Radar low-altitude target intelligent recognition method based on multi-dimensional feature fusion and dual-mode decision

Abstract

The application provides a radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and double-mode decision, and relates to the technical field of radar data processing. The method comprises: processing the radar signal, extracting the radar features and motion features of the target; through an iterative feature collinearity elimination method based on the vector angle, a high-quality and low-redundancy feature set is screened out, on the basis of which, a single-cycle identification model and a continuous-cycle identification model based on a machine learning model are respectively constructed, and the hyperparameters are optimized through Bayesian optimization, the discrimination results of the two models are integrated through a weighted fusion strategy, and the final category determination of small unmanned aerial vehicles, medium-sized unmanned aerial vehicles, birds and air balls is output. The method realizes high-precision and high-efficiency low-altitude target identification, has high trajectory identification accuracy, extremely low resource occupation, and has excellent engineering landing value.

Description

Technical Field

[0001] This application relates to the field of radar data processing technology, and in particular to a radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making. Background Technology

[0002] Currently, the low-altitude economy is rapidly emerging, encompassing multiple new business models such as urban air traffic, logistics and distribution, drone inspection, agricultural plant protection, and tourism. This trend not only drives the development of related industrial chains but also poses unprecedented challenges to the safety supervision capabilities of low-altitude airspace. Especially in complex scenarios such as cities and industrial parks, low-altitude airspace faces a large number of "low, slow, and small" flying targets, such as small multi-rotor drones, fixed-wing drones, and birds with similar characteristics. These targets generally have characteristics such as slow flight speed, small radar cross-section, and unstable flight trajectory, bringing prominent problems to traditional radar systems, including low recognition rate, slow response speed, and poor anti-interference capability, seriously restricting the supervision capabilities and safety assurance level of low-altitude airspace. Summary of the Invention

[0003] This application provides a radar low-altitude target intelligent recognition method based on multi-dimensional feature fusion and dual-mode decision-making to achieve intelligent recognition of new low-altitude radar targets (mainly targeting medium and small UAVs, birds, and air-launched balls). The core idea includes: short-time radar feature extraction, cumulative time-period radar feature extraction, motion feature extraction, and based on the extracted features, constructing a time-series model intelligent recognition algorithm that integrates single-period judgment and continuous-period judgment to achieve short-time continuous high-precision recognition of UAVs, birds, and air-launched balls.

[0004] Firstly, this application provides a radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making, including:

[0005] Radar feature extraction includes target localization, radar derived variable calculation, target point radar feature extraction, local space radar feature extraction, Doppler power spectrum feature extraction, and micro-Doppler feature extraction.

[0006] Motion feature extraction, including point trajectory acquisition and motion feature extraction;

[0007] Feature collinearity removal involves sequentially performing initial correlation calculation, feature vector standardization, angle-based collinearity checking, and an iterative process based on extracted radar and motion features to obtain a dataset after feature collinearity removal. The iterative process includes repeatedly performing initial correlation calculation, feature vector standardization, and angle-based collinearity checking using the remaining feature set until all features satisfy the angle constraints with other features in the set. The remaining feature set is the feature set for which initial correlation calculation, feature vector standardization, and angle-based collinearity checks have not been performed. The feature set includes extracted radar features and motion features.

[0008] Model construction and recognition include dataset partitioning, model training, hyperparameter optimization, and weighted fusion decision-making; wherein, the weighted fusion decision-making includes: weighted fusion of the discrimination results of the single-period model and the continuous-period model to generate the final recognition result, which is the target category label determined by fusion.

[0009] In one possible design, the target localization includes: performing a fast Fourier transform on the pulse-compressed radar IQ signal to convert the time-domain information into frequency-domain information, and determining the spatial location of the target point based on the frequency-domain information and echo power, wherein the spatial location of the target point includes the Doppler velocity and the distance to the radar;

[0010] The calculation of radar-derived variables includes: calculating radar reflectivity factor, radial velocity, and spectral width;

[0011] The target point radar feature extraction includes: extracting the original radar echo signal and radar-derived variables of the target point in a single frame observation; calculating the mean, standard deviation, maximum and minimum values ​​and extreme values ​​of the feature information of all frames in a single period as training data for the single period model; and accumulating the feature information of all frames in a continuous period and calculating the statistics as training data for the continuous period model.

[0012] The local spatial radar feature extraction includes: extracting radar variable information within the local space where the target point is located, calculating the statistical features of single-frame data, including maximum value, mean, median, standard deviation, skewness, kurtosis, total energy, extreme values, signal-to-noise ratio, and coefficient of variation; averaging the statistical features of all frames within a single period as training data for a single-period model; and accumulating and averaging the statistical features of each target sample across all frames in a continuous period as training data for a continuous-period model.

[0013] The Doppler power spectrum feature extraction includes: extracting the Doppler power spectrum distribution along the distance dimension after target point localization, and calculating the features of the global and local spectra. The features of the global and local spectra include the Doppler velocity mean, standard deviation, local symmetry, half-width, amplitude, number of local peaks, left-right half-width ratio, local energy proportion, local centroid, and local spectral width. For a single-period model, the average of the features of the global and local spectra of all frames within a set period is used as training data; for a continuous-period model, the average of the features of the global and local spectra of all frames within a cumulative period is used as training data.

[0014] The micro-Doppler feature extraction includes: extracting the original IQ signal of the target point distance dimension from all frame data within a single period, resampling after splicing, performing slow Fourier decomposition to obtain time-frequency joint information, and extracting micro-Doppler features. The micro-Doppler features include the mean, standard deviation, and range of the dominant frequency, the standard deviation of the rate of change of the dominant frequency, the average spectral bandwidth, the average energy per time segment, spectral symmetry features, texture smoothness, local contrast, and image uniformity. The single-period extraction results are used as single-period training data, and the micro-Doppler features are calculated after continuous period accumulation and used as continuous period training data.

[0015] In one possible design, the formula for calculating the radar reflectivity factor is:

[0016]

[0017] In the formula, Radar reflectivity factor Power spectrum, For calibration constant, The square of the distance the radar travels to the target.

[0018] The formula for calculating the radial velocity is:

[0019]

[0020] In the formula, Radial velocity, This refers to the IQ data value corresponding to the Doppler velocity index k in the IQ data matrix (dimensions: Doppler velocity dimension, distance bin dimension). For Doppler velocity channel index, The velocity value corresponding to the Doppler channel at index k;

[0021] The formula for calculating the spectral width is:

[0022]

[0023] In the formula, For spectral width.

[0024] In one possible design, the point track acquisition includes: performing moving target detection on the radar IQ signal, generating a range-Doppler two-dimensional spectrum, refining the point track through constant probability detection and peak extraction algorithms, calculating the precise range, radial velocity, azimuth angle, pitch angle, amplitude, signal-to-noise ratio and the number of original point tracks, and generating a stable track through track initiation, point track-track association, tracking filtering and state update;

[0025] The motion feature extraction includes: for single-period models, directly using Doppler velocity, amplitude, signal-to-noise ratio, number of original points, total velocity, X-direction velocity, Y-direction velocity, Z-direction velocity, and heading from the point track data as training data; for continuous-period models, calculating the mean and standard deviation of the point track data, and further calculating the normalized average velocity, normalized standard deviation of heading, and maneuvering factor, wherein the maneuvering factor is defined as the ratio of the normalized average velocity to the normalized standard deviation of heading.

[0026] In one possible design, the initial correlation calculation includes: calculating the absolute Pearson correlation coefficient between each original feature in the feature set and the true label vector, and selecting the original feature with the highest absolute Pearson correlation coefficient as the initial reference feature, wherein the original features include radar features and motion features;

[0027] The feature vector standardization includes: standardizing each original feature to zero mean and unit variance to obtain a standardized reference feature vector and a standardized candidate feature vector;

[0028] The angle-based collinearity check includes: calculating a standardized reference feature vector. With each of the other standardized candidate feature vectors The angle between cosine value ,like If collinearity exists between the candidate feature and the reference feature, the candidate feature is removed from the feature set.

[0029] In one possible design, the dataset partitioning includes: dividing the single-cycle and continuous-cycle datasets after collinearity removal into training set, validation set, and test set, respectively;

[0030] The model training includes: training a single-cycle model and a continuous-cycle model using machine learning algorithms;

[0031] The hyperparameter optimization includes: using Bayesian optimization to tune the hyperparameters of the single-period model, and combining a periodic update strategy to optimize the continuous-period model;

[0032] In one possible design, the single-period model is a random forest, XGBoost, or LightGBM model with a decision tree architecture, and the continuous-period model is a random forest, XGBoost, or LightGBM model with a decision tree architecture.

[0033] Secondly, this application provides a radar low-altitude target intelligent identification device based on multi-dimensional feature fusion and dual-mode decision-making, the device comprising:

[0034] The first feature extraction module is configured to perform radar feature extraction, including target localization, radar derived variable calculation, target point radar feature extraction, local space radar feature extraction, Doppler power spectrum feature extraction, and micro-Doppler feature extraction.

[0035] The second feature extraction module is configured to perform motion feature extraction, including point trajectory acquisition and motion feature extraction;

[0036] The feature commonality removal module is configured to perform feature collinearity removal, including, based on extracted radar features and motion features, sequentially performing initial correlation calculation, feature vector standardization, angle-based collinearity check, and an iterative process to obtain a dataset after feature collinearity removal; wherein, the iterative process includes: using the remaining feature set, repeatedly performing initial correlation calculation, feature vector standardization, and angle-based collinearity check until all features satisfy the angle constraint with other features in the set. The remaining feature set is the feature set for which initial correlation calculation, feature vector standardization, and angle-based collinearity checks have not been performed. The feature set includes extracted radar features and motion features.

[0037] The target intelligent recognition module is configured to perform model construction and recognition, including dataset partitioning, model training, hyperparameter optimization, and weighted fusion decision; wherein, the weighted fusion decision includes: weighted fusion of the discrimination results of the single-period model and the continuous-period model to generate the final recognition result, wherein the final recognition result is the target category label determined by fusion.

[0038] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to execute the intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making as described in the first aspect and various possible designs of the first aspect.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the intelligent radar low-altitude target identification method based on multi-dimensional feature fusion and dual-mode decision-making as described in the first aspect and various possible designs of the first aspect.

[0040] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making as described in the first aspect and various possible designs of the first aspect.

[0041] The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making provided in this application has at least the following beneficial effects:

[0042] This application addresses the core challenge of intelligent identification of "low, slow, and small" flying targets in low-altitude airspace. It constructs an intelligent detection scheme that integrates microscopic structural features and macroscopic motion patterns, aiming to improve the accuracy, response speed, and resource utilization efficiency of radar systems for identifying small targets. From target point extraction, radar multidimensional feature modeling, Doppler / micro-Doppler analysis, and temporal evolution identification, to intelligent fusion modeling and lightweight deployment, this application forms a complete technical route that is operable, scalable, and highly adaptable.

[0043] In large-scale independent testing, the proposed solution demonstrated extremely high recognition accuracy and stability, with an average trajectory recognition accuracy exceeding 95%, maintaining an accuracy of approximately 93% even in complex and extreme scenarios, and a single-cycle recognition accuracy consistently around 97%. Furthermore, the proposed solution exhibits excellent first-response capability, with an average earliest recognition cycle of 1.2 seconds, meaning that the vast majority of targets can be accurately identified within the first cycle after entering the monitoring range.

[0044] More importantly, the entire system has extremely high resource utilization efficiency. It can achieve rapid identification in a CPU environment without relying on GPU. The identification time for a single trajectory is no more than 10 seconds, the single cycle response time is less than 3 seconds, and the CPU usage is less than 10%, and the memory consumption is no more than 300MB. It has good real-time performance, lightweight design and deployability, which greatly enhances the practical value and engineering implementation potential of the low-altitude airspace monitoring system. Attached Figure Description

[0045] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0046] Figure 1 A flowchart of a radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making provided for embodiments of this application;

[0047] Figure 2 Resource usage during operation of the intelligent radar low-altitude target identification method based on multi-dimensional feature fusion and dual-mode decision-making provided in this application embodiment: the middle column is CPU usage, and the right column is memory usage;

[0048] Figure 3 This is a structural diagram of a radar low-altitude target intelligent identification device based on multi-dimensional feature fusion and dual-mode decision-making, provided in an embodiment of this application.

[0049] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0050] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0051] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0052] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0053] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0054] This application provides a radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making, such as... Figure 1 As shown, the intelligent identification method for low-altitude radar targets based on multi-dimensional feature fusion and dual-mode decision-making includes the following steps S10-S40.

[0055] S10: Radar feature extraction, including the following steps S101-S106.

[0056] S101: Target positioning.

[0057] For the radar IQ signal after pulse compression to achieve artificial target information, the time domain information is converted into frequency domain information through fast Fourier transform, and the spatial position of the target point is determined based on the frequency domain information and echo power; the spatial position of the target point includes the radial Doppler velocity of the target and its distance from the radar.

[0058] S102: Calculation of radar-derived variables.

[0059] Calculate radar reflectivity factor :

[0060]

[0061] In the formula:

[0062] Power spectrum;

[0063] Calibration constant;

[0064] : The square of the distance at which the radar reaches the target.

[0065] Calculate radial velocity (Expected value of velocity distribution (first-order statistical moment)):

[0066]

[0067] In the formula:

[0068] : The IQ data value corresponding to the Doppler velocity index k in the IQ data matrix (dimension: Doppler velocity dimension, distance bin dimension);

[0069] : Index for Doppler velocity channels;

[0070] : The velocity value corresponding to the Doppler channel when the index is k.

[0071] Calculate spectral width (Second central moment of velocity distribution):

[0072]

[0073] S103: Target point radar feature extraction.

[0074] Assuming accurate identification of low-altitude target points, the original radar echo signal of the target point is extracted from a single frame observation, and radar-derived variable information, such as spectral width, radial velocity, and radar reflectivity factor, is further extracted. The mean, standard deviation, maximum and minimum values, and extreme values ​​of the feature information from all frames within a single period (generally 2-5 frames) are calculated and used as one of the training data for single-period model construction. For continuous periods, the total number of periods for each target sample is accumulated (generally 10-25 periods per sample), meaning the nth period contains information from all periods from 1 to n. These features are then used as one of the training data for continuous later-stage model construction.

[0075] S104: Local Space Radar Feature Extraction.

[0076] For the local space where the target point is located (generally adjusted according to radar performance and target size; the optimal value in testing is 3 for the Doppler axis and 5 for the range axis, i.e., a 3×5 local space), radar variable information (such as echo intensity, radar reflectivity, radial velocity, spectral width, etc.) is extracted within this local region. The statistical characteristics of single-frame data within this region are analyzed and calculated, including: maximum value, mean, median, standard deviation, skewness, kurtosis, total energy, extreme values, signal-to-noise ratio, and coefficient of variation. The statistical characteristics of all frames within a single period are averaged and used as one of the training data for single-period model construction. For continuous periods, the average is calculated by progressively accumulating the total number of periods for each target sample, and this average is used as one of the training data for continuous later-stage models.

[0077] S105: Doppler power spectrum feature extraction.

[0078] After locating the target point, Doppler information in the distance dimension is extracted to form a Doppler power spectrum distribution. Based on this distribution, the following features are calculated for the entire spectrum and the local spectrum near the target point: Doppler velocity mean, Doppler velocity standard deviation, local symmetry (Doppler velocity symmetry within the target area), half-width (the position where power drops to half the peak value), kurtosis (total energy), number of local peaks (number of local peaks at and near the target point), aspect ratio, proportion of local energy to total window energy, local centroid (weighted velocity center near the target's main peak, more accurately representing target motion characteristics), and local spectral width (range of local velocity extension). For both single-period and continuous-period datasets, the mean of all frames within a specific period or cumulative period is used as one of the training datasets.

[0079] S106: Micro-Doppler feature extraction.

[0080] For all frames within a single period, the original IQ signal of the distance dimension after target point localization is extracted, and these signals are then stitched together frame by frame and resampled. Based on the resampled data, slow Fourier decomposition is performed to obtain joint information in the time and frequency domains, capturing the micro-motion characteristics of different targets (such as UAV rotors, bird wing flapping, etc.). Based on the micro-Doppler signal, after filtering out extreme values, the following features are extracted:

[0081] The mean of the dominant frequency reflects the target's mean Doppler velocity;

[0082] The standard deviation of the dominant frequency characterizes the degree of velocity fluctuation; a high value may indicate that the target is changing speed or rotating.

[0083] The range of the dominant frequency, quantizing the range of velocity variation, such as the amplitude of the vibrating target;

[0084] The standard deviation of the dominant frequency change rate reflects the fluctuation of acceleration and identifies non-uniform acceleration targets;

[0085] Average spectral bandwidth; a wide bandwidth may indicate more micro-motion characteristics around the target.

[0086] The average energy per time segment is used to assess the level of sustained echo.

[0087] Spectral symmetry characteristics, the energy ratio between the two regions on either side of the dominant frequency, and asymmetry may indicate unilateral motion, such as the flapping of a bird's wings;

[0088] Texture smoothness;

[0089] Local contrast (texture coarseness);

[0090] Image uniformity.

[0091] The single-cycle extraction results are used as one of the training datasets for the single-cycle model, while the continuous-cycle results are accumulated and then the above features are calculated, which are used as one of the training datasets for the continuous-cycle model.

[0092] S20: Motion feature extraction, including the following steps S201-S202.

[0093] S201: Point Track Acquisition

[0094] This part differs from the microscopic features observed by radar; it involves refining and tracking target information step by step, acquiring its point trajectory information, and extracting features. The main steps are as follows:

[0095] First, the compressed IQ signal from the radar pulse is used for moving target detection via Fast Fourier Transform (FFT), generating a range-Doppler two-dimensional spectrum to separate the moving target from stationary clutter and measure its radial velocity. Then, a constant false alarm rate (CFAR) detection technique is employed, adaptively setting the detection threshold based on the noise and clutter power of surrounding cells to identify all potential candidate targets with a constant CFAR probability. Next, a peak extraction algorithm is used to find the point of maximum power within the local region of these candidate targets, suppressing surrounding non-peak detection and achieving data refinement from a "spot" to a "point". Finally, for each refined point, its precise information is calculated: the precise range is obtained through range gate indexing and interpolation; the radial velocity is calculated using the Doppler frequency; the azimuth and elevation angles are measured using the phase difference between different channels; and the amplitude, signal-to-noise ratio, and number of original points are recorded simultaneously.

[0096] After the radar completes the condensation of the raw IQ signal into point information, these isolated and discontinuous point sequences are correlated and smoothed into a stable and continuous target motion trajectory. This process begins with track initiation. When a sequence of points conforming to the initial motion pattern appears in an adjacent spatial region within several consecutive scan cycles, a provisional track is created. Then, the core point-track correlation process begins. The system predicts a correlation gate for each existing track in its next cycle; this is an adaptively sized region centered on the predicted position. Points falling within this gate are called candidate points. The correlation algorithm (global optimal allocation) searches for the best match among these candidate points for the track, typically based on the combined similarity between the point and the predicted position in terms of distance, velocity, and angle.

[0097] After successful association, the system proceeds to tracking filtering and state updates. Techniques such as Kalman filtering smooth the trajectory, suppress measurement noise, and accurately predict the target's position at the next moment, providing a basis for the next association. Finally, successfully associated and updated tracks are confirmed or maintained; tracks that fail to associate continuously are terminated; and isolated tracks that remain unassociated may become seeds for a new provisional track. This process repeats continuously, ultimately outputting reliable tracks with unique numbers, stable and smooth, reflecting the target's real-time motion and intentions. Based on these tracks, the radar calculates the velocity and heading information for each track.

[0098] S202: Motion feature extraction.

[0099] For the single-period model, Doppler velocity, sum of amplitude, signal-to-noise ratio, number of raw points, total velocity, X-direction velocity, Y-direction velocity, Z-direction velocity, and heading are directly used to form one of the training datasets from the point track data. For the continuous-period model, Doppler velocity, sum of amplitude, signal-to-noise ratio, number of raw points, total velocity, X-direction velocity, Y-direction velocity, and Z-direction velocity are superimposed, and their mean, standard deviation, and other statistics are calculated. In addition, the normalized mean velocity is calculated and scaled to the [0,1] range, and the normalized standard deviation of the heading is also calculated and scaled to the [0,1] range. Based on this, a maneuvering factor (normalized mean velocity / normalized standard deviation of heading) is calculated to measure the target maneuverability. The smaller the value, the more drastic the target turn (high maneuverability), and the larger the value, the more linear the target moves (low maneuverability). These features are used to construct the continuous-period model.

[0100] S30: Feature collinearity removal.

[0101] A rich feature dataset was constructed based on steps S10 and S20. To ensure the independence between features used to train the model and to minimize redundancy, an iterative process was adopted to eliminate collinear features. The specific steps of this process are as follows: S301-S304.

[0102] S301: Initial correlation calculation.

[0103] Calculate each raw (unnormalized) feature (i.e., the radar feature and motion feature mentioned above). Compared to the true label vectors (pre-defined by humans, indicating whether each sample belongs to a large or small drone, bird, or boulders). The absolute Pearson correlation coefficient between them :

[0104]

[0105] In the formula, Features standard deviation Features With tags Covariance between For tags The standard deviation.

[0106] Choose the one with the highest Features As an initial reference feature.

[0107] S302: Eigenvector standardization.

[0108] Each feature vector (i.e., the initial reference features) are standardized to zero mean and unit variance to eliminate scale dependence:

[0109]

[0110] in and They represent The mean and standard deviation, This is a standardized feature vector.

[0111] S303: Angle-based collinearity check.

[0112] Calculate the standardized reference eigenvector With each of the other standardized candidate feature vectors The angle between cosine value :

[0113]

[0114] like (Right now, or If these features are collinear, then they are considered to be collinear and removed.

[0115] S304: Iterative process.

[0116] Using the next most relevant feature in the remaining feature set as the new reference feature, repeat steps S301-S303 until all retained features satisfy the angle constraint with other features in the set. .

[0117] S40: Model building and recognition, including the following steps S401-S404.

[0118] S401: Dataset partitioning.

[0119] The original radar features and motion features constructed in steps S10 and S20 are then processed using the collinearity removal method in step S30 to obtain high-quality features with low collinearity, which serve as the final training dataset. It is important to note that this single-cycle dataset and continuous-cycle dataset require separate collinearity removal. After obtaining the complete training dataset, it is randomly split into a training set, a validation set, and a test set. The training set is used to train the model, the validation set is used to optimize model performance, and the test set is used to test the model's generalization ability.

[0120] S402: Model training.

[0121] Machine learning algorithms are used to train both single-period and continuous-period models. The main reason for separating these two modeling approaches is to avoid the drawbacks of traditional deep learning solutions, such as high resource consumption, high computational requirements, and weak instantaneous judgment capabilities in engineering applications. Therefore, after obtaining the complete training dataset, traditional machine learning algorithms are chosen, such as Random Forest (with decision tree architecture), XGBoost, and LightGBM. Model selection needs to be tested in conjunction with the actual data distribution to select the optimal model. Here, the XGBoost model is used as an example.

[0122] XGBoost (Extreme Gradient Boosting) is an advanced machine learning model based on the gradient boosting framework. Its core idea is to sequentially integrate multiple weak learners (decision trees), allowing subsequent models to continuously correct the prediction errors of preceding models, thus forming a powerful ensemble model. Compared to traditional gradient boosting models, XGBoost's core innovation lies in its carefully designed objective function. It explicitly incorporates a regularization term to control model complexity and prevent overfitting, and utilizes the second-order Taylor expansion of the loss function (i.e., simultaneously using the first-order gradient and the second-order Hessian matrix) to more accurately and quickly approximate the optimal solution. Furthermore, XGBoost integrates several engineering innovations such as parallel processing, automatic handling of missing values, and efficient sparse data optimization, giving it extremely high prediction accuracy while also possessing excellent computational efficiency and generalization ability. As a result, XGBoost excels in processing structured / tabular data, has long dominated data science competitions, and has become one of the preferred tools in industry for solving classification, regression, and ranking problems.

[0123] S403: Hyperparameter optimization.

[0124] After determining to use the XGBoost model, fine-grained hyperparameter optimization for different task scenarios is a key step in improving model performance. For single-cycle tasks, Bayesian optimization is adopted as the core tuning strategy. This method constructs a probabilistic surrogate model (such as a Gaussian process) to fit the black-box functional relationship between hyperparameters and model performance (such as validation set loss), and intelligently selects the next set of hyperparameters that are expected to bring the greatest performance gain for evaluation using a sampling function (such as the desired improvement in EI). This sequential process of "evaluation-update-selection" enables Bayesian optimization to explore a complex hyperparameter space (including learning rate, maximum tree depth, regularization parameters lambda and alpha, etc.) with far greater efficiency than grid search and random search within a limited number of evaluations, ultimately locking in the globally optimal or near-optimal parameter combination, achieving a delicate balance between exploration and utilization. For continuous periodic tasks (such as models that need to be periodically updated with new data), the Bayesian optimization framework is combined with a periodic update strategy: the above methods are used to fully optimize the model in the first period; in subsequent periods, the optimal parameters of the previous period can be used as prior knowledge, or a new Bayesian optimization process can be started in a reduced parameter space around the historical optimal point. This ensures that the model can adapt to data drift while significantly improving the tuning efficiency of each model iteration, ensuring that the model always maintains the best performance throughout the entire trajectory.

[0125] S404: Weighted fusion decision.

[0126] After obtaining two optimal models—a single-cycle model optimized for static scenarios and a continuous-cycle model adapted to the temporal evolution of data—this invention adopts a weighted fusion-based decision integration strategy to achieve lightweight deployment and combine the advantages of both: the discrimination results of the single-cycle model and the continuous-cycle model are weighted and fused, and assigned weights of 0.4 and 0.6 respectively (the weights can also be optimized according to actual task testing) to generate the final recognition result of the system. This weighting scheme aims to allow the continuous periodic model, which better reflects long-term trends and temporal dependencies, to play a dominant role, while retaining the stable discrimination capability of the single-period model at specific moments as an effective supplement. This fusion approach not only maximizes the discrimination advantages of different time dimensions from an information theory perspective, effectively correcting potential local misjudgments that may exist in a single model, thus significantly optimizing the overall trajectory recognition accuracy and robustness, but more importantly, its output results naturally inherit the inherent temporal correlation and state continuity of radar observation data through the emphasis on the weights of the continuous periodic model. This makes the final recognized trajectory logically smoother and more coherent, strictly conforming to the actual observation time sequence pattern of the radar system and the dynamic evolution law in the physical world.

[0127] This embodiment evaluates the resource utilization efficiency of the method. For example... Figure 2 As shown in the test results, without relying on GPU acceleration, this method can complete the recognition process for a single trajectory within 10 seconds by relying entirely on the CPU, with a single cycle recognition time of less than 3 seconds, and the response speed meets the requirements of real-time detection. During this process, system resource consumption is extremely low: CPU utilization remains below 10%, and memory usage does not exceed 500MB. This demonstrates the lightweight and low-resource-dependency characteristics of the method's design, making it suitable for deployment on resource-constrained edge computing devices or existing low-altitude monitoring systems, and possessing good practical implementation capabilities and engineering promotion value.

[0128] This application also provides an intelligent radar target identification device based on multi-dimensional feature fusion and dual-mode decision-making, such as... Figure 3 As shown, the radar low-altitude target intelligent identification device based on multi-dimensional feature fusion and dual-mode decision-making includes:

[0129] The first feature extraction module 301 is configured to perform radar feature extraction, including target localization, radar derived variable calculation, target point radar feature extraction, local space radar feature extraction, Doppler power spectrum feature extraction, and micro-Doppler feature extraction.

[0130] The second feature extraction module 302 is configured to perform motion feature extraction, including point trajectory acquisition and motion feature extraction;

[0131] The feature commonality removal module 303 is configured to perform feature collinearity removal, including, based on the extracted radar features and motion features, sequentially performing initial correlation calculation, feature vector standardization, angle-based collinearity check, and an iterative process to obtain a dataset after feature collinearity removal; wherein, the iterative process includes: using the remaining feature set, repeatedly performing initial correlation calculation, feature vector standardization, and angle-based collinearity check until all features satisfy the angle constraint with other features in the set. The remaining feature set is the feature set for which initial correlation calculation, feature vector standardization, and angle-based collinearity checks have not been performed. The feature set includes extracted radar features and motion features.

[0132] The target intelligent recognition module 304 is configured to perform model construction and recognition, including dataset partitioning, model training, hyperparameter optimization, and weighted fusion decision; wherein, the weighted fusion decision includes: weighted fusion of the discrimination results of the single-period model and the continuous-period model to generate the final recognition result, wherein the final recognition result is the target category label determined by fusion.

[0133] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0134] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0135] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0136] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0137] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making described in the above embodiments.

[0138] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making in the above embodiments.

[0139] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0140] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0141] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0142] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0143] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0144] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0145] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0146] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0147] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0148] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A radar low-altitude target intelligent identification method based on multi-dimensional feature fusion and dual-mode decision-making, characterized in that, The method includes: Radar feature extraction includes target localization, radar derived variable calculation, target point radar feature extraction, local space radar feature extraction, Doppler power spectrum feature extraction, and micro-Doppler feature extraction. Motion feature extraction, including point trajectory acquisition and motion feature extraction; Feature collinearity removal involves sequentially performing initial correlation calculation, feature vector standardization, angle-based collinearity checking, and an iterative process based on extracted radar and motion features to obtain a dataset after feature collinearity removal. The iterative process includes repeatedly performing initial correlation calculation, feature vector standardization, and angle-based collinearity checking using the remaining feature set until all features satisfy the angle constraints with other features in the set. The remaining feature set is the feature set for which initial correlation calculation, feature vector standardization, and angle-based collinearity checks have not been performed. The feature set includes extracted radar features and motion features. Model construction and recognition include dataset partitioning, model training, hyperparameter optimization, and weighted fusion decision-making; wherein, the weighted fusion decision-making includes: weighted fusion of the discrimination results of the single-period model and the continuous-period model to generate the final recognition result, which is the target category label determined by fusion.

2. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to claim 1, characterized in that, The target localization includes: performing a fast Fourier transform on the pulse-compressed radar IQ signal to convert the time-domain information into frequency-domain information, and determining the spatial location of the target point based on the frequency-domain information and echo power. The spatial location of the target point includes the Doppler velocity and the distance to the radar. The calculation of radar-derived variables includes: calculating radar reflectivity factor, radial velocity, and spectral width; The target point radar feature extraction includes: extracting the original radar echo signal and radar-derived variables of the target point in a single frame observation; calculating the mean, standard deviation, maximum and minimum values ​​and extreme values ​​of the feature information of all frames in a single period as training data for the single period model; and accumulating the feature information of all frames in a continuous period and calculating the statistics as training data for the continuous period model. The local spatial radar feature extraction includes: extracting radar variable information within the local space where the target point is located, calculating the statistical features of single-frame data, including maximum value, mean, median, standard deviation, skewness, kurtosis, total energy, extreme values, signal-to-noise ratio, and coefficient of variation; averaging the statistical features of all frames within a single period as training data for a single-period model; and accumulating and averaging the statistical features of each target sample across all frames in a continuous period as training data for a continuous-period model. The Doppler power spectrum feature extraction includes: extracting the Doppler power spectrum distribution along the distance dimension after target point localization, and calculating the features of the global and local spectra. The features of the global and local spectra include the Doppler velocity mean, standard deviation, local symmetry, half-width, amplitude, number of local peaks, left-right half-width ratio, local energy proportion, local centroid, and local spectral width. For a single-period model, the average of the features of the global and local spectra of all frames within a set period is used as training data; for a continuous-period model, the average of the features of the global and local spectra of all frames within a cumulative period is used as training data. The micro-Doppler feature extraction includes: extracting the original IQ signal of the target point distance dimension from all frame data within a single period, resampling after splicing, performing slow Fourier decomposition to obtain time-frequency joint information, and extracting micro-Doppler features. The micro-Doppler features include the mean, standard deviation, and range of the dominant frequency, the standard deviation of the rate of change of the dominant frequency, the average spectral bandwidth, the average energy per time segment, spectral symmetry features, texture smoothness, local contrast, and image uniformity. The single-period extraction results are used as single-period training data, and the micro-Doppler features are calculated after continuous period accumulation and used as continuous period training data.

3. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to claim 2, characterized in that, The formula for calculating the radar reflectivity factor is as follows: In the formula, The radar reflectivity factor is represented by `mtd_raw`, which is the IQ data processed by Fast Fourier Transform. Power spectrum, For calibration constant, The square of the distance the radar travels to the target. The formula for calculating the radial velocity is: In the formula, Radial velocity, This refers to the IQ data value corresponding to the Doppler velocity index k in the IQ data matrix. For Doppler velocity channel index, The velocity value corresponding to the Doppler channel at index k; The formula for calculating the spectral width is: In the formula, For spectral width.

4. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to claim 1, characterized in that, The point track acquisition includes: performing moving target detection on the radar IQ signal, generating a range-Doppler two-dimensional spectrum, refining the point track through constant probability detection and peak extraction algorithms, calculating the precise range, radial velocity, azimuth angle, pitch angle, amplitude, signal-to-noise ratio and original number of point tracks, and generating a stable track through track initiation, point track-track association, tracking filtering and state update; The motion feature extraction includes: for single-period models, directly using Doppler velocity, amplitude, signal-to-noise ratio, number of original points, total velocity, X-direction velocity, Y-direction velocity, Z-direction velocity, and heading from the point track data as training data; for continuous-period models, calculating the mean and standard deviation of the point track data, and further calculating the normalized average velocity, normalized standard deviation of heading, and maneuvering factor, wherein the maneuvering factor is defined as the ratio of the normalized average velocity to the normalized standard deviation of heading.

5. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to claim 1, characterized in that, The initial correlation calculation includes: calculating the absolute Pearson correlation coefficient between each original feature in the feature set and the true label vector, and selecting the original feature with the highest absolute Pearson correlation coefficient as the initial reference feature. The original features include radar features and motion features. The feature vector standardization includes: standardizing each original feature to zero mean and unit variance to obtain a standardized reference feature vector and a standardized candidate feature vector; The angle-based collinearity check includes: calculating a standardized reference feature vector. With each of the other standardized candidate feature vectors The angle between cosine value ,like If collinearity exists between the candidate feature and the reference feature, the candidate feature is removed from the feature set.

6. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to any one of claims 1 to 5, characterized in that, The dataset partitioning includes: dividing the single-cycle and continuous-cycle datasets after collinearity removal into training sets, validation sets, and test sets, respectively; The model training includes: training a single-cycle model and a continuous-cycle model using machine learning algorithms; The hyperparameter optimization includes: using Bayesian optimization to tune the hyperparameters of the single-period model, and combining a periodic update strategy to optimize the continuous-period model.

7. The intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making according to claim 6, characterized in that, The single-period model is a random forest, XGBoost, or LightGBM model with a decision tree architecture, and the continuous-period model is a random forest, XGBoost, or LightGBM model with a decision tree architecture.

8. A radar low-altitude target intelligent identification device based on multi-dimensional feature fusion and dual-mode decision-making, characterized in that, The device includes: The first feature extraction module is configured to perform radar feature extraction, including target localization, radar derived variable calculation, target point radar feature extraction, local space radar feature extraction, Doppler power spectrum feature extraction, and micro-Doppler feature extraction. The second feature extraction module is configured to perform motion feature extraction, including point trajectory acquisition and motion feature extraction; The feature commonality removal module is configured to perform feature collinearity removal, including, based on extracted radar features and motion features, sequentially performing initial correlation calculation, feature vector standardization, angle-based collinearity check, and an iterative process to obtain a dataset after feature collinearity removal; wherein, the iterative process includes: using the remaining feature set, repeatedly performing initial correlation calculation, feature vector standardization, and angle-based collinearity check until all features satisfy the angle constraint with other features in the set. The remaining feature set is the feature set for which initial correlation calculation, feature vector standardization, and angle-based collinearity checks have not been performed. The feature set includes extracted radar features and motion features. The target intelligent recognition module is configured to perform model construction and recognition, including dataset partitioning, model training, hyperparameter optimization, and weighted fusion decision; wherein, the weighted fusion decision includes: weighted fusion of the discrimination results of the single-period model and the continuous-period model to generate the final recognition result, wherein the final recognition result is the target category label determined by fusion.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory to implement the intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the intelligent radar target identification method based on multi-dimensional feature fusion and dual-mode decision-making as described in any one of claims 1-7.

Images