SA radar sensor for automobiles

Abstract

To provide a radar sensor that needs only a small amount of memory space for storing coefficients of a transform function.SOLUTION: The radar sensor includes a high-frequency unit (10) and an electronic analysis unit (18), and has a transform module (32) configured to transform raw data by applying a transform function defined by N coefficients (c(n, k)), and a coefficient module (34) for providing the coefficients for the transform module. The coefficient module (34) includes a memory (36), in which an initial set of coefficients including less than N coefficients is stored, and a recursion module (38) for recursive calculation of the remaining coefficients.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This invention relates to a radar sensor for automobiles. [Background technology]

[0002] Radar systems for measuring the distance, relative speed, and angle of objects (such as vehicles and obstacles) are increasingly being used in automobiles for safety and comfort functions. In the automotive sector, the use of radar sensors with synthetic apertures (SAs) has been studied for several years. The principle of synthetic apertures allows for particularly accurate angle measurements by interpreting measurements at multiple different local locations as a synthetic antenna aperture (antenna plane) as the radar sensor moves and processing them accordingly. Synthetic apertures are realized by mathematically treating the transmitting and receiving antennas, which are at different local locations at the time of each measurement due to the radar's movement, as if there were a large antenna aperture along the travel trajectory. This makes it possible to achieve angle measurement resolutions at each transmitting and receiving antenna that would not be achievable with existing real-world antenna apertures.

[0003] Modern automotive radar systems typically use FMCW modulation (Fast Chirp Modulation) with high-speed ramps, where multiple linear frequency ramps (frequency-modulated pulses) with the same gradient are transmitted sequentially. By mixing the transmitted signal at each point in time with the received signal, a low-frequency signal (called the beat frequency) with a frequency proportional to the distance is generated. The system is usually designed to ignore the component of the beat frequency caused by the Doppler shift during the relative motion of objects in the radial direction. In this case, the distance information obtained is nearly unique. Subsequently, the Doppler shift (and thus the radial velocity) can be determined by observing the temporal change in the phase of the complex distance signal across the ramps. Typically, the distance and velocity are determined independently of each other using a two-dimensional fast Fourier transform (FFT).

[0004] SA radar can also use the measurement principle of high-speed chirp modulation. Distance analysis remains largely unchanged. Here, the Doppler analysis across the ramp is replaced by SA analysis. As a final result, this provides angle measurement assuming a stationary target and recognizing the movement of the vehicle, rather than relying on Doppler measurement. For this purpose, the relative velocity of the object obtained from Doppler measurement is analyzed, which involves projecting the vehicle's velocity onto the line-of-sight axis to the object.

[0005] Regarding data analysis using the synthetic aperture principle, it is necessary for the vehicle itself to move in order for the synthetic aperture to be generated. However, this movement of the vehicle causes the stationary object being measured for angle measurement to undergo apparent positional changes (migration) during the measurement cycle, which distorts the radar image and makes it difficult to calculate the distance-velocity radar image.

[0006] Several algorithms for compensating for such migration effects are known in the literature. Conceptually, two classes of algorithms can be distinguished: (a) algorithms that can handle arbitrary synthetic apertures (e.g., back projection) at the cost of higher computational complexity, and (b) others (e.g., keystone algorithms) that are limited to a specific aperture type (e.g., linear) but are more computationally efficient. In automotive applications, where real-time processing is required, the efficient computation of distance-velocity radar images and the resulting distance-angle radar images is extremely important.

[0007] To achieve high computational efficiency and the resulting real-time processing, conventional automotive radar sensors utilize accelerators for FFT calculations. For example, the core of the keystone algorithm is the so-called chirp Z transform, which can be viewed as a convolution of two functions, and an FFT accelerator can also be used for its calculation. To do this, the coefficients of the chirp Z transform included in the fast convolution must either be pre-calculated and written to memory, or they must be calculated online. The pre-calculation method requires additional memory, while the online calculation method requires complex exponential calculations in real time. [Overview of the project] [Problems that the invention aims to solve]

[0008] The present invention relates to an automotive radar sensor comprising a high-frequency unit configured to transmit a sequence of modulated radar pulses and receive corresponding radar echoes, and an electronic analysis unit configured to measure distance and angle according to the principle of synthetic aperture (SA). A sampling module configured to sample the amplitude of a received radar echo as raw data in a two-dimensional data space, wherein in the two-dimensional data space, one dimension represents the temporal change in amplitude within a radar pulse, and the other dimension represents the change in amplitude between pulses. An FFT module with dedicated hardware for performing the Fast Fourier Transform (FFT) to compute a 2D distance-velocity radar image, A transformation module configured to convert raw data into a format processable by an FFT module, while simultaneously correcting migration effects, by applying a transformation function defined by N coefficients. A coefficient module for supplying coefficients related to the conversion module and It is equipped with these features.

[0009] The object of the present invention is to provide a radar sensor that can supply the coefficients necessary for real-time calculation of radar images quickly enough, even though it requires only a small amount of memory area to store the coefficients of the conversion function. [Means for solving the problem]

[0010] This problem is solved by the present invention, where the coefficient module is A memory that stores an initial set of coefficients containing fewer than N coefficients, A recursive module to recursively calculate the remaining coefficients and This is resolved by having [something].

[0011] This invention takes advantage of the fact that the necessary coefficients of a transformation function can be recursively calculated. Since the raw data to which the transformation is applied is defined in a two-dimensional data space, the transformation function is also defined in a two-dimensional space, and therefore its coefficients c(n,k) are functions of two indices n and k. Thus, by storing, for example, the coefficients c(0,k) as an initial set, the remaining coefficients c(n,k) (where n>0) can be calculated based on the recurrence relation c(n,k)=f(c(n-1,k)). It is found that this method can calculate the coefficients significantly faster than conventional algorithms that calculate the coefficients for each pair of indices (n,k) individually and independently. In this way, it is possible to supply coefficients at the speed required to perform real-time transformations. The number of coefficients n, i.e., the size of the data space in the first dimension, is N. fast This is expressed as follows, and the number of coefficients k in the second dimension is N slow Expressed as, the memory area required for the coefficients in the method according to the present invention is N fast ×N slow It is not proportional to N fast or N slow It is proportional to one of the following: N fast and N slowThe number needs to be relatively large, such as 256 or 512, so that the SA analysis can be performed with sufficient accuracy. Moreover, since the coefficients are complex numbers, a significant amount of memory can be saved by recursively calculating the coefficients.

[0012] The present invention is not limited to data analysis by the keystone algorithm, and can generally be applied to all analysis algorithms that can recursively calculate the coefficients of the conversion function. Similarly, the present invention is not limited to high-speed chirp FMCW radars, but can generally be applied to radar sensors equipped with a modulated pulse sequence of transmitted signals, such as a phase-modulated pulse sequence.

[0013] Advantageous and developed forms of the present invention are presented in the dependent claims.

[0014] When the conversion of raw data is a mathematical convolution, dedicated hardware for fast Fourier transform existing in the radar sensor can also be used to calculate the conversion defined by the coefficients. This is done, for example, by converting the raw data and the conversion function from the time domain to the frequency domain by FFT, then multiplying the functions in the frequency domain, and then converting back to the time domain by inverse FFT, and then performing a two-dimensional Fourier transform in the FFT module to calculate the range-velocity radar image.

[0015] The initial set of coefficients to be stored does not have to be a single vector (having indices n or k as components), and for example, two or more such vectors can also be stored. N fast or N slow When N or N is very large, since the calculated coefficients form a plurality of short recursive sequences instead of a single very long sequence, there is an advantage that the accumulation of rounding errors is suppressed.

[0016] When the matrix of coefficients n, k forms one or more blocks, it may also be suitable to start the recursion with respect to the columns or rows at the center of each block in the initial set of stored coefficients, and then proceed in opposite directions within the block in two separate recursive sequences. Thereby, the length of the recursive sequence is halved, and thus not only the rounding error is reduced, but at the same time, the larger rounding error occurs at most only at the edges of the block, where in any case the data is reduced by the window function, and as a result, it is achieved that the error does not have too strong an influence.

[0017] Hereinafter, exemplary embodiments will be described in more detail based on the drawings.

Brief Description of the Drawings

[0018] [Figure 1] It is a block diagram of the SA radar sensor according to the present invention. [Figure 2] It is a timing diagram of the pulse sequence transmitted by the radar sensor. [Figure 3] It is a flowchart regarding the chirp Z transform. [Figure 4] It is a diagram showing a recursive scheme for calculating the coefficients of the chirp Z transform. [Figure 5] It is a diagram showing an example regarding a modified recursive scheme.

Mode for Carrying Out the Invention

[0019] The radar sensor shown in FIG. 1 has a high-frequency section 10, and the high-frequency section 10 is configured as, for example, a fast chirp FMCW radar, and in each measurement cycle, a sequence of frequency-modulated radar pulses or ramps is sent out via an array of antennas 12. The radar signal reflected from the object 14 is received again by the antenna 12 and mixed with the components of the signal transmitted at each point in time as is usually done with an FMCW radar, and as a result, a beat signal having a clearly low frequency (beat frequency) is obtained.

[0020] The analog-to-digital converter stage 16 forms an interface between the high-frequency section 10 and the analysis section 18. There, the digital complex amplitude of the beat signal is sampled at regular time intervals and stored as a time signal. Data storage takes place in a two-dimensional data space. That is, the amplitude A(n,k) is stored as a function of the "fast" index n and the "slow" index k.

[0021] Figure 2 schematically shows frequency-modulated pulses 22 transmitted by a high-frequency portion, also called a ramp or "chirp," on a frequency-time graph with respect to two consecutive measurement cycles. The vertical axis plots the frequency f, and the horizontal axis plots the time t. Within each chirp, the frequency f increases linearly. The center frequency of the pulse is indicated by fc, and the bandwidth is indicated by B. The fast index n counts the consecutive sampling times 24 within each pulse 22, and the slow index k counts the consecutive pulses within each measurement cycle. The number of sampling times 24 within each pulse is N. fast As shown, the number of pulses 22 in each measurement cycle is N slow This is shown by N. fast and N slow The number is clearly larger than in the simplified figure. Typical values ​​are, for example, N fast =256 and N slow = 512.

[0022] The analysis unit 18 (Figure 1) further includes an FFT module 26 for Fast Fourier Transform (FFT). The hardware of this module is designed to perform the Discrete Fourier Transform particularly quickly and efficiently. As is typically done in fast chirp FMCW radars, the FFT module performs a one-dimensional or two-dimensional Fourier Transform, in which case one dimension is a sequence of sampling times counted by index n.

[0023] In a typical high-speed char radar (without SA analysis), the amplitude A(n,k) stored in the sampling module is transmitted directly to the FFT module. The two-dimensional spectrum generated by the FFT module then represents a distance-velocity radar image 28, in which each object 14 is characterized by a peak 14' in the radar echo, and its position in the two-dimensional distance-velocity space indicates the object's distance d and its relative velocity v. Since the frequency ramp of pulse 22 is very steep, the Doppler shift in the pulse due to the relative motion of the object can be ignored, and as a result, the peak position in the first dimension provides a good approximation of the object's distance d. The relative velocity v of the object is obtained from the phase change of the signal between pulses measured at the same sampling time, and is therefore acquired by a two-dimensional Fourier transform.

[0024] However, in the SA radar described here, object 14 is not a vehicle traveling ahead whose distance and relative velocity should be measured, or at least not a vehicle traveling ahead, but rather, above all, a stationary object on the roadside whose precise position (distance and angle) should be measured. In Figure 1, vector Vf also indicates the movement of the vehicle on which the radar sensor is installed, and therefore the movement of the radar sensor itself. The line of sight from the radar sensor to object 14 forms an angle φ with vector Vf (direction of travel). Due to the movement of the radar sensor itself, a radial velocity Vr = -Vfcos(φ) is measured with respect to object 14, even though object 14 is stationary. Therefore, by solving this equation for φ, the measured radial velocity Vr of object 14 can be converted to the azimuth angle φ from which the object is seen, and as a result, the distance-velocity radar image 28 can also be interpreted as a distance-angle radar image 30, where the sign φ (object on the right or left side of the vehicle) can be determined based on the phase difference between signals received from various antennas 12.

[0025] However, the apparent positional change of object 14 during the measurement cycle results in a migration effect that leads to distortion of the distance-velocity radar image 28. To correct this distortion, a transformation module 32 is inserted between the sampling module 20 and the FFT module 26. The transformation module 32 performs a transformation on the amplitude A(n,k) stored in the sampling module 20 to correct for the migration effect, according to a specific algorithm, such as the keystone algorithm. Therefore, the FFT module 26 does not directly obtain the amplitude A(n,k) as input data, but rather obtains the transformed (migration-free) amplitude T(n,k). Furthermore, in the transformation module 32, a Fourier transform is already performed in the dimension corresponding to the sequence of pulses counted by index k.

[0026] The transformation that corrects migration effects is defined by a set of coefficients c(n,k), where c(n,k) depends on indices n and k.

[0027] For example, in the case of the keystone algorithm, the following holds true:

[0028]

number

[0029] In principle, these coefficients c(n,k) only need to be calculated once for each pair of n,k indices, and then stored in the memory of the analysis unit 18. However, in this case, N fast ×N slow Memory space is required for the complex coefficients (131,072 in the example shown). If the radar sensor has various operating modes with different center frequencies fc, for example (to avoid interference with radar sensors in other vehicles), the required memory space will increase by several times accordingly.

[0030] On the other hand, when calculating each individual coefficient according to the above formula as needed, a large number of very complex calculations must be performed during each measurement cycle, resulting in the need for a computer with high computing power.

[0031] However, the following recursive expression can be derived from equation (1) above.

[0032]

number

[0033] The constant D(k) only needs to be calculated once for each k and then stored. Furthermore, once the initial set of coefficients is stored, all coefficients can be calculated recursively, requiring only a simple multiplication for each increment of n and for each index k.

[0034]

number

[0035] The memory area is N slow Since it is only needed for the initial value c(0,k) and the constant D(k), the required memory area is clearly reduced.

[0036] In this way, a favorable compromise between memory requirements and computing power can be reached, resulting in a significant reduction in overall hardware costs.

[0037] As shown in Figure 1, the analysis unit 18 includes a coefficient module 34 that supplies coefficients c(n,k) related to the conversion module 32. The coefficient module provides N at index k. slowThe system includes a memory 36 that stores an initial value c(0,k) and an associated coefficient D(k) for each of the values ​​of n, and a recursive module 38 for recursively calculating the coefficient c(n,k) for n>0. When the recursive module 38 calculates a set of coefficients c(n-1,k) and supplies it to the conversion module 32, this set is also stored in the registers of the recursive module and then used to calculate the next set of coefficients c(n,k).

[0038] The operating modes of the transformation module 32 are shown in more detail in Figure 3. The complex amplitude A(n,k) is a function of time, more precisely, a function of fast time (index n) and slow time (index k). The chirp Z transform is mathematically a convolution of this time-dependent function with another time-dependent function that can be described as a phase factor exp(iψ) having a time-dependent phase ψ. Instead of performing this convolution directly by numerical integration, a mathematically equivalent operation is performed. This essentially involves the Fourier transform of both functions into the frequency domain, the multiplication of the two frequency-dependent functions, and then the inverse transform into the time domain. For this purpose, first, the complex amplitude A(n,k) for each pair of indices is multiplied by the phase factor exp(iψ) which depends on these indices n and k. Then, this product is subjected to the complex Fourier transform cFFT in the "slow" dimension (index k) in the FFT stage 40. slow This is performed. In parallel with this, the same Fourier transform is again applied to the phase factor exp(iψ) in a further FFT stage 42. In the frequency domain, the transformed functions are multiplied together, and then the inverse Fourier transform cIFFT is applied in a further FFT stage 44. slow The result is then converted back to the time domain using the specified method. The result is then multiplied again by the phase factor exp(iψ) for each pair of indices n,k. The product thus obtained provides the transformed amplitude T(n,k), which are then supplied as input data to the FFT module 26 for a Fourier transform in the dimension given by index n.

[0039] The procedure shown in Figure 3 has the advantage that the Fourier transforms in FFT stages 40, 42, and 44 can be performed very quickly and efficiently using dedicated hardware. In some cases, the hardware of the FFT module 32 can also be used for this purpose. In this case, the remaining operations performed in the transformation module 32 are simple multiplications that can be performed much faster than numerical convolution operations.

[0040] The multiplication with the phase factor exp(iψ) is performed in the transformation module 32 by first calculating the product over all values ​​of index k with respect to a fixed value of index n (e.g., n=0), and then moving on to the next highest value of index n. The coefficient module 34 can then supply the coefficients c(n,k) necessary for calculating the phase factor in the order required for multiplication with the phase factor. In the FFT stage 42, integration with respect to index k is also performed with respect to a fixed n. Therefore, the recursive calculation of the coefficients in the coefficient module 34 only needs to be performed once per measurement cycle.

[0041] The order in which calculations are performed in the transformation module 32 for multiple different values ​​of n is essentially arbitrary. Therefore, it is not necessary to start the recursion in the coefficient module 34 from n=0. For example, as schematically shown in Figure 4, fast It is also possible to start with a value of n close to / 2 and then proceed through two recursive sequences to smaller n and larger n. For one thing, this is possible if each of the two recursive sequences goes from n=0 to n=N fast It has the advantage of being only half the length of the previous sequence. Since rounding errors, which are unavoidable during computation, tend to accumulate during the recursive process, shortening the recursive sequence reduces the accumulation of errors.

[0042] A further advantage of this procedure is that in the FFT module 26, the transformed amplitude (T(n,k)) is typically multiplied by a window function, and the window function is applied to the edges of the target time interval (n=0 and n=N). fastThis is obtained by suppressing the amplitude at ). This windowing helps mitigate the by-products that arise from the fact that the transformation can only be performed over a finite time interval. Here, as the recursion of coefficient calculations also proceeds from the center of the time interval toward the edge, we gain the advantage that the accumulated error is also suppressed by the window function at the edge of this interval.

[0043] One way to further suppress the accumulation of errors is to divide the numerical range of index n into several blocks, and then perform recursion on each block, preferably also progressing from the center to the edges, thereby obtaining a further shortening of the recursive sequence.

[0044] Figure 5 shows an example of a recursive scheme that divides the numerical range for index n(0-255) into two blocks: 0-128 and 129-255. In this case, two sets of initial coefficients c(96,k) and c(160,k) are stored in memory 36. The recursion then proceeds from these initial values ​​to the edges of each block. However, the initial values ​​at n=96 and n=160 are not at the center of their respective blocks, but are shifted towards the center of the entire numerical range. Therefore, the recursive sequence proceeding to n=128 or n=129 is shorter than the recursive sequence proceeding to the outer edges n=0 and n=255. As a result, less error accumulates at the center of the numerical range than at the outer edges where errors are further suppressed by windowing.

[0045] Furthermore, errors can be reduced by storing multiple sets of constants instead of a single set D(k), where these sets are pre-calculated precisely for different step sizes, i.e., different increments of index n. For example, one set D_1(k) can be used for increments of length 1, and an additional set D_10(k) for increments of length 10. As a result, iterative calculations can be performed every 10 times using D_10(k) with iterative calculations using D_1(k) in between. This reduces the number of iterations to one-tenth. The number of sets D_i(k) used and their scaling can be arbitrarily selected, depending on the precision requirements, numerical representation, and available memory.

Claims

1. It is a radar sensor, A radar sensor for an automobile comprising a high-frequency unit (10) configured to transmit a sequence of modulated radar pulses (22) and receive corresponding radar echoes, and an electronic analysis unit (18) configured to measure distance and angle according to the principle of synthetic aperture (SA), A sampling module (20) configured to sample the amplitude (A(n,k)) of the received radar echo as raw data in a two-dimensional data space, wherein in the two-dimensional data space, one dimension represents the temporal change of the amplitude within a radar pulse (22), and the other dimension represents the change of the amplitude between pulses. An FFT module (26) has dedicated hardware for performing a Fast Fourier Transform (FFT) to compute a two-dimensional distance-velocity radar image (28), A conversion module (32) is configured to convert the raw data into a format that can be processed by the FFT module (26) while simultaneously correcting migration effects by applying a conversion function defined by N coefficients (c(n,k)), A coefficient module (34) for supplying the coefficients related to the conversion module and Equipped with, The coefficient module (34) is A memory (36) that stores an initial set of coefficients containing fewer than N coefficients, A recursive module (38) for recursively calculating the remaining coefficients and A radar sensor characterized by having [a certain feature].

2. The radar sensor according to claim 1, wherein the coefficient (c(n,k)) is a function of two integer indices n and k, the memory (36) includes at least one initial set of coefficients, the coefficients correspond to a fixed value of index n and all values ​​of index k.

3. The radar sensor according to claim 2, wherein the numerical range of index n is divided into a plurality of blocks, the memory (36) includes an initial set of coefficients for each block, and the recursive module (38) is configured to perform individual recursion for each block.

4. The radar sensor according to claim 3, wherein the recursive module (38) is configured to compute two recursive sequences with respect to the entire numerical range of index n or for each block, the two recursive sequences starting from a value of n within the numerical range or block and progressing in opposite directions toward the edge of the numerical range or block.

5. The radar sensor according to claim 2, wherein the recursive module (38) is configured to perform recursive execution at different increments of the index n.

6. The radar sensor according to claim 1, wherein the conversion module (32) is configured to perform convolution of two time-dependent functions by converting the functions to the frequency domain using dedicated hardware for performing a fast Fourier transform (FFT), multiplying them together in the frequency domain, and then converting the product back to the time domain using dedicated hardware for the fast Fourier transform.

7. A radar sensor according to any one of claims 1 to 6 in the form of a high-speed chirp FMCW radar.

8. The radar sensor according to claim 6, wherein the conversion function is a chirp Z-convert function.

Images