Radar target classification method and system based on spectrogram tiles and data augmentation

Abstract

The application provides a radar target classification method and system based on a spectrum block and data enhancement, the method comprising the following steps: S1: collecting radar original ADC data and performing FFT processing to generate a distance-velocity dimension and a distance-angle dimension spectrum, and based on a preset energy threshold, outlining a target contour and extracting a target spectrum block to construct a training data set; S2: inputting the target spectrum block into a generative adversarial network for training, and using the network to generate a synthetic target spectrum block to expand the training data set; S3: calculating the signal amplitude statistical value in the contour region of each target spectrum block in the expanded data set to obtain an energy feature; S4: inputting the target spectrum block and the energy feature into a classifier for training to obtain a target classification model; and S5: inputting the target feature to be classified into the model to output a classification result. The application fuses a physically constrained GAN and a spectrum block, solves data scarcity, and improves radar classification precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a radar target classification method and system based on a spectrum block and data enhancement. BACKGROUND

[0002] As a key sensor for all-weather environment perception, the target classification performance of millimeter wave radar is crucial for applications such as autonomous driving. The current technology mainly has the following fundamental bottlenecks: relying on the "target point cloud" generated at the end of the traditional signal processing chain and its derivative scalar features (such as distance, speed, spectral width, etc.) for classification decision. These features are essentially low-dimensional compressed projections of the complex scattering structure of the target, and have the inherent defect of serious lack of information dimension. Specifically, the ability to distinguish targets with similar sizes but different structures such as pedestrians and bicycles is limited, the features are distorted due to the limitation of radar physical resolution, and the features are extremely sensitive to noise and clutter in low signal-to-noise ratio environments, and the feature stability is poor.

[0003] To break through the limitations of point features, existing technologies mainly improve through two paths. One path adopts a multi-sensor fusion strategy, such as a Chinese patent application with patent number CN117093872A discloses a radar target classification model self-training method and system, which compensates for the lack of radar in contour, texture, and other information by fusing radar and visual data. The other path focuses on deep mining of radar signal features, such as a Chinese patent application with patent number CN117250594B discloses a radar target classification and recognition method, which extracts micro-Doppler features or relies on specific designed radar waveform timing to obtain more detailed target information. In addition, some technologies also try to introduce data enhancement methods, such as a Chinese patent application with patent number CN113030950A discloses a small unmanned aerial vehicle classification method based on deep learning technology, which uses a generative adversarial network to generate synthetic spectrum graphs to expand the training data set.

[0004] However, these existing processing solutions themselves raise new problems. Although the multi-sensor fusion solution enriches the feature information to some extent, its fatal defect lies in the serious dependence on external visual data, and the performance will deteriorate sharply in visual limited scenes such as fog, rain and night. At the same time, the complex multi-sensor calibration and synchronization process not only increases the system cost and integration difficulty, but also introduces new uncertainty sources such as calibration error. The scheme of deeply mining radar features is limited to specific waveforms or working modes, and has insufficient universality and flexibility, which is difficult to adapt to diversified radar systems, and the complex signal processing algorithm often sacrifices real-time performance. More importantly, these schemes have not effectively solved the fundamental problem of the scarcity of high-quality labeled data and the imbalance of data categories in the radar field, resulting in insufficient classification robustness of the model for rare targets such as pedestrians. Even the data enhancement method may generate physically unreasonable synthetic samples due to the lack of explicit constraints on the physical mechanism of radar signals, affecting the reliability and explainability of the model.

[0005] In summary, the existing technology is trapped in a vicious cycle of "insufficient information dimension-external dependence or special design-data scarcity and imbalance" when dealing with radar target classification challenges, and has not yet formed a systematic solution that can directly extract high-dimensional structural features from radar raw signals and efficiently utilize limited data to build a universal, robust and explainable classification model. SUMMARY

[0006] The technical problem to be solved by the present application is that in view of the technical problems existing in the prior art, the present application provides a radar target classification method and system based on spectrum graph blocks and data enhancement, aiming to solve the core technical problems of incomplete information, data scarcity, poor model robustness and weak environmental adaptability caused by dependence on low-dimensional point features, external sensors or specific waveforms in existing radar target classification technology.

[0007] To solve the above technical problems, the technical solution provided by the present application is:

[0008] A radar target classification method based on spectrum graph blocks and data enhancement comprises the following steps:

[0009] Step S1: Collecting radar raw analog-to-digital conversion (ADC) data, performing fast Fourier transform (FFT) processing to generate target distance-velocity spectrum graph and distance-angle spectrum graph; based on a preset energy threshold, the energy contours of each target are drawn in the distance-velocity spectrum graph and the distance-angle spectrum graph respectively, and the rectangular region bounded by the energy contours is extracted as a target spectrum graph block, thereby obtaining a training data set; the target spectrum graph block includes a distance-velocity spectrum graph block and a distance-angle spectrum graph block;

[0010] Step S2: input the extracted target spectrum block to a generative adversarial network (GAN) for training, wherein a constraint term for making the generated block conform to a physical law of energy diffusion of a radar target is included in a loss function of the GAN; use the trained GAN to generate a synthetic target spectrum block to expand the training data set, and obtain an expanded training data set; the synthetic target spectrum block includes a synthetic target range-velocity dimension spectrum block and a target range-angle dimension spectrum block;

[0011] Step S3: calculate a statistical value of signal amplitude in a contour region of each target spectrum block in the expanded training data set, and obtain an energy feature representing a comprehensive reflection characteristic of a target;

[0012] Step S4: input the range-velocity dimension spectrum block and the range-angle dimension spectrum block of a target in the expanded training set and the corresponding energy feature to a classifier for model training, and obtain a trained target classification model;

[0013] Step S5: input a corresponding feature of a target to be classified into the target classification model, and output a target classification result.

[0014] As a further improvement of the method, in step S1, the steps of generating a target range-velocity dimension spectrum and a target range-angle dimension spectrum include:

[0015] perform two-dimensional fast Fourier transform (FFT) processing on each frame of radar original ADC data in the range and velocity dimensions to form a two-dimensional matrix in the range and velocity dimensions, and perform energy direction projection on the two-dimensional matrix in the range and velocity dimensions to generate the range-velocity dimension spectrum;

[0016] perform constant false alarm rate (CFAR) detection on the two-dimensional matrix in the range and velocity dimensions to extract potential target points;

[0017] extract signals of each potential target point on a radar multi-channel, perform angle dimension FFT processing on the extracted signals to obtain a spectrum distribution of the potential target in the angle dimension, associate the angle dimension spectrum distribution of each target point to corresponding range dimension information to form a range-angle two-dimensional matrix, and perform energy direction projection on the range-angle two-dimensional matrix to generate the range-angle dimension spectrum.

[0018] As a further improvement of the method, in step S2, the method further includes converting a complex signal value in the target spectrum block into an amplitude value or a power value.

[0019] As a further improvement of the method of the present application: in step S2, a physical constraint term in the loss function of the generative adversarial network (GAN) is used to constrain the target energy in the generated target spectrum block to conform to the inherent diffusion characteristics of the radar target echo in the range dimension and the velocity dimension.

[0020] As a further improvement of the method of the present application: in step S3, the statistical value of the signal amplitude in the contour region of each target spectrum block is calculated using the following formula:

[0021]

[0022] wherein, is the region average amplitude value of the target in the contour region of the spectrum block, is the number of vertical pixel points in the contour region of the target spectrum block, is the number of horizontal pixel points in the contour region of the target spectrum block, is the amplitude value of the spectrum point in the contour region of the target spectrum block. i is the amplitude value of the spectrum point in the contour region of the target spectrum block. j is the amplitude value of the spectrum point in the contour region of the target spectrum block.

[0023] As a further improvement of the method of the present application: in step S4, the model training includes:

[0024] using the range-velocity dimension spectrum block and the range-angle dimension spectrum block of the target to preliminarily train the classifier to obtain an initial classification model;

[0025] using the energy feature as a new input parameter, inputting the range-velocity dimension spectrum block and the range-angle dimension spectrum block of the target into the initial classification model together, fine-tuning the initial classification model to obtain a trained target classification model.

[0026] As a further improvement of the method of the present application: the range-velocity dimension spectrum block, the range-angle dimension spectrum block and the energy feature are combined in a weighted sum manner and input into the classifier.

[0027] The present application also provides a radar target classification system based on spectrum block and data enhancement, comprising a microprocessor and a memory connected to each other, the microprocessor being programmed or configured to execute the radar target classification method based on spectrum block and data enhancement.

[0028] The present application also provides a computer readable storage medium, which stores computer programs / instructions programmed or configured to execute the radar target classification method based on spectrum block and data enhancement by a processor.

[0029] The present application also provides a computer program product comprising computer programs / instructions programmed or configured to perform the radar target classification method based on the spectrum tile and data enhancement by a processor.

[0030] Compared with the prior art, the present application has the following beneficial effects:

[0031] 1、The present application realizes a fundamental leap from traditional point features to high-dimensional structural features by extracting the complete spectrum tile of the target in the range-velocity dimension and the range-angle dimension as image features, significantly improving the model's ability to distinguish similar targets (such as pedestrians and bicycles) and noise resistance. At the same time, the generative adversarial network with physical constraints is used for data enhancement, which generates a large number of physically reasonable training samples at low cost by constraining the energy diffusion law of the synthesized tile, effectively solving the problems of data scarcity and class imbalance, and greatly improving the model's recognition robustness and generalization ability for rare targets.

[0032] 2、The present application further fuses the spectrum tile features and the average energy features representing the target's reflection intensity for decision-making, effectively suppresses the model's "hallucinations" by introducing interpretable physical constraints, and enhances the reliability of the classification results. The entire scheme relies only on radar raw signals for end-to-end processing, eliminating the dependence on external sensors such as vision, reducing system complexity and cost, ensuring all-weather working ability in bad weather, and the algorithm process is efficient, meeting the real-time requirements in practical applications. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The present application further fuses the spectrum tile features and the average energy features representing the target's reflection intensity for decision-making, effectively suppresses the model's "hallucinations" by introducing interpretable physical constraints, and enhances the reliability of the classification results. The entire scheme relies only on radar raw signals for end-to-end processing, eliminating the dependence on external sensors such as vision, reducing system complexity and cost, ensuring all-weather working ability in bad weather, and the algorithm process is efficient, meeting the real-time requirements in practical applications.

[0034] Figure 2 The present application further fuses the spectrum tile features and the average energy features representing the target's reflection intensity for decision-making, effectively suppresses the model's "hallucinations" by introducing interpretable physical constraints, and enhances the reliability of the classification results. The entire scheme relies only on radar raw signals for end-to-end processing, eliminating the dependence on external sensors such as vision, reducing system complexity and cost, ensuring all-weather working ability in bad weather, and the algorithm process is efficient, meeting the real-time requirements in practical applications.

[0035] Figure 3 The present application further fuses the spectrum tile features and the average energy features representing the target's reflection intensity for decision-making, effectively suppresses the model's "hallucinations" by introducing interpretable physical constraints, and enhances the reliability of the classification results. The entire scheme relies only on radar raw signals for end-to-end processing, eliminating the dependence on external sensors such as vision, reducing system complexity and cost, ensuring all-weather working ability in bad weather, and the algorithm process is efficient, meeting the real-time requirements in practical applications.

[0036] Figure 4 The present application further fuses the spectrum tile features and the average energy features representing the target's reflection intensity for decision-making, effectively suppresses the model's "hallucinations" by introducing interpretable physical constraints, and enhances the reliability of the classification results. The entire scheme relies only on radar raw signals for end-to-end processing, eliminating the dependence on external sensors such as vision, reducing system complexity and cost, ensuring all-weather working ability in bad weather, and the algorithm process is efficient, meeting the real-time requirements in practical applications. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0038] As Figure 1 shown, the embodiment provides a radar target classification method based on spectrum tile and data enhancement, comprising the following steps:

[0039] Step S1: Collecting radar original analog-to-digital conversion (ADC) data, performing fast Fourier transform (FFT) processing, generating target distance-velocity dimension spectrum and distance-angle dimension spectrum; based on a preset energy threshold, respectively drawing the energy contour of each target in the distance-velocity dimension spectrum and the distance-angle dimension spectrum, and extracting the rectangular region defined by the energy contour as the target spectrum tile, thereby obtaining the training data set; the target spectrum tile includes distance-velocity dimension spectrum tile and distance-angle dimension spectrum tile.

[0040] In this embodiment, the radar original analog-to-digital conversion (ADC) data is a digital sequence obtained by sampling and quantizing the received analog echo signal through an analog-to-digital converter, which constitutes the basis for subsequent signal processing. First, the distance dimension fast Fourier transform (FFT) is performed on a frame of ADC data to extract the distance information of the target; then, the velocity dimension FFT is performed in multiple pulse periods to analyze the Doppler velocity of the target; further, the multi-channel data of the radar array is used to perform angle dimension FFT to estimate the azimuth angle of the target. Through the above three-dimensional FFT processing, a radar data cube (three-dimensional matrix) containing distance, velocity and angle information is finally formed, providing complete signal representation for subsequent feature extraction and target classification.

[0041] In this embodiment, the step of generating the target distance-velocity dimension spectrum and the target distance-angle dimension spectrum comprises:

[0042] The distance dimension and velocity dimension two-dimensional fast Fourier transform (FFT) processing is performed on each frame of radar original ADC data to form a distance and velocity dimension two-dimensional matrix of the radar target, and the energy direction projection is performed on the distance and velocity dimension two-dimensional matrix to generate the distance-velocity dimension spectrum, as shown in Figure 2 The spectrum echo of an actual vehicle in the distance dimension and the velocity dimension is not a single frequency point, but a rectangular image region. The distance and velocity dimension spectrum is obtained by two-dimensional FFT processing of the distance and velocity dimension two-dimensional matrix, the horizontal axis represents the target velocity, the vertical axis represents the target distance, and reflects the distribution characteristics of the target in the distance-velocity domain.

[0043] A constant false alarm rate (CFAR) detection is performed on a two-dimensional matrix of distance and velocity dimensions to extract potential target points. The CFAR detection is only used to extract target points for subsequent fast Fourier transform (FFT) operation in the target angle dimension, and is not used as input of velocity information in subsequent target classification. The CFAR detection is an adaptive threshold detection algorithm that can maintain a constant false alarm rate in a clutter and noise environment, and is commonly used for preliminary radar target detection.

[0044] The signals of the potential target points on the radar multi-channel are extracted, and the extracted signals are subjected to FFT processing in the angle dimension to obtain the spectral distribution of the potential target in the angle dimension. The angle dimension spectral distribution of each target point is associated with the corresponding distance dimension information to form a distance-angle two-dimensional matrix. The distance-angle two-dimensional matrix is projected in the energy direction to generate a distance-angle dimension spectrum diagram, as shown in FIG. 6. The spectral echo of an actual vehicle in the distance dimension and the angle dimension is not a single frequency point, but a rectangular image region. The distance-angle dimension spectrum diagram is a two-dimensional matrix of distance and angle obtained by three-dimensional FFT processing, and the horizontal axis represents the target angle and the vertical axis represents the radial distance, reflecting the distribution characteristics of the target in the distance-angle domain. Figure 3

[0045] In this embodiment, the target in the spectrum diagram is profiled by a preset energy threshold. Specifically, in the distance-velocity dimension spectrum diagram, the target is profiled in the distance direction and the velocity direction based on the energy being significantly higher than the local background region. Similarly, in the distance-angle dimension spectrum diagram, the target profile is drawn in the distance direction and the angle direction based on the same energy significance principle. The area enclosed by the profile is regarded as the "area" of the target, and the signal amplitude values of the points in the area are analogous to the pixel gray values in the image, which together constitute the spectrum block of the target.

[0046] The energy significance standard for the above profile extraction can be preferably that the energy of the target is higher than the energy of the surrounding region by a predetermined decibel value (for example, 5 dB) as the judgment threshold, but this specific value is not the only limitation of the present application, and other suitable thresholds can be determined by a person skilled in the art according to the actual signal-to-noise ratio condition, the detection sensitivity requirement, or through an adaptive algorithm.

[0047] Step S2: inputting the extracted target spectrum block into a generative adversarial network (GAN) for training, wherein a constraint term for making the generated block conform to the physical law of radar target energy diffusion is included in the loss function of the GAN; using the trained GAN to generate a synthetic target spectrum block to expand the training data set, and obtaining an expanded training data set; the synthetic target spectrum block includes a synthetic target distance-velocity dimension spectrum block and a target distance-angle dimension spectrum block.

[0048] ​In this embodiment, the signal values in the target spectrum block in complex form are converted into amplitude values or power values.

[0049] Specifically, to adapt to the input requirements of the generative adversarial network, the complex spectrum data needs to be converted into real number images, and the amplitude values of all spectrum points in the distance-velocity dimension spectrum and the distance-angle dimension spectrum are calculated respectively. The calculation formula is:

[0050] (1)

[0051] Where x is the analog-to-digital conversion (ADC) data value of the echo signal, and P is the power of the echo signal, i.e., the amplitude value.

[0052] Through the above calculation, the two-dimensional spectrum matrix is converted into a single-channel grayscale image, thereby meeting the requirements of the generative adversarial network for processing real number image data.

[0053] In this embodiment, the physical constraint term in the loss function of the generative adversarial network (GAN) is used to constrain the target energy in the generated target spectrum block to conform to the inherent diffusion characteristics of the radar target echo in the distance dimension and the velocity dimension.

[0054] Specifically, in the training process of the generative adversarial network, to ensure the physical reasonableness of the generated samples, a physical constraint term is introduced into the loss function. This constraint term is designed based on the energy diffusion characteristics of the radar point target in the distance dimension and the velocity dimension (which are determined by the window function and the Fourier transform in signal processing), aiming to guide the network to generate spectrum blocks with natural energy distribution and conforming to physical laws, and to punish the generator output that generates unnatural sharp patterns or abnormally wide patterns in the distance dimension without any relation to the velocity dimension.

[0055] In this embodiment, after completing target contour extraction, signal amplitude conversion, and loss function physical constraint configuration, the preprocessed real target spectrum blocks are input into the generative adversarial network for training. Through the adversarial learning and optimization adjustment of the network, a large number of synthetic target spectrum blocks consistent with the physical characteristics of the real samples can be generated. This process effectively solves the class imbalance problem of the training dataset caused by the scarcity of samples of specific categories (such as pedestrians and bicycles) or specific dangerous scenarios (such as vehicles crossing at high speed). The generative adversarial network can synthetically generate samples of these rare categories, thereby expanding and balancing the training dataset. This method enhances the generalization ability and robustness of the subsequent target detection and classification model from the data level, enabling it to maintain stable and reliable performance when facing insufficient or unevenly distributed samples.

[0056] So far, the embodiment completes the data preparation and enhancement work in the offline training phase, obtains an expanded training data set containing real and a large number of synthetic samples, and corresponding energy features.

[0057] Step S3: Calculate the statistical value of the signal amplitude in the contour region of each target spectrum block in the expanded training data set to obtain the energy feature representing the target comprehensive reflection characteristic.

[0058] In this embodiment, after obtaining the rectangular region defined by the target contour, the average energy feature of the target in the region needs to be calculated. For the range-velocity dimension and the range-angle dimension spectrum block, the average value of the signal amplitude in the contour region is calculated respectively to represent the comprehensive reflection intensity of the target in the corresponding dimension. The specific calculation formula is:

[0059] (2)

[0060] wherein, is the region average amplitude value of the target in the contour region of the spectrum block, is the number of vertical pixel points in the contour region of the target spectrum block, is the number of horizontal pixel points in the contour region of the target spectrum block, is the amplitude value of the spectrum point in the contour region. i is the amplitude value of the spectrum point in the contour region. j

[0061] Step S4: Input the range-velocity dimension spectrum block, the range-angle dimension spectrum block and the corresponding energy feature of the target in the expanded training set into the classifier for model training to obtain the trained target classification model.

[0062] In this embodiment, the classifier can be implemented by selecting various machine learning models, including but not limited to decision tree, support vector machine (SVM), random forest or shallow neural network, etc. Among them, the support vector machine finds the optimal classification hyperplane in the high-dimensional feature space to realize the differentiation of different categories of targets, and is especially suitable for processing the high-dimensional feature data provided by the present application, and has the characteristics of stable classification performance and strong generalization ability.

[0063] In this embodiment, the model training includes:

[0064] using the range-velocity dimension spectrum block and the range-angle dimension spectrum block of the target to preliminarily train the classifier to obtain an initial classification model;

[0065] using the energy feature as a new input parameter, inputting the energy feature, the range-velocity dimension spectrum block and the range-angle dimension spectrum block of the target into the initial classification model to realize fine tuning of the initial classification model to obtain the trained target classification model. ​

[0066] In this embodiment, the distance-velocity dimension spectrum patch, the distance-angle dimension spectrum patch, and the energy feature are combined in a weighted sum manner and then input into the classifier.

[0067] Step S5: Input the corresponding features of the target to be classified into the target classification model and output the target classification result.

[0068] Specifically, during the actual operation of the radar, the system processes each frame of data acquired in real time: by projection and contour extraction, it obtains the range-velocity and range-angle spectral patches of each target in the current frame, and simultaneously calculates the average energy feature Pi corresponding to each patch. Subsequently, the two spectral patches of each target and their corresponding Pi are input into a pre-trained target classification model. The model performs type matching and discrimination based on these features, outputs the classification results, and saves them, thereby achieving high-accuracy, end-to-end real-time radar target classification.

[0069] like Figure 4 As shown, the left half illustrates the radar's detection scenario in a real-world environment, while the right half displays the real-time processing results observed on the host computer interface. In this interface, the system performs real-time classification and visual labeling of detected targets: rectangular symbols represent targets classified as "cars," while circular symbols represent targets classified as "non-motorized vehicles." This illustration intuitively demonstrates the ability of the method of this invention to effectively distinguish and output real-time results for different categories of targets in practical applications.

[0070] In summary, this embodiment constructs a new paradigm for radar target classification that integrates "feature-data-decision" collaborative optimization, achieving a systematic breakthrough from information dimension, data foundation to decision reliability.

[0071] First, at the feature representation level, this invention breaks through the limitations of traditional methods that rely on post-point features of CFAR. By treating the complete projection spectrum of the target in the distance-velocity and distance-angle dimensions as image patches and directly using them as input to the classification model, this method can retain more comprehensive and realistic two-dimensional structural reflection information of the target, fundamentally solving the problems of low dimensionality of traditional point feature information and weak ability to distinguish similar targets.

[0072] Secondly, at the data foundation level, this invention addresses the industry bottleneck of scarce radar annotation data. By introducing a physically constrained generative adversarial network (GAN) to augment the aforementioned spectral patches, a massive amount of synthetic samples consistent with real physical characteristics can be generated. This not only significantly expands the training set size but also allows for the targeted generation of samples for rare targets such as pedestrians and bicycles, effectively alleviating the problem of data class imbalance. Consequently, it greatly improves the sufficiency of model training and the generalization ability and robustness of the final classifier.

[0073] Finally, at the level of decision mechanism, the application effectively improves the reliability of classification by multi-dimensional feature fusion. While taking advantage of the rich structural information provided by the spectrum image block, the average energy of the target within the block is further calculated as a supplementary feature. This energy feature can represent the comprehensive reflection characteristics of the target in the distance, speed, and angle dimensions, providing the model with interpretable physical constraints. This dual input strategy of "image structure feature" and "physical energy feature" enables the model to make decisions based on both morphology and physical properties, effectively suppressing the "hallucination" effect of artificial intelligence models and significantly enhancing the accuracy and reliability of the classification results.

[0074] The embodiment also provides a radar target classification system based on spectrum image blocks and data enhancement, which comprises a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the radar target classification method based on spectrum image blocks and data enhancement.

[0075] The embodiment also provides a computer readable storage medium, in which a computer program / instruction is stored, and the computer program / instruction is programmed or configured to execute the radar target classification method based on spectrum image blocks and data enhancement by means of a processor.

[0076] Those skilled in the art should understand that the above-mentioned embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer readable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer usable program codes. The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks These computer program instructions can also be stored in a computer readable memory capable of causing the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product comprising instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocksFigure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. The above merely provides the preferred embodiment of the present application, but does not make any form of limitation to the present application. Although the present application has been disclosed with the preferred embodiment as above, it is not intended to limit the present application. Thus, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present application, without departing from the technical solution of the present application, shall fall within the protection scope of the technical solution of the present application.

[0077] The embodiment also provides a computer program product, comprising computer programs / instructions programmed or configured to execute the radar target classification method based on spectrum tile and data enhancement by a processor.

[0078] The above merely provides the preferred embodiment of the present application, but does not make any form of limitation to the present application. Although the present application has been disclosed with the preferred embodiment as above, it is not intended to limit the present application. Thus, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present application, without departing from the technical solution of the present application, shall fall within the protection scope of the technical solution of the present application.

Claims

1. A radar target classification method based on spectral patches and data augmentation, characterized in that, Includes the following steps: Step S1: Collect the original radar analog-to-digital converter (ADC) data, perform Fast Fourier Transform (FFT) processing, and generate the target range-velocity dimension spectrum map and the range-angle dimension spectrum map; Based on a preset energy threshold, the energy contours of each target are drawn in the distance-velocity dimension spectrum map and the distance-angle dimension spectrum map, respectively, and the rectangular regions defined by the energy contours are extracted as target spectrum patches to obtain the training dataset; the target spectrum patches include distance-velocity dimension spectrum patches and distance-angle dimension spectrum patches; Step S2: Input the extracted target spectrum patches into a Generative Adversarial Network (GAN) for training. The loss function of the GAN includes a constraint term to ensure that the generated patches conform to the physical laws of radar target energy diffusion. Use the trained GAN to generate synthetic target spectrum patches to expand the training dataset, resulting in an expanded training dataset. The synthetic target spectrum patches include synthetic target range-velocity dimension spectrum patches and target range-angle dimension spectrum patches. Step S3: Calculate the statistical values ​​of the signal amplitude within the contour region of each target spectrum patch in the expanded training dataset to obtain the energy characteristics characterizing the comprehensive reflection properties of the target; Step S4: Input the distance-velocity dimension spectrum patch, distance-angle dimension spectrum patch, and corresponding energy features of the target in the expanded training set into the classifier for model training to obtain the trained target classification model; Step S5: Input the corresponding features of the target to be classified into the target classification model and output the target classification result.

2. The radar target classification method based on spectrum patches and data augmentation according to claim 1, characterized in that, In step S1, the step of generating the target's range-velocity dimension spectrum map and range-angle dimension spectrum map includes: Each frame of raw radar ADC data is processed by a two-dimensional Fast Fourier Transform (FFT) in the range and velocity dimensions to form a two-dimensional matrix in the range and velocity dimensions. The energy direction is then projected onto the two-dimensional matrix in the range and velocity dimensions to generate the range-velocity spectrum. Constant false alarm rate (CFAR) detection is performed on the two-dimensional matrix of distance and velocity dimensions to extract potential target points; The signals of each potential target point on the radar multi-channel are extracted, and the extracted signals are processed by Fast Fourier Transform (FFT) in the angle dimension to obtain the spectral distribution of the potential target in the angle dimension. The spectral distribution of each target point in the angle dimension is associated with the corresponding range dimension information to form a two-dimensional matrix of range and angle. The range and angle two-dimensional matrix is ​​projected into the energy direction to generate the range-angle dimension spectrum map.

3. The radar target classification method based on spectrum patches and data augmentation according to claim 1, characterized in that, Step S2 also includes converting the complex signal values ​​in the target spectrum block into amplitude or power values.

4. The radar target classification method based on spectrum patches and data augmentation according to claim 1, characterized in that, In step S2, the physical constraint term in the loss function of the Generative Adversarial Network (GAN) is used to constrain the target energy in the generated target spectrum patch to conform to the inherent diffusion characteristics of radar target echo in the range and velocity dimensions.

5. The radar target classification method based on spectrum patches and data augmentation according to claim 1, characterized in that, In step S3, the statistical value of the signal amplitude within the contour area of ​​each target spectrum block is calculated using the following formula: in, This represents the regional average amplitude value of the target within the outline region of the spectrum patch. This represents the number of pixels vertically within the outline region of the target spectral patch. This represents the number of pixels in the horizontal direction within the outline region of the target spectrogram patch. For the first region within the outline area i Line number j The amplitude values ​​of the column spectrum points.

6. The radar target classification method based on spectrum patches and data augmentation according to claim 1, characterized in that, In step S4, the model training includes: The classifier is initially trained using the distance-velocity spectral patch and the distance-angle spectral patch of the target to obtain an initial classification model. The energy features are added as new input parameters, and together with the distance-velocity spectral patch and the distance-angle spectral patch of the target, they are input into the initial classification model to fine-tune the initial classification model and obtain the trained target classification model.

7. The radar target classification method based on spectrum patches and data augmentation according to claim 6, characterized in that, The distance-velocity spectral patch, the distance-angle spectral patch, and the energy feature are combined in a weighted sum and then input into the classifier.

8. A radar target classification system based on spectral patches and data augmentation, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the radar target classification method based on spectral patches and data augmentation as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute, via a processor, the radar target classification method based on spectral patches and data augmentation as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute, via a processor, the radar target classification method based on spectral patches and data augmentation as described in any one of claims 1 to 7.

Images