JP7494099B2 - Radar device and radar signal processing method - Google Patents

Description

This embodiment relates to a radar device and a radar signal processing method.

Conventional radar equipment often uses a spectrogram generated by STFT (Short-Time Fourier transform) to observe the time series Doppler changes on the slow-time axis to identify targets. However, this method has the problem that when the observation time is short (the number of hits is small), the Doppler resolution is low and the identification ability is reduced.

Pulse Compression, Yoshida, 'Revised Radar Technology', Institute of Electronics, Information and Communication Engineers, pp. 278-280 (1996) CFAR (Constant False Alarm Rate), Yoshida, 'Revised Radar Technology', Institute of Electronics, Information and Communication Engineers, pp.87-89 (1996) KR Product Array, Wing-Kin Ma, ‘DOA Estimation of Quasi-Stationary Signals With Less Sensors Than Sources and Unknown Spatial Noise Covariance: A Khatri-Rao Subspace Approach’, IEEE Trans. Signal Process, vol.58, no.4, pp.2168-2180, April (2010) Spatial Averaging Method, Kikuma, 'Adaptive Signal Processing Using Array Antennas', Science and Technology Publishing, pp.163-170, pp.336-337 (1999) Short-time FFT, Nakano, 'Signal Processing and Image Processing Using Wavelets', Kyoritsu Shuppan Co., Ltd., pp.97-100 (1999) Compressed Sensing, Toyoki Hoshikawa, ‘Performance Comparison of Compressed Sensing Algorithms for DOA Estimation of Multi-band Signals’, 2018 15TH WORKSHOP ON POSITIONING NAVIGATION AND COMMUNICATIONS(2018) MUSIC, Kikuma, 'Adaptive Antenna Technology', Ohmsha, pp.137-142 (2003) Convolutional Neural Network (CNN), Saito, 'Deep Learning from Scratch', O'Reilly Japan, pp.205-220 (2016) Recognition using CNN, Fujita, 'Implementation of Deep Learning', Ohmsha, pp.94-111 (2016)

As mentioned above, when conventional radar devices used STFT spectrograms to identify targets, if the observation time was short or the number of hits was small, the Doppler resolution was low and the identification ability was reduced.

The objective of this embodiment is to provide a radar device and a radar signal processing method that can improve target identification capabilities when the observation time is short or the number of hits is small.

In order to solve the above problem, according to an embodiment, a radar device uses a single pulse or modulated N (N≧1) pulses, performs pulse compression on the fast-time axis as necessary, and performs coherent integration processing on the slow-time axis to obtain range-Doppler data (hereinafter, RD data), detects multiple targets from the RD data using CFAR (Constant False Alarm Rate) processing, extracts and expands points on the slow-time axis for each detected target while sliding L (M≧L≧1) hits at a time, performs short-time time-series processing on the time-Doppler axis data for each of the M (M≧2) points extracted for each target, and learns the results of the short-time time-series processing using a CNN (Convolutional Neural Network) to identify whether the target is genuine or false.

FIG. 1 is a block diagram showing the configuration of a transmission system of a radar device according to a first embodiment. FIG. 2 is a block diagram showing the configuration of a receiving system of the radar device according to the first embodiment. FIG. 3 is a flowchart showing a flow of processing after reception data is acquired in the reception system of the first embodiment. FIG. 4 is a conceptual diagram for explaining the KR product processing and short-time processing in the reception system of the first embodiment. FIG. 5 is a diagram for explaining a partial correlation matrix averaging technique used in the receiving system of the first embodiment for suppressing correlation between reflection points of radar reflected waves. FIG. 6 is a block diagram showing the configuration of a receiving system of a radar device according to the second embodiment. FIG. 7 is a flowchart showing a process flow in the receiving system of the second embodiment when the configuration shown in FIG. 6 is used. FIG. 8 is a block diagram showing the configuration of a receiving system of a radar device according to the third embodiment. FIG. 9 is a flowchart showing a flow of processing after reception data is acquired in the reception system of the third embodiment. FIG. 10 is a block diagram showing the configuration of a receiving system of a radar device according to the fourth embodiment. FIG. 11 is a flowchart showing a flow of processing after reception data is acquired in the reception system of the fourth embodiment.

The following describes the embodiment with reference to the drawings.

First Embodiment
A radar device according to a first embodiment will be described below with reference to FIGS. 1 to 7. FIG.
1 and 2 are block diagrams showing the configurations of a transmission system and a reception system of a radar device according to a first embodiment, respectively, and FIG. 3 is a flowchart showing the flow of processing of the reception system shown in FIG.

In the transmission system shown in Figure 1, the signal generator 1 generates a transmission seed signal, the modulator 2 generates a modulated signal from the transmission seed signal, the frequency converter 3 converts the modulated signal into a high-frequency signal, the pulse modulator 4 pulse-modulates (pulse-compresses) the high-frequency signal to generate a transmission pulse of N (N ≥ 2) hits, and the transmission antenna 5 transmits the transmission pulse.

In the receiving system shown in Figure 2, the receiving antenna 11 receives a reflected signal of the transmitted pulse. The frequency converter 12 converts the received signal to an intermediate frequency, and the AD converter 13 converts the frequency-converted signal into a digital signal. The pulse compressor 14 pulse-compresses the digital signal (see Non-Patent Document 1), and the slow-time FFT processor 15 performs FFT processing on the slow-time axis on the pulse-compressed received signal to obtain range-Doppler data (RD data). The CFAR detector 16 uses the RD data to detect targets by CFAR processing (see Non-Patent Document 2). The target range extractor 17 obtains the range cells of each target detected by CFAR processing. The time-series short-time KR product processor 18 extracts the slow-time axis, extracts M (M≧2) points while sliding L (M≧L1) hits at a time, extracts K (K≧2) pieces of time-series time-Doppler axis data from the extracted points, and performs KR (Khatri-Rao) product processing (see Non-Patent Document 3) on this data as extended array processing. The time-series short-time processor 19 divides the KR product-processed time-series signal on the slow-time axis into short periods and performs FFT processing. The CNN classifier 20 uses CNN learning processing to classify the short-time FFT-processed signal as the true or false target. The classification result output unit 21 extracts and outputs only true targets from the CNN classification results.

That is, in the receiving system, the signal received by the receiving antenna 11 is frequency converted by the frequency converter 12 and converted to a digital signal by the AD converter 13. Next, as shown in FIG. 3, when the digitized received data is input (step S11), if it is a pulse-compressed (see Non-Patent Document 1) signal, it is pulse-compressed by the pulse compressor 14 (step S12), and the slow-time axis FFT processing is performed by the slow-time FFT processor 15 (step S13), to obtain range-Doppler data (RD data). Using this RD data, the CFAR processor 16 performs CFAR processing to detect targets (step S14), and the range extractor 17 extracts the range cells of each detected target (step S15).

Next, consider extracting the slow-time Doppler change for each range cell of the detected target. This is called a spectrogram when a short-time Fourier transform is used. In this case, there is a problem that the Doppler resolution of each Fourier transform result decreases because the time series signal on the slow-time axis is divided into short periods and then FFT is performed. To address this issue, consider using KR product processing (see non-patent document 3) as an extended array processing.

First, the time series short-time KR accumulation processor 18 extracts the slow-time axis after pulse compression 14, and extracts K (K≧2) pieces of time series time-Doppler axis data for M (M≧2) points extracted while sliding L (M≧L1) hits at a time. This is shown in Figure 4 (a). KR accumulation processing (see Non-Patent Document 3) is applied to this data (step S16). This is formulated as follows. First, the extracted signal sequence on the slow-time axis is expressed by the following equation.

Next, we perform KR multiplication using this signal X. First, the correlation matrix can be expressed as follows:

In radar reflected waves, there is a correlation between reflection points, and the correlation component of equation (2) has the problem of cross-terms occurring between reflection points. In order to suppress this correlation, a partial correlation matrix averaging method is applied, which separates the correlation matrix into submatrices and averages them, as shown in Figure 5 (see Non-Patent Document 4).

To suppress the correlation components of the correlation matrix Rxx, partial sequences of P cells are extracted from the signal length of X (M dimensions) while sliding on the slow-time axis, and each time a partial correlation matrix is calculated and the average value of the partial correlation matrix is calculated. Here, by selecting 2P-1>M, high resolution can be achieved on the Doppler axis after KR product processing. This process suppresses cross terms in the components of the correlation matrix, and suppresses correlation between multiple reflection points.

Using Rxxp, the average value of this partial correlation matrix, the leftmost and uppermost elements are vectorized by the KR product processor 18, resulting in a 2P-1-dimensional column vector.

As shown in FIG. 4(b), the signal Xkr after the KR product processing has overlapping portions on the slow-time axis of one range cell. In this case, the overlapping portions can be represented by one of the overlapping signals. Next, the signal after the KR product processing is used to perform time series short-time processing in the time series short-time processor 19 (step S17).

As a result, time-Doppler data (hereinafter referred to as TD data) showing the change in Doppler over time can be drawn for each detected range cell, as shown in FIG. 4(c). This two-dimensional TD data is input to a CNN discrimination processor 20 using a CNN (convolutional neural network, see Non-Patent Document 8) to discriminate whether the target is genuine or false (step S18). The above processes from step S14 to step S18 are repeated by changing the detection range until the target number of detections in the CFAR detection process is reached (steps S19, S20). As a discrimination method using CNN (see Non-Patent Document 9), learning data using TD data is prepared for each discrimination class, the discrimination class is trained as a teacher, and the learning result is set in the CNN discriminator 20. Actual measurement data is input, and each time TD data for each detection range cell is input to the CNN discriminator 20, the discrimination result is output by the discrimination result output unit 21 (step S21). This model setting method for CNN discrimination processing is a common method, and details are described in references (Non-Patent Document 9) and the like. Needless to say, the classification method can be a method that does not use CNN. The input data is two-dimensional TD data for each detection range cell, but the classification method can also be such that the two-dimensional TD data is rearranged into one dimension before being input.

As described above, according to this embodiment, KR integration processing is performed as a pre-processing step for time series short-time processing, which results in low resolution on the Doppler axis, so it is possible to achieve high resolution on the Doppler axis, thereby improving target identification capabilities even when the observation time is short (the number of hits is small).

Second Embodiment
In the second embodiment, a method using FFT processing (see Non-Patent Document 5) as the time-series short-time processing described in the first embodiment will be described.

Figure 6 is a block diagram showing the configuration of the receiving system of the radar device according to the second embodiment, and Figure 7 is a flowchart showing the processing flow of the receiving system shown in Figure 6. The transmitting system is the same as in the first embodiment, so a description thereof will be omitted here. Also, in Figures 6 and 7, the same parts are denoted by the same reference numerals, and duplicate descriptions will be omitted here.

In the reception system shown in FIG. 6 and the process flow shown in FIG. 7, a time series short-time FFT processor 19a is used as the time series short-time processor 19 shown in the first embodiment, and time series short-time FFT processing (step S17a) is performed as time series short-time processing (step S17). That is, a representative time series short-time processing is short-time FFT processing (see non-patent document 5). In this embodiment, when short-time FFT processing is used, KR product processing is performed as pre-processing of time series short-time FFT processing, which has low resolution on the Doppler axis, to achieve high resolution on the Doppler axis, which makes it possible to improve target identification ability even when the observation time is short (the number of hits is small).

Third Embodiment
In the second embodiment, a method will be described in which CS processing (compressed sensing, see Non-Patent Document 6) is used as the time-series short-time processing described in the first embodiment.

Figure 8 is a block diagram showing the configuration of the reception system of the radar device according to the third embodiment, and Figure 9 is a flowchart showing the processing flow of the reception system shown in Figure 8. The transmission system is the same as in the first embodiment, so a description thereof will be omitted here.

The receiving system shown in Figure 8 and the processing flow shown in Figure 9 are characterized in that a time series short-term CS processor 19b is used instead of the time series short-term processor 19 shown in the first embodiment, and time series short-term CS processing (step S17b) is performed as time series short-term processing (step S17). The other parts are the same as in Figures 2 and 3, so the same parts are denoted by the same reference numerals and their explanations are omitted here.

The CS processing applied in this embodiment can be formulated as follows. The input signal is the signal after KR integration processing (step S16) in the time-series short-time KR integration processor 18, and is the same as equation (4). This signal is subjected to FFT on the slow-time axis and is defined as the input signal Ys.

The element in the qth column of As is the signal vector on the Doppler axis when the source is at Doppler xq.

Here, by using equation (6) and taking advantage of the fact that Xs is sparse, it is possible to calculate Xs that minimizes the following equation (see non-patent document 6).

This Xs corresponds to the signal of the Doppler axis for each detection range cell, and TD data can be obtained in the same way as with short-time FFT processing. The CNN identification process (step S18) by the CNN processor 20, which uses the TD data for each detection range cell to identify the signal, is the same as in the first embodiment.

As described above, in this embodiment, KR integration processing is performed as preprocessing for time-series short-time CS processing, which results in low resolution on the Doppler axis, so it is possible to achieve high resolution on the Doppler axis, thereby improving target identification capabilities even when the observation time is short (the number of hits is low).

Fourth Embodiment
In the fourth embodiment, a method using MUSIC (Multiple Signal Classification) processing (see Non-Patent Document 7) as time-series short-time processing will be described.

Figure 10 is a block diagram showing the configuration of the reception system of a radar device according to the fourth embodiment, and Figure 11 is a flowchart showing the processing flow of the reception system shown in Figure 10. The transmission system is the same as in the first embodiment, so a description thereof will be omitted here.

The receiving system shown in FIG. 10 and the processing flow shown in FIG. 11 are characterized in that a time series short-time MUSIC processor 19c is used instead of the time series short-time processor 19 shown in the first embodiment, and time series short-time MUSIC processing (step S17c) is performed as time series short-time processing (step S17). The other parts are the same as in FIG. 2 and FIG. 3, so the same parts are denoted by the same reference numerals and their explanations are omitted here.

The MUSIC processing applied in this embodiment can be formulated as follows. The input signal is a signal after time-series short-time KR integration processing 18, and is the same as equation (4). In the second embodiment, the result of FFT processing this signal on the slow-time axis is taken as Ys, but in this embodiment, a signal similar to equation (4) before FFT processing is used as the input.

This Rxx is used to perform MUSIC processing and calculate the MUSIC spectrum using the following formula.

このSmusicが、ドップラ軸の信号に相当し、短時間FFTと同様に、TDデータを得ることができる。この検出レンジセル毎のTDデータを用いて識別するCNN識別器２０のCNN識別処理（ステップS18）は、第１の実施形態と同様である。

This Smusic corresponds to the signal on the Doppler axis, and TD data can be obtained in the same manner as in the short-time FFT. The CNN classification process (step S18) of the CNN classifier 20, which uses the TD data for each detection range cell to classify, is the same as in the first embodiment.

As described above, in this embodiment, KR integration processing is performed as preprocessing for time-series short-time MUSIC processing, which has low resolution on the Doppler axis, so it is possible to achieve high resolution on the Doppler axis, thereby improving target identification capabilities even when the observation time is short (the number of hits is low).

The above describes an embodiment in which time-series short-time KR integration processing is performed before time-series short-time processing, but a method of performing other time-series short-time processing without performing time-series short-time KR integration processing may also be used. In addition, a method of data expansion other than KR integration processing may also be used.

The present invention is not limited to the above-described embodiment as it is, and in the implementation stage, the components can be modified and embodied without departing from the gist of the invention. In addition, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, components from different embodiments may be appropriately combined.

1... signal generator, 2... modulator, 3... frequency converter, 4... pulse modulator, 5... transmitting antenna,
11...receiving antenna, 12...frequency converter, 13...AD converter, 14...pulse compressor, 15...slow-time FFT processor, 16...CFAR detector, 17...target range extractor, 18...time series short-time KR product processor, 19...time series short-time processor, 19a...time series short-time FFT processor, 19b...time series short-time CS processor, 19c...time series short-time MUSIC processor, 20...CNN classifier, 21...classification result output device.

Claims (6)

A radar device that uses a single pulse or modulated N (N≧1) pulses, performs pulse compression on a fast-time axis as necessary, and performs coherent integration on a slow-time axis to obtain range-Doppler data (hereinafter, referred to as RD data),
Detect multiple targets from the RD data using CFAR (Constant False Alarm Rate) processing,
For each detected target, the points on the slow-time axis are extracted and expanded by sliding them by L (M ≥ L ≥ 1) hits.
The time-Doppler axis data of each of the M (M≧2) points extracted for each target is subjected to short-time time series processing;
The radar device uses a Convolutional Neural Network (CNN) to learn the results of the short-term time-series processing and identify whether the target is genuine or false. The radar device of claim 1, in which the time-Doppler axis data is expanded by Khatri-Rao (KR) product processing before the time series short-time processing is performed. The radar device of claim 1, which uses FFT (Fast Fourier Transform) as the time series short-time processing. The radar device of claim 1 uses compressed sensing (CS) as the time series short-time processing. The radar device of claim 1 uses the MUSIC (Multiple signal classification) method for the short-term time series processing. It is used in radar equipment that uses a single pulse or modulated N (N ≥ 1) pulses, compresses the pulse as necessary on the fast-time axis, and performs coherent integration on the slow-time axis to obtain range-Doppler data (hereafter referred to as RD data).
Detect multiple targets from the RD data using CFAR (Constant False Alarm Rate) processing,
For each detected target, the points on the slow-time axis are extracted and expanded by sliding them by L (M ≥ L ≥ 1) hits.
The time-Doppler axis data of each of the M (M≧2) points extracted for each target is subjected to short-time time series processing;
A radar signal processing method in which the results of the short-term time-series processing are learned using a Convolutional Neural Network (CNN) to identify whether the target is genuine or false.

Images