A radar target recognition model training method, a radar target recognition method and a device

Abstract

The application discloses a radar target recognition model training method and device, and belongs to the technical field of deep learning. The method comprises the following steps: adopting a two-stage gradual sequence construction strategy to perform sample expansion construction processing on a plurality of radar track samples, so as to obtain a training sample set and a preset training category label corresponding to each training sample; inputting each training sample into a radar target recognition model to be trained to perform target category recognition processing, so as to obtain a predicted category label corresponding to each training sample; the radar target recognition model to be trained comprises a position encoding model, a speed encoding model, an attention fusion model and a classification model; and based on the predicted category label and the preset training category label corresponding to each training sample, the radar target recognition model to be trained is trained, so as to obtain a trained radar target recognition model. The technical scheme provided by the application can improve the radar target recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a radar target recognition model training, radar target recognition method and apparatus. Background Technology

[0002] Radar target identification is one of the core technologies for low-altitude airspace security. With the widespread use of small drones, the need for rapid and accurate identification of drones and non-cooperative targets such as birds in low-altitude airspace is becoming increasingly urgent. Drones and birds exhibit similar characteristics in radar echoes: both fly at low altitudes, are small in size, and move at relatively slow speeds, which poses a significant challenge to radar-based target identification.

[0003] In existing technologies, radar target identification is achieved based on traditional features and mathematical models. For example, methods based on two-time-time offset calculations compare and prioritize data by calculating the offsets of features such as height, speed, aspect ratio, and micro-Doppler shift between two time points. While these methods are computationally fast, they only use data from two time points, resulting in significant information loss and limited ability to identify UAVs with complex motion patterns. Other methods are based on unsupervised learning, such as unsupervised clustering combined with random forest classification, which can classify targets without labeled data. However, unsupervised methods have a lower accuracy ceiling, and random forests, as shallow classifiers, are weak at modeling deep patterns in radar time-series data, making it difficult to fully extract discriminative information from target trajectories.

[0004] In addition, deep learning-based methods exist for radar target recognition. While deep learning has achieved great success in fields such as image and speech recognition, existing methods still have the following shortcomings in the field of UAV recognition based on radar data: First, the fixed-length sequence construction method ignores the physical characteristics of the gradual accumulation of information about the target in the early stages of radar detection. Existing methods typically require the construction of fixed-length sequence samples, which cannot be recognized when there are insufficient available frames, and fails to fully utilize the physical process of gradual information accumulation when there are sufficient frames. Second, the multi-dimensional features of radar points contain information of different semantic types. Existing methods mix features with different semantics and input them into a single model, ignoring semantic differences and failing to fully utilize the modeling advantages of each. Third, existing multi-model fusion methods often use simple concatenation or fixed weight addition. However, different input features and different encoder branches contribute differently to the discrimination of different samples, and fixed fusion strategies easily lead to suboptimal recognition results.

[0005] Therefore, a more reliable solution is needed to fully utilize the progressive information accumulation characteristics of radar points in radar target recognition, distinguish different semantic feature types and model them separately, and adaptively learn multi-branch fusion weights. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a radar target recognition model training, radar target recognition method and device, which can improve the robustness and training effectiveness of the radar target recognition model, as well as the accuracy of radar target recognition, thereby improving the accuracy, effectiveness and robustness of target category recognition in subsequent radar target recognition applications.

[0007] To achieve the above objectives, the present invention employs the following technical solution:

[0008] On one hand, the present invention provides a radar target recognition model training method, the method comprising:

[0009] Multiple radar trajectory samples and their corresponding preset category labels are acquired. Each radar trajectory sample includes N frames of multi-dimensional features, including position features, velocity features, and attribute features. Each radar trajectory sample corresponds to a preset category label.

[0010] A two-stage progressive sequence construction strategy is adopted to expand and construct multiple radar trajectory samples to obtain a training sample set and a preset training category label corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage.

[0011] Each training sample is input into the radar target recognition model to be trained for target category recognition processing to obtain the predicted category label corresponding to each training sample; wherein, the radar target recognition model to be trained includes a position encoding model for processing the position feature sequence of each training sample, a velocity encoding model for processing the velocity and attribute feature sequence of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the output features of the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing;

[0012] Based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, the radar target recognition model to be trained is trained to obtain a trained radar target recognition model.

[0013] In some possible implementations, the two-stage progressive sequence construction strategy is used to expand and construct multiple radar trajectory samples to obtain a training sample set and a preset training category label corresponding to each training sample, including:

[0014] Iterate through each radar trajectory sample in the plurality of radar trajectory samples;

[0015] When traversing any radar trajectory sample, in the initial expansion phase, based on the current radar trajectory sample, the number of preceding target frames with several dimensions of features are sequentially acquired, gradually accumulating information from the target entering the radar detection range, to obtain the number of target prefix subsequence samples with lengths ranging from 1 to the target number, as shown in the following formula:

[0016] ;

[0017] in, This represents the k-th prefix subsequence sample. Let W represent the features of the k-th frame, W represent the fixed window length, and N represent the number of valid frames for the trajectory. This represents the minimum value of the fixed window length and the number of valid frames for the trajectory, and the number of targets is determined based on the minimum value of the fixed window length and the number of valid frames;

[0018] The number of target prefix subsequence samples inherit the preset category label of the current radar trajectory, thus obtaining the number of target preset training category labels corresponding to the number of target prefix subsequence samples;

[0019] When the number of valid frames of the trajectory is greater than the fixed window length, the system enters the fixed window sliding phase. A sliding window of fixed length slides sequentially over the current radar trajectory sample with a preset step size, generating several sliding sequence samples of equal length and fixed window length, as shown in the following formula:

[0020] ;

[0021] in, This represents the j-th sliding sequence sample; Represents several dimensions of features in the j-th frame;

[0022] Several equal-length sliding sequence samples inherit the preset category label of the current radar trajectory, thus obtaining several preset training category labels corresponding to the several equal-length sliding sequence samples;

[0023] The target number of prefix subsequence samples with preset training category labels and several equal-length sliding sequence samples are merged to form the trajectory training sample set corresponding to the current radar trajectory sample and the preset training category label corresponding to each trajectory training sample;

[0024] Multiple training sample sets corresponding to multiple radar trajectory samples are aggregated into a complete training set, and a random undersampling strategy is used to balance the samples to obtain the training sample set and the preset training category label corresponding to each training sample.

[0025] In some possible implementations, the step of inputting each training sample into the radar target recognition model to be trained for target category recognition processing, and obtaining the predicted category label corresponding to each training sample, includes:

[0026] The position feature sequence of each training sample is input into the position encoding model for position encoding processing to obtain the position encoding output feature of each training sample;

[0027] The speed and attribute feature sequence of each training sample is input into the speed encoding model for speed encoding processing to obtain the global output features of speed and attributes of each training sample.

[0028] The position encoding output features and the velocity and attribute global output features are input into the attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fusion features of each training sample;

[0029] The fused features are input into the classification model for classification processing to obtain the predicted category label.

[0030] In some possible implementations, the position encoding model is a two-layer bidirectional LSTM model, wherein inputting the position feature sequence of each training sample into the position encoding model for position encoding processing to obtain the position encoded output features of each training sample includes:

[0031] The position feature sequence is input into the two-layer bidirectional LSTM model. For each time step, the forward hidden state and backward hidden state corresponding to each time step are determined according to each position feature in the position feature sequence.

[0032] The forward hidden state and the backward hidden state are concatenated to obtain the position encoding feature corresponding to each time step, as shown in the following formula:

[0033] ;

[0034] in, This represents the positional encoding feature corresponding to time step t. This represents the forward hidden state corresponding to time step t. This represents the backward hidden state corresponding to time step t;

[0035] The location encoding feature of the last time step in the location feature sequence is used as the location encoding output feature, as shown in the following formula:

[0036] ;

[0037] in, This indicates the location-encoded output feature. This represents the positional encoding feature of the last time step L, where L represents the length of the positional feature sequence.

[0038] In some possible implementations, the velocity encoding model is a two-layer Transformer encoder. The step of inputting the velocity and attribute feature sequences of each training sample into the velocity encoding model for velocity encoding processing, to obtain the global output features of the velocity and attributes of each training sample, includes:

[0039] The velocity and attribute feature sequence is input into the two-layer Transformer encoder. For each time step, each velocity and attribute feature is processed by the self-attention encoding of the two-layer Transformer encoder to obtain each encoded velocity and attribute feature.

[0040] The average pooling process is applied to all encoded speed and attribute features to obtain the global output features of speed and attributes, as shown in the following formula:

[0041] ;

[0042] in, This indicates the speed and attribute output characteristics. Indicates Transformer encoder, Let L represent the velocity and attribute feature sequence, L represent the length of the velocity and attribute feature sequence, and t represent the t-th time step.

[0043] In some possible implementations, the attention fusion model includes a first linear projection layer, a second linear projection layer, and a bilinear attention network;

[0044] The step of inputting the position-encoded output features and the global output features of velocity and attributes into the attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fused features of each training sample includes:

[0045] The position-encoded output features and the velocity and attribute global output features are mapped to a common latent space dimension through the first linear projection layer and the second linear projection layer, respectively, to obtain position projection features and velocity and attribute projection features.

[0046] The position projection feature and the velocity and attribute projection feature are stitched together to obtain the stitched projection feature.

[0047] The attention weights of the stitched projection features are determined by the bilinear attention network to obtain the position branch weights corresponding to the position projection features and the velocity and attribute branch weights corresponding to the velocity and attribute projection features, as shown in the following formula:

[0048] ;

[0049] in, Indicates the branch weight at position. Indicates speed and attribute branch weight; Indicates the location projection features, Represents velocity and attribute projection characteristics. This indicates that the position projection features and velocity and attribute projection features are stitched together. Indicates the weights of the first layer. Indicates the output layer weights. This represents the first bias vector. This represents the second bias vector. Represents the hyperbolic tangent activation function; Used for normalization to ensure ,and ;

[0050] Based on the position branch weight and the velocity and attribute branch weight, the position projection feature and the velocity and attribute projection feature are weighted and summed to generate a fused feature, as shown in the following formula:

[0051] ;

[0052] in, This indicates the fusion feature.

[0053] In some possible implementations, the plurality of radar trajectory samples and their corresponding plurality of preset category labels are obtained in the following manner:

[0054] Acquire raw radar spot data for M frames;

[0055] Based on the trajectory history and preset threshold of each frame of original radar spot data, the M frames of original radar spot data are filtered to obtain N frames of valid radar spot data.

[0056] Extract 13-dimensional features from each frame of valid radar spot data. The 13-dimensional features of each frame include 6-dimensional position features and 7-dimensional velocity features.

[0057] Based on the 13-dimensional features of the N frames of valid radar point data, multiple radar trajectories are generated.

[0058] Generate category labels for each radar trajectory to obtain the preset category labels for each radar trajectory;

[0059] The multiple radar trajectories and their corresponding preset category labels are determined as the multiple radar trajectory samples and their corresponding preset category labels.

[0060] On the other hand, a radar target identification method is provided, the method comprising:

[0061] Obtain the radar trajectory to be identified;

[0062] The radar trajectory to be identified is input into the radar target recognition model for target category identification processing to obtain the target category label corresponding to the radar trajectory to be identified. The radar target recognition model is obtained based on the radar target recognition model training method described above.

[0063] On the other hand, a radar target recognition model training device is provided, the device comprising:

[0064] The first information acquisition module is configured to acquire multiple radar trajectory samples and corresponding multiple preset category labels. Each radar trajectory sample includes N frames of multi-dimensional features, including position features, velocity and attribute features; each radar trajectory sample corresponds to a preset category label.

[0065] The sample construction module is configured to perform sample expansion construction processing on multiple radar trajectory samples using a two-stage progressive sequence construction strategy to obtain a training sample set and a preset training category label corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage.

[0066] The category recognition processing module is configured to input each training sample into the radar target recognition model to be trained for target category recognition processing, and obtain the predicted category label corresponding to each training sample; wherein, the radar target recognition model to be trained includes a position encoding model for processing the position features of each training sample, a velocity encoding model for processing the velocity and attribute features of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the output features of the velocity encoding model and fusing the output features of the two models, and a classification model for performing classification processing;

[0067] The model training module is configured to train the radar target recognition model to be trained based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, so as to obtain a trained radar target recognition model.

[0068] On the other hand, a radar target identification device is provided, the device comprising:

[0069] The second information acquisition module is configured to acquire the radar trajectory to be identified.

[0070] The category information determination module is configured to input the radar trajectory to be identified into the radar target recognition model for target category recognition processing, and obtain the target category label corresponding to the radar trajectory to be identified. The radar target recognition model is obtained based on the radar target recognition model training method described above.

[0071] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0072] In radar target recognition scenarios, the target category label corresponding to the radar trajectory to be identified is obtained by inputting the radar trajectory to be identified into a radar target recognition model for target category recognition processing. This radar target recognition model is trained based on the predicted category labels and preset category labels corresponding to each training sample in the training sample set. The predicted category label for each training sample is obtained by inputting each training sample into the radar target recognition model to be trained for target category recognition processing. The radar target recognition model to be trained includes a position encoding model for processing the position features of each training sample, a velocity encoding model for processing the velocity and attribute features of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing. The position encoding model and velocity encoding model can process features with different semantics, avoiding information confusion caused by mixing features with different semantics, and the self-attention fusion module can adaptively... Adjusting the fusion weights of the two encoding branches improves the recognition performance of the trained model, enhancing its accuracy and robustness in target category identification. The training sample set is constructed using a two-stage progressive sequence construction strategy, expanding multiple radar trajectory samples. These two stages include an initial expansion stage and a fixed-window sliding stage. The samples established in the first stage enable the model to learn to handle target recognition under data with varying degrees of completeness, allowing it to output recognition results even with only a few valid point frames, thus improving model response speed. The second stage significantly increases the amount of data involved, achieving data augmentation to alleviate the problem of insufficient radar training samples, thereby improving the effectiveness and accuracy of model training. Furthermore, training the model with samples of diverse sequence lengths constructed in the two stages enhances the model's robustness and accuracy in recognizing input sequences of arbitrary lengths. This effectively improves the reliability of the radar target recognition model during training, further enhancing the accuracy of target category identification, and ultimately improving the accuracy and effectiveness of target category identification during radar target recognition. Attached Figure Description

[0073] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0074] Figure 1 This is a flowchart illustrating a radar target recognition model training method provided in an embodiment of the present invention;

[0075] Figure 2 This is a schematic diagram of a sample set constructed by a two-stage progressive sequence construction strategy provided in an embodiment of the present invention;

[0076] Figure 3 A schematic diagram illustrating the process of using a radar target recognition model to be trained, as provided in an embodiment of the present invention, to identify target categories.

[0077] Figure 4 This is a schematic flowchart of a radar target recognition method provided in an embodiment of the present invention;

[0078] Figure 5 This is a schematic diagram of the structure of a radar target recognition model training device provided in an embodiment of the present invention;

[0079] Figure 6 This is a schematic diagram of the structure of a radar target identification device provided in an embodiment of the present invention. Detailed Implementation

[0080] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0081] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0082] In this embodiment of the invention, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0083] Various exemplary embodiments, features, and aspects of the present invention will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0084] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0085] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0086] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without certain specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art have not been described in detail in order to highlight the spirit of the invention.

[0087] Figure 1 This is a flowchart illustrating a radar target recognition model training method according to an embodiment of the present invention. This specification provides the method operation steps as shown in the embodiments or flowcharts, but based on conventional or non-inventive methods, more or fewer operation steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only possible execution order. In actual system or server products, the method can be executed sequentially according to the embodiments or drawings, or in parallel (e.g., in a parallel processor or multi-threaded processing environment). Specifically, as shown... Figure 1 As shown, the above method may include:

[0088] S101: Acquire multiple radar trajectory samples and their corresponding preset category labels;

[0089] In one specific embodiment, each radar trajectory sample includes N frames of multi-dimensional features, including position features and velocity and attribute features; each radar trajectory sample corresponds to a preset category label. Specifically, each radar trajectory sample includes N frames of 13-dimensional features, which include 6-dimensional position features and 7-dimensional velocity and attribute features. The 6-dimensional features can reflect the spatial trajectory shape of the target; the 7-dimensional velocity and attribute features are specifically velocity and physical attribute features, which can reflect the instantaneous motion state and scattering characteristics of the target.

[0090] In an optional embodiment, the aforementioned plurality of radar trajectory samples and corresponding plurality of preset category labels can be obtained in the following manner:

[0091] Acquire raw radar spot data for M frames;

[0092] Based on the trajectory history and preset threshold of each frame of original radar spot data, M frames of original radar spot data are filtered to obtain N frames of valid radar spot data.

[0093] Extract 13-dimensional features from each frame of valid radar spot data. The 13-dimensional features of each frame include 6-dimensional position features and 7-dimensional velocity features.

[0094] Multiple radar trajectories are generated based on the 13-dimensional features of N frames of valid radar point data.

[0095] Generate category labels for each radar trajectory to obtain the preset category labels for each radar trajectory;

[0096] Multiple radar trajectories and their corresponding preset category labels are identified as multiple radar trajectory samples and their corresponding preset category labels.

[0097] In one specific embodiment, the original radar point data may contain multiple data, such as track quality, track history, target status, radial distance, point distance, radial velocity, target signal-to-noise ratio, etc. The track history in each frame of original radar point data can be used to perform quality control on the M frames of original radar point data. Optionally, if the aircraft history value in the current frame of original radar point data is greater than a preset threshold, the current frame of original radar point data is retained; otherwise, the current frame of original radar point data is discarded. The retained original radar point data can be point data that has passed quality control. Specifically, the above-mentioned filtering of M frames of original radar point data based on the track history and preset threshold of each frame of original radar point data to obtain N frames of valid radar point data may include: retaining original radar point data with track history values ​​greater than the preset threshold in the M frames of original radar point data to obtain N frames of valid radar point data; and discarding original radar point data with track history values ​​less than or equal to the preset threshold. Optionally, the N frames of valid radar spot data can be radar spot data that has passed quality control; the preset threshold can be set according to actual application requirements; the track history can represent the number of times the radar has observed the target, and the target can only be confirmed as a real observed target when it reaches the preset threshold. Data quality control through preset thresholds can effectively remove noise frames and invalid measurement frames in the radar detection process, thereby ensuring the quality of subsequent feature extraction.

[0098] In a specific embodiment, the 13-dimensional features of each frame of valid radar point data include 6-dimensional position features and 7-dimensional velocity and attribute features. Optionally, the 6-dimensional position features corresponding to each frame of valid radar point data may include: radial distance, azimuth angle, elevation angle, point distance, point azimuth, and point elevation. Specifically, the position feature vector of the t-th frame... Specifically as follows: ,in, This represents the radial distance in frame t. This represents the azimuth angle of the t-th frame. This represents the pitch angle in frame t. This represents the distance of the point in frame t. This indicates the orientation of the point in frame t. This represents the elevation of the radar trace in frame t. The 7-dimensional velocity and attribute features corresponding to each frame of valid radar trace data can include: total velocity, radial velocity, azimuth velocity, elevation velocity, radar cross section, Doppler broadening, and jet engine modulation. Specifically, the velocity and attribute feature vector of frame t... Specifically as follows: ,in, Represents the full velocity of frame t. Represents the radial velocity in frame t. Represents the azimuth velocity in frame t. This represents the pitch speed in frame t. This represents the radar cross section of the t-th frame. This indicates the Doppler broadening of the t-th frame. This represents the jet-powered transmission mechanism for frame t. Optionally, utilizing the 13-dimensional universal physical features of radar spot data can eliminate reliance on proprietary parameters specific to a particular radar model, thereby enabling the trained model to have good cross-platform transfer capabilities.

[0099] In one specific embodiment, target association can be performed on N frames of valid radar point data to generate multiple radar trajectories. Specifically, target association is performed on N frames of valid radar point data and their corresponding 13-dimensional features to generate multiple radar trajectories. Specifically as follows: Where N is the number of valid frames. Let be the 13-dimensional feature of frame t, and .

[0100] In one specific embodiment, the target category identification based on radar data of the present invention mainly targets targets with low-altitude flight, small size, and relatively slow transport speed, such as drones and birds. This embodiment specifically illustrates the identification of drones. Specifically, the above-mentioned generation of category labels for each radar trajectory can include: if the number of valid radar points corresponding to the radar trajectory is valid radar point data from the drone catalog, then the radar trajectory is marked as a positive sample (y=1, drone), that is, the preset category label corresponding to the radar trajectory is 1, and the category corresponding to the radar trajectory is drone; if the valid radar point data corresponding to the radar trajectory is valid radar point data from a non-drone catalog, then the radar trajectory is marked as a negative sample (y=0, non-drone), that is, the preset category label corresponding to the radar trajectory is 0, and the category corresponding to the radar trajectory is non-drone.

[0101] S102: A two-stage progressive sequence construction strategy is adopted to expand and construct multiple radar trajectory samples to obtain a training sample set and the preset training category labels corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage.

[0102] In a specific embodiment, the two-stage progressive sequence construction strategy includes a first stage, namely the initial expansion stage, in which prefix subsequence samples are gradually expanded starting from sequence length 1, and a second stage, namely the fixed window sliding stage, in which a fixed-length window is used to slide and truncate equal-length sequence samples on the complete trajectory.

[0103] In an optional embodiment, the two-stage progressive sequence construction strategy described above for expanding and constructing multiple radar trajectory samples to obtain a training sample set and a preset training category label corresponding to each training sample may include:

[0104] Iterate through each radar trajectory sample in multiple radar trajectory samples;

[0105] When traversing any radar trajectory sample, in the initial expansion phase, based on the current radar trajectory sample, the number of preceding target frames with several dimensions of features are sequentially acquired, gradually accumulating information from the target entering the radar detection range, to obtain the number of target prefix subsequence samples with lengths ranging from 1 to the target number, as shown in the following formula:

[0106] ;

[0107] in, This represents the k-th prefix subsequence sample. Let W represent the features of the k-th frame, W represent the fixed window length, and N represent the number of valid frames for the trajectory. This represents the minimum value of the fixed window length and the number of valid frames for the trajectory. The number of targets is determined based on the minimum value of the fixed window length and the number of valid frames.

[0108] The number of target prefix subsequence samples inherit the preset category labels of the current radar trajectory, resulting in the number of target preset training category labels corresponding to the number of target prefix subsequence samples;

[0109] If the number of valid frames of the trajectory is greater than the fixed window length, the system enters the fixed window sliding phase. A sliding window of fixed length is used to slide sequentially over the current radar trajectory sample with a preset step size, generating several sliding sequence samples of equal length and fixed window length, as shown in the following formula:

[0110] ;

[0111] in, This represents the j-th sliding sequence sample; Represents several dimensions of features in the j-th frame;

[0112] Several equal-length sliding sequence samples inherit the preset category labels of the current radar trajectory, resulting in several preset training category labels corresponding to the several equal-length sliding sequence samples;

[0113] The target number of prefix subsequence samples with preset training category labels and several equal-length sliding sequence samples are merged to form the trajectory training sample set corresponding to the current radar trajectory sample and the preset training category label corresponding to each trajectory training sample.

[0114] A plurality of trajectory training sample sets corresponding to a plurality of radar trajectory samples are aggregated into a complete training set, and a random under-sampling strategy is adopted for sample balance processing, so as to obtain a training sample set and a preset training category label corresponding to each training sample.

[0115] In a specific embodiment, Figure 2 is a schematic diagram of a sample set constructed by a two-stage progressive sequence construction strategy provided by an embodiment of the present invention. Since the information volume is gradually accumulated from less to more when a target just enters the radar detection range, in the first stage, i.e., the initial expansion stage, for the preceding frames of the current radar trajectory sample, 13-dimensional features of the target with the information volume gradually accumulating from less to more starting from entering the radar detection range may be obtained in sequence, to obtain prefix subsequence samples with lengths ranging from 1 to in number. Optionally, the target number of prefix subsequence samples means that the number of the prefix subsequence samples is the target number. Optionally, the prefix subsequence samples have a sequence length L=k, and the maximum number of samples output in this stage may be . Specifically, when W>N, N prefix subsequence samples of various lengths with lengths ranging from 1 to N are obtained; when W<N, W prefix subsequence samples of various lengths with lengths ranging from 1 to W are obtained. Optionally, the fixed window length W may be determined in combination with actual application requirements, and the valid trajectory frame number N is the actual valid frame number. Specifically, when the fixed window length W=200 and the valid trajectory frame number N=250, as shown in Figure 2 , the 13-dimensional features of the preceding 200 frames corresponding to the target with the information volume gradually accumulating from less to more starting from entering the radar detection range are obtained in sequence, so as to obtain 200 prefix subsequence samples of various lengths with lengths ranging from 1 to 200.

[0116] In a specific embodiment, when k=1, a training sample can be generated using only the radar features of the first frame, which enables the model to learn the recognition capability in an extreme cold start scenario. Optionally, when the fixed window length is less than the valid trajectory frame number, the number of prefix subsequence samples is determined according to the fixed window length, and the second stage is further performed to continue expanding and constructing samples; when the fixed window length is greater than the valid frame number, the number of prefix subsequence samples is determined according to the valid frame number.

[0117] ​In one specific embodiment, when the target has been stably tracked, the number of valid trajectory frames is greater than the fixed window length. Therefore, when the number of valid trajectory frames is greater than the fixed window length, the target has been stably tracked, and the process enters the fixed window sliding phase to expand and construct samples. Specifically, a sliding window of length W is used to slide sequentially along the complete trajectory with a preset step size, generating several sliding sequence samples of equal length and length W. Optionally, the preset step size can be set according to actual application requirements; specifically, the preset step size is 1 frame. Optionally, each sliding sequence sample... The sequence length can be fixed at W, and the sliding start index j starts from 2 to avoid duplication with the W-th sample in the first stage. In this case, the maximum number of samples produced in this stage can be NW (N>W). By generating a large number of samples over a long trajectory through a sliding window, implicit data augmentation can be achieved. Specifically, for example... Figure 2 As shown, a sliding window with a fixed length of W=200 is used to slide sequentially on the complete trajectory with a step size of 1 frame. When N=250, 50 sliding sequence samples of equal length and length 200 are generated.

[0118] In one specific embodiment, the training data can be labeled with preset training tags. The prefix subsequence samples and NW equal-length sliding sequence samples are merged to form the trajectory training sample set corresponding to the current radar trajectory sample and the preset training category label corresponding to each trajectory training sample, as shown in the following formula:

[0119] ,in, This represents the trajectory training sample set generated from the current radar trajectory sample T. The total number of samples in the trajectory training sample set is the sum of the number of prefix subsequence samples generated in the first stage and the number of equal-length sliding sequence samples generated in the second stage. Optionally, each radar trajectory sample corresponds to one trajectory training sample set.

[0120] After traversing all radar trajectory samples, the trajectory training sample set of all radar trajectory samples is summarized into a complete training set.

[0121] In the above embodiments, radar trajectory samples are constructed using a two-stage progressive sequence construction strategy. This allows for the construction of diverse long sequence data of different lengths from a small amount of point data, thereby increasing sample diversity. This enables the model to learn the ability to process inputs with different levels of information completeness during the training phase, and thus ensures the effectiveness and robustness of the model's recognition results for radar data input sequences of arbitrary lengths detected by radar during the radar target recognition process.

[0122] Optionally, the above-mentioned collection of multiple trajectory training sample sets corresponding to multiple radar trajectory samples into a complete training set, and the sample balancing process using a random undersampling strategy, can result in a training sample set that includes:

[0123] Multiple radar trajectory samples corresponding to multiple trajectory training sample sets are aggregated into a complete training set;

[0124] The complete training set is divided into a set of drone samples and a set of non-drone samples.

[0125] The number of drone sample sets is determined as the first quantity, and the number of non-drone sample sets is determined as the second quantity;

[0126] Take the minimum of the first and second quantities as the base quantity;

[0127] A baseline number of drone samples and a baseline number of non-drone samples are randomly drawn without replacement from the drone sample set and the non-drone sample set, respectively.

[0128] The baseline number of drone samples and the baseline number of non-drone samples are merged to obtain the initial training sample set;

[0129] The initial training sample set is randomly divided into a training sample set and a validation test set according to a preset ratio.

[0130] In one specific embodiment, during full training, there is often a serious imbalance in the number of drone samples and non-drone samples. Directly using imbalanced data for training can easily lead to the model being biased towards the majority class, severely degrading the recognition performance of the minority class. Therefore, a random undersampling strategy can be used to balance the samples.

[0131] In one specific embodiment, the multiple trajectory training sample sets corresponding to multiple radar trajectory samples are divided into a UAV sample set and a non-UAV sample set. Specifically, the UAV training sample set is D1, with a quantity of |D1|=n1, and the non-UAV sample set is D0, with a quantity of |D0|=n0. The minimum value between n1 and n0 is taken as the baseline quantity, i.e., the baseline quantity is... Randomly sampled without replacement from the drone sample set and the non-drone sample set respectively. A drone sample and Non-drone samples The specific formula is as follows:

[0132] ;

[0133] ;

[0134] Will A drone sample and Non-drone samples Merge to obtain the initial training sample set. The specific formula is as follows:

[0135] At this point, the ratio of positive to negative samples is 1:1. Optionally, the initial training sample set can be a balanced dataset that has undergone sample balancing.

[0136] Optionally, the initial training sample set can be randomly divided into a training sample set and a validation test set according to a preset ratio, with the random seed set to 1 to ensure reproducibility. Specifically, the preset ratio can be set according to actual application needs. A preset ratio can be 9:1, meaning that 90% of the initial training sample set can be used as the training sample set for model training, and 10% of the initial training sample set can be used as the validation test set for calculating evaluation metrics.

[0137] S103: Input each training sample into the radar target recognition model to be trained for target category recognition processing, and obtain the predicted category label corresponding to each training sample;

[0138] In one specific embodiment, the radar target recognition model to be trained includes a position encoding model for processing the position feature sequence of each training sample, a velocity encoding model for processing the velocity and attribute feature sequence of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing. Specifically, the position encoding model is used to process the 6-dimensional position feature sequence, the velocity encoding model is used to process the 7-dimensional velocity and attribute feature sequence, the attention fusion model is used to adaptively fuse the output features of the two encoding models, and the classification model is used to output the binary classification probability of UAVs and non-UAVs.

[0139] In an optional embodiment, the above-described inputting each training sample into the radar target recognition model to be trained for target category recognition processing, to obtain the predicted category label corresponding to each training sample, may include:

[0140] The position feature sequence of each training sample is input into the position encoding model for position encoding processing to obtain the position encoding output feature of each training sample.

[0141] The speed and attribute feature sequence of each training sample is input into the speed encoding model for speed encoding processing to obtain the global output features of speed and attributes of each training sample.

[0142] The positional encoding output features and the global output features of velocity and attributes are input into the attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fused features of each training sample.

[0143] The fused features are input into the classification model for classification processing to obtain the predicted category label.

[0144] In a specific embodiment, such as Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the process of using a radar target recognition model to perform target category recognition, as provided in this embodiment of the invention. Specifically, the position feature sequence of each training sample is input into a position encoding model for position encoding processing to obtain the position encoded features of each training sample, and the velocity and attribute feature sequences of each training sample are input into a velocity encoding model for velocity encoding processing to obtain the global output features of velocity and attributes of each training sample. Next, the position encoded output features and the global output features of velocity and attributes are input into an attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fused features of each training sample. Finally, the fused features are input into a classification model for classification processing to obtain the predicted category label.

[0145] In a specific embodiment, the position encoding output feature can be obtained by combining the position encoding model to perform position encoding processing on the position feature sequence; the velocity and attribute global output feature can be obtained by combining the velocity encoding model to perform velocity encoding processing on the velocity and attribute feature sequence; the fusion feature is the fusion feature information corresponding to the training sample after performing feature weighted fusion processing on the position encoding output feature and the velocity and attribute global output feature. Optionally, the fusion feature can be obtained by combining the attention fusion model to perform feature weighted fusion processing on the position encoding output feature and the velocity and attribute global output feature.

[0146] In an optional embodiment, the above-mentioned positional encoding model is a two-layer bidirectional LSTM model. The above-mentioned inputting the positional feature sequence of each training sample into the positional encoding model for positional encoding processing to obtain the positional encoding output features of each training sample may include:

[0147] The location feature sequence is input into a two-layer bidirectional LSTM model. For each time step, the forward hidden state and backward hidden state corresponding to each time step are determined based on each location feature in the location feature sequence.

[0148] The forward hidden state and the backward hidden state are concatenated to obtain the positional encoding feature corresponding to each time step, as shown in the following formula:

[0149] ;

[0150] in, This represents the positional encoding feature corresponding to time step t. This represents the forward hidden state corresponding to time step t. This represents the backward hidden state corresponding to time step t;

[0151] The location-encoded feature of the last time step in the location feature sequence is used as the location-encoded output feature, as shown in the following formula:

[0152] ;

[0153] in, This indicates the location-encoded output feature. This represents the positional encoding feature of the last time step L, where L represents the length of the positional feature sequence.

[0154] In one specific embodiment, the positional encoding model is used to process positional feature sequences. Optionally, the model structure of the positional encoding model can be set according to the actual application. Optionally, the positional encoding model can be a positional encoder, specifically a two-layer bidirectional LSTM (Long Short-Term Memory) model, whose hidden layer dimension can be [missing information]. Optionally, each training sample can be a trajectory sample of length L in the training sample set, specifically... Its positional feature sequence can be Optionally, the training sample set can be input into the model in batches for training. Since the sample sequence lengths in the training batches are inconsistent, a variable-length sequence packing encoding technique can be used. Specifically, the samples within the batch are sorted in descending order of sequence length; the samples sorted in descending order within the batch are padded, and the padded sequence is packed into a compact format using a function call and input into the bidirectional LSTM model to obtain positional encoding features; however, during training, the sequence length obtained by the model is the true length of the sequence, and the padding bits are ignored to avoid the padded 0s affecting model learning.

[0155] In a specific embodiment, for each time step t, the forward hidden state of each time step is determined based on each position feature in the position feature sequence. and backward hidden state Next, the forward hidden state and the backward hidden state are concatenated to obtain the positional encoding features at each time step. ,Right now , The positional encoding feature of the last time step in the sequence is used as the final output of the positional branch, i.e., the positional encoding output feature. That is, the output corresponding to the actual sequence length L. The positional encoding output feature of the final output contains the complete contextual information of the entire sequence.

[0156] In an optional embodiment, the above-mentioned velocity encoding model is a two-layer Transformer encoder. The process of inputting the velocity and attribute feature sequences of each training sample into the velocity encoding model for velocity encoding to obtain the global output features of velocity and attributes for each training sample may include:

[0157] The velocity and attribute feature sequences are input into a two-layer Transformer encoder. For each time step, each velocity and attribute feature is processed by the self-attention encoding of the two-layer Transformer encoder to obtain each encoded velocity and attribute feature.

[0158] The average pooling process is applied to all encoded speed and attribute features to obtain the global output features of speed and attributes, as shown in the following formula:

[0159] ;

[0160] in, This indicates the speed and attribute output characteristics. Indicates Transformer encoder, Let L represent the velocity and attribute feature sequence, L represent the length of the velocity and attribute feature sequence, and t represent the t-th time step.

[0161] In one specific embodiment, the velocity encoding model is used to process velocity and attribute feature sequences. Optionally, the model structure of the velocity encoding model can be set according to the actual application. Optionally, the velocity encoding model can be a velocity encoder, specifically a two-layer Transformer encoder, each layer being a standard Transformer encoder, whose configuration parameters can be the input / output feature dimensions. That is, the speed and attribute feature dimensions, and the number of self-attention heads. Feedforward network dimension Dropout rate = 0.1. The velocity and attribute feature sequence of each training sample can be... .

[0162] In a specific embodiment, the velocity and attribute feature sequences are input into a two-layer Transformer encoder. After self-attention encoding by the two layers of Transformer encoders, mean pooling is performed on all encoded velocity and attribute features to obtain the global output features of velocity and attributes, as shown in the following formula:

[0163] ;

[0164] Optionally, the self-attention mechanism of the two-layer Transformer encoder allows each velocity and attribute feature to pay attention to other features, thereby capturing global dependencies so that each velocity and attribute feature contains global information. Mean pooling can utilize global information more evenly across all time steps and is more robust to changes in sequence length.

[0165] In the above embodiments, the position features and velocity and attribute features can be trained independently by using an LSTM model and a Transformer encoder, respectively. This allows different models to process two semantically different features. By using LSTM to process position features, the long-term temporal dependencies of the target's motion trajectory can be captured. The Transformer encoder can use its self-attention mechanism to capture global feature interactions. The two branches are independent of each other, giving full play to their respective advantages and preventing interference between them.

[0166] In an optional embodiment, the attention fusion model described above includes a first linear projection layer, a second linear projection layer, and a bilinear attention network;

[0167] The above-mentioned input of position-encoded output features and global output features of velocity and attributes into the attention fusion model performs adaptive weight determination and feature weighted fusion processing to obtain the fused features of each training sample, which may include:

[0168] The position-encoded output features and the global output features of velocity and attributes are mapped to a common latent space dimension by the first linear projection layer and the second linear projection layer, respectively, to obtain the position projection features and the velocity and attribute projection features.

[0169] The position projection features and velocity and attribute projection features are stitched together to obtain the stitched projection features.

[0170] The stitched projection features are processed by a bilinear attention network to determine the attention weights, resulting in the position branch weights corresponding to the position projection features and the velocity and attribute branch weights corresponding to the velocity and attribute projection features, as shown in the following equation:

[0171] ;

[0172] in, Indicates the branch weight at position. Indicates speed and attribute branch weight; Indicates the location projection features, Represents velocity and attribute projection characteristics. This indicates that the position projection features and velocity and attribute projection features are stitched together. Indicates the weights of the first layer. Indicates the output layer weights. This represents the first bias vector. This represents the second bias vector. Represents the hyperbolic tangent activation function; Used for normalization to ensure ,and ;

[0173] Based on the position branch weight and the velocity and attribute branch weight, the position projection features and the velocity and attribute projection features are weighted and summed to generate the fused features, as shown in the following formula:

[0174] ;

[0175] in, This indicates the fusion feature.

[0176] In one specific embodiment, the attention fusion model is used to determine the adaptive weights of the location-encoded output features, velocity, and attribute global output features, and then weights and fuses these features. Optionally, the model structure of the attention fusion model can be configured according to the actual application; specifically, the attention fusion model may include a first linear projection layer, a second linear projection layer, and a bilinear attention network. Because the location-encoded output features output by the location-encoded model... The velocity and attribute global output features of the velocity encoding model Since the dimensions are different, the position-encoded output features and the global output features of velocity and attributes are mapped to a common latent space dimension through independent first and second linear projection layers, respectively. The position projection features and velocity and attribute projection features can be obtained as shown in the following formula:

[0177] ;

[0178] ;

[0179] in, Indicates the location projection features, Represents velocity and attribute projection characteristics; This represents the learnable projection weight matrix of the first linear projection layer. This represents the learnable projection weight matrix of the second linear projection layer; This represents the bias vector of the first linear projection layer. This represents the bias vector of the second linear projection layer;

[0180] Next, the position projection features and velocity and attribute projection features are concatenated to obtain the concatenated projection features. Then, a bilinear attention network is used to determine the attention weights of the concatenated projection features, thus determining the adaptive weights for the position branch, velocity branch, and attribute branch, i.e., the position branch weights and the velocity and attribute branch weights. Optionally, the position branch weights can characterize the relative importance of the position-encoded output features in the training samples to the final recognition decision, and the velocity and attribute branch weights can characterize the relative importance of the global output features of velocity and attributes in the training samples to the final recognition decision; optional, The normalized bi-branch weights can be used to characterize the degree of attention focus of the model;

[0181] Specifically, First layer weight First bias vector Output layer weights Second bias vector ;

[0182] Next, based on the position branch weights and velocity and attribute branch weights, the position projection features and velocity and attribute projection features are weighted and summed to generate fused features: .

[0183] In the above embodiments, the attention fusion module enables the model to discover the differentiated contributions of different feature branches to different samples and can adaptively adjust the fusion weights of the two encoding branches. Specifically, for UAV targets with obvious motion features, the model strengthens the contribution of the velocity encoding model branch; for non-UAV targets with unique spatial trajectory shapes, the model strengthens the contribution of the position encoding model branch.

[0184] Optionally, the above classification model can be a classifier. Specifically, the classifier consists of two fully connected layers. Optionally, the first fully connected layer performs feature extraction and nonlinear transformation processing on the fused features to output high-level features, as shown below:

[0185] ;

[0186] in, Indicates high-level features, This represents the first weight matrix. This represents the first classification bias vector;

[0187] Next, the high-level features are classified using a second fully connected network to output the predicted class label, as shown in the following formula:

[0188] ;

[0189] in, This represents the predicted category label, indicating the probability that the predicted training sample belongs to a drone. This represents the second weight matrix. This represents the second classification bias vector. This represents the Sigmoid activation function, here. It can correspond to 'm' in the activation function and be processed by the activation function;

[0190] Optionally, if the probability of a training sample being a drone is greater than a preset probability threshold, it is determined to be a drone; otherwise, it is considered a non-drone. The preset probability threshold can be set according to the actual application. Specifically, the preset probability threshold is 0.5, meaning that... At that time, the training sample was determined to be a drone.

[0191] S104: Based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, train the radar target recognition model to be trained to obtain the trained radar target recognition model.

[0192] In one specific embodiment, training the radar target recognition model to be trained based on the predicted category label and the preset training category label to obtain the radar target recognition model may include: determining the category recognition loss of the radar target recognition model to be trained based on the predicted category label and the preset training category label; and training the radar target recognition model to be trained based on the category recognition loss to obtain the radar target recognition model.

[0193] In one specific embodiment, the category recognition loss can be calculated using a preset loss function. Optionally, the preset loss function can be set according to actual application requirements, such as the exponential loss function, cross-entropy loss function, etc. The category recognition loss can characterize the accuracy of the current radar target recognition model being trained in class recognition. Specifically, the preset loss function can adopt the standard binary cross-entropy loss, as shown in the following formula:

[0194]

[0195] in, B represents the category recognition loss, and B represents the batch. This represents the preset training category label for the i-th training sample. This represents the predicted category label of the i-th training sample, i.e., the model's prediction of the probability that the sample belongs to a drone.

[0196] In a specific embodiment, training the radar target recognition model based on category recognition loss to obtain the radar target recognition model may include: updating the model parameters of the radar target recognition model based on category recognition loss; repeating the training iteration steps of inputting each training sample into the radar target recognition model for target category recognition processing to obtain the predicted category label corresponding to each training sample; and updating the model parameters of the radar target recognition model based on category recognition loss, until a preset convergence condition is met. The preset convergence condition may be that the category recognition loss is less than or equal to a preset loss threshold, or that the number of training iterations reaches a preset number, etc. Specifically, the preset loss threshold and the preset number of iterations can be set according to the model accuracy and training speed requirements in practical applications.

[0197] In one specific embodiment, after the model training is completed, evaluation metrics can be calculated using a validation test set to evaluate the model. Optionally, the evaluation metrics may include accuracy, precision, recall, etc.

[0198] As can be seen from the technical solutions provided in the embodiments of this specification above, this specification obtains multiple radar trajectory samples and corresponding multiple preset category labels. Each radar trajectory sample includes N frames of several-dimensional features, including position features, velocity, and attribute features. Each radar trajectory sample corresponds to a preset category label. A two-stage progressive sequence construction strategy is used to expand and construct multiple radar trajectory samples to obtain a training sample set and preset training category labels corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage. The samples established in the first stage enable the model to learn to process target recognition under data with different completeness. Even when there are only a few valid point trace frames, the recognition result can be output, which can improve the model's response speed. The second stage can significantly increase the amount of data involved, realize data augmentation, alleviate the problem of insufficient radar data training samples, and thus improve the effectiveness and accuracy of model training. Furthermore, by training the model with samples of diverse sequence lengths constructed in the two stages, the robustness and accuracy of the model in recognizing input sequences of arbitrary lengths can be improved.

[0199] Each training sample is input into the radar target recognition model to be trained for target category recognition processing, thereby obtaining the predicted category label corresponding to each training sample. The radar target recognition model to be trained includes a position encoding model for processing the position feature sequence of each training sample, a velocity encoding model for processing the velocity and attribute feature sequences of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing. The position encoding model and velocity encoding model can process features with different semantics, avoiding information confusion caused by mixing features with different semantics. Furthermore, the self-attention fusion module can adaptively adjust the fusion weights of the two encoding branches, thereby improving the recognition performance of the trained model and enhancing the accuracy and robustness of target category recognition.

[0200] Based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, the radar target recognition model to be trained is trained to obtain a trained radar target recognition model; thereby improving the effectiveness and reliability of model training and the accuracy and effectiveness of radar target category recognition.

[0201] The following describes a radar target recognition method based on the radar target recognition model trained by the radar target recognition model training method described above in this invention. Figure 4 This is a flowchart illustrating a radar target recognition method provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the method may include:

[0202] S401: Acquire the radar trajectory to be identified;

[0203] S402: Input the radar trajectory to be identified into the radar target recognition model for target category recognition processing to obtain the target category label corresponding to the radar trajectory to be identified.

[0204] In one specific embodiment, the radar trajectory to be identified can be a radar trajectory for confirming the target category. Specifically, the radar trajectory to be identified can be a radar trajectory for confirming whether the category is a drone. The target category label can be the probability prediction result of the radar target recognition model that the radar trajectory to be identified belongs to a drone. Optionally, if the probability prediction result of belonging to a drone is greater than a preset probability threshold, the training sample is determined to be a drone; otherwise, it is not a drone. The preset probability threshold can be set according to actual applications. Optionally, the radar target recognition method provided by this invention can achieve accurate and effective identification of drones or birds among low-altitude, small-sized, and relatively slow-moving targets in low-altitude scenarios, and can be applied to engineering applications such as low-altitude airspace security and airport bird strike prevention.

[0205] As can be seen from the technical solutions provided in the embodiments of this specification above, in the radar target recognition scenario, the target category label corresponding to the radar trajectory to be recognized is obtained by inputting the radar trajectory to be recognized into the radar target recognition model for target category recognition processing. This radar target recognition model is a radar target recognition model trained on the radar target recognition model to be trained based on the predicted category label and preset category label corresponding to each training sample in the training sample set. The predicted category label corresponding to each training sample is obtained by inputting each training sample into the radar target recognition model to be trained for target category recognition processing. The radar target recognition model to be trained includes a position encoding model for processing the position features of each training sample, a velocity encoding model for processing the velocity and attribute features of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing. The position encoding model and velocity encoding model can process features with different semantics, avoiding information confusion caused by mixing features with different semantics. The self-attention fusion module adaptively adjusts the fusion weights of the two encoding branches, thereby improving the recognition performance of the trained model and enhancing the accuracy and robustness of target category recognition. Furthermore, the training sample set is obtained by expanding multiple radar trajectory samples using a two-stage progressive sequence construction strategy. These two stages include an initial expansion stage and a fixed-window sliding stage. The samples established in the first stage enable the model to learn to handle target recognition under data with varying degrees of completeness, allowing it to output recognition results even with only a few valid point frames, thus improving model response speed. The second stage significantly increases the amount of data involved, achieving data augmentation to alleviate the problem of insufficient radar training samples, thereby improving the effectiveness and accuracy of model training. Moreover, training the model with samples of diverse sequence lengths constructed in the two stages improves the robustness and accuracy of the model in recognizing input sequences of arbitrary lengths. This effectively enhances the reliability of model training during the radar target recognition model training process, further improving the accuracy of target category recognition, and ultimately enhancing the accuracy and effectiveness of target category recognition in the radar target recognition process.

[0206] This invention also provides a radar target recognition model training device, accordingly, Figure 5 This is a schematic diagram of the structure of a radar target recognition model training device provided in an embodiment of the present invention; as shown below. Figure 5 As shown, the above-mentioned device includes:

[0207] The first information acquisition module 510 is configured to acquire multiple radar trajectory samples and corresponding multiple preset category labels. Each radar trajectory sample includes N frames of several-dimensional features, including position features, velocity and attribute features; each radar trajectory sample corresponds to a preset category label.

[0208] The sample construction module 520 is configured to perform sample expansion construction processing on multiple radar trajectory samples using a two-stage progressive sequence construction strategy to obtain a training sample set and a preset training category label corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage.

[0209] The category recognition processing module 530 is configured to input each training sample into the radar target recognition model to be trained for target category recognition processing, and obtain the predicted category label corresponding to each training sample; wherein, the radar target recognition model to be trained includes a position encoding model for processing the position features of each training sample, a velocity encoding model for processing the velocity and attribute features of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the output features of the velocity encoding model and fusing the output features of the two models, and a classification model for performing classification processing;

[0210] The model training module 540 is configured to train the radar target recognition model to be trained based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, so as to obtain a trained radar target recognition model.

[0211] This invention also provides a radar target identification device, accordingly, Figure 6 This is a schematic diagram of the structure of a radar target identification device provided in an embodiment of the present invention; as shown. Figure 6 As shown, the above-mentioned device includes:

[0212] The second information acquisition module 610 is configured to acquire the radar trajectory to be identified.

[0213] The category information determination module 620 is configured to input the radar trajectory to be identified into the radar target recognition model for target category recognition processing, and obtain the target category label corresponding to the radar trajectory to be identified. The radar target recognition model is obtained based on the radar target recognition model training method provided in the embodiments of the present invention.

[0214] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0215] This invention also provides an electronic device, which includes a processor and a memory. The memory stores at least one instruction, at least one program, a code set, or an instruction set. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the radar target recognition model training method or the radar target recognition method as described in this invention.

[0216] Embodiments of the present invention also provide a computer storage medium, which can be disposed in a server to store at least one instruction, at least one program, code set, or instruction set for implementing the method embodiments. The at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by the processor to implement the radar target recognition model training method or radar target recognition method as described in the embodiments of the present invention.

[0217] Optionally, in embodiments of the present invention, the storage medium may be located at at least one of a plurality of network servers in a computer network. Optionally, in embodiments of the present invention, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0218] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0219] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more flowcharts and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0220] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more flowcharts and / or boxes Figure 1 The function specified in one or more boxes.

[0221] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more flowcharts and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0222] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0223] Finally, it should be noted that the embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A radar target recognition model training method, characterized in that, The method includes: Multiple radar trajectory samples and their corresponding preset category labels are acquired. Each radar trajectory sample includes N frames of multi-dimensional features, including position features, velocity features, and attribute features. Each radar trajectory sample corresponds to a preset category label. A two-stage progressive sequence construction strategy is adopted to expand and construct multiple radar trajectory samples to obtain a training sample set and a preset training category label corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage. Each training sample is input into the radar target recognition model to be trained for target category recognition processing to obtain the predicted category label corresponding to each training sample; wherein, the radar target recognition model to be trained includes a position encoding model for processing the position feature sequence of each training sample, a velocity encoding model for processing the velocity and attribute feature sequence of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the output features of the velocity encoding model and fusing the output features of the two models, and a classification model for classification processing; Based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, the radar target recognition model to be trained is trained to obtain a trained radar target recognition model.

2. The radar target recognition model training method according to claim 1, characterized in that, The two-stage progressive sequence construction strategy is used to expand and construct multiple radar trajectory samples to obtain a training sample set and a preset training category label corresponding to each training sample, including: Iterate through each radar trajectory sample in the plurality of radar trajectory samples; When traversing any radar trajectory sample, in the initial expansion phase, based on the current radar trajectory sample, the number of preceding target frames with several dimensions of features are sequentially acquired, gradually accumulating information from the target entering the radar detection range, to obtain a number of prefix subsequence samples of length from 1 to the target number, as shown in the following formula: ； in, This represents the k-th prefix subsequence sample. Let W represent the features of the k-th frame, W represent the fixed window length, and N represent the number of valid frames for the trajectory. This represents the minimum value of the fixed window length and the number of valid frames for the trajectory, and the number of targets is determined based on the minimum value of the fixed window length and the number of valid frames; The number of target prefix subsequence samples inherit the preset category label of the current radar trajectory, thus obtaining the number of target preset training category labels corresponding to the number of target prefix subsequence samples; When the number of valid frames of the trajectory is greater than the fixed window length, the system enters the fixed window sliding phase. A sliding window of fixed length slides sequentially over the current radar trajectory sample with a preset step size, generating several sliding sequence samples of equal length and fixed window length, as shown in the following formula: ； in, This represents the j-th sliding sequence sample; Represents several dimensions of features in the j-th frame; Several equal-length sliding sequence samples inherit the preset category label of the current radar trajectory, thus obtaining several preset training category labels corresponding to the several equal-length sliding sequence samples; The target number of prefix subsequence samples with preset training category labels and several equal-length sliding sequence samples are merged to form the trajectory training sample set corresponding to the current radar trajectory sample and the preset training category label corresponding to each trajectory training sample; Multiple training sample sets corresponding to multiple radar trajectory samples are aggregated into a complete training set, and a random undersampling strategy is used to balance the samples to obtain the training sample set and the preset training category label corresponding to each training sample.

3. The radar target recognition model training method according to claim 1, characterized in that, The step of inputting each training sample into the radar target recognition model to be trained for target category recognition processing, and obtaining the predicted category label corresponding to each training sample, includes: The position feature sequence of each training sample is input into the position encoding model for position encoding processing to obtain the position encoding output feature of each training sample; The speed and attribute feature sequence of each training sample is input into the speed encoding model for speed encoding processing to obtain the global output features of speed and attributes of each training sample. The position encoding output features and the velocity and attribute global output features are input into the attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fusion features of each training sample; The fused features are input into the classification model for classification processing to obtain the predicted category label.

4. The radar target recognition model training method according to claim 3, characterized in that, The position encoding model is a two-layer bidirectional LSTM model. The step of inputting the position feature sequence of each training sample into the position encoding model for position encoding processing to obtain the position encoded output features of each training sample includes: The position feature sequence is input into the two-layer bidirectional LSTM model. For each time step, the forward hidden state and backward hidden state corresponding to each time step are determined according to each position feature in the position feature sequence. The forward hidden state and the backward hidden state are concatenated to obtain the position encoding feature corresponding to each time step, as shown in the following formula: ； in, This represents the positional encoding feature corresponding to time step t. This represents the forward hidden state corresponding to time step t. This represents the backward hidden state corresponding to time step t; The location encoding feature of the last time step in the location feature sequence is used as the location encoding output feature, as shown in the following formula: ； in, This indicates the location-encoded output feature. This represents the positional encoding feature of the last time step L, where L represents the length of the positional feature sequence.

5. The radar target recognition model training method according to claim 3, characterized in that, The velocity encoding model is a two-layer Transformer encoder. The velocity and attribute feature sequences of each training sample are input into the velocity encoding model for velocity encoding processing to obtain the global output features of the velocity and attributes of each training sample, including: The velocity and attribute feature sequence is input into the two-layer Transformer encoder. For each time step, each velocity and attribute feature is processed by the self-attention encoding of the two-layer Transformer encoder to obtain each encoded velocity and attribute feature. The average pooling process is applied to all encoded speed and attribute features to obtain the global output features of speed and attributes, as shown in the following formula: ； in, This indicates the speed and attribute output characteristics. Indicates Transformer encoder, Let L represent the velocity and attribute feature sequence, L represent the length of the velocity and attribute feature sequence, and t represent the t-th time step.

6. The radar target recognition model training method according to claim 3, characterized in that, The attention fusion model includes a first linear projection layer, a second linear projection layer, and a bilinear attention network; The step of inputting the position-encoded output features and the global output features of velocity and attributes into the attention fusion model for adaptive weight determination and feature weighted fusion processing to obtain the fused features of each training sample includes: The position-encoded output features and the velocity and attribute global output features are mapped to a common latent space dimension by the first linear projection layer and the second linear projection layer, respectively, to obtain position projection features and velocity and attribute projection features. The position projection feature and the velocity and attribute projection feature are stitched together to obtain the stitched projection feature. The attention weights of the stitched projection features are determined by the bilinear attention network to obtain the position branch weights corresponding to the position projection features and the velocity and attribute branch weights corresponding to the velocity and attribute projection features, as shown in the following formula: ； in, Indicates the branch weight at position. Indicates speed and attribute branch weight; Indicates the location projection features, Represents velocity and attribute projection characteristics. This indicates that the position projection features and velocity and attribute projection features are stitched together. Indicates the weights of the first layer. Indicates the output layer weights. This represents the first bias vector. This represents the second bias vector. Represents the hyperbolic tangent activation function; Used for normalization to ensure ,and ; Based on the position branch weight and the velocity and attribute branch weight, the position projection feature and the velocity and attribute projection feature are weighted and summed to generate a fused feature, as shown in the following formula: ； in, This indicates the fusion feature.

7. The radar target recognition model training method according to claim 1, characterized in that, The multiple radar trajectory samples and their corresponding multiple preset category labels are obtained using the following method: Acquire raw radar spot data for M frames; Based on the trajectory history and preset threshold of each frame of original radar spot data, the M frames of original radar spot data are filtered to obtain N frames of valid radar spot data. Extract 13-dimensional features from each frame of valid radar spot data. The 13-dimensional features of each frame include 6-dimensional position features and 7-dimensional velocity features. Based on the 13-dimensional features of the N frames of valid radar point data, multiple radar trajectories are generated. Generate category labels for each radar trajectory to obtain the preset category labels for each radar trajectory; The multiple radar trajectories and their corresponding preset category labels are determined as the multiple radar trajectory samples and their corresponding preset category labels.

8. A radar target identification method, characterized in that, The method includes: Obtain the radar trajectory to be identified; The radar trajectory to be identified is input into the radar target recognition model for target category recognition processing to obtain the target category label corresponding to the radar trajectory to be identified. The radar target recognition model is obtained based on the radar target recognition model training method according to any one of claims 1 to 7.

9. A radar target recognition model training device, characterized in that, The device includes: The first information acquisition module is configured to acquire multiple radar trajectory samples and corresponding multiple preset category labels. Each radar trajectory sample includes N frames of multi-dimensional features, including position features, velocity and attribute features; each radar trajectory sample corresponds to a preset category label. The sample construction module is configured to perform sample expansion construction processing on multiple radar trajectory samples using a two-stage progressive sequence construction strategy to obtain a training sample set and a preset training category label corresponding to each training sample. The two stages include an initial expansion stage and a fixed window sliding stage. The category recognition processing module is configured to input each training sample into the radar target recognition model to be trained for target category recognition processing, and obtain the predicted category label corresponding to each training sample; wherein, the radar target recognition model to be trained includes a position encoding model for processing the position features of each training sample, a velocity encoding model for processing the velocity and attribute features of each training sample, an attention fusion model for adaptively determining the weights of the output features of the position encoding model and the output features of the velocity encoding model and fusing the output features of the two models, and a classification model for performing classification processing; The model training module is configured to train the radar target recognition model to be trained based on the predicted category labels and preset training category labels corresponding to each training sample in the training sample set, so as to obtain a trained radar target recognition model.

10. A radar target identification device, characterized in that, The device includes: The second information acquisition module is configured to acquire the radar trajectory to be identified. The category information determination module is configured to input the radar trajectory to be identified into the radar target recognition model for target category recognition processing, and obtain the target category label corresponding to the radar trajectory to be identified. The radar target recognition model is obtained based on the radar target recognition model training method according to any one of claims 1 to 7.

Images