Low altitude unmanned aerial vehicle radar detection method based on clutter multi-dimensional clustering extended features

Abstract

This invention belongs to the field of radar signal processing and target detection technology, and specifically relates to a low-altitude UAV radar detection method based on clutter multidimensional clustering extended features. The method includes the following steps: pulse Doppler processing of radar echo data to obtain a range-Doppler image; designing a detection threshold based on prior knowledge of clutter amplitude distribution, performing preliminary detection on the range-Doppler image to obtain preliminary detection points; clustering the preliminary detection points to group multiple detection points belonging to the same physical target into clusters, removing discrete noise points to form target point clusters; for each clustered target point cluster, extracting a multidimensional feature vector characterizing its geometry and scattering intensity, wherein the multidimensional feature vector includes geometric features based on ellipse fitting, RCS features, and harmonic frequency features; and inputting the multidimensional feature vector into a pre-trained random forest classifier to complete the final differentiation and identification of the target and clutter.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and target detection technology, and specifically relates to a low-altitude UAV radar detection method based on clutter multidimensional clustering extended features. Background Technology

[0002] Low-altitude flying targets, especially small rotary-wing drones and birds, pose an increasingly serious threat to national defense, civil aviation, and the protection of critical infrastructure. Radar, with its all-weather, all-day operation capability, has become an important means of detecting such targets. However, the complex and high-intensity background clutter in the low-altitude environment, coupled with the small radar cross-section (RCS) and low signal-to-clutter ratio (SCR) of targets, presents traditional radar detection methods with the following severe challenges:

[0003] 1. Easy to miss: The echo signal of small targets is weak and is easily drowned out by strong clutter.

[0004] 2. High false alarm rate: The characteristics of clutter are highly variable, which can easily cause the detection system to misjudge strong clutter as the target.

[0005] 3. Difficulty in identification: Traditional detection methods often only provide point information and lack effective features for distinguishing target types (such as drones, birds, clutter).

[0006] Existing methods still rely on single features or a limited set of features. The lack of effective utilization of multiple features, such as spatial spread, amplitude characteristics, and flapping frequency, means there is still significant room for improvement in low-altitude target detection.

[0007] Therefore, there is an urgent need for an innovative method that can effectively utilize the multidimensional features of low-altitude targets in a strong cluttered background to achieve high-reliability detection and low false alarm rate identification. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a low-altitude UAV radar detection method based on clutter multidimensional clustering extended features. This method systematically analyzes and extracts the differences between targets and clutter in amplitude distribution and multidimensional spatial extension features, and combines advanced clustering and machine learning algorithms to achieve high detection rate, low false alarm rate, and accurate classification and identification of low-altitude, weak targets against strong clutter backgrounds.

[0009] The technical solution for implementing the present invention is as follows: In a first aspect, the present invention provides a low-altitude unmanned aerial vehicle (UAV) radar detection method based on clutter multidimensional clustering extended features, comprising the following steps: Step 1: Radar echo preprocessing and extended target detection Pulse Doppler processing is performed on radar echo data to obtain range-Doppler images; based on prior knowledge of clutter amplitude distribution, a detection threshold is designed to perform preliminary detection on the range-Doppler images and obtain preliminary detection points; Step 2: Extended Target Agglomeration Based on Density Clustering Clustering is performed on the initial detection points to group multiple detection points belonging to the same physical target into a cluster, and discrete noise points are removed to form target point clusters; Step 3: Extraction of Multidimensional Differential Features For each cluster of target points after clustering, a multi-dimensional feature vector describing its geometry and scattering intensity is extracted. The multi-dimensional feature vector includes geometric features based on ellipse fitting, RCS features, and harmonic frequency features. Step 4: Target classification and recognition based on ensemble learning The multidimensional feature vector is input into a pre-trained random forest classifier to complete the final distinction and recognition between the target and clutter.

[0010] Optionally, the geometric features based on ellipse fitting described in this invention include: major axis length, minor axis length, eccentricity, and deflection angle; the deflection angle is the angle between the major axis of the fitted ellipse and the horizontal axis of the distance-Doppler image plane.

[0011] Optionally, the specific process of scattering intensity feature extraction in this invention is as follows: convert the amplitude value A of each point in the cluster into an RCS value σ; calculate the RCS statistics of all points in the cluster to form two intensity features: average RCS and maximum RCS.

[0012] Optionally, the present invention converts the amplitude value A of each point in the cluster into the RCS value σ according to the following formula;

[0013] in, This is the distance from the point to the radar. This represents the peak power of the radar. The pulse width of the radar transmitted signal. For scanning loss, For the transmit antenna gain, For the equivalent antenna area, To accumulate pulse count for coherent processing, Boltzmann's constant, The system noise temperature, For system losses, For RCS, The average amplitude of the noise can be obtained by averaging over all points when statistically analyzing the clutter amplitude distribution in step one. It is the signal-to-noise ratio at that point.

[0014] Optionally, the process for extracting harmonic frequency features according to the present invention is as follows: Location Analysis Unit: For each cluster of target points obtained in step two, locate the point with the strongest scattering intensity and record the distance unit coordinates of that point. ; Extracting distance slices: Extracting from distance cells The above refers to a one-dimensional range image sequence that spans multiple coherent processing intervals or a long time window, i.e., a slow time series signal; Micro-Doppler analysis: Perform a short-time Fourier transform on the above slow time series signal to obtain its time spectrum; in the time spectrum, the periodic modulation sidebands appearing on both sides of the main lobe are the micro-Doppler features, and their modulation frequency is... This refers to the frequency of the target's flapping wings or rotor rotation. Frequency feature extraction: In the time-frequency spectrum, find the main peak and secondary peak near zero frequency, and calculate the frequency difference between the secondary peak and the main peak. ; Feature Records: The most significant frequency differences extracted As a new feature of this target point cluster, harmonic frequencies are also included.

[0015] Optionally, for each target point cluster, the present invention combines the extracted features into a 9-dimensional feature vector: [centroid distance, centroid Doppler, major axis length, minor axis length, eccentricity, deflection angle, average RCS, maximum RCS, harmonic frequency].

[0016] Optionally, in step four of this invention, a random forest model is used, and the specific training process is as follows: Autosample: To train each decision tree, a subset is randomly selected with replacement from the total training set; Random feature selection: When constructing each node of each tree, a subset of features is randomly selected first, and then the optimal split is found only in this subset; Generate decision trees: Based on the sampled data and randomly selected features, grow each decision tree completely; After training, the result is a "forest" model composed of multiple decision trees.

[0017] Optionally, in step two of the present invention, the DBSCAN density clustering algorithm is used to aggregate multiple detection points belonging to the same physical target, while sparse and discrete clutter points are marked as noise and removed.

[0018] Secondly, the present invention provides a low-altitude unmanned aerial vehicle (UAV) radar detection device based on clutter multidimensional clustering extended features, comprising: The radar echo preprocessing and extended target detection module is used to perform pulse Doppler processing on radar echo data to obtain range-Doppler images; based on prior knowledge of clutter amplitude distribution, a detection threshold is designed to perform preliminary detection on the range-Doppler images and obtain preliminary detection points; The extended target agglomeration module based on density clustering is used to perform clustering processing on the initial detection points, agglomerating multiple detection points belonging to the same physical target into a cluster, eliminating discrete noise points, and forming target point clusters; The multidimensional differential feature extraction module is used to extract multidimensional feature vectors that characterize the geometry and scattering intensity of each clustered target point cluster. The multidimensional feature vectors include geometric features based on ellipse fitting, RCS features, and harmonic frequency features. The target classification and recognition module based on ensemble learning is used to complete the final distinction and recognition of targets and clutter based on the input of the multi-dimensional feature vector. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is the overall flowchart for radar signal processing.

[0021] Figure 2 shows the fitting results of the clutter probability distribution. Among them, (a) is the probability density plot of the five fitted distributions and the original distribution, and (b) is the cumulative distribution plot of the five fitted distributions and the original distribution.

[0022] Figure 3 shows the results of target detection using the new threshold. (a) shows the results using the traditional cell average CFAR detection method, and (b) shows the detection results using the new threshold designed by the method of this invention.

[0023] Figure 4 This is a diagram showing the clustering results.

[0024] Figure 5 This is a schematic diagram for extracting the shape features of point clusters.

[0025] Figure 6 This is a diagram showing the results after classification. Detailed Implementation

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0029] like Figure 1-6 As shown in the figure, an embodiment of this application provides a low-altitude UAV radar detection method based on clutter multidimensional clustering extended features. The specific process is as follows: Step 1: Radar echo preprocessing and extended target detection The purpose of this step is to convert the raw radar echo data into a range-Doppler (RD) image that clearly shows the target, and to design a detection threshold that is superior to traditional methods by utilizing prior statistical knowledge of clutter, so as to perform preliminary detection and suppress background clutter to the maximum extent while preserving the real target.

[0030] Step 1.1 Radar echo preprocessing and RD image generation First, raw echo data is acquired using a high-resolution pulse Doppler (PD) radar system. The raw echo data is the baseband complex signal (I / Q signal) received by the radar.

[0031] The raw echo data is processed using the following sequence: (1) Matched filtering (Pulse Compression): The echo signal in each pulse is pulse compressed. The copy of the transmitted signal (reference signal) is used as the coefficient of the matched filter to perform convolution operation on the echo signal, thereby improving the resolution and signal-to-noise ratio of the distance dimension.

[0032] (2) Clutter Suppression: Using Moving Target Indication (MTI) or Moving Target Detection (MTD) filter banks to process slow time dimension (pulse dimension) data, suppressing static or slow clutter such as ground objects and meteorological data, while retaining moving target echoes.

[0033] (3) Range Walk Correction: For high-speed moving targets, their echo envelopes will span different range cells between adjacent pulses; this phenomenon is called range walk. The Keystone transform is used to correct the data. The Keystone transform eliminates the coupling between range walk and velocity by resampling in the range-frequency domain, allowing the target energy to accumulate within the same range cell during subsequent accumulation. The core of the transform formula is the introduction of a new slow-time variable. ,in For slow time, For carrier frequency, For distance frequency.

[0034] (4) Coherent Integration: The data after the above processing is subjected to Fourier transform (i.e., Doppler processing) along the slow time dimension to accumulate the signal energy in the Doppler dimension, generating the final high-resolution range-Doppler (RD) image. The horizontal axis of this image is the range cell, the vertical axis is the Doppler cell (corresponding to radial velocity), and the pixel value represents the signal energy (amplitude) in the range-velocity cell.

[0035] After the above preprocessing, the energy of low-altitude weak targets (such as drones) is concentrated, which appears as obvious bright spots on the RD image.

[0036] Step 1.2 Extended Target Detection Based on Clutter Prior Traditional constant false alarm rate (CFAR) detectors estimate local clutter levels using reference cells around the detection unit when the clutter distribution is unknown, resulting in significant performance degradation in low-altitude, high-clutter environments. This invention utilizes pre-acquired clutter prior knowledge to directly design a superior detection threshold.

[0037] The specific implementation steps are as follows: (1) Clutter Data Acquisition and Statistics: When there are no targets flying in the detection area, a large amount of radar echo data is acquired, and RD images are generated through the preprocessing steps described in step 1.1. The amplitude values ​​of all non-target cells in these RD images are recorded. This constitutes a clutter amplitude sample dataset.

[0038] (2) Clutter distribution modeling and fitting: modeling and fitting clutter amplitude samples Statistical analysis was performed. Several typical statistical distribution models (such as Rayleigh, Log-Normal, Weibull, Gamma, and Exponential) were selected to fit the distribution. The probability density functions (PDFs) of each distribution are shown in the table below: Table 1. Mathematical Forms of Five Typical Distribution Models

[0039] (3) Optimal distribution test: The Kolmogorov-Smirnov (KS) test is used as the goodness-of-fit criterion. The KS test calculates the maximum vertical distance between the empirical cumulative distribution function (ECDF) and the theoretical cumulative distribution function (CDF). As a statistic, calculate the statistic for each fitted distribution. value, The distribution with the smallest value is the optimal fit distribution. Given a significance level... If the smallest If the value is less than the corresponding critical value, the clutter data is considered to follow the optimal distribution. The test threshold is: (1) in, At the significance level, n and m Let be the sample size for each of the two samples. Five fitted distributions. The values ​​are shown in the table below. It can be seen that the Weibull distribution best fits the probability distribution of clutter.

[0040] Table 2 shows the fitting of five distributions. value

[0041] (4) Detection threshold calculation: Based on the optimal clutter distribution model obtained above and the preset expected false alarm rate of the system. Directly calculate the detection threshold T Taking the Weibull distribution as the optimal distribution as an example, its threshold calculation formula is: (2) in, and The proportions and shape parameters of the Weibull distribution are obtained through fitting.

[0042] (5) Global detection: The calculated static threshold T Apply to the entire RD image (or apply in blocks based on clutter region characteristics). Measure the signal amplitude of each cell in the RD image. XWith threshold T Compare. If X ≥ T If the condition is met, the unit is determined to be a preliminary detection point, and its distance and Doppler coordinates are recorded; otherwise, it is determined to be clutter and is removed.

[0043] Results: To evaluate the detection and recognition module built based on the collected data and to better compare traditional algorithms with the algorithm proposed in this invention, false alarm rate and recognition rate were selected as the main performance indicators. False alarm rate is the probability that clutter signals are detected as targets (both drones and birds are considered targets), expressed by the formula: (3) in, This represents the number of units whose clutter signals are detected as targets. This represents the total number of units. Depending on the steps taken, the number of units... It will constantly change. For example, after passing through a traditional CFAR detector, n false alarms refer to all clutter points that are mistakenly identified as targets. However, in this experiment, the final... This refers to the total number of clutter units that are still identified as targets after passing through a random forest classifier.

[0044] The recognition rate is the probability that the final target can be correctly identified; it is the proportion of correctly classified target point clusters, expressed by the formula: (4) in, It is the number of target point clusters that were correctly identified. It represents the total number of clusters after clustering.

[0045] This threshold detection based on prior knowledge, compared to traditional CFAR detection, can reduce the false alarm rate from... Reduce to This significantly suppresses a large amount of clutter, providing a "cleaner" set of input points for subsequent clustering processing.

[0046] Step 2: Extended Target Agglomeration Based on Density Clustering After processing in step one, we obtained a binarized detection point map. However, in low-altitude, strong clutter environments, even with a detection threshold based on prior knowledge, a large number of discrete clutter points (false alarms) still exist in the map. More importantly, real targets under high-resolution radar (such as UAVs) often have their echo energy distributed across multiple adjacent range-Doppler cells due to their physical size or motion characteristics (such as rotor rotation), forming an "extended target," which appears as a cluster of multiple detection points on the detection map.

[0047] The core objective of this step is to utilize the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) density clustering algorithm to cluster multiple detection points belonging to the same physical target, while simultaneously marking sparse, discrete clutter points as noise and removing them. This not only effectively suppresses false alarms but also transforms the discrete "point target" representation into a coherent "extended target" representation, laying the foundation for subsequent extraction of features characterizing the target's shape and structure.

[0048] The specific implementation process is as follows: Step 2.1 Algorithm Input and Parameter Settings The binary detection image output from step one is used as the input for this step. Each detected point (pixel) in the image is represented by its coordinates (…). r , ) indicates that among them r Number the distance unit. The Doppler cells are numbered. All these points constitute the two-dimensional spatial point set to be clustered.

[0049] The DBSCAN algorithm requires setting two key parameters: Neighborhood radius (Eps, ): The distance threshold used to determine whether two points are "density reachable points" of each other.

[0050] Minimum number of points (MinPts): The minimum number of core points required to form a "dense region" (i.e., a cluster).

[0051] Step 2.2 Clustering Execution and Noise Filtering The DBSCAN algorithm traverses the point set based on the concept of "density reachability," and its execution process is briefly described below: (1) The algorithm randomly selects an unvisited point.

[0052] (2) Find all points in the neighborhood of the given point Eps.

[0053] (3) If the number of points in the neighborhood is not less than MinPts, then mark the point as a core point and create a new cluster based on it. Then the algorithm will recursively search for all points that are density reachable from the core point (i.e., the extended point set of the same target) and add them to this cluster.

[0054] (4) If the number of points in the neighborhood is less than MinPts, the point is temporarily marked as a noise point.

[0055] (5) Repeat the above process until all points have been visited.

[0056] Ultimately, the algorithm's output falls into two categories: Several clusters: Each cluster contains several points that are geographically close together and have a high density, and are identified as belonging to the same physical target (such as a drone).

[0057] Noise point set: Sparse and isolated points that cannot be classified into any cluster are judged as discrete clutter false alarms and are directly eliminated.

[0058] Step 2.3 Output and Effects After DBSCAN clustering, the system no longer outputs a large number of discrete detection points, but instead outputs several clustered target point clusters. Each cluster represents a detected potential target and has its unique identifier. Following clustering, we further improved the false alarm rate to [missing information]. This further reduces clutter interference.

[0059] Step 3: Extraction of Multidimensional Differential Features After the DBSCAN clustering process in step two, each physical target (such as a drone, bird, or residual clutter) has been clustered into an independent point cluster. However, clustering alone is insufficient to distinguish the categories of these clusters. The purpose of this step is to extract a series of digital features from each target point cluster that can quantify its geometry and scattering intensity, forming a multi-dimensional feature vector to accurately characterize the essential differences between different target categories (drones, birds, clutter).

[0060] These features are mainly divided into two categories: geometric shape features and scattering intensity features. The extraction process is as follows: Step 3.1 Geometric feature extraction (based on ellipse fitting) To accurately describe the spatial distribution and shape of a point cluster in the range-Doppler plane, an ellipse fitting method is used to characterize its geometric properties. This method better reflects the actual distribution of the point cluster than a simple bounding box. Each fitted ellipse generates four key features: (1) Major Axis Length: The diameter of the major axis of the fitted ellipse. This feature directly reflects the main expansion scale of the target in the RD plane.

[0061] (2) Minor Axis Length: The diameter of the minor axis of the fitted ellipse. This feature reflects the expansion scale of the target in the minor direction and, when combined with the major axis, can characterize the overall size of the target.

[0062] (3) Eccentricity: Used to describe how close the shape of an ellipse is to a circle. The formula is as follows: This feature reflects the density of points within a cluster. A compact cluster (such as certain clutter) has an eccentricity close to 0, while a narrow cluster (such as a flapping bird) has an eccentricity close to 1.

[0063] (4) Orientation: The angle between the major axis of the fitted ellipse and the horizontal axis (usually the distance axis) of the RD plane. This feature reveals the dominant direction of the target's extension and may be related to the target's motion posture.

[0064] In addition, the centroid coordinates (range mean and Doppler mean) of the point cluster are also an important feature, providing core information about the target's average distance and average radial velocity.

[0065] Step 3.2 Scattering Intensity Feature Extraction (Based on RCS) The amplitude value of each point in the cluster contains information about the target's scattering intensity. To eliminate the influence of range and facilitate comparison between different targets, the amplitude value is converted into a more physically meaningful radar cross section (RCS).

[0066] (1) RCS Conversion: According to the radar equation, the amplitude value A of each point in the cluster is converted into the RCS value σ. The radar equation is shown below. (5) in, This is the distance from the point to the radar. This represents the peak power of the radar. The pulse width of the radar transmitted signal. For scanning loss, For the transmit antenna gain, For the equivalent antenna area, To accumulate pulse count for coherent processing, Boltzmann's constant, The system noise temperature, For system losses, For RCS, The average amplitude of the noise can be obtained by averaging over all points when statistically analyzing the clutter amplitude distribution in step one. It is the signal-to-noise ratio at that point.

[0067] (2) Feature statistics: Calculate the RCS statistics of the entire point cluster to form two intensity features. Mean RCS: The arithmetic mean of the RCS values ​​of all points within a cluster. This feature reflects the average scattering intensity of the target and is closely related to the target's size and material.

[0068] Maximum RCS: The maximum RCS value among all points within a cluster. This feature reflects the intensity of the target's strongest scattering center and is particularly important for distinguishing targets with highly reflective components (such as UAV bodies or metal motors).

[0069] Step 3.3 Harmonic Frequency Feature Extraction Flapping frequency and its harmonics are a unique and efficient feature for distinguishing biological targets (such as birds and insects) from mechanical targets (such as drones). The rotation of a drone rotor also produces similar harmonic characteristics. This feature reflects the micro-Doppler effect introduced by the periodic motion of the target component (wings or rotor).

[0070] This invention extracts this feature by analyzing the micro-Doppler in radar echoes. The specific steps are as follows: (1) Location Analysis Unit: For each target point cluster obtained in step two, first locate the point with the strongest scattering intensity (usually the point corresponding to the maximum RCS, or the centroid), and record its precise distance unit coordinates. .

[0071] (2) Range slice extraction: Taking advantage of the high temporal resolution of the radar system, range slices are extracted from the range cells. The signal is a one-dimensional range image sequence (i.e., a slow time series signal) spanning multiple coherent processing intervals (CPI) or a long time window. This signal contains phase modulation information of the target micro-motion component.

[0072] (3) Micro-Doppler analysis: Perform a short-time Fourier transform (STFT) on the above slow time series signal to obtain its time spectrum. In the time spectrum, the periodic modulation sidebands appearing on both sides of the main lobe are the micro-Doppler features, and their modulation frequency is... This refers to the frequency of the target's flapping wings or rotor rotation.

[0073] (4) Frequency Feature Extraction: In this distribution map, besides the main peak near zero frequency (representing the target's average radial velocity), look for secondary peaks. The frequency difference between these secondary peaks and the main peak... This refers to the target's fundamental frequency (wing flapping / rotation frequency) or its harmonic frequencies (usually the second harmonic is the most prominent).

[0074] (5) Feature recording: The most significant harmonic frequency value extracted is used as a new feature of the target point cluster.

[0075] Step 3.4 Feature Vector Construction For each cluster of target points, we combine the features extracted above into a 9-dimensional feature vector: [centroid distance, centroid Doppler, major axis length, minor axis length, eccentricity, deflection angle, average RCS, maximum RCS, harmonic frequency].

[0076] The feature vector set output in this step will serve as the input to the final step four—the random forest classifier—to complete the final identification of the target.

[0077] Step 4: Target classification and recognition based on ensemble learning After step three, each cluster of target points to be identified is transformed into a fixed-dimensional multi-dimensional feature vector. The core task of this step is to construct a high-performance classifier that can accurately and reliably classify targets into preset categories (such as "drones," "birds," and "clutter") based on these feature vectors. To achieve this goal, this invention uses the Random Forest ensemble learning algorithm as the classifier, leveraging its high accuracy, strong robustness, and good anti-overfitting ability to address the complexity of low-altitude detection environments.

[0078] Step 4.1 Model Training Before the algorithm is actually deployed, an offline model training phase is needed to build the classifier: (1) Constructing a training dataset: Collect a large number of radar data samples with known category labels (including various types such as UAVs, birds, and clutter). For each sample data, perform the processing steps one to three in sequence to extract feature vectors for each target point cluster and assign them real category labels (e.g., "0" represents clutter, "1" represents birds, and "2" represents UAVs). All these "feature vector-label" pairs constitute the labeled training dataset.

[0079] (2) Training the Random Forest Model: Input the above training dataset into the Random Forest algorithm for training. During the training process, the algorithm will autonomously complete the following key operations: Bootstrap Sampling: To train each decision tree, a subset is randomly selected with replacement from the total training set.

[0080] Random feature selection: When constructing each node of each tree, instead of selecting the optimal splitting feature from all features, a subset of features is randomly selected first, and then the optimal splitting feature is searched only within this subset. This step is crucial for ensuring tree diversity in the forest.

[0081] Generate decision trees: Based on the sampled data and randomly selected features, grow each decision tree completely (usually without pruning).

[0082] Ultimately, the training process yields a "forest" model composed of multiple decision trees.

[0083] Step 4.2 Target Classification and Recognition In actual radar systems, the target point clusters obtained through real-time processing are classified: (1) Feature vector input: Input the feature vector extracted for the current target point cluster in step three into the trained random forest model.

[0084] (2) Forest decision: Each decision tree in the random forest independently classifies the feature vector and gives a category prediction result.

[0085] (3) Voting and Output: The prediction results of all decision trees are statistically analyzed, and the final category label is determined by majority voting. That is, the category that receives the most votes is taken as the final identification result of the target point cluster.

[0086] (6) Thus, through the sequential processing of steps one through four, this invention completes a comprehensive low-altitude UAV radar detection and identification process, from the raw echo to the final target category output. Ultimately, the false alarm rate of the entire system is only [percentage missing]. The success rate of detection and identification is as high as 93.33%.

[0087] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-altitude UAV radar detection method based on clutter multidimensional clustering extended features, characterized in that, Includes the following steps: Step 1: Radar echo preprocessing and extended target detection Range-Doppler images are obtained by pulse Doppler processing of radar echo data; Based on prior knowledge of clutter amplitude distribution, a detection threshold is designed to perform preliminary detection on the range-Doppler image and obtain preliminary detection points. Step 2: Extended Target Agglomeration Based on Density Clustering Clustering is performed on the initial detection points to group multiple detection points belonging to the same physical target into a cluster, and discrete noise points are removed to form target point clusters; Step 3: Extraction of Multidimensional Differential Features For each cluster of target points after clustering, a multi-dimensional feature vector describing its geometry and scattering intensity is extracted. The multi-dimensional feature vector includes geometric features based on ellipse fitting, RCS features, and harmonic frequency features. Step 4: Target classification and recognition based on ensemble learning The multidimensional feature vector is input into a pre-trained random forest classifier to complete the final distinction and recognition between the target and clutter.

2. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, The geometric features based on ellipse fitting include: major axis length, minor axis length, eccentricity, and deflection angle; the deflection angle is the angle between the major axis of the fitted ellipse and the horizontal axis of the distance-Doppler image plane.

3. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 2, characterized in that, The specific process for extracting the scattering intensity features is as follows: convert the amplitude value A of each point in the cluster into the RCS value σ; calculate the RCS statistics of all points in the cluster to form two intensity features: the average RCS and the maximum RCS.

4. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 3, characterized in that, The amplitude value A of each point within the cluster is converted to the RCS value σ according to the following formula; in, This is the distance from the point to the radar. This represents the peak power of the radar. The pulse width of the radar transmitted signal. For scanning loss, For the transmit antenna gain, For the equivalent antenna area, To accumulate pulse count for coherent processing, Boltzmann's constant, The system noise temperature, For system losses, For RCS, The average amplitude of the noise.

5. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, The process for extracting the harmonic frequency features is as follows: Location Analysis Unit: For each cluster of target points obtained in step two, locate the point with the strongest scattering intensity and record the distance unit coordinates of that point. ; Extracting distance slices: Extracting from distance cells The above refers to a one-dimensional range image sequence that spans multiple coherent processing intervals or a long time window, i.e., a slow time series signal; Micro-Doppler analysis: Perform a short-time Fourier transform on the above slow time series signal to obtain its time spectrum; in the time spectrum, the periodic modulation sidebands appearing on both sides of the main lobe are the micro-Doppler features, and their modulation frequency is... This refers to the frequency of the target's flapping wings or rotor rotation. Frequency feature extraction: In the time-frequency spectrum, find the main peak and secondary peak near zero frequency, and calculate the frequency difference between the secondary peak and the main peak. ; feature Record: The most significant frequency difference extracted. As a new feature of this target point cluster, harmonic frequencies are also included.

6. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, For each cluster of target points, the extracted features are combined into a 9-dimensional feature vector: [centroid distance, centroid Doppler, major axis length, minor axis length, eccentricity, deflection angle, average RCS, maximum RCS, harmonic frequency].

7. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, Step four uses a random forest model, and the specific training process is as follows: Autosample: To train each decision tree, a subset is randomly selected with replacement from the total training set; Random feature selection: When constructing each node of each tree, a subset of features is randomly selected first, and then the optimal split is found only in this subset; Generate decision trees: Based on the sampled data and randomly selected features, grow each decision tree completely; After training, the result is a "forest" model composed of multiple decision trees.

8. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, In step two, the DBSCAN density clustering algorithm is used to cluster multiple detection points belonging to the same physical target, while sparse and discrete clutter points are marked as noise and removed.

9. A low-altitude unmanned aerial vehicle (UAV) radar detection device based on clutter multidimensional clustering extended features, characterized in that, include: The radar echo preprocessing and extended target detection module is used to perform pulse Doppler processing on radar echo data to obtain range-Doppler images; Based on prior knowledge of clutter amplitude distribution, a detection threshold is designed to perform preliminary detection on the range-Doppler image and obtain preliminary detection points. The extended target agglomeration module based on density clustering is used to perform clustering processing on the initial detection points, agglomerating multiple detection points belonging to the same physical target into a cluster, eliminating discrete noise points, and forming target point clusters; The multidimensional differential feature extraction module is used to extract multidimensional feature vectors that characterize the geometry and scattering intensity of each clustered target point cluster. The multidimensional feature vectors include geometric features based on ellipse fitting, RCS features, and harmonic frequency features. The target classification and recognition module based on ensemble learning is used to complete the final distinction and recognition between the target and clutter based on the input of the multi-dimensional feature vector.

10. The low-altitude UAV radar detection method based on clutter multidimensional clustering extended features according to claim 1, characterized in that, The geometric features based on ellipse fitting include: major axis length, minor axis length, eccentricity, and deflection angle; the deflection angle is the angle between the major axis of the fitted ellipse and the horizontal axis of the distance-Doppler image plane; the scattering intensity features include: average RCS and maximum RCS.

Images