Unmanned aerial vehicle trajectory tracking target classification method and system based on multi-band radar

Abstract

The invention relates to the technical field of trajectory tracking, in particular to an unmanned aerial vehicle trajectory tracking target classification method and system based on a multi-frequency-band radar, and the method comprises the following steps: synchronously collecting original radar echo data of a millimeter wave frequency band, a Ku frequency band and an L frequency band, and generating a multi-frequency-band data set comprising a timestamp, an azimuth angle and a pulse sequence; extracting multi-dimensional features of the target from the synchronized multi-band data set, and generating a multi-band feature vector; constructing a dynamic trajectory association matrix, and determining the space-time continuity of a target trajectory through a trajectory association algorithm based on an attention mechanism; constructing a lightweight classification model; performing online classification on the real-time multi-band feature vectors by using a lightweight classification model, and outputting a target type and confidence; according to the method, capture of the motion state of the target is enhanced, rapid response can be made to tiny motion and change of the target in a complex scene, and therefore more comprehensive and accurate time sequence context information is provided for subsequent classification tasks.

Description

Technical Field

[0001] The present invention relates to the field of trajectory tracking technology, and in particular to a method and system for classifying unmanned aerial vehicle trajectory tracking targets based on a multi-band radar. Background Art

[0002] In recent years, with the rapid development and widespread application of drone technology, traditional drone radar systems mostly rely on a single frequency band for target detection. They have certain limitations in resolution, anti-interference capability and environmental adaptability, and it is difficult to meet the requirements of accurate detection and tracking of high-speed, small, and low-observable targets (such as drones) in complex electromagnetic environments.

[0003] At present, some studies have attempted to use multi-band radar systems to collect echo data in different frequency bands (such as millimeter wave, Ku band, and L band) in order to improve target detection performance through multi-source information fusion. However, due to the problems of clock deviation and inconsistent pulse trigger timing in the signal acquisition of each frequency band, it is difficult to achieve high-precision synchronization of multi-band data, resulting in information distortion or mismatch in the subsequent feature extraction and target recognition process. In addition, the target features captured by different frequency bands are different in the physical sense. For example, the millimeter wave band is suitable for extracting subtle motion features, the Ku band reflects the target scattering cross-section characteristics, and the L band is more suitable for Doppler velocity measurement. How to effectively fuse these multi-dimensional features and realize cross-band collaborative recognition has always been one of the technical difficulties.

[0004] In addition, existing target tracking and classification methods mostly focus on the processing of single-frame or short-time series data, and lack the full utilization of the target's motion information in continuous spatiotemporal sequences. Traditional data association algorithms often rely on fixed models or simple matching rules and cannot adaptively respond to the complex spatiotemporal associations and dynamic changes between targets, which in turn affects the continuity judgment of the target trajectory and the subsequent classification accuracy. Summary of the invention

[0005] The present invention provides a method and system for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar.

[0006] The method for classifying UAV trajectory tracking targets based on multi-band radar includes the following steps:

[0007] S1: synchronously collect original radar echo data of millimeter wave band, Ku band and L band through a multi-band radar system, generate a multi-band data set including a timestamp, an azimuth angle and a pulse sequence, and perform time synchronization calibration on the multi-band data set;

[0008] S2: Extract the multi-dimensional features of the target from the synchronized multi-band data set, including the high-resolution micro-motion features of the millimeter-wave band, the scattering cross-section features of the Ku band, and the Doppler velocity features of the L band, and generate a multi-band feature vector;

[0009] S3: Based on the multi-band feature vectors, a dynamic trajectory association matrix is ​​constructed, and the spatiotemporal continuity of the target trajectory is determined by the trajectory association algorithm based on the attention mechanism;

[0010] S4: Based on the dynamic trajectory association matrix and historical classification data, a lightweight classification model is constructed, wherein the model includes a long short-term memory network (LSTM) module and an adaptive frequency band weight allocation module;

[0011] S5: Use the lightweight classification model to perform online classification on the real-time multi-band feature vector, and output the target type and confidence.

[0012] Optionally, the S1 specifically includes:

[0013] S11: configure the hardware architecture of the multi-band radar system, including a millimeter-wave radar transmitter, a Ku-band radar transmitter, an L-band radar transmitter and a common-aperture receiving antenna array, wherein the millimeter-wave radar transmitter, the Ku-band radar transmitter and the L-band radar transmitter adopt a master-slave clock synchronization mechanism, use the millimeter-wave radar clock as a reference clock source, and synchronize the pulse trigger timing of the Ku-band and L-band radars through a phase-locked loop;

[0014] S12: The pulse repetition frequency (PRF) is set independently for each frequency band, where the PRF for the millimeter wave band is 100kHz1MHz, the PRF for the Ku band is 1050kHz, and the PRF for the L band is 110kHz;

[0015] S13: At the receiving end, the original echo data is segmented in the time domain, the signals of each frequency band are separated according to the transmission timing, and the target azimuth is calculated based on the radar beamforming algorithm to generate a triplet of data with a timestamp, azimuth and corresponding frequency band pulse sequence;

[0016] S14: After aligning the triplet data of each frequency band according to the timestamp, the triplet data are merged into a unified multi-band data set, wherein each entry in the multi-band data set includes a synchronized timestamp, an azimuth angle, and a three-channel pulse sequence of a millimeter wave band, a Ku band, and an L band.

[0017] Optionally, the radar beamforming algorithm adopts a delay sum method, using a common aperture receiving antenna array, and uses the delay sum method to calculate the azimuth of the target for each separated frequency band signal;

[0018] For a certain frequency band, the common aperture receiving antenna array contains N antenna units, and the signal received by each antenna unit is xn (t), where n = 1, 2, ..., N, for the candidate target azimuth θ, by calculating the corresponding delay compensation value τ of each antenna element n (θ), the signals of each unit are weighted and summed;

[0019] The delay compensation value τ n The calculation of (θ) is determined based on the geometric structure of the antenna array. Let the projection distance of the relative position of the nth antenna element relative to the array reference element along the array direction be d n , the signal propagates at the speed of light c, then when the target is located at the candidate azimuth angle θ, the delay time that the nth antenna unit needs to compensate is:

[0020] Among them, d n is the distance projection of the nth antenna unit relative to the reference unit, θ is the azimuth of the candidate target, and c is the speed of light;

[0021] The candidate angle θ is scanned within the preset search range, and the power of the beamforming output signal at each angle is calculated to determine the actual azimuth of the target: Determine the estimated azimuth of the target by finding the angle where the output power is maximum: Where P(θ) is the signal power at the candidate angle θ, θ target is the estimated target azimuth.

[0022] Optionally, the S2 specifically includes:

[0023] S21: Signal processing technology is used for the synchronous data in the millimeter wave frequency band to perform time-frequency domain analysis on the target echo signal and extract the micro-motion feature μ of the target. mm , the micro-motion feature reflects the micro-Doppler effect caused by the slight movement of the target (such as rotation or vibration);

[0024] S22: Use amplitude and phase information to estimate the scattering cross section of the Ku-band synchronization data and calculate the scattering cross section characteristic σ of the target Ku , characterizes the target's ability to reflect radar waves;

[0025] S23: Pulse Doppler processing technology is used for the synchronous data of the L band. By performing spectrum analysis on the echo signal, the Doppler frequency shift in the target echo is calculated and the Doppler velocity characteristics v of the target are extracted. L ;

[0026] S24: The extracted millimeter wave micro-motion features, Ku-band scattering cross-section features, and L-band Doppler velocity features are fused at the feature level to generate a multi-band feature vector F including multi-dimensional information of the target, where F =

[0027] [μ mm ,σ Ku ,v L ] for subsequent trajectory tracking and target classification.

[0028] Optionally, the S3 specifically includes:

[0029] S31, multi-band feature vector definition: In each time frame t, for each detected target, a multi-band feature vector set is extracted Among them, each represents the feature vector of the i-th target in time frame t, d is the dimension of the feature vector, N t is the number of targets in time frame t;

[0030] S32, Query, key and value vector calculation: using the preset weight matrix Among them, W q is the query weight matrix, W k is the key weight matrix, W v is the value weight matrix, for each eigenvector Calculate the query vector, key vector, and value vector separately:

[0031]

[0032] S33, calculation of attention weight score: Consider adjacent time frames t and t+1, for the i-th target in time frame t and the j-th target in time frame t+1, calculate the attention weight score

[0033] S34, Construction of dynamic trajectory association matrix: According to the attention weight score, construct the dynamic trajectory association matrix Its matrix elements are defined as:

[0034]

[0035] S35, aggregation of trajectory features: using matrix M t,t+1 The trajectory of the targets in adjacent time frames is associated, and the aggregate feature vector of the i-th target in time frame t in association with time frame t+1 is calculated.

[0036] S36, calculation of spatiotemporal continuity score: To determine the spatiotemporal continuity of the target trajectory, for each target i, its original feature vector is compared With the aggregate feature vector The similarity of , using cosine similarity as the spatiotemporal continuity score, the calculation formula is: Among them, ‖·‖ represents the Euclidean norm, Score the spatiotemporal continuity of target i between adjacent frames. The closer the value is to 1, the better the continuity.

[0037] Optionally, the calculation of the attention weight score The dot product attention mechanism is adopted, which is expressed as:

[0038] in, is the dot product of the query vector of the i-th target in time frame t and the key vector of the j-th target in time frame t+1, is the normalization factor, N t+1 is the number of targets in time frame t+1, represents the association weight between target i and target j, reflecting the similarity between the two in the feature space. q represents the index variable, which represents the sequence number of all candidate targets in time frame t+1.

[0039] Optionally, S3 further includes determining the spatiotemporal continuity: according to a preset spatiotemporal continuity threshold, Compare with the spatiotemporal continuity threshold to determine whether the trajectory of target i in time frame t and t+1 has spatiotemporal continuity, and update the trajectory association information of the target accordingly.

[0040] Optionally, the S4 specifically includes:

[0041] S41 training data set construction: using the obtained multi-band feature vector set And the dynamic trajectory correlation matrix M constructed between adjacent time frames t,t+1 , combined with historical classification data in, Represents the historical classification results of target i in time frame t, and constructs a training data set;

[0042] S42, lightweight classification model composition: The constructed lightweight classification model includes:

[0043] Long short-term memory network module: used to capture the temporal changes of the target multi-band feature vector set, and its input is the serialized multi-band feature vector (It can be combined with the target association information reflected in the dynamic trajectory association matrix), the long short-term memory network module outputs the predicted target category and the corresponding confidence;

[0044] Adaptive frequency band weight allocation module: used to dynamically adjust the relative weights of the millimeter wave, Ku band and L band features during fusion to form a weighted fusion feature vector F weighted .

[0045] S43, model training and parameter update: Using the constructed training data set, the lightweight classification model is trained through an end-to-end training method, so that the long short-term memory network module and the adaptive frequency band weight allocation module can output accurate target categories and confidence levels based on the input historical features and trajectory association information. At the same time, the model parameters are adaptively updated based on the classification error using historical classification data feedback.

[0046] Optionally, the lightweight classification model inputs real-time multi-band feature vectors, and fuses the feature vectors of each band into a single feature vector F through an adaptive band weight allocation module. weighted , the fused feature vector F weighted As the input of the long short-term memory network module, the hidden state h is calculated by the long short-term memory network module. t Finally, the target classification is achieved through a fully connected layer and a Softmax layer, and its calculation is expressed as: t =softmax(W y h t +b y ), where W y is the output layer weight matrix, b y is the bias vector, and the softmax(·) function converts the output into a probability distribution of each category, y t is the classification result vector at time t, each component of which represents the confidence of the corresponding category.

[0047] The UAV trajectory tracking target classification system based on multi-band radar is used to implement the above-mentioned UAV trajectory tracking target classification method based on multi-band radar, and includes the following modules:

[0048] A multi-band radar data acquisition module, used to synchronously acquire raw radar echo data of the millimeter wave band, Ku band and L band, generate a multi-band data set including a timestamp, an azimuth angle and a pulse sequence, and perform time synchronization calibration on the data set;

[0049] A feature extraction module is used to extract multi-dimensional features of the target from the synchronized multi-band data set, wherein the multi-dimensional features include high-resolution micro-motion features of the millimeter wave band, scattering cross-section features of the Ku band, and Doppler velocity features of the L band, and generate a multi-band feature vector;

[0050] A dynamic trajectory association module, used to construct a dynamic trajectory association matrix according to the multi-band feature vector, and determine the spatiotemporal continuity of the target trajectory through a trajectory association algorithm based on an attention mechanism;

[0051] A lightweight classification module, which constructs a lightweight classification model including a long short-term memory network module and an adaptive frequency band weight allocation module based on the dynamic trajectory association matrix and historical classification data;

[0052] The online classification module is used to perform online classification on the real-time multi-band feature vector using the lightweight classification model, and output the target type and the corresponding confidence.

[0053] Beneficial effects of the present invention:

[0054] The present invention uses a trajectory association algorithm based on an attention mechanism to construct a dynamic trajectory association matrix, which can automatically calculate the similarity between targets according to multi-band feature vectors between different time frames, and thereby achieve efficient and accurate target matching. The mechanism can adaptively adjust the association weights between targets, effectively reducing the risk of misassociation caused by noise, occlusion or interference, thereby ensuring that the spatiotemporal continuity of the target trajectory is accurately judged, and improving the robustness and reliability of overall target tracking. The various target features extracted from multiple frequency bands are weighted by the attention mechanism, and the association matrix is ​​dynamically constructed, so that the features of different frequency bands and different times can be efficiently fused, which not only enhances the capture of the target motion state, but also can quickly respond to the target's tiny movements and changes in complex scenes, thereby providing more comprehensive and accurate temporal context information for subsequent classification tasks.

[0055] The lightweight classification model constructed based on the dynamic trajectory association matrix and historical classification data can realize real-time target recognition and classification under the premise of low computing resource consumption. The model continuously and adaptively adjusts the classification parameters by integrating the target's temporal characteristics and historical judgment results, thereby improving the classification accuracy and response speed. At the same time, the model design takes into account the real-time performance and accuracy of the system, and is easier to deploy and promote in practical applications. It is suitable for resource-constrained online UAV monitoring and tracking scenarios.

[0056] The lightweight classification model constructed by the present invention integrates the long short-term memory network and the adaptive frequency band weight allocation module, which can fully capture the temporal changes of the target multi-band characteristics and dynamically adjust the weight of each frequency band, thereby realizing efficient and real-time target classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 A schematic diagram of a classification method flow chart of an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of system function modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. At the same time, it is explained here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments, and those skilled in the art may also adopt other alternatives to implement some known technologies; and the accompanying drawings are only for more specific description of the embodiments, and are not intended to specifically limit the present invention.

[0061] It should be noted that the references to "one embodiment", "an embodiment", "an exemplary embodiment", "some embodiments" and the like in the specification indicate that the embodiments described may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, when a specific feature, structure or characteristic is described in conjunction with an embodiment, it should be within the knowledge of a person skilled in the art to implement such feature, structure or characteristic in conjunction with other embodiments (whether or not explicitly described).

[0062] In general, a term can be understood, at least in part, from its use in context. For example, depending, at least in part, on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending, at least in part, on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0063] like Figure 1 As shown, the UAV trajectory tracking target classification method based on multi-band radar includes the following steps:

[0064] S1: The original radar echo data of the millimeter wave band, Ku band and L band are synchronously collected through the multi-band radar system to generate a multi-band data set including timestamp, azimuth and pulse sequence, and the multi-band data set is time-synchronized and calibrated;

[0065] S2: Extract the multi-dimensional features of the target from the synchronized multi-band data set, including the high-resolution micro-motion features of the millimeter-wave band, the scattering cross-section features of the Ku band, and the Doppler velocity features of the L band, and generate a multi-band feature vector;

[0066] S3: Based on the multi-band feature vectors, a dynamic trajectory association matrix is ​​constructed, and the spatiotemporal continuity of the target trajectory is determined by the trajectory association algorithm based on the attention mechanism;

[0067] S4: Based on the dynamic trajectory association matrix and historical classification data, a lightweight classification model is constructed. The model includes a long short-term memory network (LSTM) module and an adaptive frequency band weight allocation module.

[0068] S5: Use a lightweight classification model to perform online classification on real-time multi-band feature vectors and output the target type and confidence.

[0069] S1 specifically includes:

[0070] S11: Configure the hardware architecture of the multi-band radar system, including millimeter-wave radar transmitter, Ku-band radar transmitter, L-band radar transmitter and co-aperture receiving antenna array. The millimeter-wave radar transmitter, Ku-band radar transmitter and L-band radar transmitter adopt a master-slave clock synchronization mechanism, using the millimeter-wave radar clock as the reference clock source, and synchronizing the pulse trigger timing of the Ku-band and L-band radars through a phase-locked loop;

[0071] S12: The pulse repetition frequency (PRF) is set independently for each frequency band, where the PRF for the millimeter wave band is 100kHz1MHz, the PRF for the Ku band is 1050kHz, and the PRF for the L band is 110kHz;

[0072] S13: At the receiving end, the original echo data is segmented in the time domain, the signals of each frequency band are separated according to the transmission timing, and the target azimuth is calculated based on the radar beamforming algorithm to generate a triplet of data with a timestamp, azimuth and corresponding frequency band pulse sequence;

[0073] S14: After aligning the triplet data of each frequency band according to the timestamp, the triplet data are merged into a unified multi-band data set, where each entry in the multi-band data set includes a synchronized timestamp, azimuth, and three-channel pulse sequences of the millimeter wave band, Ku band, and L band.

[0074] The radar beamforming algorithm uses the delay sum method, using the common aperture receiving antenna array, and uses the delay sum method to calculate the azimuth of the target for each separated frequency band signal;

[0075] For a certain frequency band, the common aperture receiving antenna array contains N antenna units, and the signal received by each antenna unit is x n (t), where n = 1, 2, ..., N, for the candidate target azimuth θ, by calculating the corresponding delay compensation value τ of each antenna element n (θ), the weighted summation of each unit signal is performed, and the calculation expression is: Where Y(θ) is the superimposed output signal after beamforming at the candidate azimuth angle θ, and w n is the weighting factor assigned to the nth antenna element (optionally a uniform weighting w n =1 / N), xn (t-τ n (θ)) represents the signal after delay compensation is performed on the signal received by the nth antenna unit;

[0076] Delay compensation value τ n The calculation of (θ) is determined based on the geometric structure of the antenna array. Let the projection distance of the relative position of the nth antenna element relative to the array reference element along the array direction be d n , the signal propagates at the speed of light c, then when the target is located at the candidate azimuth angle θ, the delay time that the nth antenna unit needs to compensate is:

[0077] Among them, d n is the distance projection of the nth antenna unit relative to the reference unit, θ is the azimuth of the candidate target, and c is the speed of light;

[0078] The candidate angle θ is scanned within the preset search range, and the power of the beamforming output signal at each angle is calculated to determine the actual azimuth of the target: Determine the estimated azimuth of the target by finding the angle where the output power is maximum: Where P(θ) is the signal power at the candidate angle θ, θ target is the estimated target azimuth.

[0079] After completing the above beamforming process, for the signal in each receiving time slot, we obtain:

[0080] Timestamp: The mark of the raw data collected synchronously when the pulse is emitted;

[0081] Target azimuth θtarget: calculated based on the above delay-sum beamforming algorithm;

[0082] Corresponding frequency band pulse sequence: the echo pulse sequence that belongs to this frequency band after time domain segmentation and signal separation.

[0083] Combining the above three elements generates a set of triplet data with timestamp, azimuth and corresponding frequency band pulse sequence, which is used for subsequent multi-frequency band feature extraction and tracking classification processing.

[0084] S2 specifically includes:

[0085] S21: Signal processing technology is used for the synchronous data in the millimeter wave frequency band to perform time-frequency domain analysis on the target echo signal and extract the micro-motion feature μ of the target. mm ,The micro-motion feature reflects the micro-Doppler effect produced by the subtle movement of the target (e.g., rotation or vibration);

[0086] S22: Use amplitude and phase information to estimate the scattering cross section of the Ku-band synchronization data and calculate the scattering cross section characteristic σ of the target Ku , characterizes the target's ability to reflect radar waves;

[0087] S23: Pulse Doppler processing technology is used for the synchronous data of the L band. By performing spectrum analysis on the echo signal, the Doppler frequency shift in the target echo is calculated and the Doppler velocity characteristics v of the target are extracted. L ;

[0088] S24: The extracted millimeter wave micro-motion features, Ku-band scattering cross-section features, and L-band Doppler velocity features are fused at the feature level to generate a multi-band feature vector F including multi-dimensional information of the target, where F =

[0089] [μ mm ,σ Ku ,v L ] for subsequent trajectory tracking and target classification.

[0090] High-resolution micro-motion feature extraction in the millimeter wave band: Assume that the echo signal received in the millimeter wave band is s mm (t), the time-frequency domain representation is calculated using short-time Fourier transform and is expressed as:

[0091] Where w(·) is the window function, T is the duration of the analysis window, f is the frequency variable, S mm (t,f) is the STFT value at time t corresponding to frequency f;

[0092] The target micro-motion features are extracted using the short-time Fourier transform results, and its spectrum center of gravity (spectrum center) is defined as the micro-motion feature, which is expressed as:

[0093] Among them, f min and f max are the lower and upper bounds of the set frequency, respectively, |S mm (t,f)| 2 represents the power spectral density of the signal at frequency f, μ mm (t) is the high-resolution micro-motion feature in the millimeter wave frequency band.

[0094] Ku-band scattering cross section feature extraction: Assume that the synchronization data of Ku-band is s Ku (t), whose amplitude information is:

[0095] A Ku (t)=|s Ku (t)|. According to the radar cross section estimation principle, the relationship between the echo amplitude and the system parameters is used to estimate the scattering cross section σ Ku , calculated as: Where R is the distance between the target and the radar, P t is the radar transmission power, G t and G r are the radar transmitting and receiving antenna gains, λ Ku is the Ku band wavelength, calculated as: Where c is the speed of light, f Ku is the center frequency of the Ku band, (4π) 3 is the constant factor in the radar equation.

[0096] L-band Doppler velocity feature extraction: Assume that the L-band synchronization data is s L (t), where the discrete signal s is sampled at the pulse repetition frequency PRF L (kT PRF )(k=0,1,…,N-1),T PRF =1 / PRF is the pulse repetition interval, N is the number of pulses used for Doppler processing, and the Doppler spectrum is obtained by fast Fourier transform of the discrete signal: Among them, f D is the Doppler frequency variable, S L (f D ) is the corresponding Doppler spectrum, and the frequency f with the largest amplitude in the Doppler spectrum is selected. D,target The Doppler frequency of the target is: Using the relationship between Doppler frequency and target radial velocity, the target radial velocity v is calculated. L , expressed as: Among them, λ L is the wavelength of the L band, calculated as: where f L is the center frequency of the L-band, and the denominator factor of 2 takes into account the round-trip propagation of radar waves.

[0097] S3 specifically includes:

[0098] S31, multi-band feature vector definition: In each time frame t, for each detected target, a multi-band feature vector set is extracted Among them, each represents the feature vector of the i-th target in time frame t, d is the dimension of the feature vector, N t is the number of targets in time frame t;

[0099] S32, Query, key and value vector calculation: using the preset weight matrix Among them, W q is the query weight matrix, W k is the key weight matrix, W v is the value weight matrix, for each eigenvector Calculate the query vector, key vector, and value vector separately:

[0100]

[0101] S33, calculation of attention weight score: Consider adjacent time frames t and t+1, for the i-th target in time frame t and the j-th target in time frame t+1, calculate the attention weight score

[0102] S34, Construction of dynamic trajectory association matrix: According to the attention weight score, construct the dynamic trajectory association matrix Its matrix elements are defined as:

[0103]

[0104] S35, aggregation of trajectory features: using matrix M t,t+1 The trajectory of the targets in adjacent time frames is associated, and the aggregate feature vector of the i-th target in time frame t in association with time frame t+1 is calculated.

[0105] S36, calculation of spatiotemporal continuity score: To determine the spatiotemporal continuity of the target trajectory, for each target i, its original feature vector is compared and the aggregated feature vector The similarity of , using cosine similarity as the spatiotemporal continuity score, the calculation formula is: Among them, ‖·‖ represents the Euclidean norm, Score the spatiotemporal continuity of target i between adjacent frames. The closer the value is to 1, the better the continuity.

[0106] Calculate the attention weight score The dot product attention mechanism is adopted, which is expressed as:

[0107] in, is the dot product of the query vector of the i-th target in time frame t and the key vector of the j-th target in time frame t+1, is the normalization factor, N t+1 is the number of targets in time frame t+1, represents the association weight between target i and target j, reflecting the similarity between the two in the feature space. q represents the index variable, which represents the sequence number of all candidate targets in time frame t+1.

[0108] S3 also includes a temporal and spatial continuity judgment: according to the preset temporal and spatial continuity threshold, the specific value is 0.8, that is, when the temporal and spatial continuity score is greater than or equal to 0.8, it is considered that the target has good temporal and spatial continuity between adjacent frames, and the target is considered to have good temporal and spatial continuity between adjacent frames. Compare with the spatiotemporal continuity threshold to determine whether the trajectory of target i in time frame t and t+1 has spatiotemporal continuity, and update the trajectory association information of the target accordingly.

[0109] Through the above steps, the dynamic trajectory association matrix and attention mechanism constructed based on the multi-band feature vector F not only realize the effective association of target features between different time frames, but also provide a reliable basis for subsequent target trajectory tracking and classification through the calculated spatiotemporal continuity score.

[0110] S4 specifically includes:

[0111] S41 training data set construction: using the obtained multi-band feature vector set And the dynamic trajectory correlation matrix M constructed between adjacent time frames t,t+1 , combined with historical classification data in, Represents the historical classification results of target i in time frame t, and constructs a training data set;

[0112] S42, lightweight classification model composition: The constructed lightweight classification model includes:

[0113] Long short-term memory network module: used to capture the temporal changes of the target multi-band feature vector set, and its input is the serialized multi-band feature vector (It can be combined with the target association information reflected in the dynamic trajectory association matrix), the long short-term memory network module outputs the predicted target category and the corresponding confidence;

[0114] Adaptive frequency band weight allocation module: used to dynamically adjust the relative weights of the millimeter wave, Ku band and L band features during fusion to form a weighted fusion feature vector F weighted , its calculation expression is:

[0115] F weighted =α mm ·F mm +α Ku ·F Ku +α L ·F L , where F mm ,F Ku ,F L They represent the characteristic components extracted from millimeter wave, Ku band and L band respectively, α mm ,α Ku ,α Lis the weight coefficient of the corresponding frequency band, satisfying the normalization condition α mm +α Ku +α L =1.

[0116] S43, model training and parameter update: Using the constructed training data set, the lightweight classification model is trained through an end-to-end training method, so that the long short-term memory network module and the adaptive frequency band weight allocation module can output accurate target categories and confidence levels based on the input historical features and trajectory association information. At the same time, the model parameters are adaptively updated based on the classification error using historical classification data feedback.

[0117] In the long short-term memory network module, the first input features received are passed to multiple gating units together with the hidden state of the previous time step, and the forget gate is used to determine which information in the current input should be discarded; then, the input gate decides which new information needs to be added and generates candidate memories. Subsequently, the system combines the forgotten old information with the selected new information to update the current memory state. After that, the output gate determines the proportion of information to be output based on the current input and the updated memory state. Finally, after a series of nonlinear processing, a new hidden state is generated. This state not only retains historical information, but also contains the new information of the current input, providing an effective temporal feature representation for subsequent classification tasks.

[0118] The lightweight classification model inputs real-time multi-band feature vectors, and the feature vectors of each band are fused into a single feature vector F through the adaptive band weight allocation module. weighted , the fused feature vector F weighted As the input of the long short-term memory network module, the hidden state h is calculated by the long short-term memory network module. t Finally, the target classification is achieved through a fully connected layer and a Softmax layer, and its calculation is expressed as: t =softmax(W y h t +b y ), where W y is the output layer weight matrix, b y is the bias vector, and the softmax(·) function converts the output into a probability distribution of each category, y t is the classification result vector at time t, each component of which represents the confidence of the corresponding category.

[0119] like Figure 2 As shown, the UAV trajectory tracking target classification system based on multi-band radar is used to implement the above-mentioned UAV trajectory tracking target classification method based on multi-band radar, and includes the following modules:

[0120] The multi-band radar data acquisition module is used to synchronously collect raw radar echo data of the millimeter wave band, Ku band and L band, generate a multi-band data set containing timestamps, azimuths and pulse sequences, and perform time synchronization calibration on the data set;

[0121] A feature extraction module is used to extract multi-dimensional features of the target from the synchronized multi-band data set, the multi-dimensional features including high-resolution micro-motion features in the millimeter-wave band, scattering cross-section features in the Ku band, and Doppler velocity features in the L band, and generate a multi-band feature vector;

[0122] The dynamic trajectory association module is used to construct a dynamic trajectory association matrix based on multi-band feature vectors and determine the spatiotemporal continuity of the target trajectory through a trajectory association algorithm based on the attention mechanism;

[0123] A lightweight classification module, which constructs a lightweight classification model including a long short-term memory network module and an adaptive frequency band weight allocation module based on the dynamic trajectory association matrix and historical classification data;

[0124] The online classification module is used to perform online classification on the real-time multi-band feature vector using the lightweight classification model, and output the target type and the corresponding confidence.

[0125] The present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention. In order to make the public have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, but those skilled in the art can fully understand the present invention without the description of these details. In addition, in order to avoid unnecessary confusion about the essence of the present invention, well-known methods, processes, procedures, components and circuits are not described in detail.

[0126] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for classifying UAV trajectory tracking targets based on multi-band radar, characterized in that: The following steps are involved: S1: synchronously collect original radar echo data of millimeter wave band, Ku band and L band through a multi-band radar system, generate a multi-band data set including a timestamp, an azimuth angle and a pulse sequence, and perform time synchronization calibration on the multi-band data set; S2: Extract the multi-dimensional features of the target from the synchronized multi-band data set, including the high-resolution micro-motion features of the millimeter-wave band, the scattering cross-section features of the Ku band, and the Doppler velocity features of the L band, and generate a multi-band feature vector; S3: Based on the multi-band feature vectors, a dynamic trajectory association matrix is ​​constructed, and the spatiotemporal continuity of the target trajectory is determined by the trajectory association algorithm based on the attention mechanism; S4: Based on the dynamic trajectory association matrix and historical classification data, a lightweight classification model is constructed, wherein the model includes a long short-term memory network module and an adaptive frequency band weight allocation module; S5: Use the lightweight classification model to perform online classification on the real-time multi-band feature vector, and output the target type and confidence.

2. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 1 is characterized in that: The S1 specifically includes: S11: configure the hardware architecture of the multi-band radar system, including a millimeter-wave radar transmitter, a Ku-band radar transmitter, an L-band radar transmitter and a common-aperture receiving antenna array, wherein the millimeter-wave radar transmitter, the Ku-band radar transmitter and the L-band radar transmitter adopt a master-slave clock synchronization mechanism, use the millimeter-wave radar clock as a reference clock source, and synchronize the pulse trigger timing of the Ku-band and L-band radars through a phase-locked loop; S12: The pulse repetition frequency (PRF) is set independently for each frequency band, where the PRF for the millimeter wave band is 100kHz1MHz, the PRF for the Ku band is 1050kHz, and the PRF for the L band is 110kHz; S13: At the receiving end, the original echo data is segmented in the time domain, the signals of each frequency band are separated according to the transmission timing, and the target azimuth is calculated based on the radar beamforming algorithm to generate a triplet of data with a timestamp, azimuth and corresponding frequency band pulse sequence; S14: After aligning the triplet data of each frequency band according to the timestamp, the triplet data are merged into a unified multi-band data set, wherein each entry in the multi-band data set includes a synchronized timestamp, an azimuth angle, and a three-channel pulse sequence of a millimeter wave band, a Ku band, and an L band.

3. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 2 is characterized in that: The radar beamforming algorithm adopts a delay sum method, using a common aperture receiving antenna array, and uses the delay sum method to calculate the azimuth of the target for each separated frequency band signal; For a certain frequency band, the common aperture receiving antenna array contains N antenna units, and the signal received by each antenna unit is x n (t), where n = 1, 2, ..., N, for the candidate target azimuth θ, by calculating the corresponding delay compensation value τ of each antenna element n (θ), the signals of each unit are weighted and summed; The delay compensation value τ n The calculation of (θ) is determined based on the geometric structure of the antenna array. Let the projection distance of the relative position of the nth antenna element relative to the array reference element along the array direction be d n , the signal propagates at the speed of light c, then when the target is located at the candidate azimuth angle θ, the delay time that the nth antenna unit needs to compensate is: Among them, d n is the distance projection of the nth antenna unit relative to the reference unit, θ is the azimuth of the candidate target, and c is the speed of light; The candidate angle θ is scanned within the preset search range, and the power of the beamforming output signal at each angle is calculated to determine the actual azimuth of the target: Determine the estimated azimuth of the target by finding the angle where the output power is maximum: Where P(θ) is the signal power at the candidate angle θ, θ target is the estimated target azimuth.

4. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 1, characterized in that: The S2 specifically includes: S21: Signal processing technology is used for the synchronous data in the millimeter wave frequency band to perform time-frequency domain analysis on the target echo signal and extract the micro-motion feature μ of the target. mm , the micro-motion feature reflects the micro-Doppler effect caused by the slight movement of the target; S22: Use amplitude and phase information to estimate the scattering cross section of the Ku-band synchronization data and calculate the scattering cross section characteristic σ of the target Ku , characterizes the target's ability to reflect radar waves; S23: Pulse Doppler processing technology is used for the synchronous data of the L band. By performing spectrum analysis on the echo signal, the Doppler frequency shift in the target echo is calculated and the Doppler velocity characteristics v of the target are extracted. L ; S24: The extracted millimeter wave micro-motion features, Ku-band scattering cross-section features, and L-band Doppler velocity features are fused at the feature level to generate a multi-band feature vector F including multi-dimensional information of the target, where F = [μ mm ,σ Ku ,v L ].

5. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 1, characterized in that: The S3 specifically includes: S31, multi-band feature vector definition: In each time frame t, for each detected target, a multi-band feature vector set is extracted Among them, each represents the feature vector of the i-th target in time frame t, d is the dimension of the feature vector, N t is the number of targets in time frame t; S32, Query, key and value vector calculation: using the preset weight matrix Among them, W q is the query weight matrix, W k is the key weight matrix, W v is the value weight matrix, for each eigenvector Calculate the query vector, key vector, and value vector separately: S33, calculation of attention weight score: Consider adjacent time frames t and t+1, for the i-th target in time frame t and the j-th target in time frame t+1, calculate the attention weight score S34, Construction of dynamic trajectory association matrix: According to the attention weight score, construct the dynamic trajectory association matrix Its matrix elements are defined as: S35, aggregation of trajectory features: using matrix M t,t+1 The trajectory of the targets in adjacent time frames is associated, and the aggregate feature vector of the i-th target in time frame t in association with time frame t+1 is calculated. S36, calculation of spatiotemporal continuity score: To determine the spatiotemporal continuity of the target trajectory, for each target i, its original feature vector is compared and the aggregated feature vector The similarity of , using cosine similarity as the spatiotemporal continuity score, the calculation formula is: Among them, ‖·‖ represents the Euclidean norm, Score i Score the spatiotemporal continuity of target i between adjacent frames. The closer the value is to 1, the better the continuity.

6. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 5 is characterized in that: The calculation of attention weight score The dot product attention mechanism is adopted, which is expressed as: in, is the dot product of the query vector of the i-th target in time frame t and the key vector of the j-th target in time frame t+1, is the normalization factor, N t+1 is the number of targets in time frame t+1, represents the association weight between target i and target j, reflecting the similarity between the two in the feature space. q represents the index variable, which represents the sequence number of all candidate targets in time frame t+1.

7. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 5, characterized in that: S3 also includes a spatiotemporal continuity determination: according to a preset spatiotemporal continuity threshold, Score i Compare with the spatiotemporal continuity threshold to determine whether the trajectory of target i in time frame t and t+1 has spatiotemporal continuity, and update the trajectory association information of the target accordingly.

8. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 5, characterized in that: The S4 specifically includes: S41 training data set construction: using the obtained multi-band feature vector set And the dynamic trajectory correlation matrix M constructed between adjacent time frames t,t+1 , combined with historical classification data in, Represents the historical classification results of target i in time frame t, and constructs a training data set; S42, lightweight classification model composition: The constructed lightweight classification model includes: Long short-term memory network module: used to capture the temporal changes of the target multi-band feature vector set, and its input is the serialized multi-band feature vector The long short-term memory network module outputs the predicted target category and the corresponding confidence level; Adaptive frequency band weight allocation module: used to dynamically adjust the relative weights of the millimeter wave, Ku band and L band features during fusion to form a weighted fusion feature vector F weighted ; S43, model training and parameter update: Using the constructed training data set, the lightweight classification model is trained through an end-to-end training method, so that the long short-term memory network module and the adaptive frequency band weight allocation module can output the target category and confidence according to the input historical features and trajectory association information.

9. The method for classifying unmanned aerial vehicle trajectory tracking targets based on multi-band radar according to claim 8, characterized in that: The lightweight classification model inputs real-time multi-band feature vectors, and fuses the feature vectors of each band into a single feature vector F through the adaptive band weight allocation module. weighted , the fused feature vector F weighted As the input of the long short-term memory network module, the hidden state h is calculated by the long short-term memory network module. t Finally, the target classification is achieved through a fully connected layer and a Softmax layer, and its calculation is expressed as: t =softmax(W y h t +b y ), where W y is the output layer weight matrix, b y is the bias vector, and the softmax(·) function converts the output into a probability distribution of each category, y t is the classification result vector at time t, each component of which represents the confidence of the corresponding category.

10. A UAV trajectory tracking target classification system based on multi-band radar, used to implement the UAV trajectory tracking target classification method based on multi-band radar as claimed in any one of claims 1 to 9, characterized in that: Includes the following modules: A multi-band radar data acquisition module, used to synchronously acquire raw radar echo data of the millimeter wave band, Ku band and L band, generate a multi-band data set including a timestamp, an azimuth angle and a pulse sequence, and perform time synchronization calibration on the data set; A feature extraction module is used to extract multi-dimensional features of the target from the synchronized multi-band data set, wherein the multi-dimensional features include high-resolution micro-motion features of the millimeter wave band, scattering cross-section features of the Ku band, and Doppler velocity features of the L band, and generate a multi-band feature vector; A dynamic trajectory association module, used to construct a dynamic trajectory association matrix according to the multi-band feature vector, and determine the spatiotemporal continuity of the target trajectory through a trajectory association algorithm based on an attention mechanism; A lightweight classification module, which constructs a lightweight classification model including a long short-term memory network module and an adaptive frequency band weight allocation module based on the dynamic trajectory association matrix and historical classification data; The online classification module is used to perform online classification on the real-time multi-band feature vector using the lightweight classification model, and output the target type and the corresponding confidence.

Images