KR20240001968A - FMCW radar using deep learning and method for improving of range precision the same - Google Patents

Abstract

)를 각각 구하는 제 1 및 제 2 피크 검출부; 상기 피크 인덱스(

)에 해당하는 정수배(n)와 그 일정 값(n_center)의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 시프트시키는 제 1 및 제 2 스펙트럼 시프트부; 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝을 이용하여 타겟의 소수배 거리를 추론하는 딥러닝부; 및 상기 타겟의 소수배 거리를 정수배로 복원하는 스펙트럼 복원부를 포함하는 딥러닝을 이용한 FMCW 레이더 및 거리 해상도 향상 방법이 제시된다.The present invention includes first and second radars that respectively generate intermediate frequency signals using a transmitted signal radiated to a target and a received signal reflected from the target; Peak index through peak detection from the intermediate frequencies of each of the first and second radars (

) first and second peak detection units respectively calculating; The peak index (

) Calculate the frequency difference between an integer multiple (n) and the constant value (n _center ) and shift the spectrum by the difference; A deep learning unit that inputs a spectrum-shifted frequency signal and infers the decimal distance of the target using deep learning; An FMCW radar using deep learning and a method for improving range resolution are presented, including a spectrum restoration unit that restores the fractional distance of the target to an integer multiple.

Description

FMCW radar using deep learning and method for improving range precision the same}

The present invention relates to an FMCW radar device, and in particular, to an FMCW radar and a method for improving distance resolution using deep learning that can improve distance resolution while simplifying the artificial intelligence structure using spectrum shift.

Radar is a device that detects the presence of a target and its distance by emitting electromagnetic waves and receiving reflected waves reflected by the target in the area. These radars include pulse method, FMCW (Frequency Modulated Continuous Wave) method, and Frequency Shift Keying (FSK) method depending on the modulation method of the applied transmission signal. Depending on each modulation method, The method of extracting the speed and distance of the target changes.

Here, the FMCW radar transmits a continuous wave signal and simultaneously receives a reflected wave signal from the target. At this time, the FMCW radar extracts the relative distance information and relative speed information of the target using the beat frequency, which is the difference frequency component between the transmitted signal and the received signal. Because FMCW radar uses ultra-high frequencies, it can be implemented relatively simply without being affected by the external environment. Therefore, the use of FMCW radar is gradually increasing not only in military applications but also in civilian fields such as self-driving cars.

For FMCW radar, distance resolution from the target is important. In other words, it is important for the performance of the FMCW radar to accurately detect the target's location and speed. At this time, the distance resolution of the FMCW radar can be determined by the frequency bandwidth.

Meanwhile, artificial intelligence technology can be applied to improve the distance resolution of the laser. Artificial intelligence technology can be divided into unsupervised learning and supervised learning. Unsupervised learning is a method of learning on your own by inputting data without providing the correct answer, while supervised learning is a method of learning by inputting data and the correct answer together. Radar mainly uses supervised learning to improve distance resolution. This method determines the location of the target (correct answer) and inputs the spectrum (data) measured by radar as input. This method collects spectrum information by locating the target between bins and trains it in an artificial intelligence model. The artificial intelligence finds the peak value hidden between the bins to provide a more accurate location of the target. can be found out. However, this method has the problem of having to set the distance directly. In general, the 'correct answer' is expressed as a 'label' in artificial intelligence, and this label must have data that matches the label. If the target distance resolution is small, the number of labels that exist between each bin must be large, and thus the training data also increases. Additionally, as the distance to the target increases or the detection range widens, the number of labels increases. Not only can the number of labels be infinite, but it also complicates the artificial intelligence structure, so the detection range and number of labels must be appropriately limited.

For example, to estimate the distance at 0.1m intervals with a radar with a distance resolution of 1m, 10 labels per meter are needed, and more than N pieces of training data per label must be obtained. If the maximum measurement range is 50m, the number of labels is 50 × 10 = 500, and the amount of data required for learning is 500N. In other words, the number of labels increases, thereby complicating the artificial intelligence structure.

Korean Patent Registration No. 10-1908490

The present invention provides an FMCW radar and a method for improving distance resolution using deep learning that can improve distance resolution by applying artificial intelligence.

The present invention provides an FMCW radar and a method for improving range resolution using deep learning that can simplify the artificial intelligence structure using spectrum shift.

)를 각각 구하는 제 1 및 제 2 피크 검출부; 상기 피크 인덱스(

)에 해당하는 정수배(n)와 그 일정 값(n_center)의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 시프트시키는 제 1 및 제 2 스펙트럼 시프트부; 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝을 이용하여 타겟의 소수배 거리를 추론하는 딥러닝부; 상기 타겟의 소수배 거리를 정수배로 복원하는 스펙트럼 복원부를 포함한다.An FMCW radar using deep learning according to an embodiment of the present invention includes first and second radars that respectively generate intermediate frequency signals using a transmitted signal radiated to a target and a received signal reflected from the target; Peak index through peak detection from the intermediate frequencies of each of the first and second radars (

) first and second peak detection units respectively calculating; The peak index (

) Calculate the frequency difference between an integer multiple (n) and the constant value (n _center ) and shift the spectrum by the difference; A deep learning unit that inputs a spectrum-shifted frequency signal and infers the decimal distance of the target using deep learning; It includes a spectrum restoration unit that restores the fractional distance of the target to an integer multiple.

It further includes a target estimation unit that estimates spatial coordinates of the target using the spectrum restored through the spectrum restoration unit.

Each of the first and second radars includes a transmission antenna that radiates a transmission signal of an FMCW waveform, a reception antenna that receives a signal reflected from the target, a mixer that mixes the transmission signal and the reception signal, and a fast Fourier transformer for mixing the mixed signal. It contains a fast Fourier transformer that converts and generates an intermediate frequency signal.

과 n_center의 인덱스 차이(bin_diff)를 수학식 10과 같이 구하고, 이동시켜야 할 주파수(f_diff)를 구한 후 스펙트럼 시프트된 신호를 구하며, 스펙트럼 시프트된 신호를 샘플링한 후 주파수 변환하여 주파수를 이동시킨다.The first and second spectrum shift units are

Find the index difference (bin _diff ) between n center and n _center as shown in Equation 10, find the frequency to be moved (f _diff ), find the spectrum shifted signal, sample the spectrum shifted signal, and then move the frequency by converting the frequency. I order it.

[Equation 10]

The frequency to be moved (f _diff ) is obtained as shown in Equation 11 by multiplying the index difference (bin _diff ) calculated from Equation 10 by the interval frequency between bins (fs/N).

[Equation 11]

Here, N refers to the number of FFT points, and fs is the sampling frequency.

The spectrum shifted signal is obtained using Equation 12.

[Equation 12]

Here, s _IF is an IF signal in the time domain, and s _shift is a spectrum shifted signal.

The signal sampling is obtained as in Equation 13 by applying t=nT _s to Equation 12.

[Equation 13]

Here, t is a time variable, n is a sample index ranging from 0 to (N-1), and Ts is the sampling period.

The frequency shift is obtained by converting Equation 13 to Equation 14 and then converting the frequency to Equation 15.

[Equation 14]

[Equation 15]

Here, S _shift and S _IF are the Fourier transforms of s _shift and s _IF , respectively.

The deep learning unit uses a CNN structure that includes an input layer, a convolution layer, a pooling layer, and a fully connected layer, and performs softmax and classification tasks.

The deep learning unit consists of a process in which a spectrum-shifted frequency signal is input and a label is output, and is converted to a decimal distance k through Equation 16.

[Equation 16]

Here, the trained area represents the smallest distance in the learning area, and the larger the label, the farther the target is from the radar. label _res represents the distance indicated by one label as label resolution, and the label resolution is set at intervals narrower than the radar distance resolution.

The label resolution is the distance obtained by dividing the distance resolution by the total number of labels, as shown in Equation 17.

[Equation 17]

Here, m is the total number of labels and Rres is the distance resolution.

일 때, 스펙트럼 시프트를 통해서

이 되므로

를 더해줘서 정수배 거리로 복원한다.The spectrum restoration unit determines the distance to the target.

When, through spectral shift

Because this happens

By adding , it is restored to an integer multiple distance.

)를 각각 구하는 과정; 상기 피크 인덱스(

)에 해당하는 정수배(n)와 그 일정 값(n_center)의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 시프트시키는 과정; 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝을 이용하여 타겟의 소수배 거리를 추론하는 과정; 및 상기 타겟의 소수배 거리를 정수배로 복원하는 과정을 포함한다.A method of improving the range resolution of an FMCW radar using deep learning according to another embodiment of the present invention includes the process of generating an intermediate frequency signal using a transmitted signal radiated to a target and a received signal reflected from the target; Peak index through peak detection from the intermediate frequency (

) process of calculating each; The peak index (

A process of calculating the frequency difference between an integer multiple (n) corresponding to ) and a certain value (n _center ) and shifting the spectrum by the difference; A process of inputting a spectrally shifted frequency signal and inferring the decimal distance of the target using deep learning; and a process of restoring the decimal multiple distance of the target to an integer multiple.

It further includes a process of estimating the spatial coordinates of the target using the restored spectrum.

In the process of generating the intermediate frequency, the intermediate frequency is generated by radiating a transmitted signal and receiving a signal reflected from the target, mixing the transmitted signal and the received signal, and performing fast Fourier transform.

과 n_center의 인덱스 차이(bin_diff)를 구하는 과정과, 상기 인덱스 차이(bin_diff)에 빈(bin) 사이의 간격 주파수(fs/N)를 곱하여 수학식 11과 같이 이동시켜야 할 주파수(f_diff)를 구하는 과정과, 수학식 12를 이용하여 스펙트럼 시프트된 신호를 구하는 과정과, 수학식 12에 t=nT_s를 적용해서 수학식 13과 같이 스펙트럼 시프트된 신호를 샘플링하는 과정과, 수학식 13을 수학식 14로 변환한 후 주파수 변환하여 수학식 15와 같이 주파수를 이동시키는 과정을 포함한다.The spectrum shift process is, from Equation 10, the target

The process of calculating the index difference (bin _diff ) between n _center and the index difference (bin _diff ) is multiplied by the frequency of the interval between bins (fs/N) to calculate the frequency (f _diff ) to be moved as shown in Equation 11. ), a process of obtaining a spectrum shifted signal using Equation 12, a process of sampling the spectrum shifted signal as shown in Equation 13 by applying t=nT _s to Equation 12, and Equation 13 It includes the process of converting to Equation 14, then converting the frequency, and moving the frequency as shown in Equation 15.

[Equation 10]

[Equation 11]

Here, N refers to the number of FFT points, and fs is the sampling frequency.

[Equation 12]

Here, s _IF is an IF signal in the time domain, and s _shift is a spectrum shifted signal.

[Equation 13]

Here, t is a time variable, n is a sample index ranging from 0 to (N-1), and Ts is the sampling period.

[Equation 14]

[Equation 15]

Here, S _shift and S _IF are the Fourier transforms of s _shift and s _IF , respectively.

The process of inferring the decimal distance of a target using deep learning is a process in which a spectrum-shifted frequency signal is input and a label is output, which is converted to the decimal distance k through Equation 16.

[Equation 16]

Here, the trained area represents the smallest distance in the learning area, and the larger the label, the farther the target is from the radar. label _res represents the distance indicated by one label as label resolution, and the label resolution is set at intervals narrower than the radar distance resolution.

The label resolution is the distance obtained by dividing the distance resolution by the total number of labels, as shown in Equation 17.

[Equation 17]

Here, m is the total number of labels and Rres is the distance resolution.

일 때, 스펙트럼 시프트를 통해서

이 되므로

를 더해줘서 정수배 거리로 복원한다.In the process of restoring the decimal distance of the target to an integer multiple, the distance of the target is

When, through spectral shift

Because this happens

By adding , it is restored to an integer multiple distance.

)를 구하고, 스펙트럼 시프트 후 딥러닝을 통해 소수배 거리를 구하며, 소수배 거리를 정수배 거리로 복원한 후 타겟의 공간 좌표를 추정한다. 따라서, 본 발명은 스펙트럼 시프트 후 딥러닝을 수행함으로써 모든 거리에 대해서 스펙트럼을 학습시키지 않아도 거리 해상도를 향상시킬 수 있고, 적은 라벨을 이용하여 인공지능 구조의 복잡도를 감소시킬 수 있다.The present invention peak detects the intermediate frequencies of the first and second radars spaced a predetermined distance apart and detects the largest frequency in the spectrum (i.e. peak index,

) is obtained, and after spectral shift, the decimal distance is obtained through deep learning, and the decimal multiple distance is restored to an integer multiple distance, and then the spatial coordinates of the target are estimated. Therefore, by performing deep learning after spectrum shifting, the present invention can improve distance resolution without having to learn the spectrum for all distances, and reduce the complexity of the artificial intelligence structure by using fewer labels.

Figure 1 is a chirp graph of the FMCW radar.
Figure 2 is a block diagram showing the configuration of the FMCW radar.
Figure 3 is a simulation of measuring a target located at 3.8m using a radar with a distance resolution of 0.3m.
Figure 4 is a schematic diagram illustrating a target location detection method for two radars used in the present invention.
Figures 5(a) to 5(c) are spectrum results in which the positions of each target are set to 3.5m, 4.7m, and 6.5m in a simulation of an ideal environment, respectively.
Figure 6 is a block diagram for explaining an FMCW radar using deep learning according to an embodiment of the present invention.
Figure 7 is a diagram showing the structure of CNN used in the present invention by layer.
Figure 8 is a schematic diagram showing the location of the target according to each label.
9 and 10 are flowcharts illustrating a method for improving the distance resolution of an FMCW radar using deep learning according to an embodiment of the present invention.
Figure 11 is a schematic diagram illustrating a method of obtaining the location of a target using an FMCW radar according to an embodiment of the present invention.
12 is an exemplary diagram of spectral shift according to the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings. However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms. These embodiments only serve to ensure that the disclosure of the present invention is complete and to convey the scope of the invention to those skilled in the art. It is provided for complete information.

FMCW radar uses a chirp whose frequency increases with time. Figure 1 is a chirp graph of an FMCW radar, showing the transmitted signal (Tx) and the received signal (Rx). That is, Figure 1 shows a transmitted signal transmitted from the FMCW radar and a received signal that is incident after the transmitted signal is reflected from the target. Additionally, there are parameters of beat frequency (f _beat ), time delay (τ), and frequency bandwidth (Sweep _bw ). As shown in Figure 1, a delay time (τ) occurs due to the distance between the FMCW radar and the target, and accordingly, a frequency difference exists.

Figure 2 is a block diagram showing the configuration of an FMCW radar, and is intended to explain the signal processing process for calculating the target distance from the radar. As shown in Figure 2, the FMCW laser includes a VCO (Voltage Controlled Oscillator) 11 that generates a sawtooth wave or triangle wave FMCW signal, a low noise amplifier (LNA) 12 that amplifies the FMCW signal with low noise, A transmission antenna 13 that radiates a transmission signal (Tx) through a low noise amplifier 12 toward a target, a reception antenna 14 that receives a signal (Rx) reflected from the target, and a reflected wave through the reception antenna 14. A low noise amplifier (LNA) 15 that amplifies the received signal (Rx) with low noise, and a mixer that mixes the signal reflected from the target through the low noise amplifier 15 and the FMCW signal output from the VOC 11 ( 16), an amplifier 17 that amplifies the signal mixed through the mixer 16, and a converter (Analog to Digital Converter (ADC) 18) that converts the analog signal through the amplifier 17 into a digital signal, A low pass filter (LPT) 19 that filters the signal through the converter 18, and a fast Fourier transformer (19) that converts the signal through the low pass filter 19 into an intermediate frequency (IF). It may include fast Fourier transform (FFT) (20).

Looking at the driving method of this FMCW laser, the VCO (Voltage Controlled Oscillator) 11 generates a sawtooth wave or triangle wave FMCW signal, and the transmission antenna 130 passes the FMCW signal through a low noise amplifier (LNA) 12. A signal, that is, a transmission signal (Tx), is radiated toward the target. The signal radiated through the transmitting antenna 13 is reflected by the target, and the reflected signal is delayed in time according to the distance between the radar and the target and is received through the receiving antenna 14. The signal (Rx) received through the receiving antenna 14 passes through a low noise amplifier (LNA) 15 and is mixed with the FMCW signal generated from the VOC 11 in the mixer 16. Then, the mixed signal is amplified through an amplifier 17, converted into a digital signal through a converter 18, and then filtered through a low pass filter (LPT) 19 and a fast Fourier transform (FFT). )(20) and converted to intermediate frequency (IF). The intermediate frequency has the frequencies of several targets, and the beat frequency of one target is determined by finding the bin with the highest power. Here, bin refers to one frequency bar in the spectrum graph.

When the shortest time delay that can be obtained from the radar is τ _min , the beat frequency due to this time delay can be expressed as f _b,min , and is expressed by Equation 1 as follows.

이며, fs는 샘플링 주파수이다. 일반적으로 하나의 첩 데이터를 가지고 FFT를 진행하기 때문에 Tm은

로 표시할 수 있다.Here, Sweep _bw is the bandwidth at which the frequency of FMCW changes, and Tm is the duration of one FMCW cycle. The number of points in the FFT needed to represent the target beat frequency f _b,min is

, and fs is the sampling frequency. Since FFT is generally performed with one chirp data, Tm is

It can be displayed as .

과 같이 첩의 기울기에 시간 지연을 곱하면 구할 수 있다. 여기서,

(c는 광속)이므로

이다. R_res은 레이더의 거리 해상도를 의미하며, 아래의 수학식 2로 다시 정리할 수 있다.The frequency bandwidth is

It can be obtained by multiplying the slope of the chirp by the time delay as shown. here,

(c is the speed of light), so

am. R _res means the distance resolution of the radar, and can be rearranged as Equation 2 below.

According to Equation 2, it can be seen that the range resolution of the radar is determined by the frequency bandwidth.

Generally, in order to detect the distance of a target, the frequency at which the bin value is maximum is determined as the beat frequency. This is called maximum peak detection. This can be expressed in equation 3.

은 S_f(n)의 절댓값이고

은

의 크기가 최대인 빈(bin)의 인덱스이다. 즉,

은 피크 인덱스로서, 스펙트럼에서 가장 큰 주파수를 갖는 지점이다. 이때, 빈(bin)의 인덱스는 0부터 FFT 포인트 수만큼의 이산적인 값을 갖는다. 타겟의 거리는

에 레이더의 단위 거리인 거리 해상도를 곱해주는 것으로 수학식 4와 같이 구할 수 있다.

is the absolute value of S _f (n) and

silver

is the index of the bin with the largest size. in other words,

is the peak index, which is the point with the highest frequency in the spectrum. At this time, the index of the bin has discrete values ranging from 0 to the number of FFT points. The target distance is

It can be obtained as shown in Equation 4 by multiplying the distance resolution, which is the unit distance of the radar.

To aid understanding, Figure 3 shows a simulation of measuring a target located at 3.8m with a radar with a distance resolution of 0.3m. According to Equation 4, the target distance is distance resolution × peak index, so the target distance is 3.9m. That is, since the distance resolution is 0.3m and the peak index (i.e., the largest frequency bar) in FIG. 3 is 13, the target distance is 0.3m × 13 = 3.9m according to Equation 4. At this time, an error of +0.1m occurs between the actual target position of 3.8m and the target distance according to Equation 4. The +0.1m error means that the maximum value of sinc is hidden between the 13th bin and the 14th bin. .

이다.As mentioned above, the distance obtained using maximum value-based distance detection is not accurate because there is an error. Except for the section of the signal sampled during a certain period, it can be viewed as a signal with a value of 0. These signals have the characteristics of a sinc function when FFTed. In other words, if expressed in a formula,

am.

The FMCW sawtooth wave signal whose frequency changes with time can be expressed as Equation 5 below.

를 푸리에 변환하면 아래의 수학식 6을 구할 수 있는데, τ는 2R/c로 표현되었다.s _IF (t) represents the time domain signal of the beat frequency, and the exp term is an expression of the square of a natural constant and is used together with j, which means an imaginary number, to represent a value in complex number format. A is the amplitude of the transmitted signal, σ is the amplitude attenuation, fc is the carrier frequency, τ is the time delay due to the distance of the target, Sweep _bw is the sweep bandwidth, and Tm is the chirp period. τ can be written as 2R/c, where R is the distance to the target and c is the speed of light. Π(t/Tm) is a square pulse that has a value of 1 for the duration of Tm from -Tm/2 to Tm/2 and a value of 0 in the other range.

By Fourier transforming, Equation 6 below can be obtained, where τ is expressed as 2R/c.

(n은 정수)로 기재할 수 있는데, R에 대입하면

이 된다.

는 f=n/Tm(n은 정수)마다 0의 진폭을 가지기 때문에 1/Tm만큼 이동한 신호는 원점이 움직일 뿐이며, Tm 간격으로 0의 간격을 가진다. 그런데, R이 정수배가 아니면 진폭은 무시한다고 했을 때 sinc 함수가

만큼 이동하게 되는데, 이 경우 sinc가 0인 지점이 1/Tm은 아닌 것을 알 수 있다. If R is an integer multiple of the distance resolution,

It can be written as (n is an integer), and when substituted into R,

This happens.

Since has an amplitude of 0 every f=n/Tm (n is an integer), a signal that moves by 1/Tm only moves the origin, and has an interval of 0 at Tm intervals. However, if R is not an integer multiple, the amplitude is ignored, and the sinc function

In this case, you can see that the point where sinc is 0 is not 1/Tm.

Therefore, when the target is located at an integer multiple of the distance resolution, the sidelobe becomes 0 due to the characteristics of the sinc function. Here, the sidelobe is a bin that exists near the largest value (main lobe) among the bins generated by one target and has an amplitude smaller than the peak value. However, if the target is between the distance resolutions, the orthogonality of sinc is broken and sidelobes appear. This sidelobe causes the index of the target to be obtained incorrectly. In this situation, if you can find the maximum index between those indices, the radar's range accuracy will be improved.

<Description of the present invention>

1. How to find the coordinates of a target using distance

Radars can obtain the target's location in various ways, and the present invention uses two radars to obtain the target's location. Figure 4 is a schematic diagram illustrating a target location detection method for two radars used in the present invention. That is, Figure 4 is a schematic diagram of a location detection method using the multivariate measurement method used in the present invention. As shown in FIG. 4, two radars, that is, first and second radars (Radar 1 and Radar 2), are spaced apart at a predetermined distance. At this time, each of the first and second radars has the configuration shown in FIG. 2. That is, the first and second radars each generate an FMCW signal and transmit it to the target, and generate an intermediate frequency signal (IF) from mixing and FFT of the signal reflected and received from the target and the FMCW signal. In Figure 4, r _i is the distance between the i (i=1, 2)th radar and the target (i.e., r1 and r2 are the distances between the first and second radars and the target, respectively), and dx is the first and second radar. The gap, h is the vertical distance between the target and the reference horizon, and x is the horizontal distance between the target and the radar. Here, the coordinates of the target using two radars can be obtained using Equation 7 and Equation 8. That is, the horizontal distance (x) between the target and the radar can be obtained using Equation 7, and the vertical distance (h) between the target and the reference horizon can be obtained using Equation 8.

2. Peak movement

)과 인접한 지점(

)의 크기가 서로 달라지면서 타겟 측정이 잘못될 수 있다. 그러나, 아날로그 신호로 보았을 때, 스펙트럼은 sinc 형태로 만들어지므로 샘플링된 스펙트럼에서도 sinc의 규칙성이 존재할 것이다. 레이더와 타겟 사이의 거리는 수학식 7에 의해 구할 수 있다.The peak movement proposed in the present invention uses the regularity of the intermediate frequency (IF) to solve the problems of distance limitation due to labels, increased learning data, and increased complexity. The mixed signal has discrete data lengths through A/D sampling, and when this signal is Fourier transformed, a frequency spectrum is formed in the shape of a sinc function. When frequency converting a sampled signal, one sinc function is created per time sample in the frequency domain, so it is important to maintain orthogonality. However, if the target is located between each bin, orthogonality is broken and the sidelobes have non-zero values. In this case, the point where the power of the spectrum is greatest (

) and adjacent points (

) may vary in size, causing target measurement to be incorrect. However, when viewed as an analog signal, the spectrum is created in a sinc form, so sinc regularity will exist in the sampled spectrum. The distance between the radar and the target can be obtained by Equation 7.

는 레이더와 타겟 사이의 거리이고, n은 정수배(n=1, 2,…, n_max), k는 소수배(-α/2<k<α/2)를 나타내며, α는 레이더의 거리 해상도(=R_res)이다. n은 1부터 n_max까지의 범위를 가지며 n_max는 레이더의 최대 감지 거리이다. 그리고, k는 거리 해상도(α/2)보다 작고 -α/2보다 큰 범위의 값을 가진다. 수학식 9는 타겟과의 거리를 정수배와 거리 해상도의 소수배의 합으로 표현하였다. 서로 다른 두 n에 대해서 k가 같은 경우, 스펙트럼은 n에 따라서 크기는 달라지지만 인접한 데이터들의 증감 패턴은 일정한 규칙성을 보인다. 예를 들어, 도 5(a) 내지 도 5(c)는 α가 0.3m이고, n이 각각 11, 15, 21이며, k는 0.2m인 시뮬레이션 결과이다. 즉, 도 5(a) 내지 도 5(c)는 각각 이상적인 환경의 시뮬레이션에서 각 타겟의 위치를 3.5m, 4.7m, 6.5m로 설정한 스펙트럼 결과이다. 도 5에서 볼 수 있는 바와 같이 k가 일정한 경우 스펙트럼의 증감 추이가 일정하게 유지되는 것을 그래프를 통해 확인할 수 있다. 이처럼 k 위치에 대해서 타겟 스펙트럼이 일정한 패턴을 반복한다면 n을 하나의 값으로 통일하고 k에 대해서만 추론하는 인공지능 환경을 구성하는 것이 가능하다. In equation 9:

is the distance between the radar and the target, n represents an integer multiple (n=1, 2,…, n _max ), k represents a decimal multiple (-α/2<k<α/2), and α is the distance resolution of the radar. (=R _res ). n ranges from 1 to n _max , where n _max is the maximum detection range of the radar. And, k has a value that is smaller than the distance resolution (α/2) and larger than -α/2. Equation 9 expresses the distance to the target as the sum of an integer multiple and a decimal multiple of the distance resolution. When k is the same for two different n, the size of the spectrum varies depending on n, but the increase/decrease pattern of adjacent data shows certain regularity. For example, Figures 5(a) to 5(c) are simulation results where α is 0.3m, n is 11, 15, and 21, respectively, and k is 0.2m. That is, Figures 5(a) to 5(c) are spectrum results in which the positions of each target are set to 3.5m, 4.7m, and 6.5m in a simulation of an ideal environment, respectively. As can be seen in Figure 5, when k is constant, it can be confirmed through the graph that the increase/decrease trend of the spectrum remains constant. In this way, if the target spectrum repeats a certain pattern for k positions, it is possible to configure an artificial intelligence environment that unifies n into one value and infers only about k.

This method can improve distance resolution without having to learn the spectrum for all distances, and can reduce the complexity of the artificial intelligence structure by using fewer labels.

3. Configuration of the present invention

을 각각 구하는 제 1 및 제 2 피크 검출부(210, 220)와, 수학식 9에서

에 해당하는 n과 n_center의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 이동시키는 제 1 및 제 2 스펙트럼 시프트부(310, 320)와, 제 1 및 제 2 스펙트럼 시프트부(310, 320)에 의해 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝(deep learning)을 이용하여 소수배 거리를 추론하는 딥러닝부(400)와, 스펙트럼 시프트 및 딥러닝을 통한 타겟의 소수배 거리를 정수배로 복원하는 스펙트럼 복원부(500)와, 복원된 스펙트럼을 이용하여 타겟의 공간 좌표를 추정하는 타겟 추정부(600)를 포함할 수 있다.In the present invention, the value of n is set to a constant value (n _center ) and then a pattern for the prime multiple distance k is learned. Figure 6 is a block diagram for explaining an FMCW radar using deep learning according to an embodiment of the present invention. It is a block diagram of an FMCW radar for improving distance resolution using deep learning. Referring to FIG. 6, the FMCW radar according to an embodiment of the invention is a first and second radar (110) that generates an intermediate frequency signal (IF) using a transmitted signal radiated to the target and a received signal reflected from the target, respectively. , 120) and through FFT and peak detection from the intermediate frequencies (IF) of each of the first and second radars 110 and 120.

In Equation 9, the first and second peak detection units 210 and 220 respectively calculate

By the first and second spectrum shift units 310 and 320, which calculate the frequency difference between n and n _center corresponding to and shift the spectrum by the difference, and the first and second spectrum shift units 310 and 320 A deep learning unit 400 that inputs a spectrum-shifted frequency signal and infers the fractional distance using deep learning, and spectral restoration that restores the fractional distance of the target to an integer multiple through spectrum shifting and deep learning. It may include a unit 500 and a target estimation unit 600 that estimates the spatial coordinates of the target using the restored spectrum.

3.1. radar

The first and second radars 110 and 120 are spaced apart from each other at a predetermined distance as shown in FIG. 4 . That is, the first and second radars 110 and 120 can each radiate a transmission signal toward the target at a predetermined interval and receive a signal reflected from the target. Additionally, each of the first and second radars 110 and 120 has a configuration as shown in FIG. 2. That is, as shown in FIG. 2, the first and second lasers 110 and 120 each include a VCO 11 that generates a sawtooth wave or triangle wave FMCW signal, an LNA 12 that amplifies the FMCW signal with low noise, and an LNA A transmitting antenna 13 that radiates a transmitted signal (Tx) through (12) toward the target, a receiving antenna 14 that receives a signal (Rx) reflected from the target, and a reflected wave receiving signal through the receiving antenna 14. An LNA (15) that amplifies (Rx) with low noise, a mixer (16) that mixes the signal reflected from the target through the LNA (15) and the FMCW signal output from the VOC (11), and mixing through the mixer (16) An amplifier 17 that amplifies the signal, an ADC 18 that converts the analog signal through the amplifier 17 into a digital signal, an LPT 19 that filters the signal through the ADC 18, and an LPT 19 ) may include an FFT (20) that converts the signal through IF to IF. The first and second radars 110 and 120 each generate an FMCW signal and transmit it to the target, and generate an intermediate frequency signal (IF) from mixing and FFT of the signal reflected and received from the target and the FMCW signal.

3.2. Peak detection unit

을 구한다. 즉, 제 1 피크 검출부(210)는 제 1 레이더(110)의 중간 주파수로부터 FFT 및 피크 검출을 통해

을 구하고, 제 2 피크 검출부(220)는 제 2 레이더(120)의 중간 주파수로부터 FFT 및 피크 검출을 통해

을 구한다. 여기서,

은 피크 인덱스로서, 스펙트럼에서 가장 큰 주파수를 갖는 지점으로, 수학식 3에 의해 구할 수 있다. 이때, 빈(bin)의 인덱스는 0부터 FFT 포인트 수만큼의 이산적인 값을 갖는다.The first and second and peak detection units 210 and 220 input the intermediate frequency (IF) from the first and second radars 110 and 120, respectively, and perform FFT and peak detection.

Find . That is, the first peak detection unit 210 detects the signal through FFT and peak detection from the intermediate frequency of the first radar 110.

is obtained, and the second peak detector 220 detects the signal through FFT and peak detection from the intermediate frequency of the second radar 120.

Find . here,

is the peak index, which is the point with the highest frequency in the spectrum and can be obtained by Equation 3. At this time, the index of the bin has discrete values ranging from 0 to the number of FFT points.

3.3. spectrum shift unit

에 해당하는 n과 n_center의 주파수 차이를 계산하고, 그 차이만큼 스펙트럼을 이동시킨다. 이때, 스펙트럼 시프트는 시간 도메인에서 이루어진다. 제 1 및 제 2 스펙트럼 시프트부(310, 320)는 수학식 10 내지 수학식 15에 따라 스펙트럼 시프트 과정을 수행한다. 먼저, 타겟의 스펙트럼 파워가 가장 큰 지점(

)과 n_center의 인덱스 차이를 구한다. 즉, 수학식 10과 같이 타겟의

과 n_center의 인덱스 차이로서 bin_diff를 구한다. 이어서, 이동시켜야 할 주파수를 구한다. 즉, 수학식 11의 f_diff는 인덱스 차이인 bin_diff에 bin 사이의 간격 주파수(fs/N)를 곱하여 이동시켜야 할 주파수를 구한다. 여기서, N은 FFT 포인트수를 의미하고, fs는 샘플링 주파수이다. 그리고, 수학식 12를 이용하여 스펙트럼 시프트된 신호를 구한다. 여기서, s_IF는 시간 도메인에서의 IF 신호이고, s_shift는 스펙트럼 시프트된 신호이다.The first and second spectrum shift units 310 and 320 have Equation 7:

Calculate the frequency difference between n and n _center , and move the spectrum by the difference. At this time, the spectrum shift occurs in the time domain. The first and second spectrum shift units 310 and 320 perform a spectrum shift process according to Equations 10 to 15. First, the point where the target's spectral power is greatest (

) and n _center . In other words, as shown in Equation 10, the target

Find bin _diff as the index difference between and n _center . Next, find the frequency to be moved. That is, f _diff in Equation 11 calculates the frequency to be moved by multiplying bin _diff , which is the index difference, by the interval frequency between bins (fs/N). Here, N means the number of FFT points, and fs is the sampling frequency. Then, the spectrum shifted signal is obtained using Equation 12. Here, s _IF is an IF signal in the time domain, and s _shift is a spectrum shifted signal.

Then, as shown in Equation 13, the signal is sampled by applying t=nT _s to Equation 12. t is a time variable, n is a sample index ranging from 0 to (N-1), and Ts is the sampling period.

If Equation 13 is rearranged, it can be expressed as Equation 14.

Equation 14 is expressed as Equation 15 due to frequency shift when the frequency is converted.

이 3이라면 bin_diff는 20-3=17이 된다. 수학식 15를 통과하면 피크 인덱스가 3인 스펙트럼이 bin_diff인 17만큼 오른쪽으로 이동하기 때문에 피크 인덱스가 20이 된다.At this time, S _shift and S _IF are Fourier transforms of s _shift and s _IF , respectively. After going through Equation 15, the peak of the S _IF signal comes to n _center . For example, n _center is 20,

If this is 3, bin _diff becomes 20-3=17. If Equation 15 is passed, the peak index becomes 20 because the spectrum with a peak index of 3 is shifted to the right by 17, which is bin _diff .

3.4.Deep Learning Department

The deep learning unit 400 inputs the frequency signal spectrally shifted by the first and second spectrum shift units 310 and 320 and infers the fractional distance using deep learning. Here, the deep learning unit 400 may be implemented as a deep neural network (DNN). DNN refers to a neural network structure that uses multiple hidden layers (layers other than the input/output layer), and one of its types is the Convolution Neural Network (CNN). Figure 7 shows the structure of the CNN used in the present invention by layer. The CNN structure includes an input layer, a convolution layer, a pooling layer, a fully connected neuron, and finally a flatten. After that, softmax and classification tasks are performed. As shown in Figure 7, the CNN input layer can use 256 data spectra with a maximum value normalized to 1. Convolution layers have 8, 16, and 32 channels for each layer. Each convolutional layer is a 1×1-sized mask overlapped with the number of channels. When input data of 256×1×1 size passes through a layer with 8 channels, a layer called a feature map of 256×1×8 size is created. It is output, and when it passes through a 16-channel convolution layer, 256×1×16 layer data is output. The deeper the convolutional layer, the more channels it has. After convolution, batch normalization is performed and the rectified linear unit (ReLU) activation function is passed. Batch normalization normalizes data to values from a standard normal distribution. ReLU is a type of activation function and is a filter that focuses on data with positive values by outputting positive values as is and treating negative values as 0. And Max pooling is a layer that groups data and extracts only the largest value in that group. In the present invention, stride-2 max pooling was used to select one out of two. The fc layer, that is, a fully connected layer structure, is connected to the output neuron after going through flattening. There are 11 output neurons, and the output neurons output label values for the probability of each label through softmax. Softmax and classification calculate the probability and error value (loss) of the label and finally output the label value of one of the output labels. Ultimately, the deep learning unit 400 is a CNN "inference" process in which a spectrum-shifted frequency signal is input and a label is generated, which is similar to a pattern recognition process.

The label has an integer value and represents k in Equation 9, which is converted to a decimal distance k through Equation 16.

Here, the trained area represents the smallest distance in the learning area, and the larger the label, the farther the target is from the radar. label _res represents the distance indicated by one label as label resolution, and the label resolution must be set to a narrower interval than the radar distance resolution.

Here, m is the total number of labels. Label resolution is the distance resolution divided by the total number of labels. Figure 8 shows where the target is located according to each label. If the target is located in the red square, the integer distance becomes n _center through spectral shift, and the decimal distance is obtained by multiplying the label obtained through inference by label_resolution. The learning process is carried out by a computer, and learning data is collected only in the blue trained area. The collection method placed the target on each label and measured the data 200 times each, and the CNN model learned the data for each label to infer the position of the decimal point. Meanwhile, the training data was created using MATLAB simulation. An FMCW radar with a range accuracy of 0.3m was used, the index of the center area (n _center ) was set to 20, and it had a trained area of ±0.15m. The interval between labels is 0.03m, and the learning location of the target and the label correspond 1:1. In general, the target location cannot be completely the same, so a random range equivalent to ±1.5% of the distance accuracy was added to change the measurement spectrum. In addition, Rician multipath, AWGN, and phase error are added to create data that takes into account distortions such as noise and multipath. The total data size was 200 pieces of data for each label, creating a total of 2,200 learning data sets. Additionally, in the present invention, the trained area ranges from 5.85 m to 6.15 m, and each label is arranged at intervals of 0.03 m. Therefore, it has the same effect as dividing the 0.3m interval into 10 equal parts from 5.85m to 6.15m, and 11 labels were set by index calculation. Ultimately, in order to increase the distance resolution without expanding the bandwidth of FMCW, the present invention trains the spectral pattern (learning data) corresponding to a fraction of the distance resolution together with the label information and infers the corresponding label with CNN.

3.5. Spectrum restoration unit

일 때, 스펙트럼 시프트를 통해서

이 되었기 때문에

를 더해줘서 기존의 정수배 거리로 복원해준다.The spectrum restoration unit 500 restores the fractional distance of the target to an integer multiple through spectral shift and deep learning. When the deep learning unit 400, that is, CNN, outputs one label for the spectral input, the decimal point distance has been inferred. Distance from existing target

When, through spectral shift

Because it has been

By adding , it is restored to the original integer distance.

3.6. target estimation unit

The target estimation unit 600 estimates the spatial coordinates of the target using the spectrum restored to an integer distance by the spectrum restoration unit 500. When the spectrum is restored to an integer distance through the spectrum restoration unit 500, the present invention is approximately completed, and the distance accuracy of another radar is also improved through the same method. The two distance information with improved distance accuracy according to the present invention can be obtained by Equations 7 and 8.

Meanwhile, although not shown, a spectrum normalization process may be performed after shifting the spectrum through the spectrum shift units 310 and 320 and before learning it through the deep learning unit 400. That is, a spectrum normalization unit (not shown) may be provided between the spectrum shift units 310 and 320 and the deep learning unit 400. The relationship between distance and signal strength can be obtained as Equation 18.

Here, P _r is the intensity of the received signal, P _t is the intensity of the transmitted signal, G is the antenna gain, λ is the wavelength of the transmitted signal, σ is the reflection area of the target, and R _max is the maximum detection distance. According to Equation 18, the size of the spectrum of the radar signal gradually decreases as the distance increases. However, if the peak index moves due to spectral shift, a problem may occur where the spectral sizes of the two signals, the data used for learning and the data to be inferred, do not match. Therefore, the spectrum must undergo a power normalization process after undergoing spectral shift. Power normalization divides the magnitude of the peak index across the spectrum. Through this work, it is possible to create a spectrum in which the value of the largest signal is 1, regardless of n, and the increase/decrease trend of adjacent indices is maintained.

4. Method of the present invention

9 and 10 are flowcharts illustrating a method for improving the distance resolution of an FMCW radar using deep learning according to an embodiment of the present invention. That is, FIG. 9 is a flowchart of a method for improving the range resolution of an FMCW radar using deep learning according to an embodiment of the present invention, and FIG. 10 is a flowchart for explaining the spectrum shift process. Here, since the method for improving distance resolution of the present invention uses the FMCW radar using deep learning of FIG. 6, the configuration of the FMCW radar is explained process by process. The description of redundant content is omitted and the approximate process is described as follows.

을 각각 구하는 과정(S200)과, 제 1 및 제 2 스펙트럼 시프트부(310, 320)를 이용하여

에 해당하는 n과 n_center의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 이동시키는 과정(S300)과, 딥러닝부(400)를 이용하여 제 1 및 제 2 스펙트럼 시프트부(310, 320)에 의해 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝(deep learning)을 이용하여 소수배 거리를 추론하는 과정(S400)과, 스펙트럼 복원부(500)를 이용하여 스펙트럼 시프트 및 딥러닝을 통한 타겟의 소수배 거리를 정수배로 복원하는 과정(S500)과, 타겟 추정부(600)를 이용하여 복원된 스펙트럼을 이용하여 타겟의 공간 좌표를 추정하는 과정(S600)를 포함할 수 있다. 또한, 도시되지 않았지만, 스펙트럼 시프트 과정(S300)을 거친 후 딥러닝 과정(S400) 이전에 파워 정규화 과정을 더 포함할 수 있다.As shown in FIG. 9, the method of improving the distance resolution of the FMCW radar according to an embodiment of the invention uses the first and second radars 110 and 120 to transmit a transmitted signal radiated to the target and a received signal reflected from the target. A process (S100) of generating an intermediate frequency signal (IF) using the first and second peak detectors 210 and 220, respectively, and an intermediate frequency (IF) of each of the first and second radars 110 and 120 using the first and second peak detectors 210 and 220. ) through FFT and peak detection

Using the process of obtaining each (S200) and the first and second spectrum shift units 310 and 320,

A process of calculating the frequency difference between n and n _center corresponding to and shifting the spectrum by the difference (S300), and the first and second spectrum shift units 310 and 320 using the deep learning unit 400. A process of inputting a spectrum shifted frequency signal and inferring the decimal distance using deep learning (S400), and the decimal multiple distance of the target through spectrum shifting and deep learning using the spectrum restoration unit 500. It may include a process of restoring to an integer multiple (S500) and a process of estimating the spatial coordinates of the target using the spectrum restored using the target estimation unit 600 (S600). In addition, although not shown, a power normalization process may be further included after the spectrum shift process (S300) and before the deep learning process (S400).

)과 n_center의 인덱스 차이로서 bin_diff를 구하는 과정(S310)과, 수학식 11과 같이 인덱스 차이인 bin_diff에 bin 사이의 간격 주파수(fs/N)를 곱하여 이동시켜야 할 주파수 f_diff를 구하는 과정(S320)과, 수학식 12를 이용하여 스펙트럼 시프트된 신호를 구하는 과정(S330)과, 수학식 13과 같이 수학식 12에 t=nT_s를 적용해서 신호를 샘플링하는 과정(S340)과, 수학식 14를 주파수 변환하여 수학식 15와 같이 주파수 이동시키는 과정(S350)을 포함할 수 있다.In addition, as shown in FIG. 10, the spectrum shift process (S300) is performed at the point where the target's spectral power is greatest as shown in Equation 10 (

) and the process of calculating bin _diff as the index difference between n _center (S310), and calculating the frequency f _diff to be moved by multiplying bin _diff , which is the index difference, by the frequency of the interval between bins (fs/N), as shown in Equation 11. (S320), a process of obtaining a spectrum-shifted signal using Equation 12 (S330), a process of sampling the signal by applying t=nT _s to Equation 12 as in Equation 13 (S340), and mathematics It may include a process (S350) of frequency converting Equation 14 and shifting the frequency as shown in Equation 15.

Figure 11 is a schematic diagram illustrating a method of obtaining the location of a target using an FMCW radar using deep learning to improve distance resolution in an embodiment of the present invention. First, the first and second radars are separated by dx, and each radar estimates the target's location through peak detection. As a result, the radar measures the actual location of the target in FIG. 11 as the estimated location of the target (est. target) due to a fractional error. In other words, there is a difference between the target's actual location and the estimated location due to the radar's fractional error.

로 나타낼 수 있는데, n_i의 값을 n_center로 옮겨오는 스펙트럼 시프트를 진행한다. 그러면 도 11에서 보이는 것처럼 타겟의 추정 위치가 화살표를 따라서 추론 영역의 중심점으로 들어온다. 이에 소수배 오차만큼 떨어져 있던 타겟의 실제 위치도 같이 화살표를 따라 추론 영역 안으로 들어오게 된다. 타겟의 추정 위치는 피크 검출(peak detection)로 구한 점이고, 타겟의 실제 위치는 스펙트럼 bin 사이에 숨어있는 것을 표현한 것으로 타겟의 실제 위치를 구하는 것이 목적이다. 이후 학습된 인공지능 영역에 의해서 소수배 오차인 ki를 추론한다. When the distances between the first and second radars and the target are d1 and d2, respectively, d1 and d2 are calculated according to Equation 9.

It can be expressed as, and spectral shift is performed to move the value of n _i to n _center . Then, as shown in Figure 11, the estimated position of the target follows the arrow and enters the center point of the inference area. Accordingly, the actual location of the target, which was separated by a fractional error, also follows the arrow and enters the inference area. The estimated position of the target is obtained through peak detection, and the actual position of the target is expressed as what is hidden between spectral bins, and the purpose is to obtain the actual position of the target. Afterwards, ki, the fractional error, is inferred based on the learned artificial intelligence domain.

Figure 12 is an example diagram of spectral shift according to an embodiment of the present invention. Figure 12(a) is a spectrum measured for a target of 3.2m+k, and Figure 12(b) is a spectrum shift of Figure 12(a), and it can be seen that the surrounding bin values, including the maximum value, are maintained as is. there is.

을 각각 구하고, 제 1 및 제 2 스펙트럼 시프트부(310, 320)를 이용하여

에 해당하는 n과 n_center의 주파수 차이를 계산하고 그 차이만큼 스펙트럼을 이동시킨 후 딥러닝부(400)를 이용하여 스펙트럼 시프트된 주파수 신호를 입력하고 딥러닝(deep learning)을 이용하여 소수배 거리를 추론하고, 스펙트럼 복원부(500)를 이용하여 스펙트럼 시프트 및 딥러닝을 통한 타겟의 소수배 거리를 정수배로 복원한 후 타겟 추정부(600)를 이용하여 복원된 스펙트럼을 이용하여 타겟의 공간 좌표를 추정한다. 즉, 본 발명은 소정 거리 이격된 제 1 및 제 2 레이더(110, 120)의 중간 주파수를 피크 검출하여

를 구하고, 스펙트럼 이동시킨 후 딥러닝을 통해 소수배 거리를 구하며, 소수배 거리를 정수배 거리로 복원한 후 타겟의 공간 좌표를 추정한다. 따라서, 본 발명은 스펙트럼 시프트 후 딥러닝을 수행함으로써 모든 거리에 대해서 스펙트럼을 학습시키지 않아도 거리 해상도를 향상시킬 수 있고, 적은 라벨을 이용하여 인공지능 구조의 복잡도를 감소시킬 수 있다.As described above, the apparatus and method for improving the range resolution of an FMCW radar using deep learning according to an embodiment of the present invention uses the first and second radars 110 and 120 spaced apart by a predetermined distance to generate an intermediate frequency signal (IF). After generating each, using the first and second peak detectors 210 and 220, FFT and peak detection are performed from the intermediate frequencies (IF) of the first and second radars 110 and 120, respectively.

are obtained respectively, and using the first and second spectrum shift units 310 and 320,

Calculate the frequency difference between n and n _center , move the spectrum by the difference, input the spectrum-shifted frequency signal using the deep learning unit 400, and use deep learning to calculate the decimal multiple distance. Infer, use the spectrum restoration unit 500 to restore the fractional distance of the target to an integer multiple through spectrum shift and deep learning, and then use the spectrum restored using the target estimation unit 600 to determine the spatial coordinates of the target. Estimate . That is, the present invention peak detects the intermediate frequencies of the first and second radars 110 and 120 spaced apart by a predetermined distance.

After calculating and moving the spectrum, the decimal distance is obtained through deep learning, and the spatial coordinates of the target are estimated after restoring the decimal distance to an integer distance. Therefore, by performing deep learning after spectrum shifting, the present invention can improve distance resolution without having to learn the spectrum for all distances, and reduce the complexity of the artificial intelligence structure by using fewer labels.

The technical idea of the present invention as described above has been described in detail according to the above-mentioned embodiments, but it should be noted that the above-described embodiments are for explanation and not limitation. Additionally, those skilled in the art will understand that various embodiments are possible within the scope of the technical idea of the present invention.

110, 120: 1st and 2nd radar
210, 220: first and second peak detection units
310, 320: first and second spectrum shift units
400: Deep learning unit 500: Spectrum restoration unit
600: Target estimation unit

Claims (19)

First and second radars that respectively generate intermediate frequency signals using a transmitted signal radiated to the target and a received signal reflected from the target;
Peak index through peak detection from the intermediate frequencies of each of the first and second radars (

) first and second peak detection units respectively calculating;
The peak index (

) Calculate the frequency difference between an integer multiple (n) and the constant value (n _center ) and shift the spectrum by the difference;
A deep learning unit that inputs a spectrum-shifted frequency signal and infers the decimal distance of the target using deep learning; and
An FMCW radar using deep learning including a spectrum restoration unit that restores the fractional distance of the target to an integer multiple.
The FMCW radar using deep learning according to claim 1, further comprising a target estimation unit that estimates spatial coordinates of the target using the spectrum restored through the spectrum restoration unit.
The method of claim 1 or claim 2, wherein each of the first and second radars,
A transmitting and receiving antenna that radiates a transmission signal of an FMCW waveform and receives a signal reflected from the target,
A mixer for mixing the transmitted signal and the received signal,
FMCW radar using deep learning that includes a fast Fourier transformer that generates an intermediate frequency signal by fast Fourier transforming the mixing signal.
The method of claim 3, wherein the first and second spectrum shift units of the target

Find the index difference (bin _diff ) between n center and n _center as shown in Equation 10, find the frequency to be moved (f _diff ), find the spectrum shifted signal, sample the spectrum shifted signal, and then move the frequency by converting the frequency. FMCW radar using deep learning.
[Equation 10]

The method of claim 4, wherein the frequency to be moved (f _diff ) is a dip calculated as in Equation 11 by multiplying the index difference (bin _diff ) calculated from Equation 10 by the interval frequency between bins (fs/N). FMCW radar using running.
[Equation 11]

Here, N refers to the number of FFT points, and fs is the sampling frequency.
The FMCW radar using deep learning according to claim 5, wherein the spectrum shifted signal is obtained using Equation 12.
[Equation 12]

Here, s _IF is an IF signal in the time domain, and s _shift is a spectrum shifted signal.
The method of claim 6, wherein the signal sampling is obtained as Equation 13 by applying t=nT _s to Equation 12. FMCW radar using deep learning.
[Equation 13]

Here, t is a time variable, n is a sample index ranging from 0 to (N-1), and Ts is the sampling period.
The method of claim 7, wherein the frequency shift is obtained by converting Equation 13 to Equation 14 and then frequency converting to Equation 15. FMCW radar using deep learning.
[Equation 14]

[Equation 15]

Here, S _shift and S _IF are the Fourier transforms of s _shift and s _IF , respectively.
The method of claim 8, wherein the deep learning unit includes an input layer, a convolution layer, a pooling layer, and a fully connected layer, and deep learning uses a CNN structure that performs softmax and classification tasks. FMCW radar using.
The FMCW radar using deep learning according to claim 9, wherein the deep learning unit is configured to input a spectrum-shifted frequency signal and output a label, and is converted to a decimal distance k through Equation 16.
[Equation 16]

Here, the trained area represents the smallest distance in the learning area, and label _res is the label resolution, which represents the distance represented by one label.
The FMCW radar using deep learning according to claim 10, wherein the label resolution is the distance divided by the distance resolution by the total number of labels, as shown in Equation 17.
[Equation 17]

Here, m is the total number of labels and Rres is the distance resolution.
The method of claim 11, wherein the spectrum restoration unit determines that the distance to the target is

When, through spectral shift

Because this happens

FMCW radar using deep learning that restores the distance to an integer multiple by adding .
A process of generating an intermediate frequency signal using a transmitted signal radiated to a target and a received signal reflected from the target;
Peak index through peak detection from the intermediate frequency (

) process of calculating each;
The peak index (

A process of calculating the frequency difference between an integer multiple (n) corresponding to ) and a certain value (n _center ) and shifting the spectrum by the difference;
A process of inputting a spectrally shifted frequency signal and inferring the decimal distance of the target using deep learning;
A method of improving the distance resolution of an FMCW radar using deep learning, including the process of restoring the decimal multiple distance of the target to an integer multiple.
The method of claim 13, further comprising the step of estimating the spatial coordinates of the target using the restored spectrum.
The method of claim 14, wherein the process of generating the intermediate frequency uses deep learning to radiate a transmitted signal, receive a signal reflected from the target, mix the transmitted signal and the received signal, and perform fast Fourier transform to generate the intermediate frequency. How to improve the range resolution of FMCW radar.
The method of claim 15, wherein the spectrum shift process,
From equation 10, the target

The process of calculating the index difference (bin _diff ) between and n _center ,
A process of multiplying the index difference (bin _diff ) by the interval frequency between bins (fs/N) to obtain the frequency (f _diff ) to be moved as shown in Equation 11,
The process of obtaining a spectrum shifted signal using Equation 12,
A process of sampling a spectrum-shifted signal as shown in Equation 13 by applying t=nT _s to Equation 12,
A method of improving the distance resolution of an FMCW radar using deep learning, which includes converting Equation 13 to Equation 14, then converting the frequency, and moving the frequency as shown in Equation 15.
[Equation 10]

[Equation 11]

Here, N refers to the number of FFT points, and fs is the sampling frequency.
[Equation 12]

Here, s _IF is an IF signal in the time domain, and s _shift is a spectrum shifted signal.
[Equation 13]

Here, t is a time variable, n is a sample index ranging from 0 to (N-1), and Ts is the sampling period.
[Equation 14]

[Equation 15]

Here, S _shift and S _IF are the Fourier transforms of s _shift and s _IF , respectively.
The method of claim 16, wherein the process of inferring the decimal distance of the target using deep learning is a process in which a spectrum-shifted frequency signal is input and a label is output, and deep learning is converted to the decimal distance k through Equation 16. Method for improving range resolution of FMCW radar using .
[Equation 16]

Here, the trained area represents the smallest distance in the learning area, and label _res represents the label resolution, which is the distance represented by one label.
The method of claim 17, wherein the label resolution is the distance obtained by dividing the distance resolution by the total number of labels as shown in Equation 17.
[Equation 17]

Here, m is the total number of labels and Rres is the distance resolution.
The method of claim 18, wherein the process of restoring the decimal distance of the target to an integer multiple includes:

When, through spectral shift

Because this happens

A method of improving the distance resolution of FMCW radar using deep learning, which restores the distance to an integer multiple by adding .

Images