Radar target identification method based on attitude angle division

Abstract

The invention specifically relates to a radar target identification method based on attitude angle division, and the method comprises the steps: obtaining the actual measurement data of a plurality of types of targets, each target sample comprising a time-frequency map, the real-time distance, azimuth angle and pitch angle of a target relative to a radar, and the distance, azimuth angle and pitch angle of the target in a previous time slice; calculating a course angle based on measured data; the method comprises the following steps of: preliminarily setting and dividing angle areas by utilizing a known course angle and a pitch angle on the basis of target micro-Doppler characteristics embodied by a time-frequency spectrum, and constructing a time-frequency spectrum data set of various types of targets in different angle areas; constructing a data set by using convolutional neural network training, then verifying classification precision, and judging whether a neural network model needs to be retrained by continuing an iterative division method or not according to an identification result; and determining a division boundary, training a neural network model, and finally inputting a time-frequency spectrum of a target to finish accurate classification. According to the method, the classification precision of radar target recognition is improved.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition, and more specifically to a radar target recognition method based on attitude angle division. Background Technology

[0002] Radar target identification technology is a core support for low-altitude economy and air traffic control, and its performance directly determines the ability to perceive and handle targets in complex environments. In the low-altitude domain, it is necessary to accurately identify "black flight" drones and legal general aviation aircraft; in air traffic control, it is necessary to distinguish different aircraft types such as helicopters and fixed-wing aircraft in real time. These requirements all rely on the system's accurate identification capability in complex electromagnetic environments such as strong clutter and multiple targets superimposed.

[0003] Traditional identification methods based on parameters such as radial velocity and distance can only distinguish between moving and stationary targets, failing to address the challenges of complex environments and diverse target identification needs. The micro-Doppler effect, as a direct manifestation of the micro-motion characteristics of radar targets, can capture microscopic motion information such as rotor rotation, fuselage swaying, propeller rotation, and engine blade rotation. This information is directly related to target structural attributes (such as the number of rotors and fuselage size), becoming crucial for precise radar identification. Identification techniques based on this effect, through pulse compression and clutter suppression preprocessing of echo signals, generating time-frequency spectra via short-time Fourier transform, and then using convolutional neural networks (CNNs) for classification, can distinguish between similar targets such as propeller-driven aircraft and jet aircraft, making it a hot research topic.

[0004] However, in practical applications of target classification and recognition based on micro-Doppler features, two types of problems severely restrict recognition performance: First, small sample data scenarios are common. Non-cooperative targets cannot actively cooperate with data collection, and the sample size is usually insufficient to support model training. Even for cooperative targets, the difficulty of sample collection increases significantly under special conditions such as extreme weather (heavy rain, strong winds) and complex terrain (mountains, urban canyons), resulting in insufficient effective samples for each class. Data-driven CNN models require thousands of samples for sufficient training, and overfitting is prone to occur in small sample scenarios. The training accuracy can reach over 90%, but the test accuracy drops sharply to below 65%. Second, the frequency distribution and energy accumulation pattern of the time-frequency spectrum of the same target differ significantly under different attitudes (such as tangential flight and radial flight). This cross-attitude feature variation interferes with the model's extraction of the essential features of the target, further exacerbating the decline in recognition accuracy in small sample scenarios and seriously affecting the engineering practicality of the technology.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the problem of "large differences in time-frequency spectrum features and low classification accuracy due to target attitude changes under small sample data" in radar target recognition, this invention provides a radar target recognition method based on attitude angle division. Through a closed-loop process of "heading angle calculation - region division - model training - boundary optimization", it achieves accurate classification of time-frequency spectrum of multiple types of targets.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to a first aspect of the present invention, a radar target identification method based on attitude angle division is provided, the method comprising: Acquire measured data for multiple types of targets. Each target sample includes a time-frequency spectrum, the target's real-time range, azimuth, and elevation angle relative to the radar, as well as the target's range, azimuth, and elevation angle from the previous time slice. Calculate the heading angle based on measured data; Based on the target micro-Doppler features reflected by the time-frequency spectrum, using the known heading and pitch angles, we initially set up the division of angle regions, and then constructed a multi-class target time-frequency spectrum dataset for different angle regions; A dataset is constructed by training a convolutional neural network, and the classification accuracy is then verified. Based on the recognition results, it is determined whether it is necessary to continue iterating the segmentation method and retraining the neural network model. The final heading angle boundary is determined, and the entire dataset is divided based on the fixed boundary. The final convolutional neural network model is then trained. During deployment, the target time-frequency spectrum is input, and the final recognition result is output through the convolutional neural network model.

[0009] In some exemplary embodiments, the time-frequency spectrum refers to a two-dimensional time-frequency spectrum generated by the short-time Fourier transform (STFT) after the radar system collects measured echo signals from multiple types of targets, suppresses clutter, eliminates the fuselage component, and then collects the measured echo signals.

[0010] In some exemplary embodiments, the calculation of the heading angle based on measured data uses the following formula:

[0011] in, , These are the target's real-time range, azimuth, and elevation angles relative to the radar in the current time slice. , , These are the target's distance, azimuth, and elevation angles for the previous time slice; Convert the calculation results to the range of -180° to 180°.

[0012] In some exemplary embodiments, the target micro-Doppler features based on time-frequency maps are used to initially define angular regions using known heading and pitch angles, and then a multi-class target time-frequency map dataset for different angular regions is constructed, specifically as follows: The initial angle region was set by combining the micro-Doppler characteristic discrimination of the time-frequency spectrum with the heading angle. Divide the data into m intervals: ; with pitch angle The partition can be divided into n intervals: ; Combining the two, the region is divided into For each target sample in a sub-region, based on the segmentation results, the time-frequency spectrum is assigned to the corresponding sub-dataset according to the interval to which its heading and pitch angles belong; at the same time, the category label and sample size of each sub-dataset are recorded to ensure that the distribution of the same type of target is balanced in each sub-dataset.

[0013] In some exemplary embodiments, the convolutional neural network is the AlexNet model, the input dataset is the divided dataset, and if there are K types of targets, there are a total of K*m*n sub-targets; the time-frequency spectrum size of the input layer is uniformly 224×224; features are extracted through convolutional layers and pooling layers, and the classification layer outputs the target class probability by a fully connected layer, the number of which is consistent with the number of sub-target classes.

[0014] In some exemplary embodiments, the recognition result includes a comprehensive recognition rate, a confusion matrix used to count the number of misclassifications among different sub-targets, and a global accuracy threshold. , To determine the overall recognition rate of the dataset before partitioning, the overall recognition rate of the dataset after partitioning must be greater than [a certain value]. Set a false positive threshold The false positive rate for the same type of target in different sub-regions is ≤ .

[0015] In some exemplary embodiments, the iterative optimization triggering condition is: the misclassification rate of similar targets in different sub-regions is greater than 1%. If erroneous samples are concentrated near the angle boundary, the boundary is adjusted in the direction of improving accuracy. If erroneous samples in a certain adjacent sub-region are randomly distributed, it indicates that the angle division is too fine, resulting in feature fragmentation. In this case, adjacent sub-regions with similar features are merged.

[0016] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the radar target identification method based on attitude angle division described in the first aspect above.

[0017] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the radar target recognition method based on attitude angle division described in the first aspect is implemented.

[0018] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the radar target recognition method based on attitude angle division described in the first aspect above by executing the executable instructions.

[0019] The radar target recognition method based on attitude angle division provided by the embodiments of the present invention has the following advantages compared with the prior art: 1. Improve classification accuracy in small samples: By dividing the target subclass by heading angle, the consistency of time-frequency spectrum features is enhanced, reducing the model's sensitivity to feature variation in small sample scenarios, resulting in a significant improvement in accuracy compared to traditional undivided methods; 2. Enhanced attitude robustness: The model is trained for different attitude ranges to solve the problem of "different spectrums for the same object" caused by changes in target attitude, which is especially suitable for targets with dynamic attitude changes such as low-altitude aircraft; 3. Adaptive optimization capability: The boundary can be dynamically adjusted through an iterative verification mechanism, which can adapt to the differences in the posture characteristics of different types of targets and has strong versatility.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 This is a schematic diagram of the method flow of an exemplary embodiment of the present invention. Detailed Implementation

[0023] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0024] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a radar target recognition method that improves the classification accuracy of convolutional neural networks by combining target attitude (heading angle, pitch angle) and micro-Doppler features to refine the division of time-frequency spectrum datasets.

[0026] refer to Figure 1 As shown, the specific steps may include: Step 1: Acquire measured data for multiple target types. Each target sample includes a time-frequency spectrum, the target's range, azimuth, and elevation angles relative to the radar, and the range, azimuth, and elevation angles of the target in the previous time slice. Calculate the heading angle based on the measured data. Step 2: Based on the target micro-Doppler features reflected in the time-frequency map, and using the known heading and pitch angles, initially set the division of angle regions, and then construct a multi-class target time-frequency map dataset for different angle regions; Step 3: Use a convolutional neural network to train and construct a dataset, then verify the classification accuracy. Based on the recognition results, determine whether it is necessary to continue iterating the segmentation method and retrain the neural network model. Step 4: Confirm the dividing line, train the neural network model, and finally achieve accurate classification by generating a time-frequency spectrum of a target.

[0027] Furthermore, in step one, the time-frequency spectrum contained in each target sample refers to the two-dimensional time-frequency spectrum generated by short-time Fourier transform (STFT) after the radar system collects measured echo signals of multiple types of targets, suppresses clutter, eliminates fuselage components, and then collects the measured echo signals of multiple types of targets. Furthermore, in step one, the parameters included in each target sample can be expressed as the real-time distance between the target and the radar in the current time slice. , Azimuth (Horizontal angle between the radar line of sight and the radar, ranging from 0° to 360°), elevation angle (The angle between the vertical direction and the radar line of sight, ranging from 0 to 90°, and the distance from the previous time slice) azimuth Pitch angle ; Furthermore, in step one, the formula for calculating the heading angle is:

[0028] The calculation results here are converted to the range of -180° to 180°.

[0029] Furthermore, in step two, the initial angle region is set by combining the micro-Doppler characteristic discrimination of the time-frequency spectrum (such as symmetrical frequency shift fringes generated by rotor / propeller rotation and baseline trend of fuselage motion), with the heading angle as the reference. The partition can be divided into m intervals: ; with pitch angle The partition can be divided into n intervals: Combining the two allows us to divide the region into... A sub-region, for example, the range of a certain sub-region includes (heading angle: - Pitch angle: - Based on the segmentation results, for each target sample, the time-frequency spectrum is assigned to the corresponding subset (e.g., "heading angle") according to the interval to which its heading and pitch angles belong. - Pitch angle - "Sub-datasets"); at the same time, record the category label (such as "sub-region 1-A target", "sub-region 2-A target", "sub-region 1-B target", etc.) and sample size of each sub-dataset to ensure that the distribution of the same type of target is balanced in each sub-dataset.

[0030] Furthermore, in step three, the selected convolutional neural network is the AlexNet model, and the input dataset is the partitioned dataset. If there are K classes of targets, then there are a total of Each sub-target has a uniform time-frequency map size of 224×224. Features are extracted through convolutional and pooling layers, and the classification layer outputs the target category probability from a fully connected layer, with the number of sub-target categories matching the number of sub-target categories.

[0031] Furthermore, in step three, the identification results include the overall identification rate (number of correctly identified samples / total number of samples) and the confusion matrix (statistics of the number of misclassifications among different sub-targets). A global accuracy threshold is set. , To determine the overall recognition rate of the dataset before partitioning, the overall recognition rate of the dataset after partitioning must be greater than [a certain value]. Set a false positive threshold , Based on experience, the misclassification rate is generally less than 0.15, meaning the misclassification rate of the same type of target in different sub-regions (e.g., the proportion of target A in sub-region 1 that is misclassified as target A in sub-region 2) is ≤ .

[0032] Furthermore, in step three, the iterative optimization trigger condition is: the misclassification rate of similar targets in different sub-regions is greater than [a certain value]. If erroneous samples are concentrated near the angular boundary, the boundary is adjusted in the direction of improving accuracy. If erroneous samples are randomly distributed in a certain adjacent sub-region, it indicates that the angular division is too fine, leading to feature fragmentation; therefore, adjacent sub-regions with similar features are merged.

[0033] Furthermore, in step three, the iterative process is as follows: adjust the partition boundaries, update the dataset, and retrain the AlexNet model until the accuracy requirements are met.

[0034] Furthermore, in step four, after iterative optimization in step three, the final heading angle division boundary is determined. Based on the fixed division boundary, the entire dataset is divided, and the final convolutional neural network model is trained. During deployment, the target time-frequency spectrum is input, and the final recognition result is output through the convolutional neural network model.

[0035] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0036] Example 1 Step 1: Data Acquisition and Heading Angle Calculation 1. An experimental platform is deployed in an open area to identify two types of low-altitude targets: propeller aircraft and jet aircraft. The detection range is 5km–25km, with an azimuth of 0°–360° and an elevation of 0°–60°. The experimental data sample size is 100 targets per type, for a total of 200 samples. Each sample contains a time-frequency spectrum generated from the echo signal via Short Time Fourier Transform (STFT) (specific parameters: signal length 2048, Hanning window, window length 64 points, overlap rate 25%, time-frequency spectrum size 224×224); each sample also contains the real-time distance of the target from the radar in the current time slice. , Azimuth (Horizontal angle between the radar line of sight and the radar, ranging from 0° to 360°), elevation angle (The angle between the vertical direction and the radar line of sight, ranging from 0 to 60°, and the distance from the previous time slice) azimuth Pitch angle .

[0037] 2. The heading angle of this sample is obtained based on the following calculation method. :

[0038] Step 2: Preliminary Region Division and Dataset Construction 1. Initial angle region division Due to the inconsistent angles of the main rotating components facing the radar caused by radial and tangential flight of the aircraft, the time-frequency spectra corresponding to different heading angles differ significantly. Based on this, the heading angles can be initially divided into two parts: tangential flight corresponds to (-45°, 45°), (-180°, -135°), and (135°, 180°), while radial flight corresponds to (-135°, -45°) and (45°, 135°).

[0039] Because the pitch angle of the aircraft causes the main rotating components to face the radar at different angles, the time-frequency spectra corresponding to different pitch angles differ significantly. Based on this, the pitch angle can be initially divided into two parts: low pitch angle (0°, 10°) and high pitch angle (10°, 60°).

[0040] This results in four sub-regions: the tangential low pitch region; the tangential high pitch region; the radial low pitch region; and the radial high pitch region. (Example: the tangential low pitch region is: heading angle...) Within the ranges (-45°, 45°), (-180°, -135°), and (135°, 180°), the pitch angle... Within (0°, 10°).

[0041] 2. Dataset Construction Samples were assigned based on angular constraints. For example, propeller aircraft samples with a heading angle of -65° and a pitch angle of 3° were assigned to the low radial pitch angle propeller aircraft subset. Ultimately, there were 8 subsets in total, with each subset containing 20-30 samples per region for propeller aircraft and 20-30 samples per region for jet aircraft, resulting in a balanced distribution.

[0042] Step 3: Model Training and Iterative Optimization 1. The AlexNet model input is a 224×224 grayscale time-frequency spectrum, normalized to [0,1]; training parameters: BatchSize=16, initial learning rate 0.001, training for 50 rounds, number of sub-target categories: 2 categories × 2 × 2 = 8 sub-targets.

[0043] 2. Overall recognition rate before segmentation It is 78.2%, set The initial training result after the division was 0.1. The overall recognition rate was 85% (>78.2%), but the proportion of "tangential low pitch angle propeller aircraft" being misclassified as "tangential high pitch angle propeller aircraft" reached 15% (>10%), and the wrong samples were concentrated around the pitch angle of 9°.

[0044] 3. Improved heading angle boundaries: The low pitch angle (0°, 10°) and high pitch angle (10°, 60°) were adjusted to low pitch angle (0°, 8°) and high pitch angle (8°, 60°). After retraining: the overall recognition rate improved to 89%, and the misclassification rate of similar targets across regions was ≤8% (<10%), meeting the requirements.

[0045] Step 4: Final model training and recognition.

[0046] The final region is divided into four areas: tangential low pitch angle region; heading angle region; and other regions. Within the ranges (-45°, 45°), (-180°, -135°), and (135°, 180°), the pitch angle... Within (0°, 8°); the tangential high pitch angle region is: heading angle Within the ranges (-45°, 45°), (-180°, -135°), and (135°, 180°), the pitch angle... Within (8°, 60°); the radial low pitch angle region is: heading angle Within the range of (-135°, -45°) and (45°, 135°), the pitch angle Within (0°, 8°); the radial pitch angle region is: heading angle Within the range of (-135°, -45°) and (45°, 135°), the pitch angle Within (8°, 60°).

[0047] The two target classes were divided into eight subsets. After training on the full dataset, the test set achieved an overall accuracy of 90.3%, with accuracy rates of 89.2% for propeller aircraft and 91.4% for jet aircraft. This embodiment verifies the effectiveness of the radar target recognition method based on angle division. By finely dividing the heading angle and pitch angle, it solves the problem of attitude feature variation in small sample scenarios. The recognition rate is significantly improved compared with the undivided method, which can meet the actual needs.

[0048] It should be noted that, as another aspect, this application also provides a storage medium, which may be included in an electronic device or may exist independently without being assembled into the electronic device. The storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the methods described in the following embodiments.

[0049] In one embodiment, this application provides a computer program product including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0050] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0051] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0052] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A radar target recognition method based on attitude angle division, characterized in that, The method includes: Acquire measured data for multiple types of targets. Each target sample includes a time-frequency spectrum, the target's real-time range, azimuth, and elevation angle relative to the radar, as well as the target's range, azimuth, and elevation angle from the previous time slice. Calculate the heading angle based on measured data; Based on the target micro-Doppler features reflected by the time-frequency spectrum, using the known heading and pitch angles, we initially set up the division of angle regions, and then constructed a multi-class target time-frequency spectrum dataset for different angle regions; A dataset is constructed by training a convolutional neural network, and the classification accuracy is then verified. Based on the recognition results, it is determined whether it is necessary to continue iterating the segmentation method and retraining the neural network model. The final heading angle boundary is determined, and the entire dataset is divided based on the fixed boundary. The final convolutional neural network model is then trained. During deployment, the target time-frequency spectrum is input, and the final recognition result is output through the convolutional neural network model.

2. The radar target recognition method based on attitude angle division according to claim 1, characterized in that, The time-frequency spectrum refers to the two-dimensional time-frequency spectrum generated by the short-time Fourier transform (STFT) after the radar system collects measured echo signals from multiple types of targets, suppresses clutter, eliminates the fuselage component, and then collects the measured echo signals.

3. The radar target recognition method based on attitude angle division according to claim 1, characterized in that, The formula for calculating the heading angle based on measured data is as follows: in, , These are the target's real-time range, azimuth, and elevation angles relative to the radar in the current time slice. , , These are the target's distance, azimuth, and elevation angles for the previous time slice; Convert the calculation results to the range of -180° to 180°.

4. The radar target recognition method based on attitude angle division according to claim 1, characterized in that, The target micro-Doppler features based on time-frequency maps are used to initially define angular regions using known heading and pitch angles. Then, a multi-class target time-frequency map dataset is constructed for different angular regions, specifically as follows: The initial angle region was set by combining the micro-Doppler characteristic discrimination of the time-frequency spectrum with the heading angle. Divide the data into m intervals: ; with pitch angle The partition can be divided into n intervals: ; Combining the two, the region is divided into For each target sample in a sub-region, based on the segmentation results, the time-frequency spectrum is assigned to the corresponding subset according to the interval to which its heading and pitch angles belong; at the same time, the category label and sample size of each subset are recorded to ensure that the distribution of the same type of target is balanced in each subset.

5. The radar target recognition method based on attitude angle division according to claim 1, characterized in that, The convolutional neural network is the AlexNet model. The input dataset is the divided dataset. If there are K types of targets, there are a total of K*m*n sub-targets. The time-frequency spectrum size of the input layer is uniformly 224×224. Features are extracted through convolutional layers and pooling layers. The classification layer outputs the target class probability by a fully connected layer, and the number is consistent with the number of sub-target classes.

6. The radar target recognition method based on attitude angle division according to claim 1, characterized in that, The recognition results include a comprehensive recognition rate, a confusion matrix used to count the number of misclassifications among different sub-targets, and a global accuracy threshold. , To determine the overall recognition rate of the dataset before partitioning, the overall recognition rate of the dataset after partitioning must be greater than [a certain value]. Set a false positive threshold The false positive rate for the same type of target in different sub-regions is ≤ .

7. The radar target recognition method based on attitude angle division according to claim 6, characterized in that, The iterative optimization trigger condition is: the misclassification rate of similar targets in different sub-regions is greater than [a certain value]. If erroneous samples are concentrated near the angle boundary, the boundary is adjusted in the direction of improving accuracy. If erroneous samples in a certain adjacent sub-region are randomly distributed, it indicates that the angle division is too fine, resulting in feature fragmentation. In this case, adjacent sub-regions with similar features are merged.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the radar target recognition method based on attitude angle division as described in any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the radar target recognition method based on attitude angle division as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the radar target identification method based on attitude angle division as described in any one of claims 1 to 7 by executing the executable instructions.

Images