Unmanned aerial vehicle high-precision interception method and system based on target detection algorithm

Abstract

The invention provides an unmanned aerial vehicle high-precision interception method and system based on a target detection algorithm, and relates to the technical field of unmanned aerial vehicle identification, and the method comprises the steps: collecting flight image data and radar echo data of a target unmanned aerial vehicle, and carrying out the preprocessing, and obtaining standardized image data and standardized radar data; constructing a target detection model based on a target detection algorithm; according to the target detection model, the standardized image data and the standardized radar data, converting an initial positioning result of the target unmanned aerial vehicle in an image coordinate system into initial position data of the target unmanned aerial vehicle in a geographic coordinate system; based on a Hungary algorithm, performing state correction on the initial position data of the target unmanned aerial vehicle to obtain corresponding high-precision position data; and on the basis of an improved A * algorithm, interception path analysis is carried out on the high-precision position data and the radar echo data, an optimal interception path of the target unmanned aerial vehicle is obtained, and then a corresponding interception execution instruction is executed.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) identification technology, specifically a high-precision UAV interception method and system based on target detection algorithms. Background Technology

[0002] Drone interception missions require precise positioning and efficient interception of target drones. However, the lack of an effective state correction mechanism after target positioning means that positional errors directly affect the accuracy of subsequent interception path planning. Furthermore, considering only distance factors without fully taking into account the motion characteristics of the target drone may result in paths that are not dynamically feasible or have inaccurate timing for interception, making it difficult to meet the actual needs of high-precision interception.

[0003] Therefore, a high-precision interception method and system for UAVs based on target detection algorithms is provided. Summary of the Invention

[0004] To address the aforementioned technical problems, the present invention aims to provide a high-precision interception method and system for unmanned aerial vehicles (UAVs) based on target detection algorithms.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a high-precision interception method for unmanned aerial vehicles (UAVs) based on a target detection algorithm, the method comprising: Collect flight image data and radar echo data of the target UAV within the monitoring area; preprocess the flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data; Based on the target detection algorithm, a target detection model is constructed; according to the target detection model, standardized image data, and standardized radar data, the initial positioning result of the target UAV in the image coordinate system is converted into the initial position data of the target UAV in the geographic coordinate system. Based on the Hungarian algorithm, the initial position data of the target UAV is corrected to obtain high-precision position data of the target UAV. Based on the improved A* algorithm, the interception path is analyzed using the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then the corresponding interception execution command is executed.

[0006] Furthermore, the process of constructing an object detection model based on the object detection algorithm includes: Acquire several sets of standardized image data and standardized radar data of UAVs of different models, flight attitudes, and environments with timestamp alignment to form a dataset; divide the dataset into training set and test set according to a preset ratio; Based on YOLOv8 as the basic algorithm framework, the YOLOv8 basic algorithm framework is improved to obtain an improved YOLOv8 algorithm framework. Based on the training set, the improved YOLOv8 algorithm framework is trained in three stages to obtain the trained improved YOLOv8 algorithm framework. The improved YOLOv8 algorithm framework after training was validated. If the validation was successful, the corresponding object detection model was obtained.

[0007] Furthermore, the process of improving the YOLOv8 basic algorithm framework includes: A lightweight coordinate attention mechanism module was added to the backbone feature extraction network in the YOLOv8 basic algorithm framework, and the number of feature channels in the backbone feature extraction network was adjusted. The neck feature fusion network in the YOLOv8 basic algorithm framework is optimized hierarchically, including: pyramid feature fusion structure, intermediate layer and feature fusion residual connection layer; and the number of feature channels in the neck feature fusion network is adjusted.

[0008] Furthermore, the three-stage training process for the improved YOLOv8 algorithm framework based on the training set includes: The first stage of training includes: loading the dataset, configuring data augmentation, setting hyperparameters, and completing the corresponding number of training iterations or if the detection metrics of the training set in the dataset all meet the requirements, then proceeding to the second stage of training. The second stage of training includes: increasing the proportion of drone small target samples, nighttime infrared samples, haze scene samples, and fuselage partially occluded samples in the dataset; dynamically adjusting hyperparameters and optimizing the weight saving frequency; and proceeding to the third stage of training after completing the corresponding number of training iterations or when the corresponding detection metrics of the training set and test set in the dataset meet the requirements. The third stage of training includes: loading the proportion of drone targets in the dataset that are unobstructed, have clear outlines, and complete features, and are in a single-lighting environment, and completely disabling all data augmentation configurations; performing a second step decay of the learning rate while keeping all optimizer parameters unchanged; and completing the corresponding number of training iterations and ensuring that the corresponding detection metrics of the training and test sets in the dataset meet the requirements, or triggering the early stopping mechanism, then the third stage of training is complete.

[0009] Furthermore, based on the target detection model, standardized image data, and standardized radar data, the process of converting the initial positioning result of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system includes: Standardized image data and standardized radar data are input into the trained target detection model, which outputs the bounding box coordinates of the target UAV in the image coordinate system, calculates the coordinates of the center point of the bounding box, and uses it as the initial localization result in the image coordinate system. Based on the target distance, azimuth angle, and camera installation parameters in the corresponding standardized radar data, the coordinates of the bounding box center point in the image coordinate system are converted into the initial position data in the geographic coordinate system using a coordinate transformation formula.

[0010] Furthermore, based on the Hungarian algorithm, the process of correcting the initial position data of the target UAV includes: Determine whether to perform state correction on the initial position data of the target UAV. If state correction is required, construct a cost matrix based on the Hungarian algorithm to obtain the corresponding matching cost, set a spatial distance threshold, determine the optimal correlation between the position data of the preceding and following frames based on the spatial distance threshold and the corresponding matching cost, and then perform smooth correction on the initial position data of the optimal correlation based on the corresponding radial velocity to obtain the corresponding high-precision position data.

[0011] Furthermore, based on the improved A* algorithm, the process of interception path analysis of the high-precision position data and radar echo data of the target UAV includes: Based on the geographical coordinates of the monitored area, a three-dimensional raster map is constructed. The raster size of the three-dimensional raster map is set, no-fly zones are marked as obstacle grates, and constraints are set. The radial velocity from radar echo data is introduced into the heuristic function of the A* algorithm to obtain an improved A* algorithm. Based on the improved A* algorithm, two node sets are created. Starting from the current high-precision position data of the intercepting UAV, a sequential search is performed according to priority, and constraint verification is performed according to the constraints to generate the optimal interception path of the target UAV. If all constraint verifications are satisfied, the corresponding path is the optimal interception path. If the constraint verifications are not satisfied, the parameters are adjusted and replanned. The optimal interception path is divided according to the flight state of the target UAV, and the corresponding interception execution command is executed.

[0012] A second aspect of the present invention also provides a high-precision interception system for unmanned aerial vehicles based on a target detection algorithm, comprising: a data acquisition module, a data preprocessing module, a target detection module, a target correction module, and an intelligent management module; The data acquisition module is used to collect flight image data and radar echo data of the target UAV within the monitored area; The data preprocessing module is used to preprocess flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data. The target detection module constructs a target detection model based on a target detection algorithm; and converts the initial positioning results of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system based on the target detection model, standardized image data, and standardized radar data. The target correction module, based on the Hungarian algorithm, corrects the initial position data of the target UAV to obtain high-precision position data of the target UAV. The intelligent management module, based on the improved A* algorithm, performs interception path analysis on the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then executes the corresponding interception command.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: By acquiring multi-source data from flight images and radar echoes and performing standardized preprocessing, this invention provides a high-quality data foundation for subsequent target detection, effectively avoiding the limitations of a single data dimension. The target detection model built on the improved YOLOv8 algorithm framework enhances the feature extraction capability for UAVs in different scenarios and attitudes through optimized network structure and phased training strategies, improving the reliability of initial target localization. The use of the Hungarian algorithm for state correction of initial position data effectively suppresses position jitter and random errors, ensuring the stability and accuracy of high-precision target position data. Combined with the improved A* algorithm for interception path planning, fully incorporating target motion characteristics and various constraints, the generated optimal interception path balances efficiency and feasibility, ensuring the precise execution of interception commands. The entire method achieves full-process optimization from target detection and localization correction to path planning, significantly improving the intelligence and reliability of UAV interception tasks and adapting to the single-and-multi-target interception needs in complex environments. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a schematic diagram illustrating the steps of a high-precision drone interception method based on a target detection algorithm.

[0016] Figure 2 This is a schematic diagram of a module for a high-precision UAV interception system based on a target detection algorithm. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Example 1: like Figure 1 As shown, a high-precision interception method for unmanned aerial vehicles (UAVs) based on a target detection algorithm includes the following steps: Step S1: Collect flight image data and radar echo data of the target UAV within the monitoring area; preprocess the flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data; In practical applications, multimodal sensing devices need to be deployed within a pre-defined monitoring area to achieve comprehensive data collection of the target UAV, including but not limited to flight image data and radar echo data. The flight image data is acquired through a high-definition visible light camera and an infrared thermal imager at a pre-defined acquisition frequency, such as 30 frames per second, to obtain visible light and infrared image data of the target UAV. This helps to solve complex lighting scenarios such as strong daylight and low illumination at night. The resolution of the acquired image data is standardized and can be set to 1920×1080 pixels, without specific requirements. In practical applications, the radar echo data is collected based on radar, such as pulse Doppler radar. The radar echo signal of the target UAV is collected at a preset sampling frequency, and then the target distance, radial velocity, azimuth angle, etc. of the target UAV are extracted.

[0020] In practical applications, preprocessing of the collected data is also necessary for subsequent data analysis. This includes preprocessing of flight image data and radar echo data. Flight image data preprocessing consists of four steps: denoising, image enhancement, size normalization, and pixel normalization. The denoising process uses a Gaussian filtering algorithm, with specific Gaussian kernel size and standard deviation settings, to suppress salt-and-pepper noise and Gaussian noise in both visible and infrared image data, preserving target contour details. Due to the low contrast in infrared image data and uneven illumination in visible light image data during flight, further processing is required. Image enhancement processing is performed using an adaptive histogram equalization algorithm. The corresponding visible light image data and infrared image data are divided into local regions of a preset pixel size, and histogram equalization is performed independently on each local region to avoid the loss of details caused by global equalization. The images after denoising and image enhancement are uniformly scaled to standard image data of a preset pixel size, and bilinear interpolation is used to ensure pixel continuity during the image scaling process. The image pixel values ​​are converted to the [0,1] interval, and the calculation method is pixel normalization value = (original pixel value - minimum pixel value) / (maximum pixel value - minimum pixel value) to eliminate the influence of pixel value magnitude differences on subsequent algorithms.

[0021] In practical applications, radar echo data undergoes preprocessing, including clutter suppression, data smoothing, and format normalization. The clutter suppression process employs the Constant False Alarm Rate (CFAR) algorithm, based on a sliding window. The window size can be set to 16 pulse periods. The mean and variance of the background clutter are statistically analyzed, and the false alarm probability is set to... The system adaptively generates detection thresholds to remove clutter signals such as ground reflections and meteorological interference. A moving average filtering method is used, with a window length that can be set to 5 sampling points, to smooth the target distance, radial velocity, and azimuth data. The formula is as follows: ,in, The original data for the i-th sampling point, This represents the k-th smoothed data point. The corresponding smoothed radar data is then converted into a unified JSON format, including data acquisition timestamp, target distance, radial velocity, and azimuth field, to facilitate time synchronization and fusion processing with image data.

[0022] It should be noted that the above preprocessing yields timestamp-aligned standardized image data and standardized radar data.

[0023] Step S2: Based on the target detection algorithm, construct a target detection model; according to the target detection model, standardized image data, and standardized radar data, convert the initial positioning result of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system; In practical applications, the specific steps for constructing an object detection model based on object detection algorithms are as follows: A1: Acquire several sets of standardized image data and standardized radar data of UAVs with different models, flight attitudes and environments, with timestamp alignment, to form a dataset; divide the dataset into training set and test set according to a preset ratio.

[0024] It should be noted that the dataset should cover all scenarios, including sunny days with strong light, cloudy days, low light at night, fog and haze, and light rain, to improve the system's usability. In this embodiment, the dataset consists of 4,000 visible light images and 4,000 infrared images, totaling 8,000 training samples. All images have undergone standardization preprocessing and are accurately labeled using LabelImg, with the label content being the coordinates of the UAV target bounding box (x1, y1, x2, y2). The training set samples all have scene attribute labels, with no duplicate samples, no blurred, overexposed, or heavily occluded invalid samples, and all datasets are precisely aligned with the corresponding radar echo data timestamps.

[0025] A2: Based on YOLOv8 as the basic algorithm framework, the YOLOv8 basic algorithm framework is improved to obtain an improved YOLOv8 algorithm framework. It should be noted that YOLOv8 is used as the basic algorithm framework because it balances detection speed and accuracy, with short detection time per frame, making it perfectly suited for the real-time positioning requirements of drone interception. This differs from algorithms like Faster R-CNN (Fast Region Convolutional Neural Network), which offer high accuracy but low detection speed. Faster R-CNN is a two-stage detection algorithm, while YOLOv8 is a single-stage one. In the drone field, high detection speed is often more important. This also distinguishes YOLOv5 from algorithms that offer high detection speed but low accuracy. This application does not change the algorithm framework when constructing the target detection model; it only makes in-depth improvements to the internal structure of the YOLOv8 framework to avoid a decrease in real-time performance due to algorithm changes.

[0026] A3: Based on the training set, the improved YOLOv8 algorithm framework is trained in three stages to obtain the trained improved YOLOv8 algorithm framework. A4: Validate the improved YOLOv8 algorithm framework after training, and select the optimal weights as the final training model weights. The optimal weights are those with the highest detection accuracy and lowest loss value throughout the entire training process, which are fully adapted to the drone detection scenario. Based on the test set, perform full validation of the model, focusing on the drone detection performance in small target, low light, and hazy environments, and statistically analyze the false negative rate, false positive rate, and positioning deviation. If the statistical false negative rate, false positive rate, and positioning deviation all meet the corresponding detection error thresholds, the model is considered qualified and recorded as the target detection model.

[0027] It should be noted that if the detection performance in a certain scenario does not meet the standard, for example, if the accuracy of nighttime infrared image data is 80%, then the corresponding amount of training data in that scenario will be added, and the model will be retrained until the detection accuracy of all scenarios meets the requirements, thus completing the final model optimization.

[0028] In step A2, the steps for improving the YOLOv8 basic algorithm framework are as follows: It should be noted that the conventional architecture of the YOLOv8 basic algorithm framework includes an input network, a backbone feature extraction network, a neck feature fusion network, and a head detection output network; the improvements to the YOLOv8 basic algorithm framework are reflected in the improvements to the backbone feature extraction network and the neck feature fusion network.

[0029] To further explain, the backbone feature extraction network in the YOLOv8 basic algorithm framework often uses convolutional layers + pooling layers to extract features. This results in low feature differentiation between drone targets and background interference objects in images, making it easy to misdetect interference objects such as birds, kites, and balloons as drones. At the same time, it does not extract enough small target features of drones, leading to missed detections.

[0030] The neck feature fusion network in the YOLOv8 basic algorithm framework only performs simple feature fusion of upsampling and downsampling. This results in inconsistent feature capture of the target drone and large positioning deviations when the drone is flying rapidly and changing its attitude.

[0031] In the YOLOv8 basic algorithm framework, the number of feature channels is a fixed value and cannot be adjusted according to the needs of UAV detection. This results in increased computation and decreased speed when there are too many channels, and insufficient feature capacity and decreased accuracy when there are too few channels.

[0032] A21: Two lightweight coordinate attention mechanism modules (i.e., CoordAttention) are added after the 3rd and 5th convolutional layers of the backbone feature extraction network, respectively. It should be noted that the lightweight coordinate attention mechanism module is a targeted selection, which is different from the conventional SE attention and CBAM attention. The lightweight coordinate attention mechanism can simultaneously pay attention to the spatial position features and channel features of the UAV target, and can accurately locate the coordinate position of the UAV in the image, and can also weaken background features such as sky, trees, and buildings.

[0033] The parameters of the newly added lightweight coordinate attention mechanism module are adapted, including setting the attention channel compression ratio, which can be set to 4, and using ReLU6 as the activation function to avoid overfitting caused by excessive amplification of feature values. The last ordinary convolutional layer of the backbone feature extraction network is replaced with a depthwise separable convolutional layer. The preceding connection structure of the ordinary convolutional layer is retained, namely the output of the second-to-last convolutional layer and the feature map output port after the fusion of the two coordinate attention mechanism modules. The feature map output port should be aligned with the depthwise separable convolutional layer. The following connection structure of the ordinary convolutional layer is also retained, namely the output of the ordinary convolutional layer, the batch normalization layer, the ReLU6 activation function, and the feature output interface connected to it. All of these are completely retained, with no change in position or parameters.

[0034] It should be noted that depthwise separable convolutional layers can reduce the number of network parameters and computational cost without sacrificing feature extraction capabilities, thus ensuring detection accuracy without reducing detection speed. This addresses the contradiction of increased accuracy leading to decreased speed.

[0035] It should be noted that the depth-separable convolutional layer includes depthwise convolutional sub-layers and pointwise convolutional sub-layers; For example, if the input features are 80×80 pixels with 256 channels, the kernel size of the deep convolutional sub-layer can be set to 3×3, and the number of kernels can be set to the number of channels of the input features, which is 256. The deep convolutional sub-layer only extracts features in the spatial dimension of the features, that is, it only captures the spatial distribution features of the drone's position, contour, and texture in the image, without touching the feature associations of different channels, and sets the stride to 1 to ensure that the output feature map size is still 80×80 pixels with 256 channels, with no change in size or channels.

[0036] Based on the 80×80 pixel, 256-channel UAV feature map output by the deep convolutional sub-layer, the kernel size of the pointwise convolutional sub-layer can be set to 1×1, and the number of kernels can be set to the number of output channels (512).

[0037] A22: Retain the pyramid feature fusion structure of the neck feature fusion network, and perform hierarchical optimization on the pyramid feature fusion structure. Add two convolutional layers with 1×1 kernels in the shallow feature layer (i.e., the layer responsible for small target detection) to enhance the edge features of small targets; reduce one convolutional layer with 3×3 kernels in the deep feature layer (i.e., the layer responsible for large target detection) to reduce redundant features and improve feature fusion efficiency. In the middle layer of the neck feature fusion network, a cross-modal feature fusion layer is added. This cross-modal feature fusion layer differs from the conventional YOLOv8 model. It performs weighted fusion of preprocessed standardized image data and standardized radar data. The specific fusion logic is as follows: the three core parameters of target distance, radial velocity, and azimuth angle in the standardized radar data are converted into weight coefficients of features in the standardized image data. The UAV target features in the image are weighted and enhanced. For example, the faster the radial velocity of the UAV detected by the radar, the higher the weight of the UAV feature in the image, and the model can prioritize the identification of dynamic targets.

[0038] Adding a feature fusion residual connection layer, that is, after each feature fusion layer, adding a feature fusion residual connection layer to superimpose the fused features with the original features, avoids feature loss during the fusion process, and ensures that all features of the UAV can be completely learned by the model.

[0039] A23: The number of feature channels in the backbone feature extraction network was halved, changing from 64→128→256→512 to 64→128→128→256, reducing redundant channels; the number of feature channels in the neck feature fusion network was doubled, changing from 256→128→64 to 512→256→128, increasing the capacity of fused features; after adjustment, batch normalization and activation processing were performed on all channels to ensure the stability of feature values ​​and prevent any feature distortion.

[0040] In step A3, the three-stage training process for the improved YOLOv8 algorithm framework based on the training set is as follows: A31: The first stage of training includes loading the dataset and configuring data augmentation, namely setting the random horizontal flip probability (which can be set to 0.5), random scaling ratio (which can be set to 0.8~1.2 times), random color jitter (including brightness can be set to ±0.2, contrast can be set to ±0.2, and saturation can be set to ±0.2), Gaussian blur (the convolution kernel can be set to 3×3, and the Gaussian blur probability can be set to 0.2), and random pixel value offset (the random pixel value offset can be set to ±0.05). By fully augmenting the dataset, the generalization ability of the model is improved, avoiding premature overfitting, and allowing the model to learn the basic features and shapes of the drone at different angles, lighting conditions, and scales. The hyperparameter configuration includes: setting the initial learning rate to 0.0008 and not implementing any learning rate decay strategy; using AdamW as the optimizer, with a fixed weight decay coefficient of 0.0007 and a gradient clipping threshold of 1.0; fixing the batch size to 16 to determine the balance between training efficiency and gradient stability; and setting a model weight saving rule, i.e., automatically saving a temporary model weight file once every 5 training rounds. The weight file contains parameter information for all current iterations, and the storage path is separate from the final weights to handle breakpoints in training interruptions and avoid the loss of previous training results. After each training round, the total loss value, classification loss value, and bounding box regression loss value of the training set are calculated in real time. If the loss value suddenly increases or fluctuates continuously, the learning rate is immediately reduced to 0.0006, while other parameters remain unchanged. Training continues until the loss regression stabilizes and decreases. After 40 training iterations, if the total loss value of the training set in the dataset drops below 0.25, the classification loss value is ≤0.08, and the bounding box regression loss value is ≤0.17, the current training phase can be terminated without premature termination, and the second training phase can begin.

[0041] A32: The second stage of training includes the following specific processes: Targeted emphasis on calling hard-to-distinguish samples, increasing the proportion of drone small target samples, nighttime infrared samples, haze scene samples, and samples partially occluded on the fuselage in the dataset, allowing the model to actively learn the features of hard-to-detect samples; Disabling the random scaling ratio and Gaussian blur enhancement operations, as other operations are inconvenient. It should be noted that excessive scaling and Gaussian blurring can destroy the detailed features of small targets on drones and the heat source features of infrared images.

[0042] The hyperparameters are dynamically adjusted, meaning the learning rate is decayed in a stepwise manner. For the first 29 rounds (rounds 41 to 69), the learning rate is kept constant at 0.0008. At the beginning of the 70th round of training, the learning rate is reduced to 0.0004 all at once. By reducing the learning rate, the iteration amplitude of the model parameters is reduced, thus achieving slow and deep learning of refined features. The other hyperparameters remain unchanged.

[0043] The frequency of weight saving is optimized. Temporary weights are saved once every 3 rounds of training. A filtering logic is added during saving to save only weight files with training loss values ​​lower than the previous round and test set detection accuracy higher than the previous round. Invalid weights that do not improve performance are removed to reduce storage redundancy and lock the optimal parameter iteration direction.

[0044] After each training round, calculate the total loss value of the corresponding training set. The total loss value should maintain a continuous and slow downward trend without stagnation. Calculate the detection accuracy, small target drone detection rate, and overall false detection rate of the corresponding test set. The target detection accuracy of drones on the test set should be ≥97%, the small target detection rate of drones should be ≥95%, and the overall false detection rate should be ≤1%. If the total loss value continues to decrease but the accuracy of the test set stagnates, it is considered slightly overfitting. The AdamW weight decay coefficient is increased to 0.0009 to suppress overfitting before continuing training.

[0045] Once 80 training iterations are completed, and the target detection accuracy of the UAV on the test set is stable at 97% or above, the small target detection rate of the UAV is stable at 95% or above, the overall false detection rate is stable at 1% or below, and the total loss value of the corresponding training set is reduced to below 0.12, the current training phase can be terminated, and the third training phase can begin.

[0046] A33: The third stage of training includes loading unobstructed drone targets with clear outlines and complete features into the dataset, with the scene being a single-light environment; removing extremely difficult-to-distinguish samples with severe occlusion and extreme blur; and completely disabling all data augmentation configurations. It should be noted that this third stage of training is for precise parameter convergence. Extremely difficult-to-distinguish samples will interfere with the model's optimal parameter iteration. This allows the model to focus on learning the core features of clean drones, thereby maximizing the improvement of detection accuracy.

[0047] The learning rate undergoes a double-step decay: after the 80th training round, the learning rate is immediately reduced from 0.0004 to 0.0002 and kept constant; all optimizer parameters are kept unchanged to ensure training stability.

[0048] It should be noted that a low learning rate allows for very small iterations of model parameters, enabling fine-tuning of feature weights and avoiding parameter oscillations caused by a large learning rate.

[0049] Save the weight file once after every two rounds of training, including the best weight (best.pt) and the latest weight (last.pt).

[0050] An early termination mechanism is activated, with the monitoring metric being the target detection accuracy of the test set drones. If, after eight consecutive rounds of training, the target detection accuracy of the test set does not improve and the total loss value of the training set does not show a downward trend, the training phase is terminated immediately without having to complete 120 rounds.

[0051] It should be noted that enabling the early stopping mechanism completely avoids overtraining and overfitting of the model, ensuring the model's generalization ability.

[0052] After each training round, the total loss value of the training set must be reduced to below 0.08; the target detection accuracy of the test set drones must be ≥99%; the small target detection rate of the test set drones must be ≥98%; the overall false detection rate and false negative rate of the test set must be ≤0.5% and ≤0.5%; once all these standards are met, the training can be considered successful regardless of whether 120 training rounds have been completed.

[0053] This training phase can be terminated if any of the following conditions are met: Complete 120 rounds of full training, and the total loss value of the training set, the target detection accuracy of the UAV in the test set, the small target detection rate of the UAV in the test set, the overall false detection rate and false negative rate of the test set all meet the standards. Trigger the early stop mechanism.

[0054] In practical applications, the steps for converting the initial positioning result of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system, based on the target detection model, standardized image data, and standardized radar data, are as follows: B1: Input standardized image data and standardized radar data into the trained target detection model, and output the bounding box coordinates of the target UAV in the image coordinate system. Calculate the coordinates of the center point of the bounding box. The formula is: , , And serve as the initial positioning result in the image coordinate system; B2: Based on the target distance, azimuth angle, and camera installation parameters in the corresponding standardized radar data, the coordinates of the center point of the bounding box in the image coordinate system are transformed using a coordinate transformation formula. Convert to initial location data in geographic coordinate system .

[0055] It should be noted that the camera's installation parameters include, but are not limited to, the camera's focal length, pixel size, tilt angle, and installation height.

[0056] In step B2, the initial location data is converted to a geographic coordinate system. The specific steps are as follows: B21: Transformation from image coordinate system to camera coordinate system: , , ;in, The width and height of the visible light image data; The target distance; The focal length of the camera; This refers to the pixel size of the camera.

[0057] B22: Camera coordinate system to geographic coordinate system conversion: ; ; ;in, The geographic coordinates of the unified installation reference point for radar and cameras are obtained by calibration using GPS positioning equipment; The target distance; It is the azimuth angle; The camera's tilt angle; This refers to the installation height of the camera.

[0058] Through the above transformation, the initial position data of the target UAV in the geographic coordinate system is obtained. .

[0059] To further explain, azimuth angle The camera's tilt angle .

[0060] Step S3: Based on the Hungarian algorithm, perform state correction on the initial position data of the target UAV to obtain high-precision position data of the target UAV; In practical applications, the steps for correcting the initial position data of the target UAV based on the Hungarian algorithm are as follows: C1: Determine whether to perform state correction on the initial position data of the target UAV. Continuously collect N frames of the initial position data of the target UAV to form a position data sequence. , Record the acquisition timestamp and radar radial velocity data for each frame of data; calculate the position data sequence. Statistical characteristics, including the mean and standard deviation When the deviation value of a certain frame of data, i.e. or or If the data is deemed abnormal, it needs to be corrected through correlation matching.

[0061] C2: If state correction is required, a cost matrix C is constructed based on the Hungarian algorithm. This involves using the position data of G consecutive frames as a sliding window. For adjacent frames within the sliding window, let the position data of the previous frame be the source node set A = {a1, a2, ..., am}, and the position data of the current frame be the target node set B = {b1, b2, ..., bn} (m and n represent the number of targets in the corresponding frames, and m = n = 1 indicates a single-target UAV detection scenario; m ≠ n or m = n > 1 indicates a multi-target UAV detection scenario). The element c_ij in the cost matrix C is defined as the matching cost between the i-th position point in the source node set A and the j-th position point in the target node set B. The lower the similarity, the larger the matching cost c_ij; the higher the similarity, the smaller the matching cost c_ij.

[0062] It's important to note that, based on the Hungarian algorithm, the matching relationship with the globally minimum matching cost is found in the cost matrix to determine the optimal association between the position data of the previous frame and the current frame. Essentially, the Hungarian algorithm verifies whether the initial position of the current frame is a continuous movement position of the target in the previous frame. In a single-target UAV detection scenario, the cost matrix is ​​a 1×1 square matrix, and the optimal matching solution of the Hungarian algorithm is unique. This means that a unique cost element is directly selected as the minimum matching cost, determining the source node and target node as the unique optimal matching pair, thus completing the precise association between the positions of the preceding and following frames. After solving using the Hungarian algorithm, a unique matching result is output, indicating that the corrected position of the previous frame and the initial position of the current frame are the same UAV target in a temporal sequence, establishing a stable inter-frame position association. The core value of this association is locking in the continuity of the target's motion trajectory, avoiding trajectory breaks caused by single-frame positioning errors, and providing reliable temporal constraints for subsequent smooth position correction. Compared to direct inter-frame data fusion, the optimal matching of the Hungarian algorithm can effectively eliminate false positioning points and background interference points, ensuring that the data source for each correction is the valid position of the same target, resulting in correction results without drift or misalignment.

[0063] Calculate the matching cost between the two target locations based on the source node set A and the target node set B. The matching cost for: ;in, , , Represents the three-dimensional geographic coordinates of the corrected position of the i-th UAV in the previous frame in the source node set A; , , Represents the three-dimensional geographic coordinates of the initial position of UAV j in the current frame within the source node set B; This represents the matching cost between two target locations, expressed in meters (m). It is the matching cost between the target locations corresponding to source node set A and source node set B.

[0064] Set spatial distance threshold It can be adjusted according to the drone's flight speed.

[0065] For example, in a single-target UAV detection scenario (m=n=1), the matching cost... If the cost is less than 1, it is considered a valid match cost, and the two frames represent consecutive positions of the same drone; if the cost is less than 1, it is considered a valid match cost. If the match fails, it is considered invalid and the match is discarded. The initial position of the current frame is used as the basis for correction. In a multi-target UAV detection scenario (m=3, n=2), the cost matrix is ​​a 3×2 matrix, and the matching cost includes... , , , , , This comprehensively covers all position matching combinations of the three drones in the previous frame and the two drones in the current frame; if the matching cost of the corresponding element... Mark them as candidate matching pairs; if the matching cost for all elements is... If it is determined that there are no continuous targets in the current frame, the initial position of the current frame is used as the basis for correction; if the matching cost of some elements is... Only the corresponding candidate matching pairs are retained to participate in the subsequent optimal matching.

[0066] It should be noted that after the optimal matching is completed, the initial position of the current frame has been determined to be a continuous position of the same target. At this time, there are two types of errors that need to be corrected: 1. Random position error during single-frame coordinate transformation; 2. The instantaneous positional jitter of the drone's movement.

[0067] To further explain, the matching combination with the global minimum cost is found in the m×n cost matrix to ensure that each source node matches at most one target node and each target node matches at most one source node, so as to determine the optimal association relationship between each UAV target in the previous frame and the current frame.

[0068] C3: The minimum matching cost of the cost matrix is ​​solved using the Hungarian algorithm to determine the optimal correlation between the position data of the preceding and following frames. Based on the corresponding radial velocity, the initial position data of the optimal correlation are then smoothly corrected using the following formula: ;in, This represents the high-precision position data of the UAV in the current frame after smoothing correction, i.e. ; This represents the location fusion weighting coefficient, with a value of 0.7. This represents the initial position data of the current frame after coordinate system transformation, i.e. ; This represents the initial position data of the previous frame after coordinate system transformation, i.e. ; This represents the radial velocity (m / s) of the corresponding UAV acquired by the radar. It is a three-dimensional spatial velocity, corresponding to the velocity components in the X, Y, and Z dimensions, and participates in the dimensional displacement compensation. This represents the time interval (s) between the acquisition of two consecutive frames of data. The theoretical displacement compensation amount of the UAV between frames is used to correct the position according to the actual motion state of the UAV, so that the correction result fits the actual flight trajectory of the UAV, rather than simply fusion of static data.

[0069] It should be noted that in high-precision location data, ; ; .

[0070] Step S4: Based on the improved A* algorithm, perform interception path analysis on the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then execute the corresponding interception command.

[0071] Based on the geographic coordinates of the monitored area, a 3D raster map is constructed. The raster size of the 3D raster map is set, no-fly zones (such as buildings and sensitive facilities) are marked as obstacle grids, and constraints are set, including: 1. Dynamic constraints: Maximum flight speed threshold, maximum acceleration threshold, maximum climb rate threshold, and maximum descent rate threshold for intercepting drones; 2. Safety Constraints: The minimum distance threshold between the interception path and the obstacle grid, as well as the final interception distance control range of the target drone, are preset and can be adjusted according to the effective range of the interception equipment.

[0072] It should be noted that the heuristic function of the traditional A* algorithm only considers Euclidean distance. The improved version incorporates radial velocity from the radar echo data, specifically: ;in, , These are the weighting coefficients, and , ; The geographic coordinates of the current node n are obtained by raster index mapping; Here is the radial velocity (in m / s) of the target drone in a single-target drone detection scenario, and the corresponding radial velocity of the target drone in a multi-target drone detection scenario is... ; To intercept the drone's planned speed (m / s) at the current node n, it is initially set to... The path will be dynamically adjusted as needed. The maximum flight speed threshold for intercepting drones.

[0073] It should be noted that node n refers to the basic searchable unit in a 3D raster map. Each node corresponds to a specific location in 3D space and carries the core information required for path search.

[0074] Based on the improved A* algorithm, two node sets are created: the OPEN set, which stores candidate nodes to be explored, and the CLOSED set, which stores nodes that have been explored. Starting with the current high-precision position data of the target drone, a sequential search is performed according to priority, and constraint verification is conducted based on the constraints to generate the optimal interception path for the target drone. If all constraint verifications are satisfied, the corresponding path is the optimal interception path; if the constraint verifications are not satisfied, the parameters are adjusted and the path is replanned, such as increasing the weight coefficient of the heuristic function. Alternatively, adjust the grid size until the path passes verification.

[0075] It should be noted that in multi-target UAV detection scenarios, the OPEN set and CLOSED set are initialized independently for each target, and the search is performed in parallel or sequentially according to priority.

[0076] The optimal interception path is divided according to the flight state of the target UAV into an acceleration phase, a constant speed phase, a deceleration phase, and an interception phase, and then the corresponding interception execution command is executed. It should be noted that this application reserves emergency instructions, i.e., sets emergency interruption instructions, which have the highest priority and can interrupt normal interception instructions at any time.

[0077] like Figure 2 As shown, a high-precision UAV interception system based on a target detection algorithm is disclosed. The system includes, but is not limited to, a data acquisition module, a data preprocessing module, a target detection module, a target correction module, and an intelligent management module. The data acquisition module is used to collect flight image data and radar echo data of the target UAV within the monitored area; The data preprocessing module is used to preprocess flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data. The target detection module constructs a target detection model based on a target detection algorithm; and converts the initial positioning results of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system based on the target detection model, standardized image data, and standardized radar data. The target correction module, based on the Hungarian algorithm, corrects the initial position data of the target UAV to obtain high-precision position data of the target UAV. The intelligent management module, based on the improved A* algorithm, performs interception path analysis on the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then executes the corresponding interception command.

[0078] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0079] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0080] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0081] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0082] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0085] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A high-precision interception method for unmanned aerial vehicles (UAVs) based on a target detection algorithm, characterized in that, The method includes: Collect flight image data and radar echo data of the target UAV within the monitoring area; preprocess the flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data; Based on the target detection algorithm, a target detection model is constructed; according to the target detection model, standardized image data, and standardized radar data, the initial positioning result of the target UAV in the image coordinate system is converted into the initial position data of the target UAV in the geographic coordinate system. Based on the Hungarian algorithm, the initial position data of the target UAV is corrected to obtain high-precision position data of the target UAV. Based on the improved A* algorithm, the interception path is analyzed using the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then the corresponding interception execution command is executed.

2. The high-precision UAV interception method based on target detection algorithm according to claim 1, characterized in that, The process of constructing an object detection model based on object detection algorithms includes: Acquire several sets of standardized image data and standardized radar data of UAVs of different models, flight attitudes, and environments with timestamp alignment to form a dataset; divide the dataset into training set and test set according to a preset ratio; Based on YOLOv8 as the basic algorithm framework, the YOLOv8 basic algorithm framework is improved to obtain an improved YOLOv8 algorithm framework. Based on the training set, the improved YOLOv8 algorithm framework is trained in three stages to obtain the trained improved YOLOv8 algorithm framework. The improved YOLOv8 algorithm framework after training was validated. If the validation was successful, the corresponding object detection model was obtained.

3. The high-precision UAV interception method based on target detection algorithm according to claim 2, characterized in that, The process of improving the YOLOv8 basic algorithm framework includes: A lightweight coordinate attention mechanism module was added to the backbone feature extraction network in the YOLOv8 basic algorithm framework, and the number of feature channels in the backbone feature extraction network was adjusted. The neck feature fusion network in the YOLOv8 basic algorithm framework is optimized hierarchically, including: pyramid feature fusion structure, intermediate layer and feature fusion residual connection layer; and the number of feature channels in the neck feature fusion network is adjusted.

4. The high-precision UAV interception method based on target detection algorithm according to claim 3, characterized in that, The process of training the improved YOLOv8 algorithm framework in three stages based on the training set includes: The first stage of training includes: loading the dataset, configuring data augmentation, setting hyperparameters, and completing the corresponding number of training iterations or if the detection metrics of the training set in the dataset all meet the requirements, then proceeding to the second stage of training. The second stage of training includes: increasing the proportion of drone small target samples, nighttime infrared samples, haze scene samples, and fuselage partially occluded samples in the dataset; dynamically adjusting hyperparameters and optimizing the weight saving frequency; and proceeding to the third stage of training after completing the corresponding number of training iterations or when the corresponding detection metrics of the training set and test set in the dataset meet the requirements. The third stage of training includes: loading the proportion of drone targets in the dataset that are unobstructed, have clear outlines, and complete features, and are in a single-lighting environment, and completely disabling all data augmentation configurations; performing a second step decay of the learning rate while keeping all optimizer parameters unchanged; and completing the corresponding number of training iterations and ensuring that the corresponding detection metrics of the training and test sets in the dataset meet the requirements, or triggering the early stopping mechanism, then the third stage of training is complete.

5. A high-precision UAV interception method based on a target detection algorithm according to claim 4, characterized in that, The process of converting the initial localization result of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system, based on the target detection model, standardized image data, and standardized radar data, includes: Standardized image data and standardized radar data are input into the trained target detection model, which outputs the bounding box coordinates of the target UAV in the image coordinate system, calculates the coordinates of the center point of the bounding box, and uses it as the initial localization result in the image coordinate system. Based on the target distance, azimuth angle, and camera installation parameters in the corresponding standardized radar data, the coordinates of the bounding box center point in the image coordinate system are converted into the initial position data in the geographic coordinate system using a coordinate transformation formula.

6. A high-precision UAV interception method based on a target detection algorithm according to claim 5, characterized in that, The process of correcting the initial position data of the target UAV based on the Hungarian algorithm includes: Determine whether to perform state correction on the initial position data of the target UAV. If state correction is required, construct a cost matrix based on the Hungarian algorithm to obtain the corresponding matching cost, set a spatial distance threshold, determine the optimal correlation between the position data of the preceding and following frames based on the spatial distance threshold and the corresponding matching cost, and then perform smooth correction on the initial position data of the optimal correlation based on the corresponding radial velocity to obtain the corresponding high-precision position data.

7. A high-precision UAV interception method based on a target detection algorithm according to claim 6, characterized in that, The process of interception path analysis based on the improved A* algorithm and the high-precision position data and radar echo data of the target UAV includes: Based on the geographical coordinates of the monitored area, a three-dimensional raster map is constructed. The raster size of the three-dimensional raster map is set, no-fly zones are marked as obstacle grates, and constraints are set. The radial velocity from radar echo data is introduced into the heuristic function of the A* algorithm to obtain an improved A* algorithm. Based on the improved A* algorithm, two node sets are created. Starting from the current high-precision position data of the intercepting UAV, a sequential search is performed according to priority, and constraint verification is performed according to the constraints to generate the optimal interception path of the target UAV. If all constraint verifications are satisfied, the corresponding path is the optimal interception path. If the constraint verifications are not satisfied, the parameters are adjusted and replanned. The optimal interception path is divided according to the flight state of the target UAV, and the corresponding interception execution command is executed.

8. A high-precision UAV interception system based on a target detection algorithm, implementing the high-precision UAV interception method based on a target detection algorithm as described in any one of claims 1 to 7, characterized in that, include: The system includes a data acquisition module, a data preprocessing module, a target detection module, a target correction module, and an intelligent management module. The data acquisition module is used to collect flight image data and radar echo data of the target UAV within the monitored area; The data preprocessing module is used to preprocess flight image data and radar echo data to obtain corresponding standardized image data and standardized radar data. The target detection module constructs a target detection model based on a target detection algorithm; and converts the initial positioning results of the target UAV in the image coordinate system into the initial position data of the target UAV in the geographic coordinate system based on the target detection model, standardized image data, and standardized radar data. The target correction module, based on the Hungarian algorithm, corrects the initial position data of the target UAV to obtain high-precision position data of the target UAV. The intelligent management module, based on the improved A* algorithm, performs interception path analysis on the high-precision position data and radar echo data of the target UAV to obtain the optimal interception path of the target UAV, and then executes the corresponding interception command.

Images