Adaptive radar target classification method and device, equipment and storage medium

Abstract

The invention discloses an adaptive radar target classification method and device, equipment and a storage medium. The method comprises the steps of determining a local signal-to-noise ratio and the type of feature sparsity data according to radar echo data; inputting the radar echo data into a pre-constructed target classification model, and selecting a processing branch according to a preset signal-to-noise ratio threshold range of the local signal-to-noise ratio and the type of the feature sparsity data; detecting a target category via an incremental adversarial learning module of the target classification model; when the target category is an unknown category, acquiring an unknown category confidence coefficient and a target existence probability, and determining a confidence coefficient state according to the unknown category confidence coefficient and the target existence probability; and acquiring a continuous frame number of each confidence coefficient state, when the continuous frame number of the high confidence coefficient state is detected to be greater than a preset manual annotation judgment threshold, triggering manual annotation, and inputting a target category annotation result after manual annotation into the target classification model as a training set. The method is used for realizing radar target classification.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of radar, in particular to a self-adaptive radar target classification method, device, equipment and storage medium. BACKGROUND

[0002] As a device for detecting targets by emitting electromagnetic waves and receiving echoes, radar has a wide range of applications in military, transportation, security and other fields. With the complication of application scenarios, higher requirements are put forward for the accuracy of radar target classification and the reliability of trajectory prediction.

[0003] Traditional radar target classification methods usually use fixed feature extraction and classification models. For example, traditional methods usually use manually designed features (such as RCS, Doppler shift, spectral width, etc.) combined with SVM, random forest and other shallow classifiers for identification. The model structure is fixed and difficult to adapt to different environmental interference and target characteristic changes. In complex scenarios such as low signal-to-noise ratio and sparse features, the classification performance is greatly reduced. At the same time, for unknown target categories, the traditional method lacks effective detection and learning mechanism, and cannot realize the autonomous update of the model.

[0004] In terms of trajectory prediction, existing technologies mostly use Kalman filter, particle filter or LSTM time series model to independently predict the trajectory of a single target, without considering the interaction between targets (such as vehicle following and pedestrian obstacle avoidance) and the spatio-temporal correlation of group motion patterns, resulting in large prediction errors in complex scenarios. In addition, most methods do not have self-learning ability and are difficult to dynamically optimize model parameters according to actual data, and have poor adaptability when facing new scenarios or new types of targets.

[0005] Therefore, it is of great significance to develop a radar target processing method that can adapt to complex environments, has self-learning ability, and improves the accuracy of target classification and the reliability of trajectory prediction. SUMMARY

[0006] The purpose of the present application is to at least provide a radar target classification and trajectory prediction method, device, equipment and medium, which can at least solve the technical problems of inaccurate target classification and unreliable trajectory prediction in the prior art.

[0007] To solve the above technical problems, at least one embodiment of the present application provides a self-adaptive radar target classification method, comprising: determining target point cloud feature data and original spectrum feature data of a target object according to radar echo data; determining a local signal-to-noise ratio based on the original spectrum feature data, and determining a type of feature sparsity data according to the target point cloud feature data and a preset density threshold, the type of feature sparsity data including sparse data and dense data; input the radar echo data into a pre-constructed target classification model, and select a processing branch according to a preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and a type of feature sparsity data; detect a target category via an incremental adversarial learning module of the target classification model based on feature data extracted by the processing branch, the target category including an unknown category and a known category; when the target category is the unknown category, obtain an unknown category confidence and a target existence probability, and determine a confidence state according to the unknown category confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state; obtain a continuous frame number of each confidence state, trigger manual labeling when the continuous frame number of the high confidence state is detected to be greater than a preset manual labeling determination threshold, and input the radar echo data and a target category labeling result labeled manually into the target classification model as a training set.

[0008] At least one embodiment of the present application also provides a self-adaptive radar target classification device, which includes: a feature data determination module configured to determine target point cloud feature data and original spectrum feature data of a target object according to radar echo data; a data type determination module configured to determine a local signal-to-noise ratio based on the original spectrum feature data, and determine a type of feature sparsity data according to the target point cloud feature data and a preset density threshold, the type of feature sparsity data including sparse data and dense data; a processing branch selection module configured to input the radar echo data into a pre-constructed target classification model, and select a processing branch according to a preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and a type of feature sparsity data; a target classification module configured to detect a target category via an incremental adversarial learning module of the target classification model based on feature data extracted by the processing branch, the target category including an unknown category and a known category; a confidence state determination module configured to, when the target category is the unknown category, obtain an unknown category confidence and a target existence probability, and determine a confidence state according to the unknown category confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state; a discrimination module configured to obtain a continuous frame number of each confidence state, trigger manual labeling when the continuous frame number of the high confidence state is detected to be greater than a preset manual labeling determination threshold, and input the radar echo data and a target category labeling result labeled manually into the target classification model as a training set.

[0009] At least one embodiment of the present application also provides an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the adaptive radar target classification method described above.

[0010] At least one embodiment of the present application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the adaptive radar target classification method described above.

[0011] The adaptive radar target classification method, device, electronic device and computer-readable storage medium provided by the embodiments of the present application can realize the update of the classification layer and the trajectory prediction through the calculation and processing of the local signal-to-noise ratio and the feature sparsity data, the selection of the processing branch, and the obtaining of the confidence state, can adapt to complex environments, have self-learning ability, and improve the target classification accuracy and the trajectory prediction reliability.

[0012] In some optional embodiments, the step of determining the target point cloud feature data and the original spectrum feature data of the target object according to the radar echo data comprises: performing three-dimensional point cloud clustering on the radar echo data to obtain target point cloud feature data, the target point cloud feature data comprising target point quantity and three-dimensional coordinate data corresponding to each target point; extracting original spectrum feature data based on the radar echo data, the original spectrum feature data comprising target echo power and local noise power.

[0013] In some optional embodiments, the step of determining the local signal-to-noise ratio based on the original spectrum feature data, and determining the type of feature sparsity data according to the target point cloud feature data and a preset density threshold, the type of feature sparsity data comprising sparse data and dense data, comprises: calculating the local signal-to-noise ratio according to the target echo power and the local noise power; statistically obtaining point cloud density distribution data by dividing a spatial voxel grid according to the target point quantity and the three-dimensional coordinate data corresponding to each target point; when the point cloud density distribution data is less than the preset density threshold, determining that the type of feature sparsity data is sparse data; when the point cloud density distribution data is greater than or equal to the preset density threshold, determining that the type of feature sparsity data is dense data.

[0014] In some optional embodiments, the preset signal-to-noise ratio threshold range includes a first range greater than a first preset threshold, a second range less than the first preset threshold and greater than a second preset threshold, and a third range less than the second preset threshold, wherein the first preset threshold is greater than the second preset threshold, and the step of inputting the radar echo data into the pre-constructed target classification model according to the preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and the type of feature sparsity data to select a processing branch includes: inputting the radar echo data into the pre-constructed target classification model; when the local signal-to-noise ratio is in the first range and the type of feature sparsity data is dense data, determining the processing branch as a lightweight network; when the local signal-to-noise ratio is in the second range and the type of feature sparsity data is dense data or sparse data, or the local signal-to-noise ratio is in the first range and the type of feature sparsity data is sparse data, determining the processing branch as an attention-enhanced network; when the local signal-to-noise ratio is in the third range and the type of feature sparsity data is dense data or sparse data, determining the processing branch as an anti-interference deep network.

[0015] In some optional embodiments, when the target category is an unknown category, an unknown category confidence and a target existence probability are obtained, and a confidence state is determined according to the unknown category confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state, and the step includes: when the target category is an unknown category, an unknown category confidence and a target existence probability are obtained; when the unknown category confidence is greater than or equal to a preset confidence threshold and the target existence probability is greater than a preset target existence threshold, the confidence state is determined as a high confidence state; when the unknown category confidence is greater than or equal to a preset confidence threshold and the target existence probability is less than or equal to a preset target existence threshold, the confidence state is determined as a low confidence state; when the unknown category confidence is less than a preset confidence threshold, the confidence state is determined as a low confidence state.

[0016] In some optional embodiments, the method further includes: clustering historical trajectory data of a target object and extracting a typical behavior pattern as a pattern feature vector to generate a pattern library; obtaining a current target motion feature vector and calculating a cosine similarity between the current target motion feature vector and a pattern feature vector in the pattern library; When the pre-similarity is greater than or equal to a preset trajectory calling threshold, a trajectory parameter corresponding to the mode feature vector of the current target motion feature vector is used to perform trajectory prediction, and a first prediction value is generated.

[0017] In some optional embodiments, the method further includes: performing trajectory prediction on the target object by using a pre-constructed trajectory prediction model to obtain a second prediction value; fusing the first prediction value and the second prediction value to obtain a fused prediction value; An expression of the fused prediction value is: ; wherein, the fused prediction value is, the first prediction value is, the second prediction value is; the preset mode multiplexing weight is. BRIEF DESCRIPTION OF DRAWINGS

[0018] One or more embodiments are illustrated by way of example in the figures that form a part of this patent specification. These example are not, and are not intended to be limiting, of the embodiments.

[0019] Figure 1 a flowchart of an adaptive radar target classification method provided by an embodiment of the present application; Figure 2 a flowchart of an adaptive radar target classification method provided by another embodiment of the present application; Figure 3 a flowchart of obtaining target point cloud feature data and original spectrum feature data provided by another embodiment of the present application; Figure 4 a flowchart of obtaining local signal-to-noise ratio and feature sparsity data provided by another embodiment of the present application; Figure 5 a schematic diagram of an adaptive radar target classification device provided by another embodiment of the present application; Figure 6 a structural schematic diagram of an electronic device provided by another embodiment of the present application.

[0020] In the drawings, the same components have the same reference numbers, and the drawings are not drawn according to actual proportions. DETAILED DESCRIPTION

[0021] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below with reference to the drawings. However, those skilled in the art can understand that, in the embodiments of the present application, many technical details are proposed in order to make the readers better understand the present application. However, the technical solutions claimed by the present application can be implemented even if there are no such technical details and various changes and modifications based on the following embodiments. The division of the following embodiments is for the convenience of description, and should not constitute any limitation on the specific implementation modes of the present application. The embodiments can be combined and referenced with each other without contradiction.

[0022] It should be noted that the terms "first", "second", and the like in the description, claims, and drawings of the present disclosure are used to distinguish like objects, and do not necessarily have to be described in a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0023] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that described herein.

[0024] Embodiment one: Figure 1 A flowchart of an adaptive radar target classification method provided by an embodiment of the present disclosure. As shown in Figure 1 An adaptive radar target classification method includes: Step 110, determining target point cloud feature data and original spectrum feature data of the target object according to radar echo data.

[0025] Specifically, the electromagnetic wave emitted by the radar encounters the target object, is reflected, scattered and returned to the radar receiver, and the signal name of the radar echo is obtained. By analyzing the echo, the identification and classification of the target object and the trajectory prediction can be realized. After receiving the radar echo data, a three-dimensional point cloud clustering algorithm is used to process the echo data to obtain the target point cloud feature. At the same time, the echo signal is analyzed to obtain the original spectrum feature data.

[0026] In step 120, a local signal-to-noise ratio is determined based on the original spectral feature data, and a type of feature sparsity data is determined according to the target point cloud feature data and a preset density threshold, the type of feature sparsity data including sparse data and dense data.

[0027] Specifically, the local signal-to-noise ratio and the feature sparsity data can be calculated by the target point cloud feature data and the original spectral feature data.

[0028] Further, the local signal-to-noise ratio (SNR) is a ratio of a target echo power to a local noise power, which is obtained by calculating a power difference between a target region and a background region by a sliding window; the feature sparsity can be determined by a point cloud density distribution. The preset density threshold is obtained by analyzing historical data according to an application scenario, and can be adjusted as needed by a user.

[0029] In step 130, the radar echo data is input into a pre-constructed target classification model, and a processing branch is selected according to a preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and the type of feature sparsity data.

[0030] Specifically, the preset signal-to-noise ratio threshold range is obtained by analyzing historical data, and is a threshold for distinguishing quality signals according to signal-to-noise ratio values. The processing branch includes at least two kinds, and one of the processing branches is selected according to the preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and the type of feature sparsity data.

[0031] In one example, the processing branch includes a lightweight network, an anti-interference deep network, and an attention-enhanced network. The lightweight network greatly reduces the amount of calculation and the amount of parameters under the premise of guaranteeing basic performance, and is suitable for a high signal-to-noise ratio (SNR) and feature dense scene; the anti-interference deep network maintains robustness in a low signal-to-noise ratio environment, effectively suppresses noise and interference, and is suitable for a strong interference scene; the attention-enhanced network focuses on important features and ignores redundant information in a feature sparse but key information existing scene, and improves the sensitivity to local key features.

[0032] In step 140, a target category is detected by an incremental adversarial learning module of the target classification model based on feature data extracted by the processing branch, the target category including an unknown category and a known category.

[0033] Specifically, the incremental adversarial learning module includes a generator and a discriminator. The generator adopts a DCGAN architecture, the input is random noise and radar parameters, and the output is pseudo point cloud and spectral feature. By combining spectral width and phase change rate, the physical constraint loss is optimized to ensure that the synthesized data meets: , wherein is a Doppler shift, is the relative radial velocity, λ is the radar wavelength; The discriminator is a CNN-Transformer hybrid architecture, the first half extracts features through CNN, and the second half outputs the known class probability distribution and unknown class confidence. A joint loss function is used, and the centroid coordinates and point cloud density distribution in the target point cloud features are used to assist in improving accuracy. The known class refers to the detected target category, and the unknown class refers to the undetected target category. Confidence is a core quantitative indicator for measuring the reliability of the model's classification, detection or prediction results, which can be obtained through the incremental adversarial learning module. Step 150, when the target category is an unknown class, obtaining an unknown class confidence and a target existence probability, determining a confidence state according to the unknown class confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state.

[0034] Specifically, when the detected target category is an unknown class, the unknown class confidence is obtained by processing. The unknown class confidence and the target existence probability are compared with the pre-set confidence state evaluation threshold to obtain the confidence state, including high confidence or low confidence.

[0035] Step 160, obtaining the number of consecutive frames of each confidence state, when the number of consecutive frames of the high confidence state is greater than the pre-set artificial labeling decision threshold, triggering artificial labeling, and inputting the radar echo data and the target category labeling result after artificial labeling into the target classification model as a training set.

[0036] Specifically, the number of consecutive frames of the confidence state is obtained, including the number of frames of the high confidence state and the number of frames of the low confidence state. The number of frames of the high confidence state is compared with the pre-set artificial labeling decision threshold; if the number of frames of the high confidence state is greater than the pre-set artificial labeling decision threshold, the artificial labeling process is triggered; after artificial labeling is completed, the artificial labeling result is trained by small batch gradient descent method, and the new category is synchronized.

[0037] Further, the core purpose of setting the number of consecutive frames is to filter transient interference and ensure the persistence and stability of the target. If only one frame or a few frames detect an unknown target, it may be a transient electromagnetic interference (such as a burst of clutter), a target transient characteristic change (such as point cloud deformation caused by rapid turning), or a casual misjudgment of the model. When N consecutive frames and above (each frame interval is usually 0.1-1 seconds, adjusted according to the radar frame rate) all meet the "high confidence" condition, it can be determined that the target is a new class with stable physical properties and motion characteristics (such as new types of unmanned aerial vehicles, special vehicles, etc.), and manual labeling is triggered at this time. In this embodiment, the preset value of the manual labeling determination threshold N (≥5) can be adjusted according to the application scenario: in a high-speed dynamic scene (such as a highway), N can be set to 5-8 frames; in a low-speed static scene (such as a factory monitoring), N can be appropriately increased to 10 frames to further reduce the probability of false triggering.

[0038] In this embodiment, by obtaining radar echo data, target point cloud and original spectrum feature data are extracted, local signal-to-noise ratio and feature sparsity data are calculated, a lightweight, anti-interference deep or attention enhancement network branch is selected accordingly, a target class is detected through an incremental adversarial learning module, unknown class confidence and target existence probability are obtained, confidence state is obtained through threshold comparison, and the number of frames is counted to determine manual labeling and update the classification layer. Thus, through the calculation of local signal-to-noise ratio and feature sparsity data, the selection of processing branches, and the acquisition of confidence state, the update of classification layer and target classification are realized.

[0039] In some embodiments, the step of determining target point cloud feature data and original spectrum feature data of the target object according to the radar echo data comprises: performing three-dimensional point cloud clustering on the radar echo data to obtain target point cloud feature data, wherein the target point cloud feature data comprises target point quantity and three-dimensional coordinate data corresponding to each target point; extracting original spectrum feature data based on the radar echo data, wherein the original spectrum feature data comprises target echo power and local noise power.

[0040] Specifically, after receiving the radar echo data, a three-dimensional point cloud clustering algorithm is used to process the echo data to obtain target point cloud features. The target point cloud feature data includes the number of target points and the three-dimensional coordinate data corresponding to each target point. Based on the target point cloud feature data, the target contour, centroid coordinates, size parameters (length / width / height), and point cloud density distribution can be determined. At the same time, the radar echo data is subjected to spectral analysis to extract original spectral feature data such as Doppler shift, echo amplitude spectrum peak value, spectral width, and phase change rate. The original spectral features are used to reflect the target motion state and physical properties, including target echo power, local noise power, spectral width, and phase change rate. The target echo power is the electromagnetic wave power received by the radar receiving end and reflected by the target, which is a key physical quantity for measuring the strength of the target reflection signal. The local noise power is the power of pure noise components in the local area of the target in the radar received signal, which is a key parameter for measuring the strength of the electromagnetic interference around the target. The spectral width is a physical quantity that describes the frequency distribution range of the radar echo signal, specifically the frequency span of the energy concentration region in the echo signal power spectrum, with units of hertz (Hz). The phase change rate is the rate of change of the phase of the radar echo signal with time, with units of radians per second (rad / s) or hertz (Hz, which is directly related to the frequency change).

[0041] In some embodiments, the step of determining the local signal-to-noise ratio based on the original spectral feature data and determining the type of feature sparsity data based on the target point cloud feature data and a preset density threshold includes: calculating the local signal-to-noise ratio based on the target echo power and the local noise power; obtaining point cloud density distribution data by dividing the spatial voxel grid based on the number of target points and the three-dimensional coordinate data corresponding to each target point; when the point cloud density distribution data is less than the preset density threshold, determining that the type of feature sparsity data is sparse data; when the point cloud density distribution data is greater than or equal to the preset density threshold, determining that the type of feature sparsity data is dense data.

[0042] Specifically, the local signal-to-noise ratio (SNR) is the ratio of the target echo power to the local noise power, which is obtained by calculating the power difference between the target area and the background area using a sliding window. The feature sparsity can be determined by the point cloud density distribution.

[0043] The calculation formula of the local signal-to-noise ratio is: , wherein, is the local signal-to-noise ratio, is the target echo power, is a local noise power.

[0044] The point cloud density distribution data refers to quantization data of a spatial distribution density of points in a target point cloud collected by a radar or laser sensor and a distribution rule of the points in a three-dimensional space, and is one of core features for describing a target geometric shape and a spatial structure; the preset density threshold is obtained according to an application scene and historical data analysis, and can be appropriately adjusted according to user needs; when the point cloud density distribution data is less than the preset density threshold, the type of the feature sparsity data is determined as sparse data; when the point cloud density distribution data is greater than or equal to the preset density threshold, the type of the feature sparsity data is determined as dense data.

[0045] In some embodiments, the preset signal-to-noise ratio threshold range includes a first range greater than a first preset threshold, a second range less than the first preset threshold and greater than a second preset threshold, and a third range less than the second preset threshold, wherein the first preset threshold is greater than the second preset threshold, and the step of inputting the radar echo data into the pre-constructed target classification model, and selecting a processing branch according to the preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and the type of the feature sparsity data includes: inputting the radar echo data into the pre-constructed target classification model; when the local signal-to-noise ratio is in the first range and the type of the feature sparsity data is dense data, determining that the processing branch is a lightweight network; when the local signal-to-noise ratio is in the second range and the type of the feature sparsity data is dense data or sparse data, or the local signal-to-noise ratio is in the first range and the type of the feature sparsity data is sparse data, determining that the processing branch is an attention-enhanced network; when the local signal-to-noise ratio is in the third range and the type of the feature sparsity data is dense data or sparse data, determining that the processing branch is an anti-interference deep network. Specifically, the preset signal-to-noise ratio threshold is obtained by analyzing historical data, and is a threshold for distinguishing quality signals according to signal-to-noise ratio values; in this embodiment, the first threshold is a high-quality signal threshold, the second threshold is a low-quality signal threshold, the first threshold is greater than the second threshold, and a medium-quality signal is located between the first threshold and the second threshold; the lightweight network greatly reduces the amount of calculation and the amount of parameters under the premise of guaranteeing basic performance, and is suitable for a high signal-to-noise ratio (SNR) and feature dense scene; the anti-interference deep network maintains robustness in a low signal-to-noise ratio environment, effectively suppresses noise and interference, and is suitable for a strong interference scene; the attention-enhanced network focuses on important features and ignores redundant information in a feature sparse but key information existing scene, and improves sensitivity to local key features.

[0046] The local signal-to-noise ratio is compared with the first threshold value and the second threshold value, and a processing branch is obtained by combining the feature sparsity data; if the local signal-to-noise ratio is greater than the first threshold value and the feature sparsity data is dense data, the processing branch is a lightweight network; if the local signal-to-noise ratio is less than the second threshold value, the processing branch is an anti-interference deep network; otherwise, the processing branch is an attention-enhanced network.

[0047] In some embodiments, when the target category is an unknown category, an unknown category confidence and a target existence probability are obtained, and a confidence state is determined according to the unknown category confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state, the step of determining the confidence state including: When the target category is an unknown category, an unknown category confidence and a target existence probability are obtained. When the unknown category confidence is greater than or equal to a preset confidence threshold value and the target existence probability is greater than a preset target existence threshold value, the confidence state is determined to be a high confidence state. When the unknown category confidence is greater than or equal to a preset confidence threshold value and the target existence probability is less than or equal to a preset target existence threshold value, the confidence state is determined to be a low confidence state. When the unknown category confidence is less than a preset confidence threshold value, the confidence state is determined to be a low confidence state.

[0048] Specifically, the target category is detected by an incremental adversarial learning module, including known categories and unknown categories; if it is an unknown category, an unknown category confidence is obtained; the unknown category confidence and a target existence probability are compared with preset confidence state evaluation threshold values to obtain a confidence state, including a high confidence or a low confidence; the preset confidence state evaluation threshold values include a confidence threshold value and a target existence threshold value; if the unknown category confidence is greater than or equal to the confidence threshold value and the target existence probability is greater than the target existence threshold value, the confidence state is a high confidence; otherwise, it is a low confidence.

[0049] In some embodiments, the method further includes: The historical trajectory data of the target object is clustered, and a typical behavior pattern is extracted and stored as a pattern feature vector to generate a pattern library. The current target motion feature vector is obtained, and a cosine similarity between the current target motion feature vector and the pattern feature vectors in the pattern library is calculated. When the preset similarity is greater than or equal to a preset trajectory calling threshold value, a trajectory parameter based on the pattern feature vector corresponding to the current target motion feature vector is used to perform trajectory prediction to generate a first prediction value.

[0050] Specifically, the historical trajectory clustering analysis refers to clustering the target motion trajectory data (such as the position-time sequence of an airplane or a ship) monitored by the radar in the past, and classifying the trajectories with similar motion rules into the same class. The representative motion pattern of each trajectory after clustering is an abstraction and simplification of the trajectories in the same class. The pattern feature vector converts the typical behavior pattern into a quantifiable and calculable vector form for subsequent similarity comparison. The index for measuring the directional similarity of two vectors (the current target motion feature vector and the pattern feature vector in the pattern library) is calculated by the formula: ; The cosine similarity is The pattern feature vector is The current target motion feature vector is s; the value range of s is [-1, 1], and the closer to 1, the more similar the pattern. The preset trajectory calling threshold controls the strictness of pattern matching, which is initially set by artificial or experience and then dynamically optimized by the self-learning module.

[0051] Further, the historical trajectory of the monitored target is clustered and analyzed, the typical behavior pattern is extracted and stored as a pattern feature vector, and a pattern library is generated. The current target motion feature vector is obtained, and the cosine similarity with the pattern feature vector is calculated. If the cosine similarity is greater than or equal to the preset trajectory calling threshold, the trajectory parameters of the corresponding typical behavior pattern are called for prediction.

[0052] In some embodiments, the method further comprises: predicting the trajectory of the target object by a pre-constructed trajectory prediction model to obtain a second prediction value; fusing the first prediction value and the second prediction value to obtain a fused prediction value; The expression of the fused prediction value is: ; Wherein, The fused prediction value is The first prediction value is The second prediction value is The preset mode reuse weight is

[0053] Specifically, the fused prediction value is the final prediction result obtained by the formula, which is a weighted fusion of the two prediction values, combining the advantages of the two prediction methods. In stable scenarios, the historical mode is used to ensure efficiency, and in mutation scenarios, the neural network is used to ensure accuracy. The mode reuse weight a is a parameter (a ∈ [0, 1]) that adjusts the fusion ratio of the first prediction value and the second prediction value; when a→1, it depends more on the historical mode; when a→0, it depends more on the neural network (suitable for motion mutation scenarios).

[0054] In this embodiment, by acquiring radar echo data, extracting target point cloud and original spectral features, calculating local signal-to-noise ratio and feature sparsity data, selecting lightweight, anti-interference deep or attention enhancement network branch, detecting target category through incremental adversarial learning module, obtaining unknown class confidence and target existence probability, comparing the threshold to obtain the confidence state, and counting the number of consecutive frames to determine the artificial labeling and update the classification layer, the classification layer is updated and the trajectory prediction is realized through the calculation of local signal-to-noise ratio and feature sparsity data, the selection of processing branch and the acquisition of confidence state.

[0055] Embodiment two: Another embodiment of the present application relates to an adaptive radar target classification and trajectory prediction method, as shown in Figure 2 The method comprises the following steps: S11, acquiring radar echo data, and extracting target point cloud feature data and original spectral feature data; S12, calculating local signal-to-noise ratio and feature sparsity data according to target point cloud feature data and original spectral feature data; S13, selecting processing branch according to local signal-to-noise ratio and feature sparsity data, including lightweight network, anti-interference deep network and attention enhancement network; S14, detecting target category through incremental adversarial learning module, and processing to obtain unknown class confidence and target existence probability, comparing the threshold to obtain the confidence state; S15, counting the number of consecutive frames of the confidence state, and determining the artificial labeling and updating the classification layer according to the number of confidence state frames.

[0056] It should be noted that the signal of the electromagnetic wave emitted by the radar and reflected, scattered and returned to the radar receiver after encountering the target object is called radar echo, and through analysis of the echo, identification and classification of the target object and trajectory prediction can be realized. After receiving the radar echo data, the echo data is processed by using a three-dimensional point cloud clustering algorithm to obtain target point cloud features, including target contour, centroid coordinates, size parameters (length / width / height) and point cloud density distribution; at the same time, the echo signal is subjected to spectrum analysis to extract original spectrum features such as Doppler shift, echo amplitude spectrum peak value, spectrum width and phase change rate; through the target point cloud feature data and the original spectrum feature data, local signal-to-noise ratio and feature sparsity data can be calculated and obtained, and according to the local signal-to-noise ratio and the feature sparsity, a dynamic selection processing branch can be realized, so that different data adopts different processing branches, which is beneficial to improve the data processing efficiency and quality; through the echo data detection of the incremental adversarial learning module, a target category is obtained, and after processing, an unknown class confidence target existence probability is obtained; after threshold comparison of the unknown class confidence and the target existence probability, the corresponding confidence state can be obtained; the confidence state includes high confidence or low confidence, the number of consecutive frames of the confidence state is counted, and then threshold comparison is performed to obtain whether the artificial labeling process is triggered, and then the classification layer is updated according to the artificial labeling result.

[0057] Please refer to Figure 3 , Figure 3 is a flowchart of a method for obtaining target point cloud feature data and original spectrum feature data provided by an embodiment of the present application. According to the embodiment of the present application, the radar echo data is obtained, and the target point cloud feature data and the original spectrum feature data are extracted, specifically including: S21, extracting target point cloud feature data, including target point quantity and three-dimensional coordinate data corresponding to each target point; S22, extracting original spectrum feature data, including target echo power, local noise power, spectrum width and phase change rate.

[0058] It should be noted that the target point cloud feature refers to the target contour, centroid coordinates, size parameters (length / width / height) and point cloud density distribution obtained by three-dimensional point cloud clustering. In this embodiment, the target point quantity and the three-dimensional coordinate data corresponding to the target point are obtained. The original spectrum feature is used to reflect the target motion state and physical property, including target echo power, local noise power, spectrum width and phase change rate. The target echo power is the electromagnetic wave power reflected by the target received by the radar receiving end, which is the core physical quantity for measuring the strength of the target reflection signal. The local noise power refers to the power of pure noise components in the local area of the target in the radar received signal, which is a key parameter for measuring the electromagnetic interference strength of the environment around the target. The spectrum width is a physical quantity for describing the frequency distribution range of the radar echo signal, specifically referring to the frequency span of the energy concentration area in the echo signal power spectrum, with the unit of hertz (Hz). The phase change rate refers to the rate of change of the phase of the radar echo signal with time, with the unit of radian / second (rad / s) or hertz (Hz, which is directly related to the frequency change).

[0059] Please refer to Figure 4 , Figure 4 is a flowchart of obtaining local signal-to-noise ratio and feature sparsity data provided by the adaptive radar target classification and trajectory prediction method. According to the embodiment of the present application, the local signal-to-noise ratio and feature sparsity data are calculated according to the target point cloud feature data and the original spectrum feature data, specifically including: S31, the local signal-to-noise ratio is calculated according to the target echo power and the local noise power; S32, the point cloud density distribution data is obtained by dividing the spatial voxel grid according to the target point quantity and the three-dimensional coordinate data; S33, the point cloud density distribution data is compared with the preset density threshold to obtain the feature sparsity data; S34, if the point cloud density distribution data is less than the preset density threshold, the feature sparsity data is sparse data; S35, if the point cloud density distribution data is greater than or equal to the preset density threshold, the feature sparsity data is dense data.

[0060] It should be noted that the local signal-to-noise ratio (SNR) is the ratio of the target echo power to the local noise power, which is obtained by calculating the power difference between the target area and the background area through a sliding window. The feature sparsity can be judged by the point cloud density distribution. The calculation formula of the local signal-to-noise ratio is: , wherein is the local signal-to-noise ratio, is the target echo power, The local noise power is a partial noise power; the point cloud density distribution data refers to quantitative data of a spatial distribution density of points in a target point cloud collected by a radar or laser sensor and a distribution rule of the points in a three-dimensional space, and is one of core features for describing a target geometric shape and a spatial structure; and the preset density threshold is obtained according to an application scene and historical data, and can be appropriately adjusted according to user needs.

[0061] According to the embodiment of the present application, the processing branch selected according to the local signal-to-noise ratio and the feature sparsity data includes a lightweight network, an anti-interference deep network, and an attention-enhanced network, and specifically includes: A preset signal-to-noise ratio threshold is obtained, and a first threshold and a second threshold are extracted, and the first threshold is greater than the second threshold; The local signal-to-noise ratio is compared with the first threshold and the second threshold, and the feature sparsity data is collected to obtain the processing branch; If the local signal-to-noise ratio is greater than the first threshold and the feature sparsity data is dense data, the processing branch is a lightweight network; If the local signal-to-noise ratio is less than the second threshold, the processing branch is an anti-interference deep network; Otherwise, the processing branch is an attention-enhanced network.

[0062] It should be noted that the preset signal-to-noise ratio threshold is obtained by analyzing historical data, and is a threshold for distinguishing quality signals according to signal-to-noise ratio values. In the present embodiment, the first threshold is a high-quality signal threshold, the second threshold is a low-quality signal threshold, and a medium-quality signal is located between the first threshold and the second threshold. The lightweight network greatly reduces the amount of calculation and the amount of parameters under the premise of ensuring basic performance, and is suitable for high signal-to-noise ratio (SNR) and feature dense scenes. The anti-interference deep network maintains robustness in a low signal-to-noise ratio environment, effectively suppresses noise and interference, and is suitable for strong interference scenes. The attention-enhanced network focuses on important features and ignores redundant information in scenes where feature sparsity but key information exists, thereby improving the sensitivity to local key features. Otherwise, the processing branch is an attention-enhanced network, i.e., in addition to the lightweight network and the anti-interference deep network mentioned above, it belongs to the attention-enhanced network.

[0063] According to the embodiment of the present application, the target class is detected by the incremental adversarial learning module, and unknown class confidence and target existence probability are obtained by processing, and the confidence state is obtained by comparing the unknown class confidence and the target existence probability with preset confidence state evaluation thresholds, and specifically includes: The target class is detected by the incremental adversarial learning module, including known classes and unknown classes; If it is an unknown class, the unknown class confidence is obtained by processing; The unknown class confidence and the target existence probability are compared with the preset confidence state evaluation threshold to obtain the confidence state, including high confidence or low confidence. The preset confidence state evaluation threshold value includes a confidence threshold value and a target existence threshold value; If the unknown class confidence is greater than or equal to the confidence threshold value, and the target existence probability is greater than the target existence threshold value, the confidence state is high confidence; Otherwise, the confidence state is low confidence.

[0064] It should be noted that the incremental adversarial learning module includes a generator and a discriminator. The generator adopts a DCGAN architecture, the input is random noise and radar parameters, and the output is a pseudo point cloud and a spectral feature. Combined with the spectral width and the phase change rate, the physical constraint loss is optimized to ensure that the synthesized data meets: wherein is the Doppler shift, is the relative radial velocity, and λ is the radar wavelength; The discriminator is a CNN-Transformer hybrid architecture. The front half extracts features through CNN, and the back half outputs the known class probability distribution and the unknown class confidence. A joint loss function is used, and the centroid coordinates and point cloud density distribution in the target point cloud feature are used to assist in improving the accuracy. The known class refers to the detected target category, and the unknown class refers to the undetected target category. The confidence is a core quantitative index for measuring the reliability of the model's classification, detection or prediction results, which can be obtained through the incremental adversarial learning module.

[0065] According to the embodiment of the application, the confidence state continuous frame number is obtained by statistics, and the classification layer is artificially labeled and updated according to the confidence state frame number. Specifically, it includes: The confidence state continuous frame number is obtained by statistics, including the high confidence state frame number and the low confidence state frame number.

[0066] The high confidence state frame number is compared with a preset artificial labeling judgment threshold value; If the high confidence state frame number is greater than the preset artificial labeling judgment threshold value, the artificial labeling process is triggered; After the artificial labeling is completed, the artificial labeling result is trained by the small batch gradient descent method, and the new category is synchronized.

[0067] It should be noted that the core purpose of setting the number of continuous frames is to filter transient interference and ensure the persistence and stability of the target; if only one frame or a few frames detect an unknown target, it may be a transient electromagnetic interference (such as a burst of clutter), a target transient characteristic change (such as point cloud deformation caused by rapid turning), or a casual misjudgment of the model; when N frames and above (the interval between each frame is usually 0.1-1 seconds, adjusted according to the radar frame rate) all meet the “high confidence” condition, it can be determined that the target is a new category (such as a new type of unmanned aerial vehicle, a special vehicle, etc.) with stable physical properties and motion characteristics, and at this time triggering manual labeling has practical significance. In the embodiment, the value (≥5) of the preset manual labeling determination threshold N can be adjusted according to the application scene: in a high-speed dynamic scene (such as a highway), N can be set to 5-8 frames; in a low-speed static scene (such as a factory monitoring), N can be appropriately increased to 10 frames to further reduce the probability of false triggering.

[0068] According to the embodiment of the application, further comprising: performing clustering analysis on the historical trajectory of the monitored target, extracting a typical behavior mode and storing it as a mode feature vector, and generating a mode library; obtaining a current target motion feature vector and calculating a cosine similarity with the mode feature vector; if the cosine similarity is greater than or equal to a preset trajectory calling threshold, calling the trajectory parameters of the corresponding typical behavior mode for prediction.

[0069] It should be noted that the clustering analysis of the historical trajectory means clustering the target motion trajectory data (such as the position-time sequence of an airplane or a ship) monitored by the radar in the past, and classifying the trajectories with similar motion rules into the same class; the representative motion mode of each class of trajectory after clustering is an abstraction and simplification of the same class of trajectory; the mode feature vector converts the typical behavior mode into a quantifiable and calculable vector form, which is used for subsequent similarity comparison; the index for measuring the directional similarity of two vectors (the current target motion feature vector and a mode feature vector in the mode library) is calculated by the formula: ; the cosine similarity, the mode feature vector, the current target motion feature vector; the value range of s is [-1, 1], and the closer to 1 indicates that the modes are more similar; the preset trajectory calling threshold controls the strictness of mode matching, which is initially set by manual or experience and then dynamically optimized by a self-learning module.

[0070] According to the embodiment of the application, further comprising: obtaining a real-time prediction value of a neural network for a real-time trajectory and a typical prediction value based on a typical behavior mode; The information fusion is performed on the real-time prediction value and the typical prediction value to obtain a fusion prediction value; The calculation formula of the fusion prediction value is: Among them, the fusion prediction value, the typical prediction value, the real-time prediction value; and the preset mode multiplexing weight.

[0071] It should be noted that the fusion prediction value is the final prediction result obtained by the formula, which is a weighted fusion of two prediction values, combining the advantages of two prediction methods, using historical mode to ensure efficiency in stable scenes, and using neural network to ensure accuracy in mutation scenes; the mode multiplexing weight a is a parameter (a e [0, 1]) for adjusting the fusion proportion of the “historical mode prediction value (Tmodel)” and the “neural network prediction value (Tnn)”; when a→1, it depends more on the historical mode; when a→0, it depends more on the neural network (suitable for motion mutation scenes).

[0072] The adaptive radar target classification and trajectory prediction method provided in this embodiment, by acquiring radar echo data, extracting target point cloud and original spectrum feature, calculating local signal-to-noise ratio and feature sparsity data, selecting a lightweight, anti-interference deep or attention enhancement network branch accordingly, detecting the target category through an incremental adversarial learning module, obtaining the unknown class confidence and the target existence probability, comparing the threshold to obtain the confidence state, and counting the frame number to judge the artificial labeling and update the classification layer; thereby through the calculation of the local signal-to-noise ratio and the feature sparsity data, the selection of the processing branch, and the acquisition of the confidence state, the update of the classification layer and the trajectory prediction are realized.

[0073] Embodiment three Another embodiment of the present application relates to an adaptive radar target classification device, and the implementation details of the adaptive radar target classification device of this embodiment will be specifically described below. The following content is only provided for the implementation details for easy understanding, and is not necessary for implementing the present solution. The schematic diagram of the adaptive radar target classification device of this embodiment can be as shown in Figure 5 The adaptive radar target classification device of this embodiment includes a feature data determination module 501, a data type determination module 502, a processing branch selection module 503, a target classification module 504, a confidence state determination module 505, and a discrimination module 506.

[0074] The feature data determination module 501 is configured to determine target point cloud feature data and original spectrum feature data of a target object according to radar echo data. ​The data type determination module 502 is configured to determine a local signal-to-noise ratio based on the original spectral feature data, and determine a type of feature sparsity data according to the target point cloud feature data and a preset density threshold, the type of feature sparsity data including sparse data and dense data. The processing branch selection module 503 is configured to input the radar echo data into a pre-constructed target classification model, and select a processing branch according to a preset signal-to-noise ratio threshold range in which the local signal-to-noise ratio is located and the type of feature sparsity data. The target classification module 504 is configured to detect a target category via an incremental adversarial learning module of the target classification model based on feature data extracted by the processing branch, the target category including an unknown category and a known category. The confidence state determination module 505 is configured to obtain an unknown category confidence and a target existence probability when the target category is the unknown category, and determine a confidence state according to the unknown category confidence and the target existence probability, the confidence state including a high confidence state and a low confidence state. The discrimination module 506 is configured to obtain a continuous frame number of each confidence state, and trigger manual labeling when the continuous frame number of the high confidence state is greater than a preset manual labeling determination threshold, and input the radar echo data and a target category labeling result labeled manually into the target classification model as a training set.

[0075] In some optional embodiments, the feature data determination module is further configured to perform three-dimensional point cloud clustering on the radar echo data to obtain target point cloud feature data, the target point cloud feature data including a target point number and three-dimensional coordinate data corresponding to each target point. Based on the radar echo data, original spectral feature data is extracted, the original spectral feature data including target echo power and local noise power.

[0076] In some optional embodiments, the data type determination module is further configured to calculate a local signal-to-noise ratio according to the target echo power and the local noise power. Point cloud density distribution data is statistically obtained by dividing a spatial voxel grid according to the target point number and the three-dimensional coordinate data corresponding to each target point. When the point cloud density distribution data is less than a preset density threshold, it is determined that the type of feature sparsity data is sparse data. When the point cloud density distribution data is greater than or equal to the preset density threshold, it is determined that the type of feature sparsity data is dense data.

[0077] In some optional embodiments, the preset signal-to-noise ratio threshold range includes a first range greater than a first preset threshold, a second range less than the first preset threshold and greater than a second preset threshold, and a third range less than the second preset threshold, wherein the first preset threshold is greater than the second preset threshold, and the processing branch selection module is further configured to input the radar echo data into a pre-constructed target classification model; When the local signal-to-noise ratio is in the first range and the type of the feature sparsity data is dense data, the processing branch is determined to be a lightweight network; When the local signal-to-noise ratio is in the second range and the type of the feature sparsity data is dense data or sparse data, or the local signal-to-noise ratio is in the first range and the type of the feature sparsity data is sparse data, the processing branch is determined to be an attention-enhanced network; When the local signal-to-noise ratio is in the third range, and the type of the feature sparsity data is dense data or sparse data, the processing branch is determined to be an anti-interference deep network.

[0078] In some optional embodiments, the confidence state determination module is further configured to, when the target category is an unknown category, obtain an unknown category confidence and a target existence probability; When the unknown category confidence is greater than or equal to a preset confidence threshold, and the target existence probability is greater than a preset target existence threshold, the confidence state is determined to be a high confidence state; When the unknown category confidence is greater than or equal to a preset confidence threshold, and the target existence probability is less than or equal to a preset target existence threshold, the confidence state is determined to be a low confidence state; When the unknown category confidence is less than a preset confidence threshold, the confidence state is determined to be a low confidence state.

[0079] In some optional embodiments, a first prediction value generation module is further included, and the first prediction value generation module is configured to cluster historical trajectory data of a target object, extract a typical behavior pattern as a pattern feature vector, and generate a pattern library; A current target motion feature vector is obtained, and a cosine similarity between the current target motion feature vector and a pattern feature vector in the pattern library is calculated; When the preset similarity is greater than or equal to a preset trajectory calling threshold, a trajectory parameter of a pattern feature vector corresponding to the current target motion feature vector is used to perform trajectory prediction, and a first prediction value is generated.

[0080] In some optional embodiments, a fusion prediction value generation module is further included, and the fusion prediction value generation module is configured to perform trajectory prediction on the target object by using a pre-constructed trajectory prediction model to obtain a second prediction value. fusing the first prediction value and the second prediction value to obtain a fused prediction value; An expression of the fused prediction value is: wherein, is the fused prediction value, is the first prediction value, is the second prediction value; is a preset mode multiplexing weight.

[0081] It is worth mentioning that each module involved in the embodiment is a logical module. In actual application, one logical unit can be one physical unit, or a part of one physical unit, or realized by combination of multiple physical units. In addition, in order to highlight the innovative part of the present application, the units not closely related to solving the technical problems proposed in the present application are not introduced in the embodiment, but this does not mean that there are no other units in the embodiment.

[0082] It should be noted that in the present disclosure, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements includes not only those elements, but also other elements not explicitly listed, or further includes elements inherent to such a process, method, article or apparatus. Without more limitations, the element limited by the statement "including a" does not exclude the existence of another same element in the process, method, article or apparatus including the element.

[0083] Embodiment Four: Another embodiment of the present application relates to an electronic device, such as Figure 6 as shown, comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the adaptive radar target classification method in each of the above embodiments.

[0084] ​The memory and the processor are connected via a bus. The bus can include any number of interconnecting buses and bridges depending on the specific application of the mobile terminal. The bus connects the various circuits of the memory and the processor together and mediates data communication among different components. The bus can also connect with the various other circuits such as peripheral device, voltage regulators, power management circuitry and so on, which are well known in the art, and therefore, will not be described further. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or can include a plurality of components, such as a plurality of receivers and transmitters, which are configured to transmit and receive signals over a wireless medium. The data processed by the processor is transmitted over the wireless medium via the antenna. Further, the antenna also receives data and transmits the data to the processor.

[0085] The processor is responsible for managing the bus and general processing, which can also include the functions of an operating system, such as scheduling tasks, performing memory management, power management, and / or other functions. The memory can be used for storing data used by the processor, such as intermediate and / or final results.

[0086] The processor can include, but is not limited to, one or more processors or microprocessors, application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components, for performing the methods described in the above embodiments.

[0087] Embodiment Five Another embodiment of the present application relates to a computer readable storage medium storing a computer program. The computer program, when executed by a processor, implements the adaptive radar target classification method in the above embodiments.

[0088] That is, a person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a program stored in a storage medium, including a number of instructions to make a device (which can be a single-chip microcomputer, a chip, etc.) or a processor execute all or part of the steps of the methods described in various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM for short), a random access memory (Random Access Memory, RAM for short), a magnetic disk or an optical disk, and various media that can store program codes.

[0089] The computer-readable storage medium can also store at least one computer executable program / instruction, such as computer-readable instructions. The computer-readable storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory, etc. The computer-readable storage medium may, for example, include read-only memory (ROM), a hard disk, a flash memory, etc. For example, the non-transitory computer-readable storage medium can be connected to a computing device such as a computer, and then, in the case where the computing device runs the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0090] In addition, the computer device can also include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and an input / output device (for example, a keyboard, a mouse, a speaker, etc.), etc.

[0091] The processor can communicate with an external device through a wired or wireless network through the I / O bus.

[0092] A person of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for implementing the present application, and in actual applications, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. An adaptive radar target classification method, characterized in that, include: The target point cloud feature data and raw spectrum feature data of the target object are determined based on radar echo data; The local signal-to-noise ratio is determined based on the original spectral feature data, and the type of feature sparsity data is determined based on the target point cloud feature data and a preset density threshold. The type of feature sparsity data includes sparse data and dense data. The radar echo data is input into a pre-built target classification model, and a processing branch is selected based on the preset signal-to-noise ratio threshold range where the local signal-to-noise ratio is located and the type of feature sparsity data. The target category is detected by the incremental adversarial learning module of the target classification model based on the feature data extracted from the processing branch. The target category includes unknown class and known class. When the target category is unknown, the confidence level of the unknown category and the probability of the target's existence are obtained. Based on the confidence level of the unknown category and the probability of the target's existence, the confidence level is determined. The confidence level includes a high confidence level and a low confidence level. The number of consecutive frames for each confidence state is obtained. When the number of consecutive frames for a high confidence state exceeds a preset manual annotation threshold, manual annotation is triggered. The manually annotated radar echo data and target category annotation results are then used as training sets and input into the target classification model.

2. The adaptive radar target classification method according to claim 1, characterized in that, The steps for determining the target point cloud feature data and raw spectral feature data of the target object based on radar echo data include: Three-dimensional point cloud clustering is performed on radar echo data to obtain target point cloud feature data, which includes the number of target points and the three-dimensional coordinate data corresponding to each target point. Based on the radar echo data, raw spectral feature data is extracted, which includes target echo power and local noise power.

3. The adaptive radar target classification method according to claim 2, characterized in that, The steps of determining the local signal-to-noise ratio based on the original spectral feature data and determining the type of feature sparsity data based on the target point cloud feature data and a preset density threshold, wherein the type of feature sparsity data includes sparse data and dense data, include: The local signal-to-noise ratio is calculated based on the target echo power and the local noise power. Based on the number of target points and the three-dimensional coordinate data corresponding to each target point, point cloud density distribution data is obtained by dividing the spatial voxel grid. When the point cloud density distribution data is less than a preset density threshold, the type of the feature sparsity data is determined to be sparse data. When the point cloud density distribution data is greater than or equal to a preset density threshold, the type of the feature sparsity data is determined to be dense data.

4. The adaptive radar target classification method according to claim 1, characterized in that, The preset signal-to-noise ratio (SNR) threshold range includes a first range greater than a first preset threshold, a second range less than the first preset threshold but greater than a second preset threshold, and a third range less than the second preset threshold, wherein the first preset threshold is greater than the second preset threshold. The step of inputting the radar echo data into a pre-constructed target classification model and selecting a processing branch based on the preset SNR threshold range where the local SNR is located and the type of feature sparsity data includes: The radar echo data is input into a pre-built target classification model; When the local signal-to-noise ratio is within the first range and the type of the feature sparse data is dense data, the processing branch is determined to be a lightweight network; When the local signal-to-noise ratio is in the second range and the type of the feature sparsity data is dense data or sparse data, or when the local signal-to-noise ratio is in the first range and the type of the feature sparsity data is sparse data, the processing branch is determined to be an attention enhancement network. When the local signal-to-noise ratio is within the third range, and the type of the feature sparsity data is dense data or sparse data, the processing branch is determined to be an anti-interference deep network.

5. The adaptive radar target classification method according to claim 1, characterized in that, When the target category is unknown, the steps of obtaining the confidence level of the unknown category and the probability of the target's existence, and determining the confidence state based on the confidence level of the unknown category and the probability of the target's existence, wherein the confidence state includes a high confidence state and a low confidence state, include: When the target category is unknown, obtain the confidence level of the unknown category and the probability of the target's existence. When the confidence level of the unknown class is greater than or equal to a preset confidence threshold, and the probability of the target's existence is greater than a preset target existence threshold, the confidence state is determined to be a high confidence state. When the confidence level of the unknown class is greater than or equal to a preset confidence threshold, and the probability of the target's existence is less than or equal to a preset target existence threshold, the confidence level is determined to be a low confidence level. When the confidence level of the unknown class is less than a preset confidence threshold, the confidence level is determined to be a low confidence level.

6. The adaptive radar target classification method according to claim 1, characterized in that, The method further includes: Cluster the historical trajectory data of the target object, extract typical behavior patterns and store them as pattern feature vectors to generate a pattern library; Obtain the current target motion feature vector and calculate the cosine similarity between the current target motion feature vector and the pattern feature vector in the pattern library; When the pre-similarity is greater than or equal to the preset trajectory call threshold, trajectory prediction is performed based on the trajectory parameters of the pattern feature vector corresponding to the current target motion feature vector to generate a first prediction value.

7. The adaptive radar target classification method according to claim 6, characterized in that, The method further includes: The trajectory of the target object is predicted using a pre-built trajectory prediction model to obtain a second predicted value; The first and second predicted values ​​are fused to obtain a fused predicted value. The expression for the fusion prediction value is: ； in, To merge the predicted values, The first predicted value, This is the second predicted value; The weights are reused for the preset pattern.

8. An adaptive radar target classification device, characterized in that, include: The feature data determination module is used to determine the target point cloud feature data and the original spectrum feature data of the target object based on the radar echo data. The data type determination module is used to determine the local signal-to-noise ratio based on the original spectral feature data, and to determine the type of feature sparsity data according to the target point cloud feature data and a preset density threshold. The type of feature sparsity data includes sparse data and dense data. The processing branch selection module is used to input the radar echo data into the pre-built target classification model and select the processing branch according to the preset signal-to-noise ratio threshold range where the local signal-to-noise ratio is located and the type of feature sparsity data. The target classification module is used to detect the target category based on the feature data extracted by the processing branch via the incremental adversarial learning module of the target classification model. The target category includes unknown classes and known classes. The confidence state determination module is used to obtain the confidence level of the unknown class and the probability of the target's existence when the target category is unknown, and to determine the confidence state based on the confidence level of the unknown class and the probability of the target's existence. The confidence state includes a high confidence state and a low confidence state. The discrimination module is used to obtain the number of consecutive frames for each confidence state. When the number of consecutive frames for a high confidence state exceeds a preset manual annotation threshold, manual annotation is triggered, and the manually annotated radar echo data and target category annotation results are used as training sets and input into the target classification model.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the adaptive radar target classification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the adaptive radar target classification method according to any one of claims 1 to 7.

Images