A radar target classification method and device based on time-frequency feature fusion

Abstract

This invention discloses a radar target classification method and apparatus based on time-frequency feature fusion. In this method, radar echo signals are acquired and preprocessed to obtain multiple frames of target signals. Based on pre-defined micro-Doppler physical models for different target types, time-frequency analysis is performed on the multiple frames of target signals to extract multi-dimensional feature vectors containing micro-Doppler features. Target types include pedestrians, motorcycles, cars, and buses. Variance analysis is used to select features from the multi-dimensional feature vectors, retaining a pre-defined number of key features with the strongest discriminative power to form optimized feature vectors. The optimized feature vectors are then input into a pre-trained support vector machine classification model to obtain the target classification result. This invention achieves accurate and rapid identification of millimeter-wave radar targets through precise micro-Doppler physical modeling and multi-dimensional feature fusion, improving the accuracy and robustness of target classification in complex environments.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing technology, and in particular to a radar target classification method and apparatus based on time-frequency feature fusion. Background Technology

[0002] With the rapid development of autonomous driving, security monitoring, and intelligent transportation, millimeter-wave radar, as one of the core sensing devices, possesses the characteristics of high reliability in all-weather target detection. However, traditional radar target recognition methods mainly rely on the macroscopic motion characteristics of targets, such as speed, distance, and azimuth. These characteristics have limitations in distinguishing different types of targets, such as pedestrians and vehicles, and the classification accuracy is low in complex environments.

[0003] The micro-Doppler effect can capture the minute motion characteristics of internal components of a target, providing a new technical approach for refined target recognition. However, existing target recognition methods based on micro-Doppler features still have the following problems: insufficient extraction of micro-Doppler features, making it difficult to fully reflect the motion characteristics of the target; insufficient utilization of multi-frame temporal information, resulting in insufficient algorithm robustness; complex classifier design, high computational cost, making real-time application difficult; inaccurate modeling of micro-Doppler features for different types of targets; and unoptimized feature selection methods, resulting in redundant features.

[0004] In existing research, most target classification methods based on micro-Doppler use simplified vibration models, which are difficult to accurately characterize complex micro-motion characteristics such as pedestrian gait and vehicle component movement. Furthermore, they lack systematic optimization in feature extraction and selection, resulting in limited classification performance in practical applications.

[0005] Therefore, how to achieve accurate and rapid identification of millimeter-wave radar targets and improve the accuracy and robustness of target classification in complex environments is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] This invention provides a radar target classification method and apparatus based on time-frequency feature fusion. It can achieve accurate and rapid identification of millimeter-wave radar targets through precise micro-Doppler physical modeling and multi-dimensional feature fusion, thereby improving the accuracy and robustness of targets in complex environments.

[0007] The first aspect of this invention provides a radar target classification method based on time-frequency feature fusion, comprising:

[0008] The radar echo signal is acquired, and the radar echo signal is preprocessed to obtain multiple frames of target signal; Based on the pre-defined micro-Doppler physical model for different target types, time-frequency analysis is performed on multi-frame target signals to extract multi-dimensional feature vectors containing micro-Doppler features. The target types include pedestrians, motorcycles, cars and buses. The analysis of variance method is used to select features from the multi-dimensional feature vectors, retaining a preset number of key features with the strongest discriminative power to form an optimized feature vector; The optimized feature vectors are input into a pre-trained support vector machine classification model to obtain the classification result of the target.

[0009] Optionally, the micro-Doppler physical model is a composite motion model constructed based on the physical parameters of the moving parts contained in the target type; the physical parameters include at least the frequency, amplitude, and phase of the part motion.

[0010] Optionally, the micro-Doppler physical model of a pedestrian is composed of motion components including leg swing, arm swing, and torso vibration; the micro-Doppler physical model of a motorcycle and a car includes wheel rotation components; and the micro-Doppler physical model of a bus includes engine vibration and body sway components.

[0011] Optionally, the multi-dimensional feature vector includes time-domain features composed of peak value, mean, standard deviation, root mean square value, skewness and kurtosis; frequency-domain features composed of bandwidth and center frequency; range-image features composed of range spread and peak ratio; Doppler features composed of target radial velocity; micro-Doppler features composed of micro-Doppler energy and micro-Doppler standard deviation; and multi-frame temporal features composed of multi-frame energy variance and multi-frame energy range.

[0012] Optionally, time-frequency analysis is performed on the target signal across multiple frames, including: For each frame of the multi-frame target signal, a short-time Fourier transform is performed along the slow time dimension to generate a time-frequency map, and the inter-frame energy change features are extracted based on the time-frequency map of the multi-frame signal.

[0013] Optionally, the support vector machine classification model adopts a one-to-one multi-class structure, and the kernel function type, penalty parameter, and kernel parameter of the support vector machine classification model are optimized by grid search combined with cross-validation.

[0014] Optionally, grid search optimizes the parameter combination of the support vector machine classification model. The parameter combination includes kernel function type, penalty parameter, and kernel parameter. The kernel function type includes linear kernel and radial basis kernel. The candidate value set of the penalty parameter includes 0.1, 1, and 10. The candidate value set of the kernel parameter includes 0.1, 1, and 10.

[0015] A second aspect of the present invention provides a radar target classification device based on time-frequency feature fusion, comprising: The signal processing module is used to acquire radar echo signals and process them to obtain multi-frame target signals. The feature extraction and selection module is used to perform time-frequency analysis on multi-frame target signals based on a preset micro-Doppler physical model for different target types, extract multi-dimensional feature vectors including micro-Doppler features, and use variance analysis to select features from the multi-dimensional feature vectors, retaining a preset number of key features with the strongest discriminative power to form an optimized feature vector. Target types include pedestrians, motorcycles, cars and buses. The classification module stores a pre-trained support vector machine classification model, which is used to input optimized feature vectors into the model to obtain and output the classification result of the target.

[0016] A third aspect of the present invention provides a radar target classification device based on time-frequency feature fusion, comprising: One or more processors; A memory on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the radar target classification method based on time-frequency feature fusion as described in any of the above.

[0017] A fourth aspect of the present invention provides a computer storage medium for storing a program, which, when executed, is used to implement the radar target classification method based on time-frequency feature fusion as described in any of the preceding claims.

[0018] Beneficial Effects: This invention improves feature discrimination by establishing differentiated composite motion models for different target types through precise micro-Doppler modeling based on physical principles. It employs multi-dimensional feature fusion extraction technology to comprehensively cover the time domain, frequency domain, range profile, Doppler, and micro-Doppler characteristics of the signal, providing richer information. An optimized variance analysis feature selection mechanism automatically selects the most discriminative features and removes redundant features, reducing dimensionality while maintaining classification performance. Combined with multi-frame micro-Doppler analysis technology, it utilizes temporal information to enhance feature discrimination and improves sensitivity to transient micro-motion features. An efficient SVM parameter optimization strategy is used to improve training speed while maintaining classification accuracy. The overall solution achieves accurate and rapid identification of millimeter-wave radar targets, possessing significant practical application value. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1A flowchart illustrating a radar target classification method based on time-frequency feature fusion provided in an embodiment of the present invention; Figure 2 A comparative schematic diagram of various types of micro-Doppler time-frequency maps provided for embodiments of the present invention; Figure 3 This invention provides a scatter plot of SVM feature distribution. Figure 4 A schematic diagram of a classification performance index provided in an embodiment of the present invention; Figure 5 A schematic diagram of feature importance analysis provided in an embodiment of the present invention; Figure 6 A comparative diagram of performance across different categories is provided as an embodiment of the present invention; Figure 7 A box plot of eigenvalue distribution provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the SVM decision boundary provided in an embodiment of the present invention; Figure 9 A schematic diagram of the structure of a radar target classification device based on time-frequency feature fusion provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0021] This invention provides a radar target classification method and apparatus based on time-frequency feature fusion. Through precise micro-Doppler physical modeling and multi-dimensional feature fusion, it achieves accurate and rapid identification of millimeter-wave radar targets, improving the accuracy and robustness of target classification in complex environments. It is applicable to moving target classification and identification in scenarios such as autonomous driving, security monitoring, and intelligent transportation.

[0022] See Figure 1 This figure is a flowchart illustrating a radar target classification method based on time-frequency feature fusion provided by an embodiment of the present invention. The radar target classification method based on time-frequency feature fusion provided by this embodiment of the present invention can be implemented, for example, through the following steps S101-104.

[0023] S101: Acquire radar echo signals and preprocess the radar echo signals to obtain multi-frame target signals.

[0024] In this embodiment of the invention, different types of targets are modeled from a physical model, and radar echo signals of four types of targets, including pedestrians, motorcycles, cars, and buses, are collected. To verify the robustness of the algorithm of this invention, random noise is added to the signals of each type frame by frame to construct the original signal.

[0025] Specifically, the FMCW transmit signal formula is as follows:

[0026] in, The signal is transmitted by the radar, and t is the time of a single chirp. Radar center frequency (which can be set to 77GHz in this invention). This represents the frequency modulation slope.

[0027] The formula for receiving signals is as follows:

[0028] in For radar received signals, The amplitude of the received signal, The round-trip time of the radar signal. The core of the Doppler frequency shift term.

[0029] The intermediate frequency signal is processed using the following mixing (de-skewing) formula: ; ; in, It is an intermediate frequency signal. For the conjugate of the received signal, This refers to the phase information of the intermediate frequency (IF) signal. The phase of the IF signal after mixing is the difference between the phases of the transmitted and received signals. ; Substitute into the specific expression: ; Expand and simplify (ignore) (e.g., higher order, smaller quantity) ; because Further analysis revealed: ; ; ; ; ; in, Based on Doppler, For micro Doppler, Based on the base velocity (the velocity of the target's overall center of mass). This refers to the micro-motion speed (the speed of internal components relative to the center of mass).

[0030] S102: Based on the micro-Doppler physical model preset for different target types, perform time-frequency analysis on multi-frame target signals and extract multi-dimensional feature vectors containing micro-Doppler features.

[0031] In this embodiment of the invention, for each frame of the multi-frame target signal, a short-time Fourier transform is performed along the slow time dimension to generate a time-frequency map, and the inter-frame energy change features are extracted based on the time-frequency map of the multi-frame signal.

[0032] Target types include pedestrians, motorcycles, cars, and buses. The micro-Doppler physical model is a composite motion model constructed based on the physical parameters of the moving parts included in the target type; the physical parameters include at least the frequency, amplitude, and phase of the part's motion. The pedestrian micro-Doppler physical model is composed of motion components including leg swing, arm swing, and torso vibration; the motorcycle and car micro-Doppler physical models include wheel rotation components; the bus micro-Doppler physical model includes engine vibration and body sway components. The multi-dimensional feature vector includes time-domain features composed of peak value, mean, standard deviation, root mean square value, skewness, and kurtosis; frequency-domain features composed of bandwidth and center frequency; range profile features composed of range spread and peak-to-peak ratio; Doppler features composed of target radial velocity; micro-Doppler features composed of micro-Doppler energy and micro-Doppler standard deviation; and multi-frame temporal features composed of multi-frame energy variance and multi-frame energy range. Compared to traditional methods with simple vibration models and single-component characterization models, the micro-Doppler modeling of this invention is a multi-component composite model with differentiated physical parameters. The time-frequency analysis method of this invention improves time-frequency resolution by using adaptive windows, optimizing range cells (analyzing the slow time series of each range cell and averaging the results of all range cells in the range dimension), and splicing and fusing multi-frame radar signals.

[0033] Specifically, STFT time-frequency analysis was performed on the radar signal, using a Hamming window function with a window length of 32, an overlap rate of 50%, and 128 FFT points. Based on the target type, a corresponding micro-Doppler physical model was applied to extract a 25-dimensional comprehensive feature vector. For example... Figure 2 As shown, Figure 2 A comparative schematic diagram of various types of micro-Doppler time-frequency maps provided for embodiments of the present invention; Figure 2 The micro-Doppler time-frequency maps of four typical targets, obtained through simulation or actual measurement, are presented side by side. The images clearly show that, due to significant differences in the motion characteristics of the internal components of different targets (such as pedestrian gait, wheel rotation, engine vibration, etc.), their corresponding time-frequency maps exhibit unique characteristic patterns in terms of texture, energy distribution, and frequency modulation modes. This intuitively demonstrates the effectiveness of micro-Doppler features for target classification.

[0034] The formula for the pedestrian micro-Doppler model is as follows: ; in, Step frequency (in Hz) represents the frequency at which the feet land while walking, ranging from 1.5 to 2.0 Hz, which is the normal walking frequency for adults. The leg swing amplitude (in meters) represents the radial displacement generated by the swing of the thigh, ranging from 0.3 to 0.5 meters, and is the leg length × swing angle (20-30°). The leg swing phase (in rad) represents the initial phase of the leg swing, ranging from 0 to 2π, and is a random initial condition. The arm swing amplitude (in meters) represents the radial displacement generated by the arm swing, ranging from 0.2 to 0.3 meters, approximately 0.6 to 0.8 times that of the leg swing. The arm swing phase (in rad) represents the initial phase of the arm swing, ranging from 0 to 2π, with a phase difference of π / 2 from the leg. The frequency of trunk vibration (in Hz) represents the frequency of vertical trunk vibration, ranging from 2.5 to 3.5 Hz, which is approximately 1.5 to 2 times the stride frequency. The amplitude of trunk vibration (in meters) represents the amplitude of vertical trunk vibration, ranging from 0.05 to 0.07 meters, with a vertical displacement of the center of mass of 4 to 7 centimeters. The phase of the trunk vibration (in rad) represents the initial phase of the trunk vibration, ranging from 0 to 2π, with random initial conditions. Time (in seconds) represents the observation time, ranging from 0 to the total observation time, and is an independent variable that changes with time.

[0035] The formula for the motorcycle micro-Doppler model is as follows: ; in, The wheel rotation frequency (in Hz) represents the angular frequency of wheel rotation, ranging from 15-23Hz, corresponding to vehicle speeds of 43-72km / h. This refers to the wheel radius (in meters), representing the radius of a motorcycle wheel, ranging from 0.3 to 0.4 meters, and is the standard motorcycle tire size. The wheel rotation phase (in rad) represents the initial phase of the wheel rotation, ranging from 0 to 2π, with random initial conditions. The engine vibration frequency (in Hz) represents the engine vibration frequency, ranging from 40-60Hz, corresponding to a speed of 2400-3600 RPM. Engine vibration amplitude (unit: meters) represents the engine vibration amplitude, ranging from 0.02 to 0.03 meters, and is determined by the engine mounting system. The engine vibration phase (in rad) represents the initial phase of the engine vibration, ranging from 0 to 2π, with random initial conditions.

[0036] The formula for the micro-Doppler model of a car is as follows:

[0037] ; in, Wheel rotation frequency (unit: Hz), wheel rotation angular frequency of 10-15Hz corresponds to vehicle speed of 54-90km / h. Wheel radius (in meters) indicates the radius of a car wheel, ranging from 0.35 to 0.45 meters. It is the standard size for car tires. Wheel rotation phase (in rad), representing the initial phase of wheel rotation, ranging from 0 to 2π, with random initial conditions. Vehicle body vibration frequency (unit: Hz) represents the vehicle body vibration frequency, ranging from 8-12Hz, and is the natural frequency of the suspension system. Vehicle body vibration amplitude (unit: m) represents the vibration amplitude of the vehicle body, ranging from 0.03 to 0.04 m, caused by road unevenness. Vehicle body vibration phase (in rad) represents the initial phase of vehicle body vibration, ranging from 0 to 2π, with random initial conditions.

[0038] The formula for the micro-Doppler model of a bus is as follows: ; in, Engine vibration frequency (unit: Hz) indicates the vibration frequency of a diesel engine, ranging from 4-6Hz, corresponding to speeds of 1200-1800 RPM. Engine vibration amplitude (unit: m) represents the vibration amplitude of the engine, ranging from 0.05 to 0.07 m. Larger engines experience greater vibration. Engine vibration phase (in rad) represents the initial phase of engine vibration, ranging from 0 to 2π, under random initial conditions; Body sway frequency (in Hz) indicates the low-frequency swaying frequency of the body, ranging from 1.5 to 2.0 Hz, which is caused by the long wheelbase. Vehicle body sway amplitude (in meters) indicates the amplitude of vehicle body sway, ranging from 0.08 to 0.11 meters, caused by a high center of gravity and soft suspension; Vehicle body sway phase (in rad) represents the initial phase of the vehicle body sway, ranging from 0 to 2π, with random initial conditions.

[0039] Based on the code implementation, the formula for calculating the total Doppler frequency shift is as follows: ; in, Based on Doppler frequency shift; This refers to the micro-Doppler velocity; It is the carrier frequency; It is the speed of light.

[0040] S103: Use the analysis of variance method to select features from the multi-dimensional feature vectors, retain the preset number of key features with the strongest discriminative power, and form an optimized feature vector.

[0041] In this embodiment of the invention, the F-statistic for each feature is calculated using the ANOVA method: ; in, For the first The number of samples in the group; For the first The group mean; This is the population mean; Number of groups (number of categories); This represents the total number of samples. (Selection) The 12 features with the highest values ​​are used as the final feature set to achieve feature dimensionality reduction and optimization. Features are then standardized using Z-scores. ; In the feature selection parameters, the ANOVA significance level α = 0.05, and the number of features selected is 12. This invention identifies the diversity of radar features by fusing features from multiple dimensions, deeply explores the impact of different features on radar target classification, and can form a closed-loop evaluation effect of features on the model.

[0042] See Figure 3 , Figure 3This invention provides an SVM feature distribution scatter plot, illustrating the distribution of four target classes (pedestrian, motorcycle, car, and bus) in a two-dimensional feature space after feature selection and optimization. The plot uses "Main Feature 1" and "Main Feature 2" as coordinate axes, representing the two most discriminative key feature dimensions (or two main components) extracted and selected using the method of this invention. Sample points of different categories are distinguished by different labels (such as shape or color). Sample points belonging to the same category exhibit obvious clustering in the feature space, while there are relatively clear intervals between sample groups of different categories. This visualization strongly demonstrates that the physically modeling-based micro-Doppler feature extraction and optimization method provided by this invention can effectively extract highly discriminative features, making different categories of targets highly separable in the metric space composed of these features, thus laying a solid foundation for subsequent support vector machine classifiers to achieve high-precision classification.

[0043] See Figure 5 , Figure 5 This diagram illustrates a feature importance analysis according to an embodiment of the present invention. It uses a bar chart to show the results of importance assessment and ranking of multiple initial features extracted from radar signals using the ANOVA (Analysis of Variance) method. The horizontal axis represents the "feature" category, listing the extracted multi-dimensional feature types (including micro-Doppler energy, micro-Doppler standard deviation, mean amplitude, range spread, RMS amplitude, amplitude standard deviation, and peak amplitude). The vertical axis represents the "importance score," providing a quantitative importance index for the corresponding feature. This index reflects the feature's ability to distinguish between different target categories (pedestrians, motorcycles, cars, buses); a higher value indicates stronger discriminative power.

[0044] See Figure 7 , Figure 7 This invention provides a box plot of feature value distribution. The box plot visually compares and displays the specific distribution of features selected by ANOVA on sample data of four types of targets (pedestrians, motorcycles, cars, and buses). Figure 7 In the diagram, the vertical axis represents the actual value of the feature, and the horizontal axis represents the feature name. By comparing the box plot positions, medians, and distribution ranges of different categories of targets on the same feature, we can intuitively determine the feature's ability to distinguish between different categories of targets.

[0045] S104: Input the optimized feature vector into the pre-trained support vector machine classification model to obtain the classification result of the target.

[0046] In this embodiment of the invention, the support vector machine (SVM) classification model adopts a one-to-one multi-class structure. The kernel function type, penalty parameter, and kernel parameters of the SVM classification model are optimized using a grid search combined with cross-validation. The grid search optimizes the parameter combinations of the SVM classification model, which include the kernel function type, penalty parameter, and kernel parameters. The kernel function type includes linear kernels and radial basis function kernels. The candidate value set for the penalty parameter includes 0.1, 1, and 10, and the candidate value set for the kernel parameters includes 0.1, 1, and 10.

[0047] Specifically, grid search is used to optimize the SVM parameters: kernel function, penalty parameter C, and kernel parameter γ. Five-fold cross-validation is used to evaluate model performance, and the final One-vs-One multi-class SVM model is trained. The SVM parameters range as follows: kernel function {linear, RBF}, C value {0.1, 1, 10}, and γ value {0.1, 1, 10}. The formula for the linear kernel function is as follows: ; in, Let 1 be the input vector, representing the th The feature vectors of training samples, with dimension . ; Let 2 be the input vector, representing the th The feature vectors of training samples, with dimension . ; The transpose operation represents the vector transpose operation; This is the inner product operation, representing the dot product of two vectors with scalar dimensions.

[0048] The formula for the RBF kernel function is as follows: ; in, Euclidean distance represents the squared distance between two samples, reflecting the similarity between the samples; The kernel width parameter represents the range of locality controlled by the kernel function. The larger the value, the smoother the decision boundary. The smaller the value, the more complex the decision-making boundary. This is an exponential function that maps distance to similarity, with an output range of [0,1].

[0049] The optimized SVM classifier design specifically includes the following: The formula for optimizing the objective function is as follows: ; in, Let be the weight vector, and represent the normal vector of the hyperplane. ; The bias term represents the intercept of the hyperplane, a scalar value; Let be a slack variable, representing the th The classification error of a single sample. ; The penalty parameter, determined through cross-validation, represents the strength of the penalty for controlling classification errors. The larger the value, the fewer misclassifications there are, but overfitting is possible. The smaller the value, the more misclassifications occur, and the simpler the model becomes; This is a regularization term used to control model complexity and prevent overfitting. The empirical risk term represents the sum of training errors and reflects classification performance.

[0050] Constraints: ; Decision function: ; in, For the first The true label of each sample ; For feature mapping, it means mapping the input to a high-dimensional space, which is used to solve nonlinear separable problems; The functional margin represents the signed distance from the sample to the hyperplane; a positive value indicates a correct classification. Let be the margin boundary, representing the standard margin boundary, and let the support vectors satisfy . ; Let be a slack variable, representing the allowed interval violation amount. This indicates that the sample was misclassified. This indicates that the sample was within the interval but correctly classified; For Lagrange multipliers, denote the first... The dual variables of each sample are obtained by solving the dual problem. This is a sign function that outputs the classification result. ; For support vectors, it represents The corresponding samples are the key samples that determine the decision boundary; The kernel function value represents the similarity between the test sample and the support vectors, calculated using the selected kernel function. The bias term is the decision threshold. .

[0051] Duality problem: ; Constraints: ; Multi-classification strategy (One-vs-One): For Each category, build There are 12 binary classifiers. Each classifier votes during testing, and the class with the most votes is the final classification result.

[0052] in, These are Lagrange multipliers, which are dual variables. For KKT conditions; The upper bound of the penalty parameter is an upper bound constraint. To limit the impact of individual samples; The dual feasibility condition is a linear constraint that ensures the rationality of the solution. For Gram matrix elements, there are kernel matrix items that store the similarity of all sample pairs; To maximize the margin, it is the first term of the objective function and is positively correlated with the margin size; To minimize complexity, it is the second term of the objective function, which controls the model complexity; The number of categories represents the total number of target categories. It can be 4; The number of binary classifiers represents the number of classifiers that need to be trained, with one classifier trained for each class pair; the voting mechanism represents the final classification decision, where each classifier votes, and the one with the most votes wins.

[0053] Parameter optimization goal: Minimize cross-validation error using grid search: ; in, It is a loss function and a performance evaluation metric. 0-1 loss: L = number of misclassified samples / total number of samples; It is the first A validation set is used to evaluate model performance; Cross-validation folds represent the number of parts to which the training set is divided, typically 5 or 10. For SVM models, use parameters The trained model is the object that needs to be optimized; grid search is a parameter optimization method. Search for the optimal combination on the grid.

[0054] See Figure 4 , Figure 4 This diagram illustrates a classification performance index provided by an embodiment of the present invention. It visually displays, in bar chart form, the key performance evaluation indicators obtained after classifying four types of targets (pedestrians, motorcycles, cars, and buses) using the method of the present invention. Figure 4 This demonstrates that the method of the present invention can achieve high performance levels in key indicators such as precision and recall, thereby meeting the requirements for the accuracy and reliability of target classification in practical applications.

[0055] See Figure 8 , Figure 8The schematic diagram of the SVM decision boundary provided in this embodiment of the invention intuitively illustrates how the trained SVM classifier classifies different categories of targets in a two-dimensional space composed of two of the most discriminative features (Main Feature 1 and Main Feature 2). In the diagram, the horizontal and vertical axes represent "Main Feature 1" and "Main Feature 2," respectively, and the scale range of both axes is [-1, 2]. Sample points from four target categories (pedestrian, motorcycle, car, bus) are distributed in this feature space as scatter points of different colors or shapes according to their category. The decision boundary learned by the SVM is represented as a curve or a polygonal region composed of multiple straight line segments. These boundaries divide the feature plane into different regions, each corresponding to a target category. Figure 8 As can be clearly observed, although there may be some overlap or noise in the distribution of sample points, SVM successfully finds the complex nonlinear boundary that can maximally separate samples of different categories through its optimized parameters (such as the kernel function and penalty parameters determined by grid search). This visualization result proves that based on the optimized feature vector extracted in this invention, the SVM model can effectively learn a robust decision function, thereby achieving high-precision classification of four target classes in the feature space.

[0056] In one implementation of this invention, a stratified sampling method is used to divide the training set and the test set, ensuring a consistent proportion of samples in each category. The training set comprises 70% of the sample data, and the test set comprises 30%. Classification prediction is performed on the test set samples, and performance metrics such as accuracy, precision, recall, and F1-score are calculated. The classification results and confidence scores are then output. The classification performance metrics are: accuracy > 97% and average F1-score > 0.96.

[0057] See Figure 6 , Figure 6 This is a performance comparison diagram of various categories provided in an embodiment of the present invention. It uses a grouped bar chart to compare and show the specific performance of four types of targets in three key performance indicators: precision, recall, and F1-Score. Figure 6 In the graph, the vertical axis represents the "score," indicating the quantified values ​​of various performance indicators; the horizontal axis lists four target categories: pedestrians, motorcycles, cars, and buses. For each target category, a set of parallel bar charts represents its precision, recall, and F1-Score values, distinguished by legends. This invention's method maintains high precision and recall across different target categories, with relatively balanced values, resulting in a high overall F1-Score.

[0058] Through practical testing, the algorithm of this invention performs excellently in four types of target recognition tasks. This invention achieves accurate and rapid identification of millimeter-wave radar targets through precise micro-Doppler physical modeling, comprehensive feature extraction, and optimized SVM classifier design, and has important practical application value.

[0059] Beneficial effects: Through actual simulation tests, the algorithm of this invention performs well in four types of target recognition tasks, with an overall recognition accuracy of 98.5%; the recognition accuracy for pedestrians is 97.8%, for motorcycles it is 97.5%, for cars it is 98.8%, and for buses it is 99.7%.

[0060] This invention employs precise micro-Doppler modeling based on physical principles, more closely resembling the motion characteristics of real targets and improving the feature discrimination of different target types. Through multi-dimensional, multi-domain feature fusion extraction, it comprehensively covers the multi-dimensional characteristics of signals, providing richer information compared to traditional methods. The optimized ANOVA feature selection mechanism automatically selects features with the highest F-values, removing redundant features and reducing dimensionality while maintaining classification performance, making it more objective and efficient than manual feature selection. Multi-frame micro-Doppler analysis technology is used, dividing the observation time into multiple frames for processing, extracting inter-frame energy change features, and combining temporal information to enhance feature discrimination and improve sensitivity to transient micro-motion features. An efficient SVM parameter optimization strategy and simplified grid search are employed, significantly improving training speed while maintaining classification accuracy. Physical modeling, feature engineering, and machine learning are organically combined to form a complete technical solution. Through precise micro-Doppler physical modeling, comprehensive feature extraction, and optimized SVM classifier design, accurate and rapid identification of millimeter-wave radar targets is achieved, demonstrating significant practical application value.

[0061] Based on the methods provided in the above embodiments, this invention also provides a radar target classification device based on time-frequency feature fusion. The radar target classification device based on time-frequency feature fusion is described below with reference to the accompanying drawings.

[0062] See Figure 9 The figure is a schematic diagram of a radar target classification device based on time-frequency feature fusion provided in an embodiment of the present invention.

[0063] The radar target classification device 900 based on time-frequency feature fusion provided in this embodiment of the invention includes: a signal processing module 901, a feature extraction and selection module 902, and a classification module 903.

[0064] The signal processing module 901 is used to acquire radar echo signals and process the radar echo signals to obtain multi-frame target signals. The feature extraction and selection module 902 is used to perform time-frequency analysis on multi-frame target signals based on a preset micro-Doppler physical model for different target types, extract multi-dimensional feature vectors including micro-Doppler features, and use variance analysis to select features from the multi-dimensional feature vectors, retaining a preset number of key features with the strongest discriminative power to form an optimized feature vector. The target types include pedestrians, motorcycles, cars and buses. The classification module 903 stores a pre-trained support vector machine classification model, which is used to input optimized feature vectors into the model to obtain and output the classification result of the target.

[0065] In one possible implementation, the micro-Doppler physical model is a composite motion model constructed based on the physical parameters of the moving parts contained in the target type; the physical parameters include at least the frequency, amplitude, and phase of the part motion.

[0066] In one possible implementation, the micro-Doppler physical model of a pedestrian is composed of motion components including leg swing, arm swing, and torso vibration; the micro-Doppler physical model of a motorcycle and a car includes wheel rotation components; and the micro-Doppler physical model of a bus includes engine vibration and body sway components.

[0067] In one possible implementation, the multi-dimensional feature vector includes time-domain features consisting of peak value, mean, standard deviation, root mean square value, skewness, and kurtosis; frequency-domain features consisting of bandwidth and center frequency; range-image features consisting of range spread and peak ratio; Doppler features consisting of target radial velocity; micro-Doppler features consisting of micro-Doppler energy and micro-Doppler standard deviation; and multi-frame temporal features consisting of multi-frame energy variance and multi-frame energy range.

[0068] In one possible implementation, the feature extraction and selection module 902 has the following functions: For each frame of the multi-frame target signal, a short-time Fourier transform is performed along the slow time dimension to generate a time-frequency map, and the inter-frame energy change features are extracted based on the time-frequency map of the multi-frame signal.

[0069] In one possible implementation, the support vector machine classification model adopts a one-to-one multi-class structure, and the kernel function type, penalty parameter, and kernel parameter of the support vector machine classification model are optimized by combining grid search with cross-validation.

[0070] In one possible implementation, grid search optimizes the parameter combination of the support vector machine classification model, which includes kernel function type, penalty parameter, and kernel parameter. The kernel function type includes linear kernel and radial basis kernel, the candidate value set of the penalty parameter includes 0.1, 1, and 10, and the candidate value set of the kernel parameter includes 0.1, 1, and 10.

[0071] Since the radar target classification device 900 based on time-frequency feature fusion is a device corresponding to the radar target classification method based on time-frequency feature fusion provided in the above method embodiments, the specific implementation of each unit of the radar target classification device 900 based on time-frequency feature fusion is based on the same concept as in the above method embodiments. Therefore, for the specific implementation of each unit of the radar target classification device 900 based on time-frequency feature fusion, please refer to the description of the radar target classification method based on time-frequency feature fusion in the above method embodiments, and it will not be repeated here.

[0072] This invention also provides a radar target classification device based on time-frequency feature fusion, the device comprising: a processor and a memory; The memory is used to store instructions; The processor is used to execute the instructions in the memory to perform the radar target classification method based on time-frequency feature fusion mentioned in the above embodiments.

[0073] It should be noted that the hardware structure of the radar target classification device based on time-frequency feature fusion provided in the embodiments of the present invention can be as follows: Figure 10 The structure shown, Figure 10 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention.

[0074] Please see Figure 10 As shown, device 1000 includes: a processor 1010, a communication interface 1020, and a memory 1030. The number of processors 1010 in device 1000 can be one or more. Figure 10 Taking a processor as an example, in this embodiment of the invention, the processor 1010, communication interface 1020, and memory 1030 can be connected via a bus system or other means. Figure 10 Taking the connection between China and Israel via the 1040 bus system as an example.

[0075] Processor 1010 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 1010 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0076] The memory 1030 may include volatile memory, such as random-access memory (RAM); the memory 1030 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 1030 may also include a combination of the above types of memory.

[0077] Optionally, the memory 1030 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic services and handling hardware-based tasks. The processor 1010 can read the programs in the memory 1030 to implement the radar target classification method based on time-frequency feature fusion provided in this embodiment of the invention.

[0078] The bus system 1040 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system 1040 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 10The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0079] This invention also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the radar target classification method based on time-frequency feature fusion mentioned in the above embodiments.

[0080] This invention also provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the radar target classification method based on time-frequency feature fusion mentioned in the above embodiments.

[0081] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A radar target classification method based on time-frequency feature fusion, characterized in that, include: Acquire radar echo signals and preprocess the radar echo signals to obtain multiple frames of target signals; Based on the micro-Doppler physical model preset for different target types, time-frequency analysis is performed on the multi-frame target signals to extract multi-dimensional feature vectors containing micro-Doppler features. The target types include pedestrians, motorcycles, cars and buses. The multidimensional feature vector is selected using the analysis of variance method, and the preset number of key features with the strongest discriminative power are retained to form an optimized feature vector. The optimized feature vector is input into a pre-trained support vector machine classification model to obtain the classification result of the target.

2. The method according to claim 1, characterized in that, The micro-Doppler physical model is a composite motion model constructed based on the physical parameters of the moving parts contained in the target type; the physical parameters include at least the frequency, amplitude and phase of the component motion.

3. The method according to claim 2, characterized in that, The pedestrian's micro-Doppler physical model is composed of motion components including leg swing, arm swing, and torso vibration; the motorcycle and the car's micro-Doppler physical models include wheel rotation components; and the bus's micro-Doppler physical model includes engine vibration and vehicle body sway components.

4. The method according to claim 1, characterized in that, The multi-dimensional feature vector includes time-domain features composed of peak value, mean, standard deviation, root mean square value, skewness, and kurtosis; frequency-domain features composed of bandwidth and center frequency; range-image features composed of range spread and peak ratio; Doppler features composed of target radial velocity; micro-Doppler features composed of micro-Doppler energy and micro-Doppler standard deviation; and multi-frame temporal features composed of multi-frame energy variance and multi-frame energy range.

5. The method according to claim 1, characterized in that, The time-frequency analysis of the multi-frame target signal includes: For each frame of the multi-frame target signal, a short-time Fourier transform is performed along the slow time dimension to generate a time-frequency map, and the inter-frame energy change features are extracted based on the time-frequency map of the multi-frame signal.

6. The method according to claim 1, characterized in that, The support vector machine classification model adopts a one-to-one multi-class structure. The kernel function type, penalty parameter, and kernel parameter of the support vector machine classification model are optimized by grid search combined with cross-validation.

7. The method according to claim 6, characterized in that, The grid search optimizes the parameter combination of the support vector machine classification model. The parameter combination includes kernel function type, penalty parameter, and kernel parameter. The kernel function type includes linear kernel and radial basis kernel. The candidate value set of the penalty parameter includes 0.1, 1, and 10. The candidate value set of the kernel parameter includes 0.1, 1, and 10.

8. A radar target classification device based on time-frequency feature fusion, characterized in that, include: The signal processing module is used to acquire radar echo signals and process the radar echo signals to obtain multiple frames of target signals. The feature extraction and selection module is used to perform time-frequency analysis on the multi-frame target signals based on the preset micro-Doppler physical model for different target types, extract multi-dimensional feature vectors including micro-Doppler features, and use the variance analysis method to select features from the multi-dimensional feature vectors, retaining a preset number of key features with the strongest discriminative power to form an optimized feature vector. The target types include pedestrians, motorcycles, cars and buses. The classification module stores a pre-trained support vector machine classification model, which is used to input the optimized feature vector into the model to obtain and output the classification result of the target.

9. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-7 above.

Images