the distance from the core point to the outer periphery of the gate) ε and the number of border point observation values in the gate to form a cluster that satisfies the condition. In the vicinity of the range-doppler where the target exists, although there are a plurality of reflection points depending on the target intensity, there are few points in the erroneous detection that appears randomly. Using this, it is possible to discriminate between a target and a false detection by analyzing clusters as shown in FIG. 12 with respect to range-Doppler axis data after cells are specified by multilayer NN. .

  As an input of this cluster analysis, it is assumed that the cell detected by the multilayer NN is 1 and the other is 0, and the range-Doppler two-dimensional all-cell data is input. The cluster analysis may be performed by a method other than the DBSAN method. Further, in the present embodiment, although the case of the radar performing pulse compression has been described, the present invention can also be applied to the case where pulse compression is not performed.

  As described above, the radar apparatus according to the present embodiment uses the detected cell data (observed value coordinates) to plot in two-dimensional coordinates (range-doppler axis), and the cluster using the two-dimensional plot coordinates as an input. By performing the analysis, the target detection can be determined by the density difference. That is, after extracting a block of a target candidate, cluster analysis is performed at the time of cell specification in the candidate block to extract a cell, so that discrimination between a target and a false detection is made by density difference even if false detection occurs. And the target cell can be identified.

  The present invention is not limited to the above embodiment as it is, and at the implementation stage, the constituent elements can be modified and embodied without departing from the scope of the invention. In addition, various inventions can be formed by appropriate combinations of a plurality of constituent elements disclosed in the above embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, components in different embodiments may be combined as appropriate.

  DESCRIPTION OF SYMBOLS 1 ... Transmission signal production | generation part, 2 ... Frequency conversion part, 3 ... Pulse modulation part, 4 ... Transmission / reception isolation part, 5 ... Antenna part, 11 ... Frequency conversion part, 12 ... AD conversion part, 13 ... Pulse compression part, 14 ... FFT (Fast Fourier Transform) processing unit, 15 ... pseudo high resolution processing unit, 16 ... RD data block division unit, 17 ... histogram generation unit, 18 ... multilayer NN block extraction unit, 19 ... block vectorization processing unit, 20 ... Multilayer NN cell identification unit, 21 ... CNN cell identification unit, 22 ... range / speed detection unit, 23 ... cluster analysis unit, 31 ... input layer, 32 ... convolution layer, 33 ... pooling layer, 34 ... convolution layer, 35 ... pooling Layers 36 up-sampling layer 37 total bonding layer.

Claims (7)

Generation means for converting the reception signal in the pulse repetition period of the transmission pulse of N (N ≧ 1) hits into the frequency domain of the Doppler axis to generate range-Doppler data;
A division unit configured to divide the range-Doppler data generated by the generation unit into a plurality of blocks along a range-Doppler axis;
Generation means for generating a histogram of amplitude intensity for each of the plurality of blocks divided by the division means;
Extraction means for identifying the presence or absence of a target for each divided block by a multilayer neural network from the histogram of each of the plurality of blocks generated by the generation means, and extracting a block of target detection candidates from the identification result;
Vectorization means for converting range-doppler data of the block range of the target detection candidate block extracted by the extraction means into a one-dimensional vector;
Specifying means for specifying a target detection cell of a range-doppler axis from the one-dimensional vector obtained by the vectorization means by a multilayer neural network;
A radar device comprising: calculating means for calculating at least one of a range and a speed from the target detection cell specified by the specifying means. Generation means for converting the reception signal in the pulse repetition period of the transmission pulse of N (N ≧ 1) hits into the frequency domain of the Doppler axis to generate range-Doppler data;
A pseudo high resolution means for performing an inverse Fourier transform on the range axis and the Doppler axis, zero-filling each axis on the range-Doppler data generated by the generation means, and performing a Fourier transform on the pseudo-resolution When,
A division unit configured to divide the range-Doppler data pseudo-high-resolutioned by the pseudo high-resolution means into a plurality of blocks along a range-Doppler axis;
Generation means for generating a histogram of amplitude intensity for each of the plurality of blocks divided by the division means;
Extraction means for identifying the presence or absence of a target for each divided block by a multilayer neural network from the histogram of each of the plurality of blocks generated by the generation means, and extracting a block of target detection candidates from the identification result;
Vectorization means for converting range-doppler data of the block range of the target detection candidate block extracted by the extraction means into a one-dimensional vector;
Specifying means for specifying a target detection cell of a range-doppler axis from the one-dimensional vector obtained by the vectorization means by a multilayer neural network;
A radar device comprising: calculating means for calculating at least one of a range and a speed from the target detection cell specified by the specifying means. Generation means for converting the reception signal in the pulse repetition period of the transmission pulse of N (N ≧ 1) hits into the frequency domain of the Doppler axis to generate range-Doppler data;
A pseudo high resolution means for performing an inverse Fourier transform on the range axis and the Doppler axis, zero-filling each axis on the range-Doppler data generated by the generation means, and performing a Fourier transform on the pseudo-resolution When,
A division unit configured to divide the range-Doppler data pseudo-high-resolutioned by the pseudo high-resolution means into a plurality of blocks along a range-Doppler axis;
Generation means for generating a histogram of amplitude intensity for each of the plurality of blocks divided by the division means;
Extraction means for identifying the presence or absence of a target for each divided block by a multilayer neural network from the histogram of each of the plurality of blocks generated by the generation means, and extracting a block of target detection candidates from the identification result;
Specifying means for specifying a target detection cell of a range-doppler axis by using a convolution neural network with range-doppler data of the block range of the target detection candidate block extracted by the extraction means;
A radar device comprising: calculating means for calculating at least one of a range and a speed from the target detection cell specified by the specifying means.   Furthermore, it plots in the two-dimensional coordinates of the range-doppler axis using the observation value coordinates of the target detection cell specified by the specifying means, performs cluster analysis of the plot coordinates, and detects the target based on the density difference of the plot points. The radar apparatus according to any one of claims 1 to 3, further comprising: a cluster analysis unit that determines the output and outputs the result to the calculation unit. Converting a received signal in a pulse repetition period in a transmission pulse of N (N ≧ 1) hits into a frequency domain of Doppler axis to generate range-Doppler data;
Dividing the range-Doppler data into a plurality of blocks by a range-Doppler axis;
Generating a histogram of amplitude magnitudes for each of the plurality of blocks;
The multi-layered neural network identifies the presence or absence of a target for each divided block from the histogram of each of the plurality of blocks, and extracts blocks of target detection candidates from the identification result,
Converting the range-Doppler data of the block range of the target detection candidate block into a one-dimensional vector,
Identifying a target detection cell of a range-doppler axis from the one-dimensional vector by a multilayer neural network;
A radar signal processing method for a radar device, which calculates at least one of range and velocity from the target detection cell. Converting a received signal in a pulse repetition period in a transmission pulse of N (N ≧ 1) hits into a frequency domain of Doppler axis to generate range-Doppler data;
For the range-Doppler data, the range axis and the Doppler axis are inverse Fourier transformed, zero-padded to each axis, and then Fourier transformed to make a pseudo high resolution,
The pseudo high resolution range-Doppler data is divided into a plurality of blocks by a range-Doppler axis,
Generation means for generating a histogram of amplitude intensity for each of the plurality of blocks;
The multi-layered neural network identifies the presence or absence of a target for each divided block from the histogram of each of the plurality of blocks, and extracts blocks of target detection candidates from the identification result,
Converting the range-Doppler data of the block range of the target detection candidate block into a one-dimensional vector,
Identifying a target detection cell of a range-doppler axis from the one-dimensional vector by a multilayer neural network;
A radar signal processing method for a radar device, which calculates at least one of range and velocity from the target detection cell. Converting a received signal in a pulse repetition period in a transmission pulse of N (N ≧ 1) hits into a frequency domain of Doppler axis to generate range-Doppler data;
For the range-Doppler data, the range axis and the Doppler axis are inverse Fourier transformed, zero-padded to each axis, and then Fourier transformed to make a pseudo high resolution,
The pseudo high resolution range-Doppler data is divided into a plurality of blocks by a range-Doppler axis,
Generating a histogram of amplitude magnitudes for each of the plurality of blocks;
The multi-layered neural network identifies the presence or absence of a target for each divided block from the histogram of each of the plurality of blocks, and extracts blocks of target detection candidates from the identification result,
The range detection data of the block range of the target detection candidate block is identified by a convolutional neural network, and the target detection cell of the range detection axis is specified;
A radar signal processing method for a radar device, which calculates at least one of range and velocity from the target detection cell.

Images