Radar target identification method and system

Abstract

The invention relates to the technical field of radar target recognition, and discloses a radar target recognition method and system. The method comprises the following steps: acquiring an original echo signal sequence acquired by a radar sensor, and performing noise suppression and pulse compression processing on the original echo signal sequence by using a signal preprocessing unit to generate preprocessed radar echo data; based on the preprocessed data, fusing to generate a comprehensive feature description vector of the target; a dynamically updated target feature association network is constructed, network nodes represent feature sets of different targets, network edges represent space-time association or feature similarity among the targets, and an online learning mechanism is adopted to continuously optimize a network topology structure; according to the real-time comprehensive feature description vector, gradually positioning a potential target area in the target feature association network by adopting a hierarchical focusing mechanism; and in the potential target area, a candidate target set is screened through a multi-dimensional similarity measurement algorithm, and candidate targets are sorted and output according to a feature consistency criterion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target recognition, in particular to a radar target recognition method and system. BACKGROUND

[0002] In the fields of modern air defense, air traffic management, ocean monitoring, etc., as the core detection equipment, radar needs to realize target recognition through the analysis of target echo signals, and the recognition accuracy and efficiency directly affect the operation performance of related systems. At present, in the process of radar target recognition, the original echo signal is often affected by factors such as complex electromagnetic environment, background noise and multi-target interference, resulting in poor signal quality. The traditional signal preprocessing method mostly uses fixed parameter filtering algorithm and pulse compression technology, which is difficult to dynamically adjust the processing strategy according to the real-time noise characteristics, and is prone to problems such as incomplete noise suppression or loss of signal details, thereby affecting the accuracy of subsequent feature extraction. For example, in the low-altitude small target detection scene, the echo features of micro unmanned aerial vehicles cannot be accurately captured due to noise residue or loss of signal details, resulting in target missed detection, which directly threatens low-altitude security.

[0003] In the feature extraction link, the existing technology often focuses on single-dimensional features, such as only extracting time-frequency features or scattering features, without fully integrating multi-dimensional feature information. This single feature extraction mode is difficult to fully describe the physical properties and motion state of the target, resulting in insufficient representation ability of the generated feature description vector for the target, especially when facing similar-shaped or complex motion mode targets, which is prone to feature confusion, increasing the difficulty of subsequent recognition. For example, in the aviation control scene, only relying on time-frequency features cannot distinguish between civil passenger aircraft and large cargo aircraft, which may lead to air traffic scheduling errors and cause flight delays or flight conflict risks.

[0004] In terms of target feature association, traditional methods mostly use static network structure to store target feature information, and once the network topology is determined, it is difficult to adjust. When new targets appear or the motion state of targets changes, the static network cannot update the association relationship of nodes and edges in time, resulting in poor adaptability of the network to dynamic targets. At the same time, there is a lack of effective online learning mechanism, which cannot continuously optimize the network structure using historical data, so that the feature association ability of the network is difficult to improve with the increase of use time, and the recognition performance is prone to decay in long-term continuous monitoring scenarios. In the marine ship monitoring scene, the feature association information of the ship newly entering the sea area cannot be updated in time, resulting in interruption of tracking of the ship, and inability to timely discover illegal crossing, illegal fishing, etc.

[0005] In the process of potential target area positioning and candidate target screening, the traditional method often adopts a global search strategy, which needs to traverse the entire feature space, resulting in low positioning efficiency. Especially in a multi-target dense scene, the search time is too long. In addition, the similarity measurement often uses a single-dimensional similarity calculation method, which does not consider the similarity of multi-dimensional features, and the candidate target sorting is only based on a single matching degree index, ignoring key factors such as feature stability and time freshness, resulting in insufficient matching accuracy of the screened candidate target and the real target, affecting the reliability of the final recognition result. In the air defense early warning scene, the time-consuming global search will delay the positioning opportunity of the high-speed incoming missile, and the single-dimensional similarity measurement and matching sorting may misjudge the decoy missile as the real missile, causing air defense resource mismatch and defense loopholes. SUMMARY

[0006] The purpose of the present application is to provide a radar target recognition method and system to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides a radar target recognition method, which comprises: Obtaining the original echo signal sequence collected by the radar sensor, using a signal preprocessing unit to perform noise suppression and pulse compression processing on the original echo signal sequence to generate preprocessed radar echo data; Based on the preprocessed radar echo data, the time-frequency features, scattering features and motion trajectory features of the target are calculated by a multi-dimensional feature extraction unit, and the integrated feature description vector of the target is generated by fusion; A dynamically updated target feature association network is constructed, wherein the network nodes represent the feature sets of different targets, the network edges represent the spatio-temporal association or feature similarity between targets, and an online learning mechanism is used to continuously optimize the network topology structure; According to the real-time obtained integrated feature description vector, a hierarchical focusing mechanism is used in the target feature association network to gradually locate the potential target area; In the potential target area, a multi-dimensional similarity measurement algorithm is used to screen the candidate target set, and the candidate targets are sorted and output according to the feature consistency criterion.

[0008] Preferably, the original echo signal sequence collected by the radar sensor is obtained, and the signal preprocessing unit is used to perform noise suppression and pulse compression processing on the original echo signal sequence to generate preprocessed radar echo data, which comprises: Adaptive filtering algorithm is used to eliminate background noise of the original echo signal sequence, and pulse compression processing is used to enhance signal resolution; The processed signal is analyzed in time and frequency domain to extract the energy distribution characteristics and phase modulation characteristics of the signal; The preprocessed radar echo data containing signal strength, frequency offset features and phase information are generated.

[0009] Preferably, the preprocessed radar echo data are used to calculate time-frequency features, scattering features and motion trajectory features of the target by a multi-dimensional feature extraction unit, and a comprehensive feature description vector of the target is generated by fusion, including: The time-frequency transform features of the target, including instantaneous frequency features and spectral envelope features, are extracted from the preprocessed radar echo data; The radar scattering cross-section features and polarization scattering matrix features of the target are calculated; The motion trajectory parameters of the target, including velocity vector and acceleration features, are analyzed by multi-frame data correlation; The time-frequency features, scattering features and motion trajectory features are normalized and fused into a comprehensive feature description vector of fixed dimension.

[0010] Preferably, the dynamic and updated target feature correlation network is constructed, wherein the network nodes represent the feature sets of different targets, the network edges represent the spatio-temporal correlation or feature similarity between targets, and an online learning mechanism is used to continuously optimize the network topology structure, including: The comprehensive feature description vector of each target is used as the node feature representation in the target feature correlation network; The connection edges between nodes are established according to the time sequence and spatial distribution relationship of target appearance, and the weight coefficients of the edges are calculated based on feature similarity; An incremental update strategy is used to dynamically adjust the network structure, and when new target data is added, only the affected nodes and their associated edges are updated; The network connection weights and topological relationships are optimized according to the historical data interaction mode by an online learning mechanism.

[0011] Preferably, the hierarchical focusing mechanism is used in the target feature correlation network to gradually locate the potential target area according to the real-time comprehensive feature description vector, including: The real-time comprehensive feature description vector is mapped to the feature space of the target feature correlation network; The initial node set that best matches the real-time vector in the network is searched as the starting point of the search; The initial node is expanded to adjacent nodes, and relevant nodes are selected layer by layer according to the feature similarity threshold; When the expansion layer reaches the preset depth or the node similarity is lower than the critical value, the expansion is stopped, and the sub-network composed of the selected nodes is used as the potential target area.

[0012] Preferably, the multi-dimensional similarity measurement algorithm is used to screen the candidate target set in the potential target area, and the candidate targets are sorted and output according to the feature consistency criterion, including: calculating multi-dimension similarity between each node feature in potential target area and real-time comprehensive feature description vector, including time-frequency feature similarity, scattering feature similarity and motion feature similarity; generating overall matching degree score of each node by using weighted fusion algorithm to integrate multi-dimension similarity scores; selecting nodes with overall matching degree score higher than preset threshold as candidate target set; sorting candidate target set according to matching degree score, feature stability and time freshness.

[0013] Preferably, the generating overall matching degree score of each node by using weighted fusion algorithm to integrate multi-dimension similarity scores comprises: assigning dynamic weight coefficients to time-frequency feature similarity, scattering feature similarity and motion feature similarity respectively; adjusting weight distribution of each dimension similarity according to target type and environment characteristics; calculating overall matching degree score of each node through linear weighting.

[0014] Preferably, the sorting candidate target set according to matching degree score, feature stability and time freshness comprises: calculating feature stability index of each candidate target node, reflecting variation degree of historical features of the node; evaluating time freshness index of each candidate target node, reflecting interval between latest update time of the node and current time; generating final sorting score by integrating matching degree score, feature stability index and time freshness index; the evaluating time freshness index of each candidate target node, reflecting interval between latest update time of the node and current time, comprises: recording last update time stamp of each node; calculating time difference between current time and last update time; converting time difference into freshness score by using exponential decay function.

[0015] Preferably, the calculating feature stability index of each candidate target node, reflecting variation degree of historical features of the node, comprises: extracting feature change sequence of the node within past time window; calculating variance and coefficient of variation of feature sequence; generating stability quantitative index by combining feature change trend analysis.

[0016] Preferably, the present application also includes a radar target recognition system comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the steps of the radar target recognition method described above when executing the computer program.

[0017] Compared with the prior art, the present application has the following advantages: The radar target recognition method of the present application adopts targeted noise suppression and pulse compression processing in the signal preprocessing link, which can effectively process noise interference in the original echo signal, retain target feature details in the signal, and generate high-quality preprocessed radar echo data. Compared with the traditional fixed parameter preprocessing method, this link can better adapt to the signal characteristics in different electromagnetic environments, providing a better data basis for subsequent feature extraction, and helping to reduce the feature extraction errors caused by poor data quality. In the multi-dimensional feature extraction link, the time-frequency features, scattering features and motion trajectory features of the target are calculated, and a comprehensive feature description vector is generated by fusion, which can comprehensively depict the physical properties and motion state of the target from multiple dimensions. The time-frequency features can reflect the distribution characteristics of the target echo signal in the time and frequency domains, the scattering features can reflect the shape structure and material properties of the target, and the motion trajectory features can describe the dynamic information such as the motion speed and acceleration of the target. The fusion of multi-dimensional features makes the comprehensive feature description vector more comprehensive in representing the target, which can effectively distinguish targets with similar shapes or similar motion patterns, and enhance the recognition ability of different types of targets. A dynamically updated target feature association network is constructed, the network nodes are defined as the feature sets of different targets, the network edges represent the spatio-temporal association or feature similarity between targets, and an online learning mechanism is introduced, which can realize the continuous optimization of the network topology structure. The dynamic updating feature enables the network to timely accept the feature information of new targets, adjust the association relationship of nodes and edges, and adapt to the dynamic changes of the number and motion state of targets. The online learning mechanism can continuously optimize the network connection weight and topology structure using historical data, improve the association ability of the network to target features, and maintain stable association performance even in long-term continuous monitoring scenarios, providing accurate feature association basis for subsequent potential target area positioning. A hierarchical focusing mechanism is used to locate the potential target area in the target feature association network. The real-time comprehensive feature description vector is first mapped to the feature space to determine the initial node set as the search starting point, and then the initial node is taken as the center to expand and screen layer by layer, which can avoid the inefficiency of traditional global search, greatly reduce the search range, and improve the positioning efficiency. Especially in a multi-target dense scene, the region that may contain the target can be quickly focused, reducing the time consumption caused by invalid search, and saving time cost for subsequent candidate target screening. In the potential target area, the candidate target set is screened by a multi-dimensional similarity measurement algorithm, and the candidate targets are sorted in combination with a feature consistency criterion, comprehensively considering the similarity of time-frequency features, scattering features and motion features, so that the matching degree of node features and real-time comprehensive feature description vectors can be comprehensively evaluated, and the limitations of single-dimensional similarity measurement can be reduced. At the same time, the factors of feature stability and time freshness are included in the sorting process, so that the candidate targets with stable feature performance and relatively new update time can be preferentially selected, so that the selected candidate targets are closer to the real targets, the reliability of the final recognition result is improved, the good recognition performance can be maintained in a complex multi-target scene, and the needs of radar target recognition in different application scenarios are met. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A working principle diagram of the radar target recognition method is described. Figure 2 A flowchart of radar echo signal preprocessing is described. Figure 3 A flowchart of dynamic target feature association network construction and optimization is described. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] Please refer to Figure 1The application provides a radar target recognition method and system, which includes integrated signal processing, feature extraction and network analysis to achieve efficient target recognition. The method starts with the original echo signal sequence collected by the radar sensor. The signal preprocessing unit performs noise suppression and pulse compression processing on the original sequence to generate preprocessed radar echo data. Based on the preprocessed data, the multi-dimensional feature extraction unit calculates the time-frequency features, scattering features and motion trajectory features of the target. These features are normalized and fused into a fixed-dimensional comprehensive feature description vector. A dynamically updated target feature correlation network is constructed, with network nodes representing different target feature sets and network edges representing the spatio-temporal correlation or feature similarity between targets. An online learning mechanism continuously optimizes the network topology to adapt to real-time data streams. In the recognition stage, the comprehensive feature description vector obtained in real time is mapped into the network, and a hierarchical focusing mechanism is used to gradually locate the potential target area. The multi-dimensional similarity measurement algorithm is used to select the candidate target set within the located area, and the candidate targets are sorted and output according to the feature consistency criterion, considering the matching degree score, feature stability and time freshness index.

[0021] Embodiment 1: refer to Figure 2 The original echo signal sequence collected by the radar sensor is the basis, which usually contains reflected signals from targets and complex environmental noise and clutter interference. The signal preprocessing unit then starts working, and its primary task is to perform noise suppression processing on the original sequence. The adaptive filtering algorithm used here can dynamically adjust its filter coefficients and structure parameters according to the statistical characteristics of the background noise collected in real time. This algorithm estimates the power spectral density or autocorrelation function of the noise continuously, and generates a filter response that matches it, effectively separating and eliminating the background noise component from the original echo. This process significantly improves the signal-to-noise ratio without causing significant distortion to the target signal. The pulse compression processing follows, which is achieved through a filter matched with the transmitted signal waveform. The core purpose is to enhance the resolution of the signal in the distance dimension, so that the originally widened pulse echo is compressed into a narrow pulse, allowing multiple targets that are very close in distance to be distinguished more clearly. The impulse response of the matched filter is carefully designed to match the conjugate of the transmitted signal that is reversed in time. In this way, when the echo signal passes through the filter, the components in the signal that match the transmitted waveform will be coherently superimposed, while the noise and non-matching components will be suppressed, and the main lobe width of the final output signal will be narrowed and the side lobe level will be reduced.

[0022] After the noise suppression and pulse compression, the signal needs to be analyzed in time and frequency domain to extract its intrinsic characteristics. Time domain analysis focuses on the energy distribution of the signal, and by calculating the envelope or amplitude of the signal over time, key parameters such as the peak intensity of the pulse, the duration of the pulse, the rising and falling edges, and the time can be obtained. These parameters intuitively reflect the strength and basic time structure of the target reflection signal. Frequency domain analysis converts the signal from time domain to frequency domain through fast Fourier transform, so that the frequency spectrum composition of the signal can be clearly observed, and important frequency offset features such as the shift of the signal center frequency relative to the transmission frequency can be extracted. This shift is usually caused by the Doppler effect caused by the relative motion between the target and the radar, and the phase modulation characteristics are also carefully analyzed. The phase information contains clues to the subtle movement or structural changes of the target. By integrating the results of time and frequency domain analysis, the system finally generates a set of structured radar echo data after preprocessing, which contains signal strength, frequency offset characteristics, and phase information.

[0023] Based on the radar echo data generated after preprocessing, the multi-dimensional feature extraction unit starts to run, and its task is to calculate a variety of features that can fully describe the properties of the target. The first step is to extract time-frequency features. In order to capture the dynamic behavior of the signal frequency components over time, it is common to use time-frequency analysis tools such as short-time Fourier transform or wavelet transform. Short-time Fourier transform performs segmented Fourier analysis of the signal through a sliding time window, thereby obtaining the instantaneous frequency characteristics of the signal, i.e. the trajectory of the frequency evolution over time, while the spectral envelope feature describes the overall shape and width of the signal spectrum, reflecting the overall scattering characteristics of the target. These time-frequency features are particularly important for identifying targets with specific micro-motion patterns. Next, the scattering characteristics of the target are calculated, where the radar scattering cross section is a core parameter that is derived by comparing the echo power with the power incident on the target, quantitatively describing the size of the target's ability to scatter electromagnetic waves in the direction of the radar. The polarization scattering matrix feature is extracted from the different polarization receiving channel data of the radar, which records the scattering response of the target to different polarization incident waves, revealing detailed information about the target's geometry, material, and orientation.

[0024] The acquisition of motion trajectory features requires the correlation analysis of consecutive radar echo data, and the motion parameters of the target can be estimated by tracking the position changes of the same target at different times. The data association process usually uses algorithms such as nearest neighbor method or joint probability data association to ensure that the measurement points in different frames can be correctly associated with the same target trajectory. Once the trajectory is confirmed, the velocity vector and acceleration feature can be calculated, the velocity vector indicates the speed and direction of the target motion, and the acceleration feature reflects the rate of change of the target motion state. The estimation of these motion parameters often relies on tracking filters such as Kalman filter or particle filter, which can effectively smooth the errors caused by measurement noise and provide predictions of the future motion state of the target. After all these time-frequency features, scattering features and motion trajectory features are successfully extracted, they need to go through a normalization step, because the physical meaning and numerical range of different features may differ greatly, the purpose of normalization is to eliminate these differences in dimension and scale, so that different features can be fused on a fair basis. Common normalization methods include min-max scaling, which linearly transforms the feature values to a fixed interval, and Z-score standardization, which is based on the mean and standard deviation of the feature.

[0025] These normalized heterogeneous features are fused into a fixed-dimensional comprehensive feature description vector, and the fusion strategy can use a simple vector concatenation method, that is, all feature values are arranged in order to form a long vector, but this method may result in high vector dimension, so sometimes principal component analysis or other methods are used to reduce the dimension of the features, remove redundant information and retain the most discriminative feature components, and then combine them into the final comprehensive feature description vector. This comprehensive vector serves as a compact and comprehensive digital signature, uniquely representing the target properties observed by the radar at a specific time, and providing standardized and structured input data for subsequent construction of dynamic target feature association networks and efficient target recognition matching. The whole process from preprocessing to feature extraction and fusion constitutes a complete front-end processing chain, which fully considers the characteristics of actual radar signals and the needs of subsequent processing modules.

[0026] Embodiment 2: see Figure 3, the definition of network node structure is the starting point of constructing a dynamic updating target feature association network. Each node corresponds to a target entity observed by the radar system and its features are extracted. The core feature representation of the node is directly derived from the comprehensive feature description vector generated by the multi-dimensional feature extraction unit. This vector contains the quantitative representation of multi-dimensional information such as time-frequency, scattering and motion trajectory. The dimension is fixed in the system design phase to maintain the consistency of the network structure. The appearance of each new target will instantiate a new node in the network, which is initialized using its corresponding comprehensive feature description vector. Network nodes do not exist in isolation, and their connection relationship is represented by network edges. The establishment of network edges is mainly based on the two logical relationships of spatio-temporal association or feature similarity. The calculation of spatio-temporal association depends on the time series information and spatial coordinate data of the detected target. The system checks the proximity of different target nodes in time and the proximity of spatial location. For example, if two targets appear in each other's adjacent physical area in consecutive time frames, they are more likely to establish an edge representing spatio-temporal association.

[0027] Feature similarity measures the internal connection between nodes from another dimension. It quantifies the similarity between the comprehensive feature description vectors carried by two nodes by calculating the similarity between the two vectors. Common similarity measures include cosine similarity, which focuses on the difference in direction between two vectors, or Euclidean distance, which reflects the absolute distance in space. The weight coefficient of each edge is assigned according to the calculated similarity value or association strength. The weight value intuitively reflects the closeness of the relationship between the two connected nodes represented by the targets. In order to make the network adapt to the dynamic changes of targets in the radar detection environment, such as the appearance of new targets, the disappearance of old targets or the evolution of target features, the system adopts an incremental updating strategy to dynamically adjust the network structure. When new target data flows into the system, the update process is not to reconstruct the entire network, but to make targeted local adjustments. The system first converts the new target data into a new node, then calculates the association degree between the new node and the existing nodes in the network, and only when the new node and the existing nodes meet the pre-set association conditions will a new connection edge be established. At the same time, the introduction of the new node may change the association in the local area, and the system will re-evaluate and adjust the weight of the existing connection edges in the affected area.

[0028] The online learning mechanism, as the core driver of the network's continuous self-optimization, relies heavily on the accumulation of historical data and the analysis of interaction patterns. This mechanism continuously optimizes the connection weights and topology of the entire network by analyzing the stability of historical connection relationships between nodes, the regularity of weight changes, and the patterns of cluster formation. The process of optimizing connection weights can draw on the gradient descent idea in machine learning, where the system adjusts the weight coefficients of edges based on feedback signals from historical interactions, making the weight values better reflect the true correlation strength. For example, an edge that is frequently activated and has highly consistent target feature associations may have its weight gradually strengthened. In terms of topology optimization, the system may use complex network science algorithms, such as community detection algorithms, to automatically discover naturally formed target subgroups or communities in the network. These communities often correspond to target groups with high correlation in space, time, or features, and the identification of community structure helps understand the macro relationships between targets. When the network structure no longer adapts to the current environment over time, the online learning mechanism can trigger topology restructuring, such as merging highly overlapping communities or splitting communities with weakened associations.

[0029] The dynamic updating process of the entire network is a continuous and cyclic closed loop. The influx of new data triggers incremental updates, and the changes produced by incremental updates are recorded and integrated into historical data. The online learning mechanism analyzes these historical interaction patterns to guide the next round of weight and topology optimization. This design makes the target feature association network no longer a static database, but a dynamic knowledge system with flexibility and adaptability. It can automatically adjust its internal structure to follow the evolution of the detection scene and the changes in target behavior patterns, providing a high-quality relationship map that is always synchronized with the current environment for subsequent hierarchical focused search and candidate target screening. Each node in the network maintains its own metadata, such as creation timestamp, last update timestamp, and feature history snapshot. These metadata provide necessary information support for evaluating the freshness of nodes and the stability of features. The maintenance of edges is also important, as each edge records its establishment time, last weight update time, and weight change history. These information helps the online learning mechanism evaluate the reliability and timeliness of connection relationships. Through this detailed dynamic management, the target feature association network effectively integrates isolated target observation data into a networked knowledge model that is interconnected and contains rich semantic information.

[0030] Example 3: After the system obtains the comprehensive feature description vector generated by real-time acquisition and processing, the next step in the radar target recognition process is to perform an accurate search and match in the dynamically constructed target feature correlation network. This step is achieved through a hierarchical focusing mechanism to locate the potential target area. The hierarchical focusing mechanism begins by mapping the real-time vector to the high-dimensional feature space of the network. The mapping process is completed by calculating the similarity between the real-time vector and the feature vector of each existing node in the network. This similarity measure usually uses cosine similarity or the inverse of multidimensional Euclidean distance, etc. Based on the similarity calculation results, the system selects a group of initial nodes that best match the real-time vector. The selection criteria can be an absolute similarity threshold or a relative threshold adjusted dynamically according to the scene, such as selecting the top N nodes in terms of similarity. These initial nodes form the starting point and anchor for subsequent hierarchical search. With the initial node set as the center, the search range begins to expand layer by layer outward. The expansion process follows the pre-established connection edges in the target feature correlation network, which represent the spatiotemporal correlation or feature similarity between targets. In each layer of expansion, the system accesses the direct neighbor nodes of all nodes in the current layer and calculates the feature similarity between these newly touched nodes and the real-time vector. Then, this similarity is compared with a threshold set for the current layer. The design of this hierarchical threshold is critical to the mechanism. It is usually gradually increased with the increase of search depth (i.e., the number of layers). The threshold in the initial layers is set relatively loose, allowing more nodes to be included to explore a wider range of possibilities and prevent premature convergence. The threshold in the deep layers is gradually tightened to focus on the core area highly related to the real-time vector and avoid wasting computational resources on irrelevant nodes. The expansion process continues iteratively, with the output of each layer becoming the input for the next layer of expansion, and the search range spreads like a wave in an orderly manner.

[0031] The expansion process will terminate when the preset stopping conditions are met. The stopping conditions usually include two aspects: one is that the number of expansion layers has reached the maximum depth limit set by the system , which controls the breadth of the search to prevent it from going on indefinitely; the second is that the similarity value of the newly discovered nodes in the current expansion layer to the real-time vector has fallen below a critical threshold set for the layer , indicating that the possibility of further expanding the discovery of highly relevant nodes has been very low. When the search stops, all the nodes that are visited and whose similarity reaches the threshold at the corresponding layer jointly constitute a connected subnetwork, which is defined as the potential target area of this search, representing the part of the target nodes in the network that is most relevant to the current observation. After the potential target area is determined, the next step is to filter out the candidate target set within the area through a multi-dimensional similarity measurement algorithm. This algorithm does not perform a single, general similarity comparison, but rather decomposes different feature dimensions from the comprehensive feature description vector and performs fine-grained comparison on each dimension. The time-frequency feature similarity focuses on the sub-vector representing the time-frequency characteristics of the target in the vector, such as comparing the profile of the short-time Fourier transform spectrum or the distribution of the wavelet coefficients; the scattering feature similarity focuses on evaluating the trend of the size change of the radar scattering cross section and the matching of the polarization scattering matrix elements; the motion feature similarity mainly measures the consistency of the direction and size of the velocity vector, acceleration, and other kinematic parameters.

[0032] To integrate these different physical meanings and dimensional similarities into a unified evaluation standard, the system uses a weighted fusion algorithm. The core of this algorithm is to assign appropriate weights to the similarity of each dimension, then perform linear weighted fusion, and thus generate the overall matching degree score of each node. This overall matching degree score can be formally expressed as:

[0033] wherein the symbol represents the overall matching degree score of the kth candidate node, which is a scalar value between 0 and 1, and the higher the value, the better the overall matching degree of the node with the real-time vector. The symbol represents the total number of feature dimensions used, and in this embodiment, D = 3, corresponding to the time-frequency, scattering, and motion dimensions. The symbol represents the weight coefficient assigned to the dth feature dimension, and all weight coefficients satisfy the normalization condition that their sum is 1. The specific value of the weight coefficient can be dynamically adjusted according to the current task focus (such as focusing on classification or tracking) and environmental conditions. The symbol represents the similarity score of the kth candidate node in the dth feature dimension, which is standardized to make the scores of different dimensions comparable.

[0034] After the overall matching score of each node is calculated, the system sets a preset threshold, only the nodes with a score higher than the threshold are selected into the final candidate target set, and this threshold plays a role in preliminary quality control. For the candidate targets entering the set, the system sorts them according to the feature consistency criterion, and the sorting not only depends on the overall matching score, but also considers the feature stability of the node and the time freshness of its data. By assigning weights to the three indicators, a final sorting score is calculated, and an ordered list ranked by matching possibility from high to low is output.

[0035] In the operation of the radar target recognition system, after the potential target area is determined, the fine evaluation and sorting of the candidate nodes in the area become the key, and the core of the process is to generate an overall matching score by using a weighted fusion algorithm and to sort according to multiple indicators. Assuming that the system is currently processing radar observation data in a highway scene at the edge of a city, the real-time acquired comprehensive feature description vector corresponds to a point target moving at high speed, and the potential target area contains several historically tracked target nodes, which may correspond to different types of targets such as small cars and large trucks. The weighted fusion algorithm starts to work, which needs to calculate an overall matching score for each candidate node. This score is fused from multiple dimensions of similarity components, mainly including time-frequency feature similarity, scattering feature similarity, and motion feature similarity. The time-frequency feature similarity measures the matching degree of the target echo signal features in the joint time-frequency domain, such as the similarity of signal micro-Doppler features; the scattering feature similarity evaluates the consistency of the target radar scattering cross section and polarization scattering characteristics; and the motion feature similarity compares the proximity of motion parameters such as speed, heading, and acceleration. The importance of each similarity dimension in the fusion process is not fixed, but is reflected by dynamic weight coefficients. The system assigns an adjustable weight value to the time-frequency feature similarity, the scattering feature similarity, and the motion feature similarity. The allocation strategy of the weight coefficients depends closely on the current target type hypothesis and external environmental characteristics. For example, in the above highway scene, since the targets are generally moving at high speed, and vehicle classification is not the primary task, the system may dynamically adjust the weight of the motion feature similarity to be 0.5, appropriately reduce the weight of the scattering feature similarity to 0.3, and set the weight of the time-frequency feature similarity to 0.2. This weight allocation makes the motion characteristics dominant in the matching score, which is more in line with the actual needs of distinguishing different vehicles (such as distinguishing fast lane cars from slow lane trucks) in the current scene. The fusion calculation itself is completed by linear weighted summation, that is, the overall matching score is equal to the sum of the product of each dimension similarity score and its corresponding weight. The calculation process is simple and efficient, and is suitable for real-time processing.

[0036] Table 1: Candidate target node matching score calculation table

[0037] After obtaining the overall matching score of each candidate node, the system performs a preliminary screening according to a preset threshold, for example, setting the threshold value to 0.80, then Node_3581 in the above table will be excluded, and Node_1024, Node_5762 and Node_8813 enter the final candidate target set. The system sorts this set according to the feature consistency criterion, and the sorting not only includes the matching score, but also deeply integrates the feature stability and time freshness two key indicators. The feature stability indicator is used to evaluate the reliability of the target represented by the candidate node in its historical trajectory. A node with frequent feature changes, even if its current instantaneous matching degree is high, its reliability is doubtful. When calculating this indicator, the system will extract the feature change sequence of the node in a set time window in the past. This sequence contains a series of snapshots of the node's historical comprehensive feature description vector.

[0038] The sequence is analyzed to calculate its variance, and the size of the variance value directly reflects the intensity of the fluctuation of the feature value around its average level. The larger the variance, the more unstable the feature. Further, the system will calculate the coefficient of variation, which is the ratio of the standard deviation to the average value. This is a dimensionless quantity, which is convenient for comparing the stability of features of different magnitudes. In addition to these static statistics, the feature change trend is also taken into account, for example, by linear fitting analysis of the long-term trend of the feature value on the time axis is stable, rising or falling. A feature sequence with a clear drift trend will have a lower stability evaluation. Finally, these analysis results are integrated into a quantitative stability indicator, which is usually normalized to the range of 0 to 1. The higher the indicator value, the more stable the historical features of the node.

[0039] The time freshness indicator evaluates the value of the node data from the perspective of timeliness, which reflects the interval between the last update time of the target information corresponding to the node and the current processing time. The system maintains a last update timestamp for each node in the network, which records the time when the node features were last successfully updated or confirmed. When calculating freshness, the system obtains the current system time and calculates the difference between the last update timestamp, which directly reflects the degree of data obsolescence. In order to convert the time difference into an easy-to-compare score, the system usually uses an exponential decay function, which features that the closer the update to the current time, the slower the freshness score decays, while the more distant the update, the score drops sharply, which conforms to the general law of information value decay over time. The freshness score is also normalized. The final sorting process is a trio of matching degree score, feature stability indicator and time freshness indicator. The system assigns importance weights to these three indicators, and then calculates the final sorting score of each candidate node. Assuming that in a certain sorting, the matching degree score weight is 0.6, the feature stability weight is 0.25, and the time freshness weight is 0.15, then for Node_1024, if its stability indicator is 0.9 and its freshness indicator is 0.8, its final sorting score is: Through this comprehensive sorting method, the candidate target list output by the system not only considers the matching degree at the current time, but also takes into account the historical reliability of the target identity and the timeliness of the data, making the recognition result more robust and reliable.

[0040] Example 5: Calculation of feature stability indicator starts with extracting the historical feature sequence of the specified node from the target feature association network, which consists of the comprehensive feature description vector of the node continuously updated in the past a preset time window. The length of the time window is set according to the evaluation needs of stability in specific application scenarios, for example, when tracking high-speed maneuvering targets, the window may be short to be sensitive to changes, while when monitoring uniform motion targets, the window can be appropriately lengthened to observe long-term trends. The system retrieves all valid feature vector snapshots of the node at all time points within the window from the metadata storage of the network node, which are arranged in chronological order to form a feature value sequence, each data point in the sequence represents the complete feature state of the target at a certain historical time.

[0041] The extracted historical feature sequences need to be analyzed from multiple angles. Variance calculation is the most basic quantitative analysis method, and the variance value reflects the size of the fluctuation of each dimension of the feature vector around its historical average value. The calculation process is performed for each dimension of the feature vector. First, the average value of the dimension in all historical snapshots is calculated, then the deviation square of the dimension value of each snapshot from the average value is calculated, and finally the average of these squared differences is calculated. A larger variance value indicates that the feature dimension is unstable in historical observations, which may be caused by changes in the target state, measurement noise or environmental interference. However, variance is an absolute value with dimensions, which is not convenient for comparison between different orders of magnitude of feature dimensions, so it is necessary to further calculate the coefficient of variation. The coefficient of variation is obtained by dividing the standard deviation of each feature dimension by its average value, which is a dimensionless relative index that eliminates the influence of the size of the feature value on the dispersion degree judgment. For example, a feature dimension with an average value of 100 and a standard deviation of 10 has the same coefficient of variation as a feature dimension with an average value of 10 and a standard deviation of 1, which indicates that they have a relatively consistent dispersion degree. The system will calculate the coefficient of variation for each feature dimension of the node, thereby obtaining a detailed portrait of the stability of each feature of the node. However, relying only on static statistics such as variance and coefficient of variation is still insufficient to fully evaluate stability, and analysis of feature change trends provides a dynamic perspective.

[0042] Trend analysis focuses on the long-term trend of feature values over time. The system uses linear fitting method to fit the historical sequence of each feature dimension, and obtains a straight line representing the change trend. The slope of the straight line has important implications. A slope close to zero indicates that the feature dimension remains stable during the observation period without significant systematic drift. A significantly positive or negative slope indicates that the feature has a persistent upward or downward trend, which may be due to changes in the target state (such as acceleration, turning) or slow changes in sensor measurement conditions. In addition to linear trends, the system also focuses on whether there are mutation points in the sequence, i.e. sharp jumps in feature values within a short period of time. Mutation points often indicate significant changes in the target state or measurement errors.

[0043] After the variance, coefficient of variation and trend analysis are completed, the system needs to synthesize these scattered indicators into a single, quantitative characteristic stability indicator. The synthesis process first calculates a preliminary stability score for each characteristic dimension, which is usually a function of the inverse of the coefficient of variation and the absolute value of the trend slope. The smaller the coefficient of variation and the closer the trend slope to zero, the higher the stability score of the dimension. Considering that different dimensions in the comprehensive feature description vector may have different importance for target recognition, the system assigns a weight to each dimension, for example, the weight of motion features in the tracking task may be higher. Finally, by weighted averaging all feature dimensions and possibly introducing a penalty term for the number of abrupt changes, a final stability indicator between zero and one is generated. The higher the indicator value, the more stable and reliable the target represented by the node performs in the observation period, and the higher the credibility of the current matching result. This stability indicator, together with the matching degree score and the time freshness indicator, constitutes the basis for a comprehensive evaluation of the candidate target, so that the recognition result is not only based on the current matching degree, but also fully considers the behavior consistency in the past period of time.

[0044] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A radar target identification method, characterized in that, The method includes: The original echo signal sequence collected by the radar sensor is acquired, and the signal preprocessing unit is used to perform noise suppression and pulse compression on the original echo signal sequence to generate preprocessed radar echo data. Based on the preprocessed radar echo data, the time-frequency features, scattering features and motion trajectory features of the target are calculated by a multi-dimensional feature extraction unit, and then fused to generate a comprehensive feature description vector of the target. A dynamically updated target feature association network is constructed, where network nodes represent feature sets of different targets, network edges represent spatiotemporal correlations or feature similarities between targets, and an online learning mechanism is used to continuously optimize the network topology. Based on the comprehensive feature description vector acquired in real time, a hierarchical focusing mechanism is used in the target feature association network to gradually locate potential target regions; Within the potential target region, a set of candidate targets is selected using a multi-dimensional similarity measurement algorithm, and the candidate targets are sorted and output according to the feature consistency criterion.

2. The radar target identification method according to claim 1, characterized in that, The process of acquiring the raw echo signal sequence collected by the radar sensor, and using a signal preprocessing unit to perform noise suppression and pulse compression on the raw echo signal sequence to generate preprocessed radar echo data includes: An adaptive filtering algorithm is used to eliminate background noise in the original echo signal sequence, and pulse compression processing is used to enhance signal resolution. The processed signal is analyzed in the time and frequency domains to extract its energy distribution characteristics and phase modulation characteristics; Generate preprocessed radar echo data containing signal strength, frequency offset characteristics, and phase information.

3. The radar target identification method according to claim 2, characterized in that, The preprocessed radar echo data is used to calculate the target's time-frequency features, scattering features, and trajectory features through a multi-dimensional feature extraction unit, and these features are then fused to generate a comprehensive feature description vector of the target, including: Extract the time-frequency transformation features of the target from the preprocessed radar echo data, including instantaneous frequency features and spectral envelope features; Calculate the radar cross-section characteristics and polarization scattering matrix characteristics of the target; The motion trajectory parameters of the target, including velocity vector and acceleration characteristics, are analyzed by correlation analysis of multi-frame data. After normalizing the time-frequency features, scattering features, and motion trajectory features, they are fused into a fixed-dimensional comprehensive feature description vector.

4. The radar target identification method according to claim 3, characterized in that, The construction of a dynamically updated target feature association network, where network nodes represent feature sets of different targets, network edges represent spatiotemporal correlations or feature similarities between targets, and an online learning mechanism is used to continuously optimize the network topology, includes: The comprehensive feature description vector of each target is used as the node feature representation in the target feature association network; Establish connection edges between nodes based on the time sequence and spatial distribution of the target's appearance, and calculate the weight coefficients of the edges based on feature similarity; An incremental update strategy is adopted to dynamically adjust the network structure. When new target data is added, only the affected nodes and their associated edges are updated. The network connection weights and topology are optimized based on historical data interaction patterns through an online learning mechanism.

5. The radar target identification method according to claim 4, characterized in that, The step of gradually locating potential target regions in the target feature association network using a hierarchical focusing mechanism based on the comprehensive feature description vector acquired in real time includes: Map the real-time comprehensive feature description vector to the feature space of the target feature association network; The initial set of nodes in the network that best matches the real-time vector is used as the starting point for the search; Expand from the initial node to adjacent nodes, and filter relevant nodes layer by layer according to the feature similarity threshold; When the number of expansion layers reaches the preset depth or the node similarity is below the threshold, the expansion stops, and the sub-network formed by the selected nodes is taken as the potential target region.

6. The radar target identification method according to claim 5, characterized in that, Within the potential target region, a candidate target set is selected using a multi-dimensional similarity measurement algorithm, and the candidate targets are ranked and output according to a feature consistency criterion, including: Calculate the multi-dimensional similarity between the features of each node within the potential target region and the real-time integrated feature description vector, including time-frequency feature similarity, scattering feature similarity, and motion feature similarity; A weighted fusion algorithm is used to combine multi-dimensional similarity scores to generate an overall matching score for each node; Select nodes with an overall matching score higher than a preset threshold as the candidate target set; The candidate target set is sorted according to the matching score, feature stability, and time freshness.

7. The radar target identification method according to claim 6, characterized in that, The weighted fusion algorithm is used to integrate multi-dimensional similarity scores to generate an overall matching score for each node, including: Dynamic weighting coefficients are assigned to time-frequency feature similarity, scattering feature similarity, and motion feature similarity, respectively; The weighting of similarity across dimensions is adjusted based on target type and environmental characteristics; The overall matching score for each node is calculated using linear weighting.

8. The radar target identification method according to claim 7, characterized in that, The process of sorting the candidate target set based on matching score, feature stability, and time freshness includes: Calculate the feature stability index for each candidate target node to reflect the degree of variation in the node's historical features; Evaluate the time freshness index of each candidate target node, which reflects the interval between the node's most recent update time and the current time; The final ranking score is generated by combining the matching degree score, feature stability index, and time freshness index. The time freshness metric for evaluating each candidate target node, reflecting the interval between the node's most recent update time and the current time, includes: Record the last update timestamp for each node; Calculate the time difference between the current time and the last update time; An exponential decay function is used to convert the time difference into a freshness score.

9. A radar target identification method according to claim 8, characterized in that, The calculation of the feature stability index for each candidate target node, reflecting the degree of variation in the node's historical features, includes: Extract the feature change sequence of the node within the past time window; Calculate the variance and coefficient of variation of the characteristic sequence; Stability quantification indicators are generated by combining characteristic change trend analysis.

10. A radar target identification system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the radar target identification method according to any one of claims 1 to 9.

Images