A radar trajectory target classification method and system based on machine learning
By extracting and standardizing multi-dimensional features from radar trajectory data and combining them with the LightGBM model, the problems of low accuracy and poor real-time performance in radar target recognition technology have been solved. This has enabled efficient identification and differentiation of various types of targets, such as drones, birds, and cars, and improved the robustness and real-time performance of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WISESOFT CO LTD
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-05
AI Technical Summary
Existing radar target recognition technologies suffer from low recognition accuracy, poor real-time performance, and high false alarm rate when facing complex electromagnetic environments and diverse targets. They are difficult to adapt to the needs of low-altitude security, have limited feature engineering dimensions, lack the ability to classify multiple categories into fine granularities, and have inadequate data preprocessing.
By acquiring and preprocessing time-series trajectory data monitored by radar, extracting multi-dimensional feature sets and standardizing them, constructing a LightGBM gradient boosting tree model for classification, and using the multi-dimensional feature sets as input, combined with data preprocessing module, feature extraction module, feature standardization module and target classification module, the identification of multi-category targets is achieved.
It improves the model's adaptability to different scenarios and interferences, has strong robustness, meets real-time monitoring needs, and provides intuitive classification effect charts to facilitate user tuning and implementation effect analysis.
Smart Images
Figure CN122153602A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radar target recognition technology, and specifically relates to a radar trajectory target classification method and system based on machine learning. Background Technology
[0002] In short, with the booming development of the low-altitude economy, the application of low-altitude aircraft such as drones is becoming increasingly widespread. However, disorderly flight also brings serious airspace security risks, such as interference with civil aviation routes and threats to the security of critical infrastructure. Radar, with its all-weather, all-time, and long-range detection advantages, has become the core means of low-altitude target monitoring, capable of acquiring real-time temporal trajectory data of targets. However, existing radar target recognition technologies still face many challenges in practical applications. Traditional recognition methods mainly rely on threshold rules set by human experts or simple kinematic principles. These methods suffer from low accuracy, poor real-time performance, and high false alarm rates when facing complex electromagnetic environments and diverse targets, making them difficult to adapt to the increasingly complex low-altitude security needs.
[0003] In recent years, some existing technologies have attempted to introduce machine learning algorithms to improve recognition performance, but the following significant technical bottlenecks remain:
[0004] Feature engineering suffers from a lack of depth. Existing methods mostly extract only the target's basic motion parameters (such as instantaneous velocity and distance), lacking in-depth mining of trajectory temporal evolution patterns and spatial topological features. This singular feature representation makes it difficult for models to capture subtle differences in target motion patterns, limiting further improvements in classification accuracy.
[0005] The existing models lack multi-category, fine-grained classification capabilities. Most existing solutions are limited to binary classification tasks (e.g., distinguishing only between "drones" and "birds"), failing to effectively cover scenarios involving multiple typical low-altitude targets such as drones, birds, clutter, and vehicles. Furthermore, when faced with complex real-world environments containing various interference sources, the existing models exhibit weak generalization capabilities and limited adaptability.
[0006] The data preprocessing process is inadequate. Existing technologies often lack systematic preprocessing and standardization steps when processing raw radar time-series data. Due to issues such as high noise, inconsistent sampling frequencies, and large differences in dimensions in radar trajectory data, directly inputting features into the model without effective decoupling and standardization can easily lead to difficulties in model convergence or getting trapped in local optima, severely affecting the robustness of the system.
[0007] Therefore, it is an urgent need to develop a method and system that can efficiently extract multi-dimensional decoupled features from radar time-series trajectories, has a standardized preprocessing mechanism, and combines advanced machine learning algorithms to achieve multi-class classification. Summary of the Invention
[0008] This invention aims to solve the problems of low recognition accuracy, poor real-time performance, and high false alarm rate in existing radar target recognition methods, and to achieve the recognition and differentiation of multiple types of targets such as drones, birds, clutter, and vehicles. It provides a radar trajectory target classification method and system based on machine learning.
[0009] In a first aspect, the present invention provides a radar trajectory target classification method based on machine learning, comprising the following steps: S1. Acquire time-series trajectory data monitored by radar, and preprocess and manually label the time-series trajectory data to generate category labels; the preprocessing includes grouping by target category and sorting based on timestamp.
[0010] Preferably, the time-series trajectory data is point data or flight track data, including at least target identifier, timestamp, two-dimensional plane coordinates, relative altitude, instantaneous velocity, acceleration, and label.
[0011] Preferably, the category includes drones, birds, cars, and interference clutter.
[0012] S2. Extract features from the preprocessed time-series trajectory data and decouple to obtain a multi-dimensional feature set of the target; the multi-dimensional feature set includes: time-series statistical features, difference and trend features, time-domain physical features, geometric and kinematic features, and higher-order change features.
[0013] Preferably, the time-series statistical features include the mean, standard deviation, extreme values, range, and median of spatial location, instantaneous velocity, and acceleration, used to quantify the stability and distribution range of the target motion; The difference and trend features include the first-order difference statistical features of each parameter and the slope calculated based on linear regression, which are used to capture the dynamic evolution trend of the target motion. The temporal physical characteristics, including sequence length, duration, and sampling rate, characterize the dwell time sequence of the target within the radar detection field. The geometric and kinematic features include total trajectory distance, average velocity, median acceleration, distribution of directional changes, and statistical values of trajectory curvature. The distribution of directional changes and trajectory curvature are further calculated to characterize the spatial complexity and mechanistic features of the target trajectory. The higher-order variation features, including the mean, standard deviation, and variance of acceleration and height changes, are used to capture the higher-order fluctuation differences of the target in the vertical and dynamic dimensions.
[0014] S3. Standardize the multi-dimensional feature set to eliminate the numerical differences in physical dimensions and obtain standard features.
[0015] Preferably, the standardization process adopts Z-Score, which maps each dimension feature to a standard normal distribution feature with a mean of 0 and a standard deviation of 1, so as to eliminate the numerical difference between different physical units, prevent weight bias during model training, and thus significantly improve the convergence speed and recognition accuracy of the multi-class discriminator.
[0016] S4. Construct a target classification model, set classification parameters, take the multi-dimensional standard features as input, take the category label as the supervision target, and train the target classification model until convergence using a preset training dataset; Preferably, the target classification model adopts the LightGBM gradient boosting tree model; the model uses an ensemble decision tree and a negative gradient descent algorithm to iteratively optimize the weights, so that the classification error converges quickly.
[0017] Preferably, the preset training dataset consists of a training set, a validation set, and a test set in a ratio of 6:2:2; the classification parameters include at least the number of iterations, the learning rate, and the number of leaf nodes.
[0018] S5. Input the radar trajectory data to be classified into the target classification model and output the classification and recognition results.
[0019] Preferably, the classification and recognition results include a classification effect comparison chart generated based on the results; the classification effect comparison chart includes a 2D trajectory comparison chart, a 3D trajectory comparison chart, and a detection statistics chart, which intuitively displays the classification effect.
[0020] In a second aspect, the present invention provides a radar trajectory target classification system based on machine learning, comprising: The data acquisition module is used to acquire time-series trajectory data monitored by the radar; The data preprocessing module is used to group, sort by time, and label the time-series trajectory data. The feature extraction module is used to extract a multi-dimensional feature set reflecting the target's motion mechanism from the preprocessed time-series trajectory data; The feature standardization module is used to standardize the multi-dimensional feature set to obtain a standard feature vector; The model training module is used to construct a gradient boosting tree classification model and train the classification model using the training set to obtain a trained target classification model. The target classification module is used to input the data to be classified into the trained target classification model and output the predicted category of the target. The results analysis module is used to generate visual analysis charts based on the prediction results.
[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention provides a radar trajectory target classification method based on machine learning. Through a multi-feature fusion strategy, it improves the model's adaptability to different scenarios and interference, and has strong robustness.
[0022] 2. This invention uses a lightweight classification model for training and classification, resulting in fast prediction speed and meeting the needs of real-time monitoring; 3. This invention provides intuitive classification effect charts with good visualization, making it easy for users to optimize the model and analyze the implementation effect.
[0023] 4. This invention provides a radar trajectory target classification system based on machine learning, which can easily extend the classification method to other types of radar target recognition scenarios. Attached Figure Description
[0024] Figure 1 A flowchart of the target classification method provided in this embodiment; Figure 2 This is a schematic diagram of feature extraction in this embodiment; Figure 3 To illustrate the 2D trajectories of all targets in the comparison test case; Figure 4 A 2D trajectory diagram of the drone in the comparison test example; Figure 5 To provide a comparative 3D trajectory diagram of all targets in the test case; Figure 6 A 3D trajectory diagram of the drone in the comparison test example; Figure 7 To compare the drone detection statistics in the test cases; Figure 8 This is a schematic diagram of the confusion matrix in this embodiment. Detailed Implementation
[0025] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.
[0026] Unless otherwise specified, the use of terms such as "upper," "lower," "left," "right," "center," "inner," and "outer" to indicate orientation or positional relationships in the description of specific embodiments of the present invention is based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is typically placed during use. These terms are merely for the purpose of facilitating the description of the present invention or simplifying the description in specific embodiments, enabling those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on the present invention.
[0027] Furthermore, the use of terms such as "horizontal," "vertical," "suspended," and "parallel" does not imply that the corresponding device / component / element must be absolutely horizontal, vertical, suspended, or parallel, but rather that it can be slightly tilted or have a deviation. For example, "horizontal" merely means that its direction is more horizontal relative to "vertical," not that the structure must be completely horizontal, but that it can be slightly tilted. Alternatively, it can be simplified to mean that the corresponding device / component / element, when set in a "horizontal," "vertical," "suspended," or "parallel" direction, can have an error / deviation of ±10% relative to the corresponding direction, more preferably within ±8%, more preferably within ±6%, more preferably within ±5%, and more preferably within ±4%. As long as the corresponding device / component / element is within the error / deviation range, it can still achieve its function in the present invention.
[0028] Furthermore, the use of terms such as "first," "second," and "third" in terminology is merely for distinguishing descriptions of identical or similar components and should not be interpreted as emphasizing or implying the relative importance of a particular component.
[0029] Furthermore, in the description of the embodiments of the present invention, "several", "more than", and "a number of" represent at least two. The number can be any number, such as 2, 3, 4, 5, 6, 7, 8, or 9, and can even exceed nine.
[0030] Furthermore, in the description of the technical solution of this invention, unless otherwise explicitly specified / limited / restricted, the terms "set up," "install," "connect," "link," "provided with," "laid out," and "arranged" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to common connection methods in the art, such as welding, riveting, bolting, and threaded connections. Such connections can be mechanical, electrical, or communication connections; they can be direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components.
[0031] Example 1 This embodiment provides a radar trajectory target classification method based on machine learning, the process of which is as follows: Figure 1 As shown, the specific steps include: S1. Acquire time-series trajectory data monitored by radar, and preprocess and manually label the time-series trajectory data to generate category labels; the preprocessing includes grouping by target category and sorting based on timestamp.
[0032] Preferably, the time-series trajectory data is point data or flight track data, including at least target identifier, timestamp, two-dimensional plane coordinates, relative altitude, instantaneous velocity, acceleration, and label.
[0033] S11. Real-time acquisition of time-series trajectory data during the training period via a low-altitude radar system. The data format is a CSV file, containing fields such as target ID (sid), time (time), mapped two-dimensional plane coordinates (x_axis, y_axis), relative altitude (alt), velocity (speed), acceleration (acc), and label. A pre-defined data parsing component reads the trajectory files generated by radar monitoring and converts them into a structured data table recognizable by the system. Data is grouped by target ID (sid), and each group is sorted by time to ensure the temporal sequence of the data.
[0034] S12. Manually label the collected raw data to establish category label values corresponding to physical targets (drones, birds, cars, and interference clutter), forming a complete supervised learning sample library.
[0035] S13. After clustering the trajectory data by target ID (sid), group the data and perform time increment sorting on each group. Divide the traces belonging to the same detection target into independent trajectory groups. For each trajectory group, extract the corresponding timestamp (time) parameter and perform incremental sorting to generate a trajectory sequence with strict time order. This provides an ordered data foundation for subsequent feature extraction and ensures the continuity and consistency of the data in the spatiotemporal dimension.
[0036] S14. Use a preset filtering algorithm (such as Kalman filtering) to interpolate and smooth the trajectory to eliminate random noise interference introduced by the radar measurement process.
[0037] S2. Extract features from the preprocessed time-series trajectory data and decouple to obtain a multi-dimensional feature set of the target; the multi-dimensional feature set includes: time-series statistical features, difference and trend features, time-domain physical features, geometric and kinematic features, and higher-order change features.
[0038] Feature extraction illustration as follows Figure 2As shown in the figure, this diagram illustrates the logic and process of extracting multi-dimensional features from radar trajectory data. These multi-dimensional features include basic statistical features, motion features, trajectory curvature features, and so on.
[0039] Preferably, the time-series statistical features include the mean, standard deviation, extreme values, range, and median of spatial location, instantaneous velocity, and acceleration, used to quantify the stability and distribution range of the target motion; The difference and trend features include the first-order difference statistical features of each parameter and the slope calculated based on linear regression, which are used to capture the dynamic evolution trend of the target motion. The temporal physical characteristics, including sequence length, duration, and sampling rate, characterize the dwell time sequence of the target within the radar detection field. The geometric and kinematic features include total trajectory distance, average velocity, median acceleration, distribution of directional changes, and statistical values of trajectory curvature. The distribution of directional changes and trajectory curvature are further calculated to characterize the spatial complexity and mechanistic features of the target trajectory. The higher-order variation features, including the mean, standard deviation, and variance of acceleration and height changes, are used to capture the higher-order fluctuation differences of the target in the vertical and dynamic dimensions.
[0040] Specifically, the feature extraction process is as follows: S21. For physical quantities such as spatial coordinates (x_axis, y_axis, alt), instantaneous velocity (speed), and acceleration (acc) in a time-series trajectory sequence, extract their distribution characteristics using a pre-defined statistical calculation logic. Specifically, this includes calculating the arithmetic mean (mean), standard deviation (std), minimum (min), maximum (max), numerical range (range), and median of each physical quantity within the observation time window, used to characterize the central tendency and dispersion of the target's motion state.
[0041] S22. By calculating the skewness and kurtosis parameters of the physical quantities, the distribution symmetry and peak steepness of the target trajectory data are quantified, thereby assisting in the identification of the characteristics of UAVs and clutter targets through the differences in probability distribution morphology.
[0042] S23. Integrate the above parameters and combine them with preset differential features, trajectory curvature and motion complexity indicators to encapsulate and generate feature vectors that represent the multi-dimensional information of the target, which serve as the input benchmark for the subsequent classification discriminator.
[0043] S3. Standardize the multi-dimensional feature set to eliminate the numerical differences in physical dimensions and obtain standard features.
[0044] S31. Based on the original feature vector output by the feature extraction module, calculate the arithmetic mean and standard deviation of each dimension feature (such as speed, height, curvature, etc.) in the training sample set, and use them as the benchmark reference parameters for standardization.
[0045] S32. Using a preset standardization algorithm, perform a linear transformation on the original feature sample values by subtracting the corresponding mean and dividing by the standard deviation. Through linear offset and scaling, map each physical feature to a standard normal distribution with a mean of 0 and a standard deviation of 1.
[0046] Preferably, the standardization process employs Z-Score, mapping each dimension of features to a standard normal distribution with a mean of 0 and a standard deviation of 1. This eliminates numerical differences between different physical units, prevents weight bias during model training, and significantly improves the convergence speed and recognition accuracy of the multi-class discriminator. The Z-score standardization formula is as follows:
[0047] in: χ represents the original feature sampling value; This is the arithmetic mean of the feature in this dimension across the sample set; s is the standard deviation of this feature dimension in the sample set; χ' is the target feature value after standard processing.
[0048] S33. Perform dimensional alignment on the processed target feature vector to ensure that the features of each dimension are comparable in numerical range, thereby eliminating the influence of dimensional differences on the weight convergence process of the multi-classification model, improving the discriminator's recognition stability of complex radar trajectory signals, eliminating the differences between different physical dimensions such as velocity (m / s), height (m) and curvature (1 / m), and preventing the multi-classification model from generating weight bias during calculation.
[0049] S4. Construct a target classification model, set classification parameters, take the multi-dimensional standard features as input, take the category label as the supervision target, and train the target classification model until convergence using a preset training dataset; Preferably, the target classification model adopts the LightGBM gradient boosting tree model; the model uses an ensemble decision tree and a negative gradient descent algorithm to iteratively optimize the weights, enabling the classification error to converge quickly. The specific training steps are as follows: S41. Sample set partitioning: The standardized multidimensional features are divided into training set, validation set and test set according to a preset ratio of 6:2:2.
[0050] S42. Model Parameter Configuration: Construct a discriminator based on the LightGBM gradient boosting tree architecture, and set multi-class classification parameters. Key parameters include: Number of iterations (n_estimators): The default value is 100 to ensure the model fits the trajectory features to a certain depth. Experiments showed that when the step size was 100 and the number of iterations was less than 100, the model’s false negative rate for small targets at low altitudes increased significantly. However, when the number of iterations exceeded 100 (e.g., 200, 300), the marginal gain in accuracy was extremely low (less than 0.5%), but the inference delay increased linearly, making it impossible to meet the real-time requirements of radar signal processing cycles.
[0051] Learning rate: The default value is 0.1, which is used to dynamically adjust the step size of parameter updates and convergence stability; The learning rate is related to the number of iterations. If the learning rate is set to less than 0.1, the model converges too slowly and requires a significant increase in the number of iterations (such as more than 500) to achieve the same accuracy. This will directly lead to the delay of single-frame signal processing exceeding the standard. If the learning rate is set to be greater than 0.1, although training becomes faster, the model is prone to skipping the global optimal solution of the loss function, which will cause prediction oscillations when faced with radar noise and reduce the certainty of recognition. Therefore, setting the learning rate to 0.1 is an empirical threshold that can suppress the fluctuations in recognition and prediction to the greatest extent while ensuring millisecond-level response.
[0052] Number of leaf nodes (num_leaves): 15 by default; The number of leaf nodes is set to 15, based on a balance between model expressiveness and computational efficiency; If the number of leaf nodes is less than 15, the tree structure is too simple and cannot fully exploit the nonlinear correlation of radar points in micro-motion and spatial topology, resulting in insufficient differentiation of strong interference targets such as birds and drones. If the number of leaf nodes is greater than 15, the complexity of the tree increases exponentially, which does not meet the deployment requirements of low power consumption and high real-time performance.
[0053] Target category labels (num_class=4): This invention distinguishes four target categories (0-birds, 1-drones, 2-clutter, 3-cars). Multiclass objective function (objective='multiclass'): The weights of the internal weak classifiers are optimized by superimposing the distribution of the negative gradient information of the loss function until the residual between the predicted probability distribution of the model output and the true label reaches the preset convergence threshold.
[0054] S43. Using the training set as input, iterative optimization is achieved by repeatedly adjusting the weights of the internal weak classifiers through the negative gradient descent algorithm. The loss function is monitored in real-time on the validation set to prevent overfitting, until the model performance converges. After training, the decision tree topology and weight parameters within the model are fixed, generating classification instruction units that can be called in real-time, and the trained classification model is output to the target classification module.
[0055] S5. Input the radar trajectory data to be classified into the target classification model and output the classification and recognition results.
[0056] A classification model with trained and fixed parameters is used to classify and identify real-time input radar trajectory features online and output the classification results. Preferably, the classification results include a classification result comparison chart generated based on the results; the classification result comparison chart includes a 2D trajectory comparison chart, a 3D trajectory comparison chart, and a detection statistics chart, which intuitively displays the classification results.
[0057] S51. Input the dataset to be tested into the discriminator with fixed parameters, and the model automatically performs feature mapping and logistic regression operations to output the predicted category of the target.
[0058] S52. Quantitative Performance Analysis: Based on the prediction results of the test set, calculate precision, recall, and F1 score.
[0059] S53. Validation: Through comparative experiments with fully connected networks and random forest models, the technical advantages of this solution in terms of UAV identification recall rate and overall accuracy (e.g., 89.8%) are verified.
[0060] S54. Results Filtering and Visualization Evaluation Based on the discrimination instructions, specific targets (such as drones) are filtered out, and 2D / 3D trajectory comparison maps and confusion matrices are generated. The model's ability to distinguish targets in complex environments is evaluated intuitively through graphical means.
[0061] To verify the performance advantages of the radar trajectory target classification method provided in this embodiment after adopting the LightGBM model in radar trajectory target classification tasks, the following multi-model comparison experiment was designed, including the Random Forest model and the fully connected network. The UAV was used as the core detection target to carry out quantitative performance analysis. The specific experimental design and performance indicators are as follows.
[0062] Experimental Design: Using the same training, validation, and test sets, fully connected network models, random forest models, and the LightGBM model of this invention were trained respectively. The parameters of each comparison model were set according to industry standards and optimized to eliminate the influence of parameter differences on the experimental results, ensuring fairness in the comparison with the LightGBM model of this invention. The test set data was input into the three trained models to obtain the target prediction results y for each model. pred Compare it with the real labels y in the test set test A sample-by-sample comparison was conducted, and relevant evaluation parameters were statistically analyzed.
[0063] Evaluation parameter definition: Taking UAVs as the core detection target, three core statistical parameters are statistically analyzed: true positive (TP), false positive (FP, false detections), and false negative (FN, missed detections), which serve as the basis for calculating performance indicators; Core performance metrics calculation: Based on the above statistical parameters, the three core performance metrics—precision, recall, and F1 score—are calculated. The calculation formulas and physical meanings of each metric are as follows: Precision measures the accuracy of a model's predictions of drone targets. It reflects the proportion of samples predicted as drones that are actually drones. The formula is as follows:
[0064] Recall measures the model's coverage of drone targets, reflecting the percentage of samples that are correctly identified as drones out of all actual drones. The formula is:
[0065] The F1 score is a harmonic mean that balances precision and recall, providing a comprehensive evaluation of the model's overall classification performance for drone targets and avoiding the limitations of a single metric. The calculation formula is as follows:
[0066] The overall performance results of the comparative experiment are shown in Table 1. In this embodiment, the LightGBM model has a classification accuracy of 89.80%, the Random Forest model has an accuracy of 83.67%, and the fully connected network model has an accuracy of 85.71%. The LightGBM model has a high recall rate and a significantly higher F1 score, proving that it has the strongest ability to balance precision and recall. In summary, LightGBM outperforms other models in the current task.
[0067] Table 1 Overall performance comparison test results Model Name accuracy Accuracy Recall rate F1 Fraction LightGBM 0.8980 0.9100 0.8980 0.8749 Random Forest 0.8367 0.7546 0.8367 0.7934 Fully connected network 0.8571 0.8409 0.8571 0.8412 The category prediction results of different models are shown in Tables 2, 3, and 4, respectively. Experimental results show that each model exhibits excellent recognition performance for both drones and vehicles. Among them, the LightGBM model performs the most significantly, achieving 100% recall and an F1 score exceeding 0.92 in both categories. This superior performance stems from the LightGBM model's ability to accurately extract and map the key multi-dimensional features of these two target types. For vehicles, the model effectively learns their low-altitude, relatively fixed motion patterns and trajectory characteristics constrained by the road; for drones, the model successfully captures their motion stability and low-acceleration variation characteristics during cruising.
[0068] The comprehensive comparison shows that LightGBM is the optimal model in this experiment. Especially in scenarios involving imbalanced data (i.e., the number of drone samples is significantly greater than the number of bird samples), this model demonstrates stronger generalization ability and robustness.
[0069] Table 2. Class Prediction Results of LightGBM Model category Accuracy Recall rate F1 Fraction Bird 1 0.2 0.3333 Drone 0.9429 1 0.9706 Clutter 0.6667 0.8 0.7273 Car 0.8571 1 0.9231 Weighted average 0.91 0.898 0.8749 Table 3. Random Forest Model Class Prediction Results category Accuracy Recall rate F1 Fraction Bird 0 0 0 Drone 0.8889 0.9697 0.9275 Clutter 0.5 0.6 0.5455 Car 0.8571 1 0.9231 Weighted average 0.7546 0.8367 0.7934 Table 4. Class prediction results of the fully linked model category Accuracy Recall rate F1 Fraction Bird 0.5 0.2 0.2857 Drone 0.9412 0.9697 0.9552 Clutter 0.5 0.6 0.5455 Car 0.8571 1 0.9231 Weighted average 0.8409 0.8571 0.8412 Based on the prediction results, the target category of drones is filtered out, and a comparison image before and after filtering is generated: The 2D trajectory comparison chart shows the 2D trajectories before and after filtering, where... Figure 3 For the 2D trajectory of all targets, Figure 4 To predict only the 2D trajectory of the drone.
[0070] The 3D trajectory comparison chart shows the 3D trajectories before and after filtering, where... Figure 5 3D trajectories for all targets, Figure 6 To predict only the 3D trajectory of the drone.
[0071] Drone detection statistics chart as follows Figure 7 As shown, the statistics on drone detection are displayed, including indicators such as total number of targets, actual number of drones, predicted number of drones, number of correctly detected drones, number of false positives, and number of missed detections.
[0072] Figure 8The confusion matrix clearly demonstrates the model's classification results for all categories of birds, drones, cars, and other clutter. Observing the confusion matrix, we can conclude that the model's recall, especially for the drone category, is quite high, proving the model's robustness in drone classification.
[0073] Example 2 This embodiment provides a radar trajectory target classification system based on machine learning, used to implement the method described in Example 1. The system includes the following modules: The data acquisition module is used to acquire time-series trajectory data monitored by radar.
[0074] The data preprocessing module is used to group, sort by time, and label the time-series trajectory data to ensure the logical consistency and temporal consistency of the data. It includes a data reading unit, a target grouping unit, and a time-series reconstruction unit.
[0075] The feature extraction module is used to extract a multi-dimensional feature set reflecting the target's motion mechanism from the preprocessed time-series trajectory data. It includes a basic statistical feature extraction unit, a distribution pattern recognition unit, and a multi-dimensional feature integration unit.
[0076] The feature standardization module is used to standardize the multi-dimensional feature set to obtain a standard feature vector; it includes a statistical feature learning unit, a linear mapping transformation unit, and a dimensional alignment processing unit.
[0077] The model training module is used to construct a gradient boosting tree classification model and train the classification model using the training set to obtain a trained target classification model; it includes a parameter configuration unit, an unbalanced sample compensation unit, a negative gradient iterative optimization unit, and a model solidification and output unit.
[0078] The target classification module is used to input the data to be classified into the trained target classification model and output the predicted category of the target; the real-time feature standardization unit, the classification instruction generation unit, and the recognition result output unit are also included.
[0079] The results analysis module generates visual analysis charts based on the prediction results, enabling quantitative evaluation of model performance and intuitive display of classification effectiveness. It includes the following functional units: classification performance evaluation unit, multi-dimensional spatial trajectory presentation unit, and object detection statistics unit.
[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A radar trajectory target classification method based on machine learning, characterized in that, Includes the following steps: S1. Acquire time-series trajectory data monitored by radar, and preprocess and manually categorize the time-series trajectory data to generate category labels; the preprocessing includes grouping by target category and sorting based on timestamp; S2. Extract features from the preprocessed time-series trajectory data and decouple to obtain a multi-dimensional feature set of the target; The multi-dimensional feature set includes: time-series statistical features, difference and trend features, time-domain physical features, geometric and kinematic features, and higher-order variation features; S3. Standardize the multi-dimensional feature set to eliminate the numerical differences in physical dimensions and obtain standard features; S4. Construct a target classification model, set classification parameters, take the multi-dimensional standard features as input, take the category label as the supervision target, and train the target classification model until convergence using a preset training dataset; S5. Input the radar trajectory data to be classified into the target classification model and output the classification and recognition results.
2. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The time-series trajectory data monitored by the radar in step S1 is point trace and track data, which includes at least target identifier, timestamp, two-dimensional plane coordinates, relative altitude, instantaneous velocity, acceleration and tag.
3. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The categories mentioned in step S1 include drones, birds, cars, and interference clutter.
4. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, In step S2, The time-series statistical features include the mean, standard deviation, extreme values, range, and median of spatial location, instantaneous velocity, and acceleration; The difference and trend characteristics, This includes the first-order difference statistical characteristics of each parameter and the slope calculated based on linear regression; The temporal physical characteristics include sequence length, duration, and sampling rate; The geometric and kinematic characteristics include total trajectory distance, average velocity, median acceleration, distribution of directional change, and statistical values of trajectory curvature. The higher-order variation characteristics include the mean, standard deviation, and variance of the acceleration and height changes.
5. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The standardization process described in step S3 uses Z-Score, which maps each dimension of the feature to a standard normal distribution feature with a mean of 0 and a standard deviation of 1.
6. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The target classification model in step S4 adopts the LightGBM gradient boosting tree model; the model uses an ensemble decision tree to iteratively optimize the weights using the negative gradient descent algorithm.
7. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The preset training dataset mentioned in step S4 consists of a training set, a validation set, and a test set, in a ratio of 6:2:
2.
8. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The classification parameters mentioned in step S4 include at least the number of iterations, the learning rate, and the number of leaf nodes.
9. The radar trajectory target classification method based on machine learning according to claim 1, characterized in that, The classification and recognition results in step S5 include a classification performance comparison chart generated based on the results; the classification performance comparison chart includes a 2D trajectory comparison chart, a 3D trajectory comparison chart, and a detection statistics chart.
10. A radar trajectory target classification system based on machine learning, characterized in that, include: The data acquisition module is used to acquire time-series trajectory data monitored by the radar; The data preprocessing module is used to group, sort by time, and label the time-series trajectory data. The feature extraction module is used to extract a multi-dimensional feature set reflecting the target's motion mechanism from the preprocessed time-series trajectory data; The feature standardization module is used to standardize the multi-dimensional feature set to obtain a standard feature vector; The model training module is used to construct a gradient boosting tree classification model and train the classification model using the training set to obtain a trained target classification model. The target classification module is used to input the data to be classified into the trained target classification model and output the predicted category of the target. The results analysis module is used to generate visual analysis charts based on the prediction results.