Drone identification and tracking method combined with low-altitude radar AI algorithm

Abstract

The present invention discloses a method for identifying and tracking unmanned aerial vehicles (UAVs) in combination with a low-altitude radar AI algorithm, and relates to the field of radio positioning technology. The method comprises: determining the radar frequency band of the target scene, performing radar detection through a phased array antenna, and receiving radar echo information. Performing time-frequency domain conversion to obtain target frequency domain data, and extracting intuitive feature data of the UAV therefrom. Restore the image of the target UAV to obtain a target radar image. Perform feature engineering processing on the radar image using a preset image feature indicator set to obtain image feature data. Construct and train a UAV identification model based on an AI algorithm, input the intuitive feature data and image feature data into the model for target identification, and obtain an identification result. Based on the identification result, the UAV is continuously tracked to generate a tracking trajectory of the target UAV, including an experienced trajectory and an estimated trajectory. Thereby achieving the technical effect of improving the accuracy of UAV identification and tracking stability.

Description

Technical Field

[0001] The present invention relates to the field of radio positioning technology, and in particular to a method for identifying and tracking unmanned aerial vehicles (UAVs) in combination with a low-altitude radar AI algorithm. Background Art

[0002] With the rapid development of drone technology and the increase in the number of drones, the identification and analysis of drones is crucial to low-altitude safety. As an effective detection device, low-altitude radar can efficiently detect drones and other flying objects in low-altitude environments and provide real-time data. However, existing radar detection technology can only provide basic information such as the location and speed of objects, and there is a technical problem of lacking the ability to accurately identify and track drones in complex environments. Summary of the invention

[0003] The present invention provides a drone identification and tracking method combined with a low-altitude radar AI algorithm to solve the technical problem that the prior art lacks the ability to accurately identify and track drones in complex environments, and achieves the technical effect of improving drone identification accuracy and tracking stability.

[0004] The drone identification and tracking method combined with the low-altitude radar AI algorithm provided by the present invention includes:

[0005] The target radar frequency band of the target scene is determined, and radar detection is performed with the target radar frequency band through a phased array antenna to receive radar echo information.

[0006] The radar echo information is converted into time-frequency domain to obtain target frequency domain data, and the target frequency domain data is analyzed to extract intuitive features to obtain intuitive feature data of the target UAV.

[0007] The target UAV image is restored based on the intuitive feature data to obtain a target radar image, and the target radar image is subjected to feature engineering processing based on a preset image feature indicator set to obtain image feature data.

[0008] Construct and train a drone recognition model based on an AI algorithm, and input the intuitive feature data and the image feature data into the trained drone recognition model for target recognition to obtain a drone recognition result.

[0009] Continuously identify the target drone according to the drone identification result to obtain the tracking trajectory of the target drone, wherein the tracking trajectory includes an experienced trajectory and an estimated trajectory.

[0010] In a feasible implementation, determining a target radar frequency band of a target scene includes:

[0011] Interact with the drone management system of the target area to obtain regional drone inventory information.

[0012] Clustering is performed based on the drone inventory information in the area to obtain multiple drone sample information clusters.

[0013] The drone identification and tracking requirements in the target area are analyzed, the feature information of multiple drone sample information clusters is called as the target feature information, and the frequency band screening based on the resolution and detection range is performed to determine the target radar frequency band.

[0014] In a feasible implementation, the radar echo information is converted into a time-frequency domain to obtain target frequency domain data, and the target frequency domain data is analyzed to extract intuitive features to obtain intuitive feature data of the target UAV, including:

[0015] The radar echo information is subjected to matched filtering through a matched filter to obtain the echo time and calculate the distance characteristics of the target UAV.

[0016] Based on Fourier transform, the radar echo information is converted into time-frequency domain to obtain the target frequency domain data, analyze and extract the frequency deviation data of the target frequency domain data, and convert the frequency deviation data into the speed characteristics of the target UAV in combination with the Doppler effect.

[0017] The phase difference of the radar echo information is extracted to calculate and obtain the directional characteristics of the target UAV.

[0018] According to the matched filtering result, the echo power of the radar echo information is obtained, and the reflection intensity characteristics of the target UAV are calculated and obtained by combining the echo power with the direction characteristics.

[0019] In a feasible implementation, the image of the target UAV is restored based on the intuitive feature data to obtain a target radar image, and feature engineering processing is performed on the target radar image based on a preset image feature indicator set to obtain image feature data, including:

[0020] Based on image reconstruction technology, image restoration is performed according to the distance feature, the speed feature, the direction feature and the reflection intensity feature of the intuitive feature data to obtain a three-dimensional image and / or a two-dimensional image of the target UAV, which is output as the target radar image.

[0021] The target radar image is enhanced, and features are extracted based on the enhanced target radar image to generate the image feature data, wherein the image feature data at least includes texture features, shape features, and space features.

[0022] In one feasible implementation, a drone recognition model based on an AI algorithm is constructed and trained, including:

[0023] Taking the drone inventory information in the area as the matching target, homologous identification records are collected, and the homologous identification records are marked with true values ​​to obtain a sample data set, wherein the sample data set includes associated sample intuitive feature data, sample image feature data and sample drone identification results.

[0024] The sample data set is used as a sample space for random extraction to obtain multiple sample data subsets, wherein the multiple sample data subsets are true subsets of the sample data set.

[0025] Based on the multiple sample data subsets, multiple deep learning-based weak recognition networks are constructed and trained, wherein each of the weak recognition networks is trained by a randomly selected sample data subset.

[0026] The plurality of weak recognition networks are trained and optimized through an integrated learning method to obtain the drone recognition model.

[0027] In one feasible implementation, a drone recognition model based on an AI algorithm is built and trained, and then:

[0028] Based on the principal component analysis method, the contribution of the plurality of weak recognition networks to the correct recognition of records is analyzed and determined, and a contribution set is obtained, wherein the contribution association mark is marked with a recognition task category mark.

[0029] The recognition task category is marked as a clustering constraint, the contribution degree set is clustered and divided to obtain a plurality of working condition clusters.

[0030] The plurality of operating condition clusters are serialized, and a plurality of operating condition identification network sequences are constructed according to the serialization results, wherein the operating condition identification network sequence includes a plurality of the weak identification networks in descending order of contribution.

[0031] Traversing the plurality of working condition identification network sequences to perform cumulative contribution calculation, and extracting the plurality of weak identification networks that meet the cumulative contribution threshold, and outputting the plurality of classifier call forms.

[0032] In a feasible implementation, the method further includes:

[0033] Obtain model verification record data of the drone identification model.

[0034] Based on the model verification record data, principal component analysis is performed on the image feature index set to obtain key feature groups corresponding to multiple recognition task types, and the key feature groups are output as multiple feature index call forms.

[0035] Based on the plurality of feature indicator calling forms, feature calling optimization of the image feature indicator set is performed.

[0036] In a feasible implementation, the method further includes:

[0037] Based on the obtained drone identification result, the drone management system is accessed to extract the intrinsic parameter information of the target drone.

[0038] According to the intrinsic parameter information, the beam width and scanning frequency of the phased array antenna are dynamically optimized, and the target UAV is continuously identified according to the optimized beam width and scanning frequency.

[0039] The present invention discloses a method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm, including: determining the target radar frequency band of the target scene, and performing radar detection with the target radar frequency band through a phased array antenna to receive radar echo information; performing time-frequency domain conversion on the radar echo information to obtain target frequency domain data, and analyzing the target frequency domain data to perform intuitive feature extraction to obtain intuitive feature data of the target unmanned aerial vehicle; performing image restoration of the target unmanned aerial vehicle based on the intuitive feature data to obtain the target radar image, and performing feature engineering processing on the target radar image based on a preset image feature index set to obtain image feature data; constructing and training an unmanned aerial vehicle recognition model based on an AI algorithm, and inputting the intuitive feature data and the image feature data into the trained unmanned aerial vehicle recognition model for target recognition to obtain unmanned aerial vehicle recognition results; performing continuous recognition of the target unmanned aerial vehicle according to the unmanned aerial vehicle recognition results to obtain the tracking trajectory of the target unmanned aerial vehicle, wherein the tracking trajectory includes the experienced trajectory and the estimated trajectory. The unmanned aerial vehicle recognition and tracking method in combination with a low-altitude radar AI algorithm disclosed by the present invention solves the technical problem of lacking the ability to accurately identify and track unmanned aerial vehicles in complex environments, and achieves the technical effect of improving the accuracy of unmanned aerial vehicle recognition and the stability of tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the drone identification and tracking method combined with the low-altitude radar AI algorithm of the present invention;

[0041] Figure 2 The figure is a flow chart of constructing and training a drone recognition model in the drone recognition and tracking method combined with a low-altitude radar AI algorithm of the present invention. DETAILED DESCRIPTION

[0042] The technical solution provided in the embodiments of the present invention is to solve the technical problem that the prior art lacks the ability to accurately identify and track drones in complex environments. The overall idea adopted is as follows:

[0043] First, determine the target radar frequency band of the target scene, and perform radar detection with the target radar frequency band through a phased array antenna to receive radar echo information; then, perform time-frequency domain conversion on the radar echo information to obtain target frequency domain data, and analyze the target frequency domain data to perform intuitive feature extraction to obtain intuitive feature data of the target UAV; then, perform image restoration of the target UAV based on the intuitive feature data to obtain a target radar image, and perform feature engineering processing on the target radar image based on a preset image feature indicator set to obtain image feature data; then, construct and train a UAV recognition model based on an AI algorithm, and input the intuitive feature data and the image feature data into the trained UAV recognition model to perform target recognition to obtain a UAV recognition result; then, perform continuous recognition of the target UAV based on the UAV recognition result to obtain a tracking trajectory of the target UAV, wherein the tracking trajectory includes an experienced trajectory and an estimated trajectory.

[0044] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention are shown in the drawings, rather than all of them.

[0045] Figure 1 The figure is a flow chart of a method for identifying and tracking a drone using a low-altitude radar AI algorithm according to the present invention, wherein the method comprises:

[0046] The target radar frequency band of the target scene is determined, and radar detection is performed with the target radar frequency band through a phased array antenna to receive radar echo information.

[0047] Specifically, firstly, according to the recognition requirements of the target scene and the characteristics of the drone, the appropriate radar frequency band is selected to provide better resolution and target recognition capabilities. Among them, the radar system works in different frequency bands, which are defined by specific wavelengths and frequency characteristics, including ultra-high frequency (UHF), very high frequency (VHF), L-band, S-band, C-band, X-band and K-band, etc. Different bands have different resolutions.

[0048] Specifically, by controlling the phases of multiple antenna elements in the phased array antenna of the phased array radar, the direction of the radar beam can be quickly changed without physically moving the antenna, thereby achieving rapid scanning of the target scene and multi-target tracking. During radar detection, the phased array antenna emits electromagnetic waves at a predetermined target radar frequency band, which are reflected to form radar echoes when encountering target drones. The radar echoes are collected by the receiving part of the phased array radar and converted into electrical signals, namely radar echo information. The radar echo information contains important information about the target, such as its distance, speed, direction, and size.

[0049] In some embodiments, determining a target radar frequency band of a target scene includes:

[0050] Interact with the drone management system of the target area to obtain the regional drone inventory information. Perform clustering based on the regional drone inventory information to obtain multiple drone sample information clusters. Analyze the drone identification and tracking requirements in the target area, call the feature information of multiple drone sample information clusters as target feature information, perform frequency band screening based on resolution and detection range, and determine the target radar frequency band.

[0051] Specifically, first, the drone management system is interacted with through its API or other interfaces to obtain drone inventory information in the target area, including information such as the number, type, size, shape, and flight mode of drones. Then, the drones are clustered according to their size, shape, type, and other characteristics to obtain multiple drone sample information clusters, each of which contains information about multiple drones with similar shapes and behaviors.

[0052] Furthermore, based on the requirements for drone identification and tracking in the target area and the characteristic information of the drone sample information cluster, the selection of a suitable radar frequency band involves consideration of resolution and detection range. For example, if the drones in the target area are mainly small drones, a higher frequency band (such as K-band) is selected to provide high resolution. On the other hand, if the demand for detection range is more important than resolution, a radar frequency band with a larger detection range is selected.

[0053] Specifically, the selected drone sample information clusters are analyzed and extracted to obtain representative target feature information, including geometric shape features, reflection characteristics, motion pattern characteristics, electromagnetic signal characteristics, etc., to determine the appropriate radar band. The above process combines the physical characteristics of the radio frequency band with the requirements of target characteristics to ensure that the most appropriate frequency band can be selected for target identification and tracking in different environments and mission requirements.

[0054] The radar echo information is converted into time-frequency domain to obtain target frequency domain data, and the target frequency domain data is analyzed to extract intuitive features to obtain intuitive feature data of the target UAV.

[0055] Specifically, the intuitive feature data of the target UAV refers to the intuitive kinematic characteristics of the UAV, and preferably includes the distance characteristics, speed characteristics, direction characteristics and reflection intensity characteristics of the target UAV.

[0056] In some embodiments, the radar echo information is converted into time-frequency domain to obtain target frequency domain data, and the target frequency domain data is analyzed to extract intuitive features to obtain intuitive feature data of the target UAV, including:

[0057] The radar echo information is matched filtered by a matched filter to obtain the echo time and calculate the distance characteristics of the target UAV. The radar echo information is converted into the time-frequency domain based on Fourier transform to obtain the target frequency domain data, and the frequency deviation data of the target frequency domain data is analyzed and extracted, and the frequency deviation data is converted into the speed characteristics of the target UAV in combination with the Doppler effect. The phase difference of the radar echo information is extracted to calculate the direction characteristics of the target UAV. According to the matched filtering result, the echo power of the radar echo information is obtained, and the reflection intensity characteristics of the target UAV are calculated by combining the echo power and the direction characteristics.

[0058] Specifically, through matched filtering, the signal-to-noise ratio of the radar echo can be optimized, thereby improving the performance of the radar system. Among them, the matched filter is used to match the echo signal according to the known shape of the transmitted signal, and locate the echo information from the radar echo information containing the noise signal by calculating the correlation, thereby improving the signal-to-noise ratio and accurately extracting the time information of the echo. Through matched filtering, the echo time of the echo signal can be accurately calculated, that is, the time from the signal being transmitted by the radar to the time it is reflected back by the target;

[0059] Specifically, the radar echo signal is converted into the time-frequency domain based on the Fourier transform to obtain the frequency domain data of the target. In the radar signal, the Doppler effect will be generated due to the movement of the target, that is, the relative movement of the target causes the signal frequency to shift, and the frequency deviation data is directly related to the speed of the target. Furthermore, the frequency deviation data can be further analyzed in the frequency domain to infer the speed of the target. Exemplarily, first, a long-time Fourier transform is applied to the echo signal to obtain the spectrum of the signal to show the distribution of the signal under different frequency components; then, the frequency deviation data is extracted based on the acquired spectrum, and combined with the Doppler effect, the speed characteristics of the target drone are calculated, where the relationship between the Doppler frequency shift and the radial velocity of the target is expressed by the following formula:

[0060] ;

[0061] Among them, Δf is the frequency deviation, f0 is the radar's transmitting frequency, c is the speed of light, and v is the target's radial velocity (the component of the target's velocity toward the radar).

[0062] Specifically, the phased array antenna uses multiple antennas or receivers to receive echo signals, and calculates the phase difference of the echo signals received by different antennas, combined with the geometric information of the radar array, to infer the position and direction of the target. Among them, the phase difference of the echo signals received by different antennas reflects the difference in the signal propagation path, and then reveals the azimuth of the target, that is, the angular position of the target in the horizontal or vertical direction.

[0063] Specifically, according to the result of matched filtering, the echo power of the echo signal can be further extracted and combined with the directional characteristics of the target UAV to calculate the reflection intensity characteristics of the target. Among them, the echo power is a measure of the amplitude and signal energy of the echo signal, reflecting the intensity of the target reflected echo. The larger the echo power, the larger the reflection area of ​​the target, or the closer the target is to the radar. Among them, the reflection intensity characteristics are used to provide information for the physical properties of the target (such as the size, shape and surface characteristics of the target). Preferably, the radar cross section (RCS) of the target UAV is used as the reflection intensity characteristic.

[0064] Through the above steps, the radar echo signal is processed by matching filtering, Fourier transform, Doppler effect analysis, phase difference calculation, etc., and intuitive feature data such as the distance, speed, direction and reflection intensity of the target UAV are extracted, providing rich information for further target identification, classification and tracking.

[0065] The target UAV image is restored based on the intuitive feature data to obtain a target radar image, and the target radar image is subjected to feature engineering processing based on a preset image feature indicator set to obtain image feature data.

[0066] Specifically, by restoring the image of the target drone through image reconstruction technology and applying feature engineering processing, richer and more detailed feature information can be extracted, thereby improving the accuracy of target recognition and classification. At the same time, an intuitive target image can be provided to help users understand and interpret the output results of the radar system.

[0067] In some embodiments, image restoration of the target UAV is performed based on the intuitive feature data to obtain a target radar image, and feature engineering processing is performed on the target radar image based on a preset image feature indicator set to obtain image feature data, including:

[0068] Based on the image reconstruction technology, image restoration is performed according to the distance feature, the speed feature, the direction feature and the reflection intensity feature of the intuitive feature data to obtain a three-dimensional image and / or a two-dimensional image of the target UAV, which is output as the target radar image. The target radar image is enhanced, and feature extraction is performed based on the enhanced target radar image to generate the image feature data, wherein the image feature data at least includes texture features, shape features and spatial features.

[0069] Specifically, intuitive feature data (including distance features, speed features, direction features and reflection intensity features) are used to restore the image. Among them, the image reconstruction technology includes an inversion algorithm or an image reconstruction model to convert the signal (echo information) received by the radar into a visual image. Among them, the distance feature is used to determine the spatial position of the target and the target surface, the speed feature helps to identify the movement state of the target, the direction feature determines the direction of the target in space, and the reflection intensity feature can reveal the reflection characteristics of the target surface.

[0070] Optionally, the target radar image may be a two-dimensional plane image and / or a three-dimensional stereo image, which is specifically determined based on the processing power of the target scene and the resolution of the radar.

[0071] Specifically, the obtained target radar image is enhanced, including contrast enhancement, sharpening, denoising, etc. The purpose of image enhancement is to improve the quality of the image, making the edges and details of the target clearer for subsequent feature extraction. Exemplarily, the enhancement process is performed using a deep learning network, a traditional filtering method (such as Gaussian filtering, mean filtering, edge enhancement, etc.) or a mathematical method based on wavelet transform, Laplace transform, etc.

[0072] Specifically, after image enhancement, the target radar image is further extracted for feature extraction, and the acquired image feature data mainly include: Texture features, which are used to characterize the distribution and changes of surface texture in the image, such as gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), Haralick features, etc. Shape features, which are obtained through geometric analysis of the image's contour, boundary, and shape, including the contour shape, area, perimeter, rectangularity, and other features of the target drone, help to effectively distinguish the morphology of different target objects. Spatial features refer to the features of the spatial position, arrangement, relative distance, etc. of objects in the image, which are extracted through the spatial distribution of the image, local area density, and other methods.

[0073] The above process combines image reconstruction, image enhancement and feature extraction techniques to restore the radar image of the target UAV and extract the texture, shape and spatial features that are crucial for UAV identification and analysis, thereby providing richer data support for subsequent automated processing.

[0074] Construct and train a drone recognition model based on an AI algorithm, and input the intuitive feature data and the image feature data into the trained drone recognition model for target recognition to obtain a drone recognition result.

[0075] Specifically, the drone recognition model is used to identify drones based on intuitive feature data and image feature data. Exemplarily, the drone recognition results include the existence of drones, and the type, size, model, quantity, etc. of drones. The drone recognition model based on AI algorithms helps to fully explore and construct the complex mapping relationship between feature data and drone recognition results, thereby effectively achieving the target recognition task.

[0076] In some embodiments, Figure 2 As shown, a drone recognition model based on AI algorithm is built and trained, including:

[0077] Taking the regional drone inventory information as the matching target, collect homologous identification records, and mark the homologous identification records with true values ​​to obtain a sample data set, wherein the sample data set includes associated sample intuitive feature data, sample image feature data, and sample drone identification results. Randomly extract the sample data set as the sample space to obtain multiple sample data subsets, wherein the multiple sample data subsets are true subsets of the sample data set. Based on the multiple sample data subsets, construct and train multiple weak recognition networks based on deep learning, wherein each of the weak recognition networks is trained by the randomly selected sample data subset. Through an ensemble learning method, train and optimize the multiple weak recognition networks to obtain the drone identification model.

[0078] Specifically, first, based on the regional drone inventory information, identification records in the same or similar environment are collected, and then the collected identification records are marked with true values, that is, the true label of each sample data is calibrated. For example, the type of drone (such as fixed-wing, rotary wing, etc.), whether it is a real target, and other important features (such as size, flight status, etc.) are marked. The generated sample data set also includes corresponding intuitive feature data (such as speed, direction, distance and reflection intensity) and image feature data. In other words, the sample data set contains homologous identification records with true value labels, and the associated sample intuitive feature data, sample image feature data and sample drone identification results, which are the basis for subsequent model training. Exemplarily, homologous identification records include identification records of the same radar identification scene and the same type / series of radars.

[0079] Specifically, multiple sample data subsets are randomly extracted from the sample data set. The multiple sample data subsets are true subsets of the original sample data set, and each sample data subset contains different quantities and types of sample data. For example, some subsets are mainly concentrated on samples with similar characteristics (such as whether the target is a drone), while other subsets cover more complex samples (such as different types of drones). In this way, the training data can be effectively distributed to different weak recognition networks, increasing the diversity of the data, thereby improving the generalization ability of the model. At the same time, sample data subsets with smaller data volumes also help improve training efficiency.

[0080] Specifically, each weak recognition network is trained based on a randomly extracted subset of sample data, and the design of the weak recognition network is usually simpler, such as using a smaller deep neural network (such as a shallower convolutional neural network CNN or a multi-layer perceptron MLP) or a more streamlined network architecture. Each weak recognition network only processes part of the data and learns local features in the data. Optionally, multiple weak recognition networks are built based on different network architectures, such as convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), etc.

[0081] Specifically, the multiple weak recognition networks obtained through training have different mapping relationships and are suitable for different recognition tasks. For example, some weak recognition networks perform better on binary classification problems (identifying whether it is a drone), while other weak recognition networks are more suitable for solving multi-classification problems (distinguishing between different types of drones).

[0082] Furthermore, based on ensemble learning, multiple weak recognition networks are combined to improve the overall recognition performance through multi-model voting or weighted fusion. The model optimized by ensemble learning usually has higher recognition accuracy, robustness and generalization ability, and can accurately identify drones of different types and in different environments, and effectively reduce overfitting.

[0083] Specifically, the trained drone recognition model can be used to identify drones in real time and quickly obtain drone recognition results. By building and training a drone recognition model based on an AI algorithm, using multiple weak recognition networks and integrated learning methods, the accuracy and robustness of drone recognition can be effectively improved.

[0084] In some embodiments, a drone identification model based on an AI algorithm is constructed and trained, and then, the following steps are further included:

[0085] Based on the principal component analysis method, the contribution of multiple weak recognition networks to the correct recognition of records is analyzed and determined to obtain a contribution set, wherein the contribution is associated with an identification task category tag. The contribution set is clustered using the identification task category tag as a clustering constraint to obtain multiple operating condition clusters. Multiple operating condition clusters are serialized, and multiple operating condition recognition network sequences are constructed based on the serialization results, wherein the operating condition recognition network sequence includes multiple weak recognition networks in descending order of contribution. Multiple operating condition recognition network sequences are traversed to perform cumulative contribution calculations, and multiple weak recognition networks that meet the cumulative contribution threshold are extracted, and output as multiple classifier call forms.

[0086] Optionally, principal component analysis (PCA) is used to analyze the contribution of multiple weak recognition networks to the correct recognition of records, determine the influence of each weak recognition network on the recognition task, quantify the role of each network in the entire model, and then adaptively sort and optimize the weak recognition networks.

[0087] Specifically, first, the performance of each weak recognition network in recognizing the test data set is collected. The performance can be measured by standard evaluation indicators (such as accuracy, recall, F1 value, etc.). Then, based on the standard evaluation indicators, the principal component analysis (PCA) method is used to analyze the contribution value of each weak recognition network and associate it with the corresponding recognition task category label to form a contribution set.

[0088] Specifically, based on the contribution set of the weak recognition network, multiple weak recognition networks are divided into multiple working condition clusters through cluster analysis. Each cluster represents a group of weak recognition networks with similar performance characteristics. Preferably, the category label of the recognition task is used as a constraint condition for clustering, that is, the weak recognition networks are grouped according to different tasks (such as binary classification tasks, multi-classification tasks, etc.). Each category label represents a specific type of task requirement, and the clustering results will reflect which weak recognition networks have better performance in similar recognition tasks.

[0089] Specifically, for different working condition clusters, the weak recognition networks are arranged in descending order of contribution, and multiple working condition recognition network sequences are constructed to ensure that the weak recognition networks with high contribution are used for recognition tasks first.

[0090] Furthermore, multiple working condition recognition network sequences are traversed, and the contributions of the weak recognition networks are accumulated one by one from top to bottom to calculate the cumulative contribution. The cumulative contribution represents the comprehensive contribution of the sequence to the task completion. When the cumulative contribution reaches the cumulative contribution threshold (such as 80%), it is considered that the multiple accumulated weak recognition networks corresponding to the cumulative contribution have sufficient recognition capabilities, and the output is a classifier call form. The classifier call form is a set of classifiers that can be called directly, which helps to select the optimal network for calling according to task requirements in actual applications, thereby improving recognition accuracy and efficiency, while improving adaptability to different tasks, and facilitating subsequent targeted optimization (such as enhanced training of weak recognition networks).

[0091] Specifically, the above steps combine multiple weak recognition networks, which helps to call and work together for models of different applicable tasks, thereby improving the overall performance and recognition ability of the model. For example, for scenarios where it is necessary to quickly determine whether a drone exists (such as security monitoring, airport management, etc.), the integrated learning framework can quickly call the weak recognition network corresponding to the binary classification problem, efficiently filter out drones, and ensure a low false alarm rate. In scenarios where it is necessary to identify the type of drone (such as drone control, etc.), the multi-classification network can be called accordingly to achieve accurate distinction between different types of drones.

[0092] Continuously identify the target drone according to the drone identification result to obtain the tracking trajectory of the target drone, wherein the tracking trajectory includes an experienced trajectory and an estimated trajectory.

[0093] Specifically, the state of the target UAV will change over time, so through regular or continuous radar scanning and identification, the real-time position and state update of the target UAV can be obtained to track the dynamic changes of the UAV.

[0094] Specifically, in each scanning cycle, the drone identification model continuously identifies and updates the location and status of the target drone through radar echoes, and updates the identification result of the current drone based on the latest data compared with historical data.

[0095] Specifically, the experienced trajectory is the historical trajectory of the target UAV since the tracking was started. It is constructed by recording the position of the UAV at each recognition and associating it with the previous recognition result to form trajectory points. The experienced trajectory is an accurate reflection of the actual movement path of the UAV and is a continuously updated historical data set; the estimated trajectory predicts the possible future position of the target UAV based on the existing recognition data and historical trajectory. Optionally, the trajectory prediction is performed through a time-based mathematical model (such as Kalman filtering, particle filtering, etc.) to obtain an estimated trajectory.

[0096] Through the above steps, the tracking trajectory obtained, including the experienced trajectory and the estimated trajectory, not only provides past location information, but also predicts the future movement direction, which helps the operator to control the movement of the drone in real time.

[0097] In some embodiments, the method further comprises:

[0098] The model verification record data of the drone recognition model is obtained. Based on the model verification record data, a principal component analysis is performed on the image feature index set to obtain key feature groups corresponding to multiple recognition task types, and output as multiple feature index call forms. Based on the multiple feature index call forms, feature call optimization of the image feature index set is performed.

[0099] Optionally, the image feature index set can be further optimized through model verification record data to ensure that the extracted features have a high contribution to the recognition task, thereby improving the performance of the recognition model. The model verification record data includes the results of each verification stage in the training process, such as accuracy, recall rate, F1 score, false alarm rate, etc. The model verification record data can help analyze which features have a greater impact on the performance improvement of the model and which features may introduce redundancy or noise. In other words, it is possible to further explore which features play an important role in different recognition tasks.

[0100] Specifically, PCA analysis is performed on the model validation record data to identify which features contribute more to the performance of the model on the validation dataset and which features are crucial to the successful completion of the task, thereby obtaining key feature groups corresponding to multiple recognition task types (such as drone detection, drone classification, etc.). This key feature group can reflect which image features are the most discriminative in different recognition tasks.

[0101] Specifically, different recognition tasks may focus on different image features. For example, for drone detection tasks, features such as speed, reflection intensity, and shape may be particularly important, while for type classification tasks, texture and spatial features may be more critical. By obtaining the key feature group, it is helpful to select the optimal feature set for each recognition task.

[0102] Furthermore, the key feature groups are output as multiple feature index call forms. The feature index call forms contain the main feature indexes that need to be called under different recognition tasks, which helps to focus on the most contributing features and reduce redundant or irrelevant features during the feature extraction process.

[0103] Specifically, the feature group obtained from PCA is used for screening in the image feature extraction stage, and only the features with higher contribution in the task are extracted, while the features with less impact on the model performance are ignored, so as to reduce redundant features and improve the efficiency and accuracy of the model.

[0104] In some embodiments, the method further comprises:

[0105] Based on the obtained drone identification results, the drone management system is accessed to extract the intrinsic parameter information of the target drone. According to the intrinsic parameter information, the beam width and scanning frequency of the phased array antenna are dynamically optimized, and the target drone is continuously identified according to the optimized beam width and scanning frequency.

[0106] Specifically, the UAV management system is queried through the identification results to obtain the intrinsic parameter information of the target UAV, including the type of UAV, the corresponding size, performance data, etc. Then, the radar parameters, including beam width and scanning frequency, are dynamically adjusted according to the intrinsic parameter information of the UAV. For example, if the identified UAV is fast, the scanning frequency is adaptively increased to track the dynamic position of the UAV; if the identified UAV is very small, the beam width needs to be reduced to improve the target resolution. According to the optimized radar parameters, the target UAV is continuously identified, so as to improve the radar's tracking ability and recognition accuracy of the target UAV.

[0107] In summary, the drone identification and tracking method combined with the low-altitude radar AI algorithm provided by the present invention has the following technical effects:

[0108] By determining the target radar frequency band of the target scene, and performing radar detection with the target radar frequency band through the phased array antenna, the radar echo information is received; the radar echo information is converted into the time-frequency domain to obtain the target frequency domain data, and the target frequency domain data is analyzed for intuitive feature extraction to obtain the intuitive feature data of the target UAV; the image of the target UAV is restored based on the intuitive feature data to obtain the target radar image, and the target radar image is processed by feature engineering based on the preset image feature index set to obtain the image feature data; a UAV recognition model based on the AI ​​algorithm is constructed and trained, and the intuitive feature data and image feature data are input into the trained UAV recognition model for target recognition to obtain the UAV recognition result; the target UAV is continuously recognized according to the UAV recognition result to obtain the tracking trajectory of the target UAV, wherein the tracking trajectory includes the experienced trajectory and the estimated trajectory. Thus, the technical effect of improving the accuracy of UAV recognition and the tracking stability is achieved.

[0109] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the above-mentioned embodiments. It should be understood that those skilled in the art can still modify the technical solutions recorded in the above-mentioned embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. The drone identification and tracking method combined with low-altitude radar AI algorithm is characterized by: The method comprises: Determine a target radar frequency band of the target scene, and perform radar detection with the target radar frequency band through a phased array antenna to receive radar echo information; Performing time-frequency domain conversion on the radar echo information to obtain target frequency domain data, and analyzing the target frequency domain data to perform intuitive feature extraction to obtain intuitive feature data of the target UAV; Performing image restoration of the target UAV based on the intuitive feature data to obtain a target radar image, and performing feature engineering processing on the target radar image based on a preset image feature indicator set to obtain image feature data; Constructing and training a drone recognition model based on an AI algorithm, and inputting the intuitive feature data and the image feature data into the trained drone recognition model for target recognition, and obtaining a drone recognition result; Continuously identifying the target UAV according to the UAV identification result, and obtaining the tracking trajectory of the target UAV, wherein the tracking trajectory includes an experienced trajectory and an estimated trajectory; Determine the target radar band of the target scene, including: Interact with the drone management system in the target area to obtain regional drone inventory information; Perform clustering according to the drone inventory information in the area to obtain multiple drone sample information clusters; The drone identification and tracking requirements in the target area are analyzed, the feature information of multiple drone sample information clusters is called as the target feature information, and the frequency band screening based on the resolution and detection range is performed to determine the target radar frequency band.

2. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 1, characterized in that: The radar echo information is converted into a time-frequency domain to obtain target frequency domain data, and the target frequency domain data is analyzed to extract intuitive features to obtain intuitive feature data of the target UAV, including: Performing matched filtering on the radar echo information through a matched filter to obtain the echo time and calculate the distance characteristics of the target UAV; Based on Fourier transform, the radar echo information is converted into time-frequency domain to obtain the target frequency domain data, analyze and extract the frequency deviation data of the target frequency domain data, and convert the frequency deviation data into the speed characteristics of the target UAV in combination with the Doppler effect; Extracting the phase difference of the radar echo information and calculating and obtaining the directional characteristics of the target UAV; According to the matched filtering result, the echo power of the radar echo information is obtained, and the reflection intensity characteristics of the target UAV are calculated and obtained by combining the echo power with the direction characteristics.

3. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 2, characterized in that: The target UAV image is restored based on the intuitive feature data to obtain a target radar image, and the target radar image is subjected to feature engineering processing based on a preset image feature indicator set to obtain image feature data, including: Based on the image reconstruction technology, image restoration is performed according to the distance feature, the speed feature, the direction feature and the reflection intensity feature of the intuitive feature data to obtain a three-dimensional image and / or a two-dimensional image of the target UAV, and output it as the target radar image; The target radar image is enhanced, and features are extracted based on the enhanced target radar image to generate the image feature data, wherein the image feature data at least includes texture features, shape features, and space features.

4. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 3, characterized in that: Build and train a drone recognition model based on AI algorithms, including: Taking the regional drone inventory information as the matching target, collecting homologous identification records, and marking the homologous identification records with true values ​​to obtain a sample data set, wherein the sample data set includes associated sample intuitive feature data, sample image feature data, and sample drone identification results; Taking the sample data set as a sample space, performing random extraction to obtain a plurality of sample data subsets, wherein the plurality of sample data subsets are true subsets of the sample data set; Based on the plurality of sample data subsets, construct and train a plurality of weak recognition networks based on deep learning, wherein each of the weak recognition networks is trained by a randomly selected subset of the sample data; The plurality of weak recognition networks are trained and optimized through an integrated learning method to obtain the drone recognition model.

5. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 4, characterized in that: Build and train a drone recognition model based on AI algorithms, and then: Based on the principal component analysis method, the contribution of the plurality of weak recognition networks to the correct recognition of the records is analyzed and determined, and a contribution set is obtained, wherein the contribution association mark is marked with a recognition task category mark; Using the identified task category mark as a clustering constraint, clustering the contribution set to obtain a plurality of working condition clusters; Sequencing a plurality of the operating condition clusters, and constructing a plurality of operating condition identification network sequences according to the serialization results, wherein the operating condition identification network sequence includes a plurality of the weak identification networks in descending order of contribution; Traversing the plurality of working condition identification network sequences to perform cumulative contribution calculation, and extracting the plurality of weak identification networks that meet the cumulative contribution threshold, and outputting the plurality of classifier call forms.

6. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 1, characterized in that: The method further comprises: Obtaining model verification record data of the drone identification model; Based on the model verification record data, principal component analysis is performed on the image feature index set to obtain key feature groups corresponding to multiple recognition task types, and output as multiple feature index call forms; Based on the plurality of feature indicator calling forms, feature calling optimization of the image feature indicator set is performed.

7. The method for identifying and tracking unmanned aerial vehicles in combination with a low-altitude radar AI algorithm as claimed in claim 1, characterized in that: The method further comprises: Based on the obtained drone identification result, access the drone management system to extract the intrinsic parameter information of the target drone; According to the intrinsic parameter information, the beam width and scanning frequency of the phased array antenna are dynamically optimized, and the target UAV is continuously identified according to the optimized beam width and scanning frequency.

Images