A radar target type adaptive recognition method, device and electronic equipment

Abstract

This invention discloses an adaptive radar target type identification method, device, and electronic device, belonging to the field of radar signal processing technology. The method includes: acquiring and preprocessing radar echo signals; independently measuring the Doppler initial velocity, range, and angle of the target; establishing a physical constraint model of the generalized linear Chirplet transform frequency modulation parameters to determine the coarse search range; searching for the optimal frequency modulation parameters using the energy concentration criterion; performing a generalized linear Chirplet transform to obtain the time spectrum; extracting the fused feature vector; calculating its feature distance to various centers in a pre-stored feature library; and outputting the identification result based on the relationship between distance and a threshold. This invention guides parameter optimization through physical constraints, reducing computational complexity, making it suitable for real-time operation on low-computing-power edge devices. It also supports known target identification, potential new target caching, and online incremental learning of unknown targets, offering advantages such as high accuracy, anti-interference, and low deployment cost.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and target recognition technology, and specifically relates to a radar target type adaptive recognition method, device and electronic equipment. Background Technology

[0002] Target identification is a key technology for enhancing the intelligence and application value of radar systems. In practical applications, radar needs to accurately distinguish between different types of targets within its detection range (such as pedestrians, cars, electric bicycles, motorcycles, swaying trees, swaying billboards, etc.) to provide a reliable basis for subsequent decision-making (such as security early warning, environmental monitoring, autonomous driving, etc.). However, the radar detection environment is complex and variable, and target echo signals often exhibit characteristics such as non-stationarity, low signal-to-noise ratio, and similar micro-motion features, posing a significant challenge to the accurate identification of target types.

[0003] Currently, radar target type identification methods can be mainly classified into the following categories: (1) Identification method based on conventional parameters This type of method utilizes conventional parameters such as the amplitude, range, and velocity of the target echo to determine the target type by setting a threshold or constructing a simple classifier. This method is computationally simple and easy to implement, and was widely used in early radar systems.

[0004] The drawbacks are that conventional parameters are greatly affected by environmental clutter, occlusion, multipath effects, etc., and it is difficult to effectively distinguish targets with similar micro-motion characteristics (such as pedestrians and electric bicycles), resulting in low recognition accuracy and poor adaptability.

[0005] (2) Deep learning-based recognition methods In recent years, with the development of deep learning technology, this type of method has become a research hotspot for radar target recognition. A typical approach involves first extracting the target's time-frequency map using time-frequency analysis methods such as Short-Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT), and then inputting the time-frequency map into a deep learning model such as Convolutional Neural Network (CNN) or Transformer for classification and recognition. This type of method has achieved good recognition results in closed-world scenarios and on high-computing platforms.

[0006] The drawbacks are: deep learning models have a large number of parameters (usually greater than 1MB), and the inference process relies on a large number of matrix multiplication operations, making it difficult to achieve real-time operation on low-computing-power edge chips such as the ARM Cortex-M series; in addition, this type of method relies on a large number of labeled samples for training, has poor adaptability to new scenes and new targets, and cannot effectively handle unknown target types in open worlds.

[0007] (3) Identification method based on time-frequency analysis This type of method utilizes time-frequency analysis tools such as Wigner-Ville distribution, wavelet transform, and Chirplet transform to extract non-stationary time-frequency features of radar echoes, and then combines them with traditional classifiers for target recognition. Among them, Chirplet transform, as a generalized time-frequency analysis method, can more flexibly match non-stationary signals whose frequency changes linearly with time by introducing a linear frequency modulation parameter (Reference: Mann S, Haykin S. "The chirplet transform: Physical considerations." IEEE Transactions on Signal Processing, 1995, 43(11): 2745-2761, Translated by Mann S, Haykin S. "Chirplet transform: Physical considerations", IEEE Transactions on Signal Processing, 1995, Vol. 43, No. 11, pp. 2745-2761).

[0008] The drawbacks are: the standard STFT has a fixed time-frequency resolution and cannot adaptively match the instantaneous frequency changes of the signal; although the standard Chirplet transform has parameter adaptation capabilities, its parameter optimization usually adopts gradient descent or brute-force search strategies, resulting in high computational complexity (the computational load of parameter optimization is proportional to the number of search points), making it difficult to meet the real-time processing requirements of edge computing. The Generalized Linear Chirplet Transform (GLCT), as a generalization of the Chirplet transform, further expands the adaptability of the basis functions, but current GLCT applications are mainly concentrated on high-computing-power platforms (such as rotating machinery fault diagnosis relying on PC-based real-time calculations) or offline analysis scenarios (such as post-processing analysis of radar micro-Doppler features), and have not yet covered online target recognition and low-computing-power edge computing adaptation.

[0009] In summary, the existing technologies still have the following key problems that have not been effectively resolved: Insufficient adaptability: Existing methods struggle to adaptively detect and identify unknown new targets in open-world scenarios, typically only able to identify a limited set of predefined categories; Time-frequency parameter matching is difficult: the instantaneous frequency variation law of radar echo signal is unknown, and existing methods are unable to adaptively select the optimal time-frequency analysis parameters, resulting in limited accuracy of time-frequency feature extraction; Edge computing power constraints: The high computational complexity of parameter optimization in advanced time-frequency analysis methods such as GLCT poses a significant contradiction to the real-time processing requirements of edge devices with low computing power and low power consumption. Insufficient utilization of physical priors: Existing pure mathematical transformation methods lack in-depth utilization of the physical processes of radar detection, resulting in high blindness in parameter search and poor noise resistance. Fixed-point precision guarantee: Edge devices mostly use fixed-point arithmetic. Existing methods cannot effectively guarantee the orthogonality of GLCT basis functions in fixed-point implementation, which is prone to time-spectrum artifacts.

[0010] Therefore, there is an urgent need for an adaptive radar target type identification method that can adaptively match the characteristics of radar echo signals, unknown targets in the open world, and edge computing power constraints. Summary of the Invention

[0011] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a radar target type adaptive identification method, device and electronic equipment that reduces computational complexity by guiding parameter optimization through physical constraints, and has the advantages of high precision, anti-interference and low deployment cost.

[0012] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: An adaptive radar target type identification method includes the following steps: Step S1: Acquire radar echo signals and preprocess the radar echo signals; Step S2: Independently measure the physical parameters of the target, including the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion ; Step S3: Establish the frequency modulation parameters of the generalized linear Chirplet transform based on the physical parameters. The physical constraint model is used to determine the frequency modulation parameters. The broad search range; Step S4: Within the coarse search range, using the energy concentration criterion as the optimization objective, search for and obtain the optimal frequency modulation parameter. ; Step S5: Utilize the optimal tuning frequency parameters The time spectrum is obtained by performing a generalized linear Chirplet transform on the preprocessed radar echo signal. Step S6: Extract the fused feature vector from the time spectrum; Step S7: Calculate the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and output the target recognition result based on the relationship between the feature distance and the preset threshold.

[0013] As a further improvement to the above technical solution: In step S5, the generalized linear Chirplet transform uses a Gaussian window-modulated linear frequency modulation basis function, which is defined as: in, The width of the Gaussian window. Centered on time The initial frequency, For optimal frequency modulation parameters, It is the imaginary unit.

[0014] In step S3, the specific method for establishing the physical constraint model is as follows: calculate the fundamental frequency modulation frequency using the following formula. : Then, calculate the comprehensive error coefficient based on the speed measurement error and the distance measurement error. Thus, the coarse search range is determined as follows: in, For radar wavelength, To ensure a long observation window, For equipment calibration coefficients, For speed measurement error, This represents the distance measurement error.

[0015] In step S4, the optimization objective based on the energy concentration criterion is specifically as follows: Calculate candidate frequency modulation parameters The second Rayleigh entropy of the spectrum corresponding to the generalized linear Chirplet transform. : Choose to The minimum frequency modulation parameter is taken as the optimal parameter. ,in Indicates the frequency modulation parameter is The generalized linear Chirplet transform time spectrum.

[0016] The search process in step S4 is divided into two stages: Phase 1: Within the coarse search range, perform the search using a preset step size. M candidate parameter points are generated, where M≤5. The pre-computed basis functions are obtained by looking up a table. The time-spectral energy concentration of each candidate point is calculated, and the parameter point with the highest concentration is selected as the coarse search result. ; Phase Two: Within the Interval Within, dynamic step size is used. Sub-iteration fine-tuning, where Determine the final optimal tuning frequency parameters. .

[0017] In step S6, extracting the fusion feature vector includes: Texture feature extraction: Uniform local binary pattern features are extracted from the gray-scale time-frequency spectrum obtained by the generalized linear Chirplet transform, resulting in... 3D texture feature vector ,in ; Geometric feature extraction: Calculate the energy centroid trajectory of the time-frequency spectrum to obtain the instantaneous frequency sequence. And the sequence is reduced in dimensionality to obtain 3D geometric eigenvectors ,in ; Feature fusion: concatenating texture features and geometric features into a fused feature vector. .

[0018] The instantaneous frequency sequence of the energy centroid trajectory The calculation method is as follows: in, For the first Each time unit For the first One frequency unit, For the time spectrum in time unit and frequency unit The amplitude value at that point.

[0019] In step S6, after extracting the fused feature vector, a micro-motion feature consistency verification step is also included: Periodicity measure of the extracted spectrum and stability measures ,in: For the energy center of gravity time series, It is the autocorrelation function; That is, the variance of the energy centroid sequence difference, where This represents variance calculation. This represents the first-order difference; Will , As part of the output features, it participates in subsequent target matching.

[0020] In step S7 The pre-stored feature library is constructed as follows: training samples of known target categories are collected, multiple fused feature samples are extracted for each target category, and the feature centers of that target category are calculated. Maximum distance within a class minimum distance between classes ;in, Based on the above statistics, set the identification threshold: Threshold threshold ,in The number of samples for each target class. For the first Class 1 The fused feature vector of each sample, Represents the feature distance. This is the distance fluctuation margin caused by noise.

[0021] The feature distance Weighted Euclidean distance: in, and These are the first and second feature vectors of the test sample and the reference feature vector, respectively. One portion, For the first The weights of each component are determined by the intra-class scatter matrix and inter-class scatter matrix of the target known during the training phase.

[0022] In step S7, the output of the target recognition result based on the relationship between the feature distance and the preset threshold specifically includes: Calculate the fusion features of test samples With various centers Feature distance ,make ; like If so, it is determined to be a known target, and the target type is... and output the confidence level. ; like If it is identified as a potential new target, its features are stored in a temporary cache. like And echo energy Greater than or equal to the energy threshold And continuous If a frame appears, it is determined to be an unknown target, triggering online incremental learning; Otherwise, it is judged as clutter or noise and discarded; in, Energy threshold The threshold for the number of consecutive frames.

[0023] The energy threshold The consecutive frame number threshold is determined by environmental noise floor statistics. It is a preset positive integer.

[0024] The online incremental learning specifically includes: When the number of the same type of unknown target features in the temporary cache reaches the clustering threshold During the idle time between frames, perform the following operations asynchronously: Calculate the distance between cached features and all known class centers. If more than 50% of the features are less than the clustering threshold for a given known class center, then... If the sample is identified as a known target mutation, then the updated class centers of the target mutation sample are: ; Otherwise, cluster the cache features; if the intra-cluster distance is less than 1 / 3, then... And the distance between classes is greater than If so, a new target class center is generated, added to the feature library, and marked as a new target to be confirmed; in, The minimum number of features required for clustering. To update the coefficients, This indicates the operation of calculating the average.

[0025] It also includes feature library capacity management and anti-drift mechanisms: The feature library stores the most There are 1 target categories, among which ; Sort by most recent access time, eliminating those older than [time]. Categories where time did not appear, among which ; Periodically perform incremental updates on the known class centers. The updated class centers after the anti-drift update are as follows: in To prevent drift, update the coefficients.

[0026] It also includes privacy protection mechanisms: Local persistent storage of fused feature vectors and class center The original radar echo signal is only temporarily stored in the on-chip buffer of the real-time signal processing pipeline. It does not participate in non-volatile storage, nor is it transmitted to the outside. The temporary buffer is released immediately after processing, so it is impossible to deduce the original echo information from the residual data. When synchronizing to the cloud, only the feature library update log is uploaded, and the update log includes changes in class centers. .

[0027] It also includes a fixed-point optimization step: The Gaussian window exponent term in the generalized linear Chirplet transform is pre-calculated using a lookup table method. Stored as an 8-bit or 16-bit fixed-point number; The CORDIC algorithm is used to calculate the linear frequency modulation term. The complex exponent in the number of iterations ; A fixed-point fast Fourier transform is used to achieve fast computation of the generalized linear Chirplet transform.

[0028] It also includes the processing of auxiliary sensor signals: Simultaneously acquire auxiliary sensor signals for scene classification and adaptive correction of the search step size during the fine-tuning stage of the generalized linear Chirplet transform; When an auxiliary sensor failure is detected, the system automatically switches to single radar mode: the fine-tuning stage search step size is fixed to the default value, the number of coarse search points is expanded to a maximum of 20, and the processing frame rate is reduced.

[0029] In the single radar mode, the processing frame rate is reduced from 12fps to 8fps.

[0030] The preprocessing includes bandpass filtering, noise reduction, and amplification; the independent measurement of the Doppler initial velocity... The Doppler frequency shift was extracted from the original echo signal using the FFT spectral peak detection method and then converted.

[0031] An adaptive radar target type identification device, comprising: A signal preprocessing module is used to acquire radar echo signals and preprocess the radar echo signals; The physical constraint calculation module is used to independently measure the physical parameters of the target, including the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion And based on the physical parameters, establish the frequency modulation parameters of the generalized linear Chirplet transform. The physical constraint model is used to determine the frequency modulation parameters. The broad search range; The GLCT transform module is used to search for the optimal frequency modulation parameter within the coarse search range, using the energy concentration criterion as the optimization objective. And using the optimal tuning frequency parameter The time spectrum is obtained by performing a generalized linear Chirplet transform on the preprocessed radar echo signal. The feature extraction module is used to extract the fused feature vector from the time spectrum; The dynamic matching module is used to calculate the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and output the target recognition result according to the relationship between the feature distance and the preset threshold. The feature library management module is used to store and update the local feature library.

[0032] An electronic device includes a memory and a processor, the processor being configured to execute a computer program stored in the memory to implement the radar target type adaptive identification method as described above.

[0033] Compared with the prior art, the advantages of the present invention are as follows: The radar target type adaptive identification method of the present invention identifies targets by independently measuring the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion Physical parameters are used to establish the frequency modulation parameters in the generalized linear Chirplet transform. The physical constraint model is used to determine... The coarse search range—this feature transforms the search for frequency modulation parameters from a blind global search without prior guidance to a locally constrained search with physical constraints. This reduces the number of candidate parameters that need to be evaluated and lowers the computational complexity of parameter optimization. As a result, this method can adapt time-frequency transformation parameters with lower computational overhead on edge devices with limited computing power.

[0034] Secondly, within the coarse search range, the optimal frequency modulation parameter is obtained by searching using the energy concentration criterion as the optimization objective. —This feature ensures that the selected parameters can make the energy distribution of the time spectrum more concentrated, thereby adaptively matching the instantaneous frequency change pattern of the radar echo signal. Compared with the short-time Fourier transform with fixed time-frequency resolution, it can more accurately characterize the local time-frequency features of non-stationary signals and improve the feature extraction accuracy of weak targets.

[0035] Secondly, utilize The time spectrum is obtained by performing a generalized linear Chirplet transform, and a fused feature vector is extracted from it. Since the fused feature vector integrates information from multiple dimensions in the time spectrum, it can more robustly characterize the motion pattern of the target and reduce noise and clutter interference compared to methods that rely on only a single feature.

[0036] Finally, by calculating the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and outputting the recognition result based on the relationship between the feature distance and a preset threshold, this mechanism does not rely on large-scale training of deep neural networks, nor does it require knowing all possible target categories. It can autonomously complete target classification at the edge, and can detect signals different from known categories when the feature distance exceeds the preset threshold range, thus possessing the ability to distinguish unknown targets in the open world. In summary, this invention comprehensively realizes a real-time, high-precision, interference-resistant radar target type recognition method with open-world adaptive recognition capabilities on low-computing-power edge devices.

[0037] The radar target type adaptive identification device of this invention achieves the same technical effects as corresponding methods through the collaborative work of a signal preprocessing module, a physical constraint calculation module, a GLCT transformation module, a feature extraction module, a dynamic matching module, and a feature library management module: Physical constraint-guided adaptive optimization of GLCT parameters reduces computational complexity, enabling the device to run in real-time on low-computing-power edge platforms; the combination of high-focus spectrum and fusion feature extraction improves identification accuracy and noise resistance; the dynamic matching module and feature library management module support known target identification in open worlds, caching of potential new targets, and online incremental learning triggered by unknown targets, achieving dynamic expansion of the local feature library without pre-training unknown samples, while locally storing only feature vectors and class centers to protect data privacy. An asynchronous processing mechanism between modules ensures that the real-time identification main process is not blocked.

[0038] The electronic device of this invention implements the aforementioned adaptive radar target type recognition method by executing program instructions stored in its memory. Compared with general-purpose computing devices, this electronic device embeds the physical constraint model, GLCT transformation, feature extraction, and incremental learning algorithms into the processor execution environment, enabling high-precision target recognition with lower computational overhead. It is particularly suitable for embedded scenarios with limited power consumption and computing power. The device supports online updates and category expansion of the feature library during long-term operation, solving the network latency and privacy risks associated with traditional cloud-based solutions. Attached Figure Description

[0039] Figure 1 This is a system architecture diagram of an embodiment of the radar target type adaptive identification method of the present invention.

[0040] Figure 2This is a comparison diagram of the time spectrum of pedestrian targets by the present invention and existing methods. (a) is the time spectrum of STFT, which has low time-frequency resolution and micro-Doppler blur; (b) is the time spectrum of physical constraint GLCT, which has high resolution and clear micro-motion characteristics. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment takes the identification of pedestrians, cars, electric bicycles and other targets in a park setting by a 24GHz FMCW security radar as an example, but the implementation of the present invention is not limited to this.

[0042] An embodiment of the radar target type adaptive identification method of the present invention includes the following system deployment and parameter configuration: Radar: 24GHz FMCW radar, transmission frequency 24.125GHz, wavelength λ=12.4mm, sampling rate 1000Hz.

[0043] Auxiliary sensor: camera, used for scene classification.

[0044] Main controller: ARM Cortex-M7 architecture microcontroller, with hardware FPU and DSP instruction set enabled.

[0045] GLCT parameters: 256 frequency layers, 128 sliding window length, and 50% window overlap.

[0046] LUT configuration: 16-bit quantization, lookup interval 0.001, storage size 32KB.

[0047] CORDIC iteration count: 10.

[0048] Feature dimensions: =59, =16, fusion features 75 dimensions.

[0049] The system architecture of this invention is as follows: Figure 1 As shown: The system includes radar equipment and optional auxiliary sensors. After the radar echo signal is filtered and denoised by the signal preprocessing module, the Fast Physical Constraint (GLCT) module, guided by independently measured physical parameters such as Doppler initial velocity, target distance, and angle, performs parameter optimization and time-frequency transformation through coarse search and local fine-tuning to generate a high-focus time spectrum. The low-dimensional robust feature extraction module extracts fused feature vectors from this spectrum. The dynamic feature library matching module calculates feature distances and combines them with preset thresholds to complete known target recognition, potential new target caching, or unknown target determination. The feature space dynamic expansion module asynchronously performs online incremental learning and updates the local feature library when trigger conditions are met. Auxiliary sensor signals are used for scene classification and adaptive correction of the search step size; when these fail, the system automatically degrades to single-radar mode. The system ultimately outputs the recognition result.

[0050] The feature library construction process during the training phase: During the closed-world training phase, samples of known targets (pedestrians, cars, electric bicycles, motorcycles, shaking trees, shaking billboards, etc.) were collected, with 500 samples per category, covering distances of 1-200m, angles of 0°-60°, and signal-to-noise ratios of 5-20dB.

[0051] Perform the following operations for each type of objective: 1. Extraction A fusion feature sample.

[0052] 2. Computational Center: .

[0053] 3. Calculate the maximum intra-class distance: .

[0054] 4. Calculate the minimum distance between classes: .

[0055] The threshold is set as follows: threshold : ,in This is the distance fluctuation margin caused by noise.

[0056] threshold (Unknown target detection): .

[0057] Clustering threshold : .

[0058] During the online operation phase, the process of identification in the open world: Step 1: Multi-source sensing and signal preprocessing Radar echo acquisition: The radar equipment transmits continuous waves into the detection area and receives the echo signals reflected by the target.

[0059] Auxiliary sensor synchronous acquisition (optional channel): Synchronously reads camera signals for scene classification (strong clutter / weak clutter determination) and adaptive correction of the search step size in the GLCT fine-tuning stage.

[0060] Signal preprocessing: Bandpass filtering, noise reduction, and amplification are performed on the radar echo signal; simple filtering is performed on the auxiliary sensor signal to extract scene category labels.

[0061] Sensor failure detection and switching: Real-time monitoring of the validity of auxiliary sensor data; if a failure is detected, the system automatically switches to single radar mode: the search step size is fixed at the default value during the fine-tuning stage, the maximum number of search points is expanded to 20, and the frame rate is reduced to 8fps.

[0062] Step 2: Fast GLCT Time-Frequency Feature Extraction Based on Physical Constraints (1) Physical constraint model: The GLCT transform kernel is defined as the linear frequency modulation basis function of Gaussian window modulation: in, The width of the Gaussian window. Centered on time The initial frequency, For the optimal tuning frequency parameters, It is the imaginary unit.

[0063] Establish Correlation model with radar-measurable physical quantities: in: The initial Doppler velocity is independently estimated from the coarse velocity measurement stage of the radar echo signal. The FFT spectral peak detection method is used to extract the Doppler frequency shift from the original echo signal, which is then converted into a velocity estimate. This process relies solely on the frequency difference between the radar transmitted signal and the echo signal, requiring no prior information about the target type, and is therefore a type-independent independent measurement.

[0064] Target distance, obtained independently by the radar ranging module.

[0065] The angle between the radar beam and the target's direction of motion is determined by the radar antenna beam pointing angle.

[0066] Radar wavelength (system calibration constant).

[0067] Equipment calibration coefficient (factory calibration).

[0068] The frequency modulation correction term caused by micro-motion characterizes the micro-motion characteristics of the target.

[0069] Notice: , , It is an independent measurement, not dependent on the target type identification result; The features to be identified reflect the subtle differences in the target's movements; physical constraints only lock the target. It refers to the basic range, not the exact value.

[0070] (2) Parameter optimization strategy (two stages): Phase 1 (Coarse search of physical constraints, table lookup phase): Based on independent measurements , , Calculate the fundamental frequency: Considering measurement error , Determine the scope of the initial search: , in .

[0071] The coarse search range is used to cover the frequency modulation changes caused by the overall translation of the target, while the micro-motion features of the target will be adaptively captured through the subsequent local fine-tuning stage (stage two). This scheme decouples translation and micro-motion processing, which not only utilizes physical priors to reduce the search space, but also retains the ability to finely characterize the details of micro-motion.

[0072] Within the coarse search range, according to the preset step size M candidate parameter points (M≤5) are generated. The pre-calculated basis functions are directly obtained using a lookup table. The temporal-spectral energy concentration (second-order Rayleigh entropy) of each candidate point is calculated, and a point is selected that... The smallest parameter point is used as the coarse search result. : Phase Two (Local Fine-tuning): Within the range Within, dynamic step size is used. Sub-iteration fine-tuning Determine the final optimal tuning frequency .

[0073] Phase 3 (Micro-motion feature consistency verification, post-processing): In obtaining Then, the micro-motion feature measure of the time spectrum is extracted: Periodicity measurement: ,in For the energy center of gravity time series, This is the autocorrelation function.

[0074] Stationarity measure: ,Right now Variance of sequence differences.

[0075] Will , As part of the output features, it participates in subsequent target matching.

[0076] This approach does not rely on high-precision velocity measurements, but instead uses a rough range of velocities to narrow down the search space. The error has been Coverage, subsequent local fine-tuning mechanisms can Nearby error correction ensures robustness in parameter optimization.

[0077] Using optimal frequency modulation parameters The preprocessed radar echo signal is subjected to a generalized linear Chirplet transform to obtain the time-frequency spectrum. For example... Figure 2 As shown in (b), the GLCT time spectrum of the present invention has high time-frequency resolution and clear micro-motion characteristics; in contrast, Figure 2 (a) is the STFT time spectrum, which has low time-frequency resolution and micro-Doppler blurring.

[0078] (3) Complexity analysis: Let the signal length be Traditional GLCT brute-force search requires traversing... One frequency modulation parameter point (usually) The total computational load is This scheme requires only M+P searches for the physical constraint GLCT (M≤5, P≤5), with a total computational cost of [missing information]. .because The computational complexity for parameter optimization is reduced by about an order of magnitude compared to M+P.

[0079] Storage overhead: LUT pre-storage differs (σ, , The Gaussian window exponent term under the combination has a storage requirement that depends on the quantization precision (e.g., 8-bit interval 0.01, 16-bit interval 0.001, LUT size is 16KB and 32KB respectively), and is adapted to the Flash resources of Cortex M7.

[0080] (4) Fixed-point optimization: LUT optimization: Pre-compute Gaussian window exponent term Stored as a fixed-point lookup table.

[0081] Lookup table error .

[0082] For 8-bit quantization = 0.01, ≈ 0.0018; For 16-bit quantization = 0.001, ≈ 0.00018.

[0083] CORDIC optimization: linear frequency modulation term The phase error after n iterations is achieved using the CORDIC algorithm. When n=10, It satisfies the requirement of orthogonality of basis functions.

[0084] Orthogonality verification: Define the orthogonality error of basis functions. Experimental verification: 8-bit fixed-point <0.05, 16-bit fixed-point <0.01, no visible artifacts in the time spectrum.

[0085] Step 3: Robust Feature Extraction (1) Uniform LBP texture features dimension): Uniform Local Binary Patterns (UBPs) are extracted from the GLCT grayscale time-frequency image. Physical significance: The local texture structure of the time-frequency image reflects the time-frequency energy distribution characteristics of the target's micro-motion patterns. For example, pedestrian gait exhibits periodic diagonal textures on the time-frequency image (corresponding to micro-Doppler modulation caused by limb swing); vehicle engines exhibit stable horizontal texture bands (corresponding to rigid body translational dominance with small micro-motion amplitudes). Uniform LBP reduces the dimensionality of the original 256-dimensional LBP to [the desired value]. =59 dimensions, reducing storage and computing overhead.

[0086] (2) Geometric features of ECT dimension): The Energy Centroid Trajectory (ECT) is defined as the energy-weighted average frequency sequence of the time spectrum over the time dimension. right Dimensionality reduction of the sequence (e.g., piecewise averaging, PCA) to =16 dimensions, forming ECT trajectory features. Physical meaning: The ECT trajectory reflects the instantaneous frequency change trend of the target's overall motion, has an averaging effect on noise, and has strong anti-interference ability.

[0087] (3) Fusion features and feature distance: Fusion feature vectors: .

[0088] Feature distance (used for matching): ,in The feature dimension weights are determined by the intra-class / inter-class scatter matrices of the target known during the training phase.

[0089] Step 4: Dynamic Feature Library Matching and Online Incremental Learning (1) Feature matching and decision-making: Calculate the fusion features of test samples With various centers Feature distance ,make .

[0090] like If so, it is determined to be a known target, and the target type is... and output the confidence level. .

[0091] like If the target is identified as a potential new target, its features will be stored in a temporary cache.

[0092] like And echo energy Greater than or equal to the energy threshold And continuous If a frame appears, it is determined to be an unknown target, triggering online incremental learning.

[0093] Otherwise, it is judged as noise or clutter and discarded.

[0094] (2) Online incremental learning: Triggering conditions (must be met simultaneously): System decision is "unknown target"; echo energy ≥ (Noise excluded); the target appears continuously. Frame (excluding transient interference); Number of features of this type of unknown target in the temporary buffer ≥ (Ensure the reliability of clustering).

[0095] Asynchronous execution mechanism: The main radar frame processing stream maintains the original frame rate to complete identification without blocking; the incremental learning stream is executed during the idle time between frames, with a lower priority than real-time identification.

[0096] Clustering and Update Algorithms (Based on Distance Threshold Determination + K-means Validation): Input: A set of unknown target features in a temporary buffer { , , ..., } Output: New target class center or known class update Calculate the distance between cache features and all known class centers. .

[0097] If it exists < If more than 50% of the features are valid: determine it as a known target variant sample, and update the corresponding class center: .

[0098] Otherwise: Perform K-means clustering on the cache features (K=2, to verify distinguishability). If the intra-cluster distance is less than... And the distance between classes is greater than Then a new target class center is generated. = mean(Class 1), feature library expansion, marked as "new target Class X (to be confirmed)"; otherwise continue caching and do not update for now.

[0099] (3) Feature library capacity management and aging mechanism: Maximum capacity: The feature library can store a maximum of [number] files. Target categories (typical) =20), and the aging strategy is triggered when this is exceeded.

[0100] Aging strategy: Sort by "most recent access time" and eliminate those that have been accessed for more than [time period]. Category where time did not appear (typical) =30 days), retain the active category.

[0101] Anti-drift mechanism: Incrementally update the known class centers periodically (e.g., weekly). in To prevent drift in updating coefficients and avoid misjudging old targets due to feature space drift.

[0102] (4) Privacy protection mechanism: Local storage: Only persistent storage of fused feature vectors and class center The original radar echo signal is only temporarily stored in the on-chip buffer of the real-time signal processing pipeline. It does not participate in non-volatile storage, nor is it transmitted externally. The temporary buffer is released immediately after processing, so it is impossible to deduce the original echo information from the residual data.

[0103] Optional cloud synchronization: Only upload feature library update logs (class center change Δ) ), which does not contain any original data or feature vectors.

[0104] Comparison and verification process: Comparative experimental platform: ARM Cortex-M7, 480MHz, 2MB Flash, 1MB RAM.

[0105] The dataset is a self-built dataset and includes: Training set: 6 known target classes (pedestrians, cars, electric bicycles, motorcycles, shaking trees, shaking billboards), 500 samples per class, covering distances of 1-200m, angles of 0°-60°, and signal-to-noise ratios of 5-20dB.

[0106] Test set: 200 samples of each known target class + 100 samples of each of two unknown targets (scooter and pet dog).

[0107] Comparison method: STFT+ Peak Characteristics: Traditional method, as baseline.

[0108] STFT+LBP+ECT: Verifying the necessity of GLCT.

[0109] This solution (GLCT+LBP+ECT): a complete solution.

[0110] Lightweight convolutional neural network (MobileNetV2 architecture, width coefficient 0.35): 1.4MB of parameters, used as a comparison for deep learning.

[0111] Experimental results: Key conclusions: This solution achieves accuracy close to that of a lightweight CNN (with a difference of less than 3%), but offers a 4x increase in frame rate, a 60% reduction in power consumption, and a 95% reduction in storage.

[0112] Compared to the STFT scheme, GLCT's optimization of physical constraint parameters significantly improves the recognition rate (+6.6%).

[0113] The detection rate of unknown targets was 78.4%, and the false detection rate (misclassifying known targets as unknown) was less than 5%.

[0114] Fixed-point accuracy verification: Performance testing in security monitoring scenarios: Single frame processing time: 83ms (12fps).

[0115] Power consumption: 110mW (operation mode), 15mW (low power monitoring mode).

[0116] Recognition latency: <100ms, meeting the real-time requirements of security.

[0117] Embodiments of the radar target type adaptive identification device of the present invention: The radar target type adaptive identification device in this embodiment includes: Signal preprocessing module: Acquires radar echo signals and preprocesses them.

[0118] The physics constraint calculation module is used to independently measure the physical parameters of the target, including the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion And establish the frequency modulation parameters of the generalized linear Chirplet transform based on the physical parameters. The physical constraint model is used to determine the frequency modulation parameters. The coarse search range.

[0119] The GLCT transform module is used to search for the optimal frequency modulation parameter within a coarse search range, using the energy concentration criterion as the optimization objective. And using the optimal frequency modulation parameters The time spectrum is obtained by performing a generalized linear Chirplet transform on the preprocessed radar echo signal.

[0120] Feature extraction module: Extracts fused feature vectors from the time-frequency spectrum.

[0121] Dynamic matching module: Calculates the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and outputs the target recognition result based on the relationship between the feature distance and the preset threshold.

[0122] Feature library management module: Stores and updates the local feature library.

[0123] Electronic device embodiment of the present invention: The electronic device of this embodiment includes a memory and a processor. The processor is used to execute a computer program stored in the memory to implement the above-described adaptive radar target type identification method.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive radar target type identification method, characterized in that, Includes the following steps: Step S1: Acquire radar echo signals and preprocess the radar echo signals; Step S2: Independently measure the physical parameters of the target, including the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion ; Step S3: Establish the frequency modulation parameters of the generalized linear Chirplet transform based on the physical parameters. The physical constraint model is used to determine the frequency modulation parameters. The broad search range; Step S4: Within the coarse search range, using the energy concentration criterion as the optimization objective, search for and obtain the optimal frequency modulation parameter. ; Step S5: Utilize the optimal tuning frequency parameters The time spectrum is obtained by performing a generalized linear Chirplet transform on the preprocessed radar echo signal. Step S6: Extract the fused feature vector from the time spectrum; Step S7: Calculate the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and output the target recognition result based on the relationship between the feature distance and the preset threshold.

2. The adaptive radar target type identification method according to claim 1, characterized in that, In step S5, the generalized linear Chirplet transform uses a Gaussian window-modulated linear frequency modulation basis function, which is defined as: in, The width of the Gaussian window. Centered on time The initial frequency, For optimal frequency modulation parameters, It is the imaginary unit.

3. The adaptive radar target type identification method according to claim 1, characterized in that, In step S3, the specific method for establishing the physical constraint model is as follows: calculate the fundamental frequency modulation frequency using the following formula. : Then, calculate the comprehensive error coefficient based on the speed measurement error and the distance measurement error. Thus, the coarse search range is determined as follows: in, For radar wavelength, To ensure a long observation window, For equipment calibration coefficients, For speed measurement error, This represents the distance measurement error.

4. The adaptive radar target type identification method according to claim 1 or 2, characterized in that, In step S4, the optimization objective based on the energy concentration criterion is specifically as follows: Calculate candidate frequency modulation parameters The second Rayleigh entropy of the spectrum corresponding to the generalized linear Chirplet transform. : Choose to The minimum frequency modulation parameter is taken as the optimal parameter. ,in Indicates the frequency modulation parameter is The generalized linear Chirplet transform time spectrum.

5. The adaptive radar target type identification method according to claim 1, characterized in that, The search process in step S4 is divided into two stages: Phase 1: Within the coarse search range, perform the search using a preset step size. M candidate parameter points are generated, where M≤5. The pre-computed basis functions are obtained by looking up a table. The time-spectral energy concentration of each candidate point is calculated, and the parameter point with the highest concentration is selected as the coarse search result. ; Phase Two: Within the Interval Within, dynamic step size is used. Sub-iteration fine-tuning, where Determine the final optimal tuning frequency parameters. .

6. The adaptive radar target type identification method according to claim 1, characterized in that, In step S6, extracting the fusion feature vector includes: Texture feature extraction: Uniform local binary pattern features are extracted from the gray-scale time-frequency spectrum obtained by the generalized linear Chirplet transform, resulting in... 3D texture feature vector ,in ; Geometric feature extraction: Calculate the energy centroid trajectory of the time-frequency spectrum to obtain the instantaneous frequency sequence. And the sequence is reduced in dimensionality to obtain 3D geometric eigenvectors ,in ; Feature fusion: concatenating texture features and geometric features into a fused feature vector. .

7. The adaptive radar target type identification method according to claim 6, characterized in that, The instantaneous frequency sequence of the energy centroid trajectory The calculation method is as follows: in, For the first Each time unit For the first One frequency unit, For the time spectrum in time unit and frequency unit The amplitude value at that point.

8. The adaptive radar target type identification method according to claim 6, characterized in that, In step S6, after extracting the fused feature vector, a micro-motion feature consistency verification step is also included: Periodicity measure of the extracted spectrum and stability measures ,in: For the energy center of gravity time series, It is the autocorrelation function; That is, the variance of the energy centroid sequence difference, where This represents variance calculation. This represents the first-order difference; Will , As part of the output features, it participates in subsequent target matching.

9. The adaptive radar target type identification method according to claim 1, characterized in that, In step S7 The pre-stored feature library is constructed as follows: training samples of known target categories are collected, multiple fused feature samples are extracted for each target category, and the feature centers of that target category are calculated. Maximum distance within a class minimum distance between classes ;in, Based on the above statistics, set the identification threshold: Threshold threshold ,in The number of samples for each target class. For the first Class 1 The fused feature vector of each sample, Represents the feature distance. This is the distance fluctuation margin caused by noise.

10. The adaptive radar target type identification method according to claim 9, characterized in that, The feature distance Weighted Euclidean distance: in, and These are the first and second feature vectors of the test sample and the reference feature vector, respectively. One portion, For the first The weights of each component are determined by the intra-class scatter matrix and inter-class scatter matrix of the target known during the training phase.

11. The adaptive radar target type identification method according to claim 10, characterized in that, In step S7, the output of the target recognition result based on the relationship between the feature distance and the preset threshold specifically includes: Calculate the fusion features of test samples With various centers Feature distance ,make ; like If so, it is determined to be a known target, and the target type is... and output the confidence level. ; like If it is identified as a potential new target, its features are stored in a temporary cache. like And echo energy Greater than or equal to the energy threshold And continuous If a frame appears, it is determined to be an unknown target, triggering online incremental learning; Otherwise, it is judged as clutter or noise and discarded; in, Energy threshold The threshold for the number of consecutive frames.

12. The radar target type adaptive identification method according to claim 11, characterized in that, The energy threshold The consecutive frame number threshold is determined by environmental noise floor statistics. It is a preset positive integer.

13. The radar target type adaptive identification method according to claim 11, characterized in that, The online incremental learning specifically includes: When the number of the same type of unknown target features in the temporary cache reaches the clustering threshold During the idle time between frames, perform the following operations asynchronously: Calculate the distance between cached features and all known class centers. If more than 50% of the features are less than the clustering threshold for a given known class center, then... If the sample is identified as a known target mutation, then the updated class centers of the target mutation sample are: ； Otherwise, cluster the cache features; if the intra-cluster distance is less than 1 / 3, then... And the distance between classes is greater than If so, a new target class center is generated, added to the feature library, and marked as a new target to be confirmed; in, The minimum number of features required for clustering. To update the coefficients, This indicates the operation of calculating the average.

14. The radar target type adaptive identification method according to claim 13, characterized in that, It also includes feature library capacity management and anti-drift mechanisms: The feature library stores the most There are 1 target categories, among which ; Sort by most recent access time, eliminating those older than [time]. Categories where time did not appear, among which ; Periodically perform incremental updates on the known class centers. The updated class centers after the anti-drift update are as follows: in To prevent drift, update the coefficients.

15. The radar target type adaptive identification method according to claim 13, characterized in that, It also includes privacy protection mechanisms: Local persistent storage of fused feature vectors and class center The original radar echo signal is only temporarily stored in the on-chip buffer of the real-time signal processing pipeline. It does not participate in non-volatile storage, nor is it transmitted to the outside. The temporary buffer is released immediately after processing, so it is impossible to deduce the original echo information from the residual data. When synchronizing to the cloud, only the feature library update log is uploaded, and the update log includes changes in class centers. .

16. The radar target type adaptive identification method according to claim 1, characterized in that, It also includes a fixed-point optimization step: The Gaussian window exponent term in the generalized linear Chirplet transform is pre-calculated using a lookup table method. Stored as an 8-bit or 16-bit fixed-point number; The CORDIC algorithm is used to calculate the linear frequency modulation term. The complex exponent in the number of iterations ; A fixed-point fast Fourier transform is used to achieve fast computation of the generalized linear Chirplet transform.

17. The adaptive radar target type identification method according to claim 1, characterized in that, It also includes the processing of auxiliary sensor signals: Simultaneously acquire auxiliary sensor signals for scene classification and adaptive correction of the search step size during the fine-tuning stage of the generalized linear Chirplet transform; When an auxiliary sensor failure is detected, the system automatically switches to single radar mode: the fine-tuning stage search step size is fixed to the default value, the number of coarse search points is expanded to a maximum of 20, and the processing frame rate is reduced.

18. The radar target type adaptive identification method according to claim 17, characterized in that, In the single radar mode, the processing frame rate is reduced from 12fps to 8fps.

19. The adaptive radar target type identification method according to claim 1, characterized in that, The preprocessing includes bandpass filtering, noise reduction, and amplification; the independent measurement of the Doppler initial velocity... The Doppler frequency shift was extracted from the original echo signal using the FFT spectral peak detection method and then converted.

20. A radar target type adaptive identification device, characterized in that, include: A signal preprocessing module is used to acquire radar echo signals and preprocess the radar echo signals; The physical constraint calculation module is used to independently measure the physical parameters of the target, including the Doppler initial velocity. Target distance and the angle between the radar beam and the target's direction of motion And based on the physical parameters, establish the frequency modulation parameters of the generalized linear Chirplet transform. The physical constraint model is used to determine the frequency modulation parameters. The broad search range; The GLCT transform module is used to search for the optimal frequency modulation parameter within the coarse search range, using the energy concentration criterion as the optimization objective. And using the optimal tuning frequency parameter The time spectrum is obtained by performing a generalized linear Chirplet transform on the preprocessed radar echo signal. The feature extraction module is used to extract the fused feature vector from the time spectrum; The dynamic matching module is used to calculate the feature distance between the fused feature vector and the center of each target class in the pre-stored feature library, and output the target recognition result according to the relationship between the feature distance and the preset threshold. The feature library management module is used to store and update the local feature library.

21. An electronic device comprising a memory and a processor, characterized in that, The processor is used to execute the computer program stored in the memory to implement the radar target type adaptive identification method as described in any one of claims 1 to 19.

Images