Radar device and signal processing method

Abstract

To improve the accuracy of target detection. [Solution] The radar device comprises a receiving unit, a detection processing unit, an angle measurement processing unit, and a correlation processing unit. The receiving unit digitizes the received signal of the observed radio waves to obtain range-Doppler data. The detection processing unit detects a target from the range-Doppler data. The angle measurement processing unit obtains observed values ​​relating to the detected target, including at least distance and direction. The correlation processing unit determines whether the currently detected target is the same as any previously detected target based on the results of correlation processing using the observed values ​​of the detected target. The correlation processing unit comprises a learning unit and an identification unit. The learning unit generates a trained model by providing teaching data including the observed values ​​and the results of correlation processing to a model in which the observed values ​​of the target are explanatory variables and the results of correlation processing are the dependent variable. The identification unit provides the generated trained model with the correlation results of past time series and the observed values ​​of the detected target to obtain the results of correlation processing relating to the detected target.

Description

[Technical Field]

[0001] Embodiments of the present invention relate to a radar apparatus and a signal processing method. [Background Art]

[0002] In the radar technical field, there are roughly two types of "correlation processing". One is correlation processing between radar pulses, and the other is correlation processing between detected targets. The latter correlation processing is for determining whether a latest detected target is the same as a previously detected target or not, and is an important technology for example in TWS (Track While Scan) type radars.

[0003] In correlation processing related to targets, gating processing (or gate processing) is widely used. Gating processing is processing in which a gate, which is a region of a certain size, is set centering on the predicted position of a target, and it is determined whether a target detected by a radar sensor unit exists within the gate, and the detected target is associated. [Prior Art Documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Unexamined Patent Application Publication No. 2021-015079 [Non-Patent Documents]

[0005] [Non-Patent Document 1] Supervised by Takashi Yoshida, "Revised Radar Technology", The Institute of Electronics, Information and Communication Engineers, October 1, 1996 [Summary of the Invention] [Problem to be Solved by the Invention]

[0006] In the gating process, if the predicted target position is incorrect, it will be correlated and matched with a detection result that is different from the original. Since this mismatch propagates to all subsequent processes and significantly reduces the accuracy of target detection, it is desirable to address this issue.

[0007] Therefore, the objective is to provide a radar device and a signal processing method that improve the accuracy of target detection. [Means for solving the problem]

[0008] According to the embodiment, the radar device comprises an antenna unit, a receiving unit, a detection processing unit, an angle measurement processing unit, and a correlation processing unit. The receiving unit digitizes the received signal of radio waves observed by the antenna unit to obtain range-Doppler data. The detection processing unit detects a target from the range-Doppler data. The angle measurement processing unit obtains observed values ​​relating to the detected target, including at least distance and direction. The correlation processing unit determines whether the currently detected target is the same as any previously detected target based on the results of correlation processing using the observed values ​​of the detected target. The correlation processing unit comprises a learning unit and an identification unit. The learning unit generates a trained model by repeatedly providing teaching data, including the observed values ​​and the results of the correlation processing, to a model in which the observed values ​​of the target are explanatory variables and the results of the correlation processing are the dependent variable. The identification unit provides the generated trained model with the correlation results of past time series and the observed values ​​of the detected target to obtain the results of correlation processing relating to the detected target. [Brief explanation of the drawing]

[0009] [Figure 1] Figure 1 is a functional block diagram showing an example of a radar device according to the embodiment. [Figure 2] Figure 2 is a functional block diagram showing an example of the correlation processing unit 17. [Figure 3] Figure 3 is a flowchart showing an example of a processing procedure in a radar device according to the embodiment. [Figure 4] Figure 4 is a diagram illustrating machine learning-based correlation processing. [Figure 5] Figure 5 illustrates the integrated image of correlation processing and detection. [Modes for carrying out the invention]

[0010] Figure 1 is a functional block diagram showing an example of a radar device according to the embodiment. The radar device in Figure 1 comprises an antenna unit 10, a frequency converter 11, an AD (analog-to-digital) converter 12, a pulse compression processing unit 13, an integration processing unit 14, a detection processing unit 15, an angle measurement processing unit 16, a correlation processing unit 17, and a target management processor 18.

[0011] In Figure 1, the antenna section 10 is, for example, an electronically scanned array antenna (AESA: Active Electronically Steered Array) equipped with multiple antenna elements arranged in an array. The received signal observed by this antenna section 10 is down-converted by the frequency converter 11 and converted into a digital signal by the AD converter 12. After pulse compression (range compression) processing is performed on this digital signal by the pulse compression processing section 13, a slow-time axis FFT (Fast Fourier Transform) processing is performed by the integration processing section 14. As a result, integrated range-Doppler data is obtained.

[0012] The detection processing unit 15 performs CFAR (Constant False Alarm Rate) processing on the range-Doppler data to detect targets. The angle measurement processing unit 16 performs distance measurement and angle measurement processing on the detected targets to obtain observed values ​​of the detected targets with distance and direction (elevation angle and Doppler may also be added) as axes. The correlation processing unit 17 performs correlation processing on the observed values ​​of the detected targets and, based on the results, determines whether the currently detected target is the same as any previously detected target. The result of the determination is provided to the target management processor 18, and each target is indexed and managed.

[0013] Figure 2 is a functional block diagram showing an example of a correlation processing unit 17. The correlation processing unit 17 is, for example, a computer with a physical hardware entity, and performs machine learning-based correlation processing. The correlation processing unit 17 comprises a processor 21, a storage unit 24, a ROM 22, a RAM 23, and a communication unit 25.

[0014] ROM22 is a non-volatile memory that functions as program memory. ROM22 stores programs and control data executed by the processor 21. RAM23 functions as working memory to temporarily hold data. It holds data that programs are loaded or that the processor 21 is processing. RAM23 also functions as buffer memory to temporarily hold data and other information. The memory unit 24 is a non-volatile memory such as an SSD (Solid State Drive). The memory unit 24 stores the program 24a executed by the processor 21, the data 24b used by the processor 21, and the trained model 24c.

[0015] The processor 21 is, for example, a computing device such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). The processor 21 reads programs and data stored in the ROM 22 or memory unit 24 into the RAM 23 and executes them to realize the processing functions related to this embodiment.

[0016] The communication unit 25 is connected to the angle measurement processing unit 16 (Figure 1) and acquires the angle measurement value of the detected target. The communication unit 25 is also connected to the target management processor 18 (Figure 1) and outputs the results of the correlation processing related to the target to the target management processor 18.

[0017] Incidentally, the processor 21 includes a learning unit 21a and an identification unit 21b as processing functions related to this embodiment. These are processing functions that are realized when the processor 21 executes instructions described in the program 24a.

[0018] The learning unit 21a repeatedly supplies teaching data including target observation values and correlation processing results to an unlearned model to cause the model to learn, thereby generating a trained model 24c. The generated trained model is stored as the trained model 24c in a storage unit.

[0019] Here, the unlearned model is a model that uses target observation values as explanatory variables and uses the results of correlation processing as objective variables. For example, a trained model can be generated by repeatedly supplying teaching data to a Convolutional Neural Network (CNN). This type of machine learning is known as deep learning. That is, in the embodiment, the learning unit 21a generates the trained model 24c by supervised machine learning using past target observation values and correlation processing results as teaching data.

[0020] The identification unit 21b supplies past time-series correlation results and the detected target observation values to the trained model 24c, and obtains the result of correlation processing related to the target detected this time.

[0021] FIG. 3 is a flowchart showing an example of a processing procedure in the radar apparatus according to the embodiment. In FIG. 3, after the pulse compression processing unit 13 performs pulse compression processing on a digital signal from an AD converter 12 (FIG. 1) (step S1), the radar apparatus performs integration processing by an integration processing unit 14 (step S2) to obtain range-Doppler data.

[0022] Next, the radar apparatus causes a detection processing unit 15 to execute target detection processing such as CFAR processing on the range-Doppler data (step S3) to detect a target. Next, the radar apparatus causes an angle measurement processing unit 16 to perform distance measurement and angle measurement processing on the detected target (step S4) to obtain an observation value of the detected target. This angle measurement value is input to a correlation processing unit 17. In the embodiment, the correlation processing unit 17 executes machine learning-based correlation processing (step S5).

[0023] Figure 4 illustrates the machine learning-based correlation process. The machine learning-based correlation process performs target classification on the data after detection. The input consists of historical time-series correlation results with maps of targets 1, ..., N, and the latest detection results. The output is a multidimensional data structure with range, orientation, and target (new target, target 1, ..., N) as axes. Newly appearing targets are assigned target numbers (1, ..., N) by the target management processor 18. Note that false detections due to clutter, for example, may be added to the target axis.

[0024] Machine learning-based correlation processing takes the past T correlation results and the currently detected target observations as input and outputs correlation results using a trained model. The past correlation results are 3D structured data with distance and direction as axes and target number, and data indicating the presence of the target is set at the distance and direction coordinates of the corresponding target number. For example, data can be used in which the amplitude data of the corresponding coordinate is copied to the coordinate where the target was detected beyond the detection threshold.

[0025] The radar reflection signal from a target has a spread depending on the spatial size of the target, for example, in the distance and azimuth directions. Furthermore, the effective reflection area is determined by the target's unique radar wave reflection characteristics, which manifests as a trend in the magnitude of the received signal strength.

[0026] For example, a small target will result in a received signal with a small spread in distance and direction, and a low signal strength. On the other hand, ground clutter will result in a received signal with a large spread in distance and direction, and a high signal strength. By inputting data that includes such spatial size and signal strength as target features, in addition to the coordinates of the detected signal, into a machine learning-based correlation process, the accuracy of the correlation can be improved.

[0027] The data input to the machine learning-based correlation process is a combination of the past T correlation results and the most recent detection result. In Figure 4, distance, direction, and target (number: index) are used as the output axes. The targets are, for example, target numbers 1...N, with N+1 new targets corresponding to each. The processing result is obtained, for example, as a two-dimensional likelihood map of distance and direction for each target correlation. The target with the highest likelihood is adopted as the correlation result.

[0028] Furthermore, between the detection processing unit and the machine learning-based correlation processing unit, processing such as target identification (identification of whether it is a small or large aircraft, fixed-wing or rotary-wing, etc.) may be performed, and the results may be included as input.

[0029] The target information management process by the target management processor 18 is described below. If a correlation is found with target number m, the result is added to target m as a past correlation result during the next correlation process. If a new target is determined to exist, the target information management process assigns a new target number n, and the result is added to target n as a past correlation result during the next correlation process. For targets for which a correlation result could not be obtained, for example, the movement speed is calculated from past results to obtain the current predicted value. If a correlation cannot be obtained for a predetermined number of consecutive times, that target number is released, and the addition of correlation results to the data for that target number is stopped during the next correlation process.

[0030] Referring to Figure 5, the integration of correlation and detection processing will be explained. As shown in Figure 5, in machine learning-based correlation processing, target detection and correlation processing can be performed simultaneously by using pulse-compressed and integrated data as the "current detection data". In existing technologies, target detection and correlation processing are usually performed separately, but according to this embodiment, target detection and correlation processing can be performed in a single process, which promotes faster processing.

[0031] This section describes the integration of smoothing processing. When using machine learning to train a correlation process, the training data can be the smoothed data, allowing the process to be integrated with the smoothing process. In existing technologies, the results of the correlation process are smoothed using, for example, past smoothed values ​​or predicted values. In this embodiment, a correlation process that has been pre-trained to perform smoothing is applied. In other words, when using machine learning to train a correlation process, the smoothed data can be used as the training data, allowing the correlation process to be integrated with the smoothing process.

[0032] This section explains how to change the target number. The correlation processing results are presented to the user, allowing the user to change the results of the target management process. For example, if the user determines that a target processed as Target 1 is a false positive, then the next time the data is entered for correlation processing, the information for Target 1 will be cleared, preventing the detected target from being classified as Target 1.

[0033] Furthermore, suppose that while continuously detecting Goal 1, the correlation process mistakenly determines it to be a new goal, and the goal management process registers it as Goal 2. In this case, Goal 1 is not correlated, and the goal management process outputs a predicted value. Therefore, if the goal management process outputs a predicted value without correlation, presenting this to the user allows the user to recognize that Goal 1 was not correlated and that a new Goal 2 has appeared. If the user determines that the correlation process has mistakenly identified Goal 1 as a new goal, they can specify the goal number to correct it. In other words, in this example, by instructing the system to correct Goal 2 to Goal 1, the result registered as Goal 2 is re-registered as Goal 1.

[0034] Furthermore, changes to the target number made by the user can be reflected in the correlation process. In other words, based on the results of changes made by the user in the target management process, the correlation processing unit learns parameters that have been corrected and then reflected in the correlation process.

[0035] As described above, in this embodiment, current detection data that has a strong correlation with time-series data of past detection data is determined by machine learning. By providing data such as the target coordinates, signal strength, and spatial size, and performing correlation processing using these as features, accuracy can be improved. Thus, according to this embodiment, it is possible to provide a radar device and a signal processing method that improve the accuracy of target detection.

[0036] It should be noted that this invention is not limited to the embodiments described above. For example, the correlation processing unit 17 shown in Figure 2 can also be implemented as software. In this case, the correlation processing unit 17 can be understood as a program (process) that is loaded into memory and executed by the processor.

[0037] Alternatively, the user may specify the range (domain) of data to which machine learning-based correlation processing should be performed, while existing correlation processing is applied to the remaining range. Furthermore, the system may have multiple parameters learned using different methods, allowing the user to specify which parameters to apply to each domain.

[0038] While embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims of the invention and its equivalents. [Explanation of Symbols]

[0039] 10...Antenna unit, 11...Frequency converter, 12...AD converter, 13...Pulse compression unit, 14...Integration unit, 15...Detection unit, 16...Angle measurement unit, 17...Correlation unit, 18...Target management processor, 21...Processor, 21a...Learning unit, 21b...Identification unit, 22...ROM, 23...RAM, 24...Storage unit, 24a...Program, 24b...Data, 24c...Trained model, 25...Communication unit.

Claims

1. Antenna section, A receiving unit that digitizes the received signal of radio waves observed by the antenna unit to obtain range-Doppler data, A detection processing unit for detecting a target from the aforementioned range-Doppler data, An angle measurement processing unit that obtains observational values ​​including at least distance and direction for the detected target, The system comprises a correlation processing unit that determines whether the target detected this time is the same as any previously detected target, based on the results of a correlation process using the observed values ​​of the detected target, The correlation processing unit, A learning unit generates a trained model by repeatedly providing training data containing the observed values ​​and the results of the correlation process to a model in which the target observed values ​​are the explanatory variables and the results of the correlation process are the dependent variables. A radar device comprising an identification unit that provides the generated trained model with the correlation results of past time series and the observed values ​​of the detected target to obtain the result of the correlation processing relating to the detected target.

2. The radar device according to claim 1, further comprising a target management processor that manages each target by assigning an index to it based on the results of the correlation processing relating to the aforementioned targets.

3. The radar apparatus according to claim 1, wherein the correlation processing unit provides the observed data, after pulse compression and integration processing, to the trained model.

4. The radar device according to claim 1, wherein the learning unit provides the model with the smoothed teaching data to train it.

5. The radar device according to claim 2, wherein the target management processor accepts user changes to the results of target management processing.

6. The radar device according to claim 5, wherein the correlation processing unit reflects the results of the changes in the correlation processing.

7. A signal processing method for a radar device comprising an antenna unit, a receiving unit that digitizes the received signal of radio waves observed by the antenna unit to obtain range-Doppler data, a detection processing unit that detects a target from the range-Doppler data, and an angle measurement processing unit that obtains observed values ​​related to the detected target, including at least distance and direction, A process by which a computer determines, based on the results of a correlation process using the observed values ​​of the detected target, whether the target detected this time is the same as any previously detected target, The process of generating a trained model involves repeatedly feeding training data containing the observed values ​​and the results of the correlation process to a model that uses the target observed values ​​as explanatory variables and the results of the correlation process as the dependent variable, and then training the model. A signal processing method comprising the steps of providing the generated trained model with the correlation results of past time series and the observed values ​​of the detected target to obtain the result of the correlation processing relating to the detected target.

Images