AI image recognition and high-speed tracking collaborative method and system

Abstract

The application provides an AI image recognition and high-speed tracking cooperation method and system, relates to the technical field of image recognition and tracking, and comprises the following steps: a radar guides a photoelectric turntable to shoot an unknown target, and acquires an AI image; based on an environment situation awareness baseline established adaptively, a white list target is removed from the image, and a target data set with a situation energy label is obtained; a cosine similarity of a pixel point and a preset feature is calculated, and a heat map image is generated; a target area is located in the heat map, a multi-scale feature vector is generated by using a multi-scale feature generation algorithm, and a feature matching degree is obtained by comparing the multi-scale feature vector with a target template library in layers; whether the target is a target to be intercepted is determined based on the matching degree, and a future track of the target is predicted; according to the predicted track and a threat level, an interception device is distributed, and a 3D interception network is formed through inter-machine communication, so that the target enters a capture range. The application can realize dynamic tracking and efficient interception, and improve the response speed and accuracy of airspace safety protection.

Description

Technical Field

[0001] This application relates to the technical field of image recognition and tracking, and in particular to an AI image recognition and high-speed tracking collaborative method and system. Background Technology

[0002] In modern low-altitude security scenarios, important areas such as large event sites, nuclear power facilities, and airport perimeters face potential threats from low-altitude aircraft such as multi-rotor drones and high-speed racing drones. These targets, with their high maneuverability and stealth, pose a challenge to public safety. Therefore, there is an urgent need for a technological solution capable of rapidly identifying unknown targets, accurately tracking their trajectories, and efficiently allocating interception resources in complex airspace environments to meet the practical needs of real-time prevention and precise response.

[0003] Currently, existing technologies typically employ a single type of detection equipment, such as radar or optoelectronic devices, to detect targets entering the protected area. By accessing historical target location data, linear fitting or uniform motion models are used to predict the trajectory, i.e., the position within a short period of time is estimated based on the current speed. When planning the interception path, the interception equipment is controlled to move at a fixed speed based on the straight-line distance between the predicted target position and the current position of the interception equipment. Simultaneously, feature comparison is performed based on the original image using a single scale, and tracking control is implemented according to fixed parameters.

[0004] However, the main drawback of existing technologies is that they are difficult to adapt to the changing motion states and diverse characteristics of targets in complex low-altitude environments. This results in limitations in target identification accuracy, trajectory prediction precision, and tracking control flexibility, making it impossible to efficiently meet the needs of accurate identification and dynamic tracking in complex scenarios. Summary of the Invention

[0005] The purpose of this application is to provide an AI image recognition and high-speed tracking collaborative method and system to solve the problems of low image recognition accuracy and poor tracking flexibility in the prior art.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides an AI image recognition and high-speed tracking collaborative method, comprising:

[0007] After determining that the radar has detected an unknown target based on the radar echo image, the photoelectric turntable is guided to take pictures to obtain an AI image containing at least one unknown target in the airspace.

[0008] By using the environmental situation awareness baseline corresponding to the airspace, data of existing targets in the preset whitelist are removed from the AI ​​image to obtain a target data set carrying situational energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located.

[0009] Based on the target dataset, the similarity between each pixel in the AI ​​image and the target feature template is calculated using the cosine similarity formula. Based on the similarity, a heat map is generated.

[0010] Based on the spatial location and morphological parameters of the unknown target, the region to be analyzed containing the unknown target is located in the thermal image. A multi-scale feature generation algorithm is introduced to generate multi-scale feature vectors by taking the depth features of the region to be analyzed at different proportions. The multi-scale feature vectors are then compared hierarchically with existing vector templates in a preset multi-type target template library to obtain the feature matching degree.

[0011] Based on the feature matching degree, it is determined whether the unknown target is a target to be intercepted. If the unknown target is a target to be intercepted, the predicted flight path of the target to be intercepted is predicted within a future preset time period.

[0012] Based on the predicted flight path and the threat level of the target to be intercepted, corresponding interception equipment is allocated, and the interception equipment is controlled to move and communicate with each other in real time to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

[0013] Optionally, the step of removing data of existing targets in a preset whitelist from the AI ​​image using the environmental situational awareness baseline corresponding to the airspace to obtain a target data set carrying situational energy markers includes:

[0014] The brightness data of the airspace environment is obtained. If the brightness data is lower than the preset brightness threshold, night brightness enhancement technology is introduced to enhance the brightness of the AI ​​image. The AI ​​image with enhanced brightness is then adaptively adjusted in resolution to obtain the AI ​​image with adjusted resolution.

[0015] Based on the regional deployment architecture of the airspace and the environment in which the airspace is located, an environmental situation awareness baseline corresponding to the airspace is adaptively established.

[0016] Based on the environmental situation awareness baseline, the airspace boundary and whitelist of the protected area are determined. Image data exceeding the airspace boundary are removed to obtain the first optimized AI image. Data of existing targets in the preset whitelist are removed from the first optimized AI image to obtain the second optimized AI image. Situation energy markers are determined based on the second optimized AI image and the environmental situation awareness baseline.

[0017] The data in the AI ​​image after secondary optimization is structured, and the structured data and the tagging information are combined to obtain the target data set. The tagging information includes: situational energy tag, ground detection equipment identifier, interception equipment status and unknown target temporary identifier.

[0018] Secondly, this application provides an AI image recognition and high-speed tracking collaborative system, including:

[0019] The acquisition module is used to guide the photoelectric turntable to take pictures after determining that the radar has detected an unknown target based on the radar echo image, so as to acquire an AI image containing at least one unknown target in the airspace.

[0020] The elimination module is used to remove data of existing targets in the preset whitelist from the AI ​​image through the environmental situation awareness baseline corresponding to the airspace, so as to obtain a set of target data carrying situational energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located.

[0021] The calculation module is used to calculate the similarity between each pixel in the AI ​​image and the target feature template using the cosine similarity formula based on the target dataset, and to generate a heat map based on the similarity.

[0022] The positioning and comparison module is used to locate the region to be analyzed containing the unknown target in the thermal image based on the spatial location and morphological parameters of the unknown target. A multi-scale feature generation algorithm is introduced to generate multi-scale feature vectors by taking the depth features of the region to be analyzed at different proportions. The multi-scale feature vectors are then compared hierarchically with existing vector templates in a preset multi-type target template library to obtain the feature matching degree.

[0023] The determination module is used to determine whether the unknown target is a target to be intercepted based on the feature matching degree, and if the unknown target is a target to be intercepted, to predict the predicted flight path of the target to be intercepted within a future preset time period;

[0024] The allocation control module is used to allocate corresponding interception equipment based on the predicted flight trajectory and the threat level of the target to be intercepted, and to control the interception equipment to move and communicate with each other in real time to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

[0025] Thirdly, this application provides an electronic device, comprising:

[0026] Memory, used to store computer programs;

[0027] A processor for executing the computer program to implement the steps of the AI ​​image recognition and high-speed tracking collaborative method as described in the first aspect above.

[0028] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the AI ​​image recognition and high-speed tracking collaborative method described in the first aspect above.

[0029] The AI ​​image recognition and high-speed tracking collaborative method provided in this application has the following effects: By using radar and photoelectric collaborative detection, it ensures that the acquired AI images accurately contain unknown targets, providing a reliable data source for subsequent analysis; With the help of an adaptively established environmental situational awareness baseline, it can intelligently remove known target data from a preset whitelist, greatly reducing data processing redundancy and allowing computing resources to focus on real unknown threats; By using cosine similarity calculation combined with an adaptive weighting strategy, it generates intuitive thermal images, which can quickly and globally locate potential target areas, improving the efficiency of preliminary screening of suspicious targets;

[0030] Furthermore, by introducing a multi-scale feature generation algorithm and performing hierarchical comparison with a classification template library, the difficulty of target identification caused by changes in scale and attitude is effectively overcome, significantly improving the comprehensiveness and accuracy of feature matching, thereby enhancing the robust identification capability for unknown targets. Threat determination is based on high-confidence feature matching results, and future trajectories are predicted by combining historical motion data, achieving a leap from static identification to dynamic threat assessment, providing a critical time window and precise guidance for interception decisions. Finally, based on the predicted trajectory and threat level, interception resources are intelligently scheduled to form a coordinated 3D interception network, realizing a closed loop of perception, judgment, decision-making, and action, improving the success rate of coordinated interception of high-speed maneuvering targets and the overall effectiveness of regional protection.

[0031] Furthermore, this application ensures image quality under different lighting conditions through adaptive brightness enhancement and resolution adjustment; by integrating environmental, geographical, and deployment information to establish a dynamic baseline, it can accurately define protection boundaries and identify legitimate targets; based on this baseline, it performs two-level optimization of image data through boundary screening and whitelist removal, effectively filtering out background interference and known targets, and significantly improving the focus on unknown and suspicious targets; the generated situational energy markers endow targets with dynamic environmental association attributes, and the structured target data set provides a high-quality, information-complete, and highly focused standardized data foundation for subsequent feature extraction and threat determination.

[0032] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating an AI image recognition and high-speed tracking collaborative method provided in this application embodiment;

[0035] Figure 2 This is a schematic diagram illustrating a specific implementation of an AI image recognition and high-speed tracking collaborative method provided in an embodiment of this application;

[0036] Figure 3 A schematic diagram of the structure of an AI image recognition and high-speed tracking collaborative system provided in this application embodiment;

[0037] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0038] In current low-altitude security scenarios, existing target identification and tracking technologies are insufficient to meet the precise control requirements in complex environments. The core shortcomings are concentrated in three aspects: First, data acquisition is limited to a single dimension, relying heavily on radar or optoelectronic equipment to acquire target information. This fails to comprehensively capture the multi-dimensional characteristics of targets, easily leading to identification errors due to incomplete data. Second, trajectory prediction and target determination methods are simplistic, often employing linear fitting or uniform motion models to calculate target trajectories. This is ill-suited to the flexible and ever-changing motion states of low-altitude targets, and relying solely on single-scale feature comparisons to determine target attributes results in insufficient accuracy. Third, tracking and control flexibility is lacking. The allocation of interception equipment does not fully consider the target threat level, and there is a lack of effective relocation methods after target loss, easily leading to tracking interruptions. This makes it difficult to effectively address the security challenges posed by highly concealed and maneuverable targets such as multi-rotor drones and high-speed racing drones.

[0039] To address the aforementioned issues, this application proposes a collaborative method for AI image recognition and high-speed tracking. This method first overcomes the limitations of single-sensor information by fusing multi-source data such as radar echo images, infrared radiation images, and digital images. Then, through the adaptive establishment of an environmental situational awareness baseline, it effectively eliminates known targets from a whitelist and adds situational energy markers, achieving efficient focusing and situational awareness of unknown targets. Subsequently, by extracting depth features from the target data set and generating thermal images based on morphological and infrared correlation features, it achieves in-depth feature mining and intuitive visualization of targets.

[0040] Furthermore, a multi-scale feature generation algorithm is introduced and compared hierarchically with a multi-type target template library. Combined with reinforcement learning for dynamic optimization, the accuracy and adaptability of feature matching are improved. Based on the optimized feature matching, the target to be intercepted is determined and its flight trajectory is predicted in seconds, improving the reliability of threat assessment and the proactiveness of interception preparation. Finally, interception equipment is allocated according to the threat level, and a 3D interception network is formed through real-time inter-aircraft communication, realizing the collaborative tracking and precise interception of high-speed maneuvering targets. This solution forms a technical closed loop from perception, identification, decision-making to interception, comprehensively improving the response speed, accuracy, and overall effectiveness of airspace security protection.

[0041] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] The core of this application is to provide a collaborative method for AI image recognition and high-speed tracking, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0043] S101. After determining that the radar has detected an unknown target based on the radar echo image, guide the photoelectric turntable to take pictures to obtain an AI image containing at least one unknown target in the airspace.

[0044] In the above scheme, radar echo image refers to the image formed by the radar system receiving the echo signal reflected by the target after transmitting electromagnetic waves into the airspace and converting it. It contains basic information such as the approximate spatial location of the target and the reflection intensity, and is used to initially detect unknown targets.

[0045] Unknown targets refer to airspace targets that cannot be matched in the preset known target database in radar echo images, or whose identity and threat attributes are not yet clear.

[0046] An optoelectronic turntable is an adjustable device equipped with an infrared camera and a high-definition digital camera. It has the ability to adjust the angle in the horizontal and vertical directions and is used to accurately capture images of a specified airspace.

[0047] AI images refer to image data specifically designed for artificial intelligence recognition and processing. In this step, for example, the AI ​​images include infrared radiation images and digital images. Infrared radiation images are generated by capturing the infrared energy distribution of the target, which can clearly present the thermal characteristics of the target and is not affected by lighting conditions. Digital images refer to visible light images captured by high-definition cameras, which can accurately restore the spatial features of the target, such as its shape outline and structural details.

[0048] The aforementioned S101 achieves rapid detection of unknown targets in the airspace through preliminary detection and target determination by the radar system, providing target clues for subsequent accurate identification; it converts radar coordinates into control parameters of the photoelectric turntable, ensuring the accuracy of the shooting direction and avoiding resource waste caused by blind shooting; at the same time, it acquires infrared radiation images and digital images, with the two images focusing on the thermal features and spatial details of the target respectively, forming data complementarity and effectively making up for the recognition limitations of a single image type in complex environments.

[0049] S102. Using the environmental situation awareness baseline corresponding to the airspace, remove the data of existing targets in the preset whitelist from the AI ​​image to obtain a target data set carrying situational energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located.

[0050] Among them, the regional deployment architecture of the airspace refers to the basic information set of all ground detection equipment and interception equipment launch points within the protected airspace. This architecture can include ground detection equipment such as radar stations and electro-optical turntables, as well as the geographical coordinates, actual coverage area, and equipment orientation angle of the interception equipment launch points, providing a basis for equipment deployment for baseline establishment;

[0051] Situational energy labeling refers to an identifier generated by combining environmental situation and target status, reflecting the threat-related characteristics of the target in the current environment.

[0052] In one specific implementation, S102, using the environmental situational awareness baseline corresponding to the airspace, data of existing targets in a preset whitelist are removed from the AI ​​image to obtain a target data set carrying situational energy markers, including:

[0053] Step 1021: Obtain the brightness data of the environment in the airspace. If the brightness data is lower than the preset brightness threshold, introduce nighttime brightness enhancement technology to enhance the brightness of the AI ​​image. Perform adaptive resolution adjustment on the AI ​​image after brightness enhancement to obtain the AI ​​image after resolution adjustment.

[0054] Among them, brightness data refers to the actual light intensity parameters of the spatial environment, which are used to determine whether AI images need enhancement processing; the preset brightness threshold is the standard value for determining whether the light is sufficient, and if it is lower than the value, image enhancement needs to be activated.

[0055] Nighttime brightness enhancement technology is a technique that improves the clarity of low-light images through pixel brightness adjustment algorithms, used to improve image quality in dark environments; adaptive resolution adjustment refers to a processing method that dynamically optimizes resolution based on the complexity of image content, balancing image clarity and processing efficiency; this embodiment does not specifically limit the algorithm type of the above enhancement technology or the algorithm type used for adjustment.

[0056] Step 1022: Based on the regional deployment architecture of the airspace and the environment in which the airspace is located, adaptively establish the environmental situation awareness baseline corresponding to the airspace.

[0057] Among them, the environmental situation awareness baseline refers to the core reference standard that is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located. This baseline includes three parts: temporary whitelist rules, performance attenuation coefficient, and corrected airspace boundary, which are used to accurately filter effective target data in AI images.

[0058] As a specific implementation method, step 1022 may specifically include the following steps:

[0059] Step a1: Obtain digital map data of the protected area and import the regional deployment architecture of the airspace.

[0060] Among them, digital map data refers to the collection of geographic information of the protected area, which may include terrain elevation, location and height of fixed buildings, providing a geographic basis for building an airspace model; the regional deployment architecture includes the geographic coordinates, coverage area and orientation angle of the launch points of all ground detection equipment and interception equipment.

[0061] Step a2: Obtain environmental data of the airspace in real time by connecting to meteorological sensors or preset meteorological data interfaces.

[0062] The environmental data includes visibility, ambient brightness, wind speed, wind direction, cloud base height, and precipitation data, which are used to influence the dynamic adaptability of the baseline.

[0063] Step a3: Integrate the digital map data, the regional deployment architecture, and the environmental data of the airspace to obtain a three-dimensional airspace basic model.

[0064] Among them, the three-dimensional airspace basic model refers to the three-dimensional airspace model formed by integrating digital map data, regional deployment architecture and environmental data, which intuitively presents the overall airspace situation. Different types of data can be integrated using weighted fusion algorithms, such as digital map data, deployment architecture and environmental data with weights of 40%, 30% and 30% respectively.

[0065] Step a4: Based on the environmental data, use a preset empirical model to calculate the performance attenuation coefficient of different detection methods under the current environment. On the basis of the preset fixed airspace boundary in the three-dimensional airspace basic model, perform dynamic offset correction according to real-time wind speed and direction to obtain the corrected airspace boundary.

[0066] Among them, the performance degradation coefficient refers to the quantitative value of the impact of the current environment on the performance of the detection method. The more severe the environment, the larger the coefficient. Moreover, this coefficient can be obtained by fitting historical environment and detection performance data. The training data of the empirical model is the actual performance degradation record of the detection equipment under different environments. The dynamic offset correction is to determine the boundary offset direction based on the real-time wind direction, and calculate and correct the offset of the fixed airspace boundary by combining the wind speed and airspace protection experience.

[0067] Step a5: Combine the temporary whitelist rules dynamically generated based on environmental data, the performance degradation coefficient, and the corrected airspace boundary into an environmental situation awareness baseline.

[0068] The temporary whitelist rule is used to distinguish between legitimate and illegitimate targets. This application embodiment does not limit the content of the rule and can be set according to the actual situation.

[0069] Step 1023: Determine the airspace boundary and whitelist of the protected area based on the environmental situation awareness baseline, remove image data that exceeds the airspace boundary to obtain the first optimized AI image, remove data of existing targets in the preset whitelist from the first optimized AI image to obtain the second optimized AI image, and determine the situation energy marker based on the second optimized AI image and the environmental situation awareness baseline.

[0070] The preset whitelist is a list containing information about legitimate targets, used to remove targets that do not need to be blocked.

[0071] Step 1024: Perform structured processing on the data in the AI ​​image after secondary optimization, and combine the structured data and the tagging information to obtain the target data set. The tagging information includes: situational energy tag, ground detection equipment identifier, interception equipment status and unknown target temporary identifier.

[0072] Among them, situational awareness marking is a feature identifier that provides core reference for subsequent target threat level determination and interception strategy formulation; structured processing refers to the process of converting image data into a standardized format to facilitate subsequent data retrieval.

[0073] In practical applications, in low-altitude airspace safety protection scenarios, for AI images containing unknown targets acquired by the photoelectric turntable in S101, firstly, the brightness sensor on the turntable detects that the current airspace brightness is 18 lux, while the preset brightness threshold is 50 lux, indicating insufficient illumination. Based on this, nighttime brightness enhancement technology is activated, and the gamma correction algorithm is used to adjust the image gamma value to 0.8. Then, through an adaptive resolution adjustment algorithm, because the outline of the unknown target in the image is clear but the details are few, the resolution of the digital image is optimized to 1280×720 to reduce redundant data, while the resolution of the infrared radiation image remains unchanged at 640×512, resulting in the preprocessed AI image.

[0074] Secondly, airspace digital map data was retrieved from the geographic information system, including terrain elevation up to 150m, and the locations and heights of three fixed buildings, corresponding to coordinates (8000m, 12000m) at 120m height, (15000m, 9000m) at 80m height, and (22000m, 16000m) at 150m height. The regional deployment architecture was imported, including the coordinates of three detection device launch points (5000m, 8000m), (25000m, 15000m), and (15000m, 5000m), each covering a range of 10km, with orientation angles of 120°, 60°, and 90° respectively. Environmental data was obtained through meteorological sensors and a meteorological data interface: visibility V=4.5km, brightness 18lux, wind speed W=2.8m / s, wind direction northeast, cloud base height 800m, and precipitation 0.

[0075] The three types of data are then used to generate a 3D spatial domain basic model through a weighted fusion algorithm, with weights allocated as follows: digital map 40%, deployment architecture 30%, and environmental data 30%, ensuring the model takes into account geographical, device, and environmental characteristics. Subsequently, an empirical model is used to calculate the performance degradation coefficient based on the environmental data. The formula for this coefficient is as follows: Where K is the performance degradation coefficient, V is the visibility, and W is the wind speed; for example, when V=4.5 and W=2.8, .

[0076] The preset fixed airspace boundaries are x∈[10000m, 30000m] and y∈[5000m, 20000m]. Based on the northeast wind direction and wind speed of 2.8m / s, and combined with airspace protection experience, the southwest offset of the boundary is calculated to be 100m, resulting in the corrected airspace boundaries x∈[10100m, 30000m] and y∈[5100m, 20000m]. Combined with temporary whitelist rules, which allow civil aviation passenger aircraft and registered drones to pass, an environmental situational awareness baseline is formed.

[0077] Next, based on the corrected airspace boundary in the baseline, ground scene data at x=9800m that exceeds the range x∈[10100m, 30000m] are removed from the image to obtain the initial optimized AI image. Then, the initial optimized image is compared with the whitelist using a feature matching algorithm to remove the successfully matched UAV data, whose corresponding coordinates (18000m, 11000m, 550m), resulting in a secondary optimized AI image containing only unknown targets. Combining the environmental data with a performance degradation coefficient K=1.19 and a visibility of 4.5km, the target's flight speed is stable and there are no abnormal fluctuations in infrared radiation intensity, indicating that it is less affected by the environment, and a state energy label of "medium state energy - strong environmental adaptability" is generated.

[0078] Finally, structured target data was extracted using an image structuring algorithm, including three-dimensional coordinates (20000m, 12000m, 600m), infrared radiation intensity of 1250W / m², and external contour parameters (length 3m, width 1.5m). This data was then combined with labeling information, including the status energy label "medium status energy - strong environmental adaptability," the detection equipment identifier "electro-optical turntable," the interception equipment status "3 interceptors on standby," and the temporary identifier for unknown targets "T001," forming a complete target data set.

[0079] The S102 scheme described above improves image quality through image preprocessing and adjusts the resolution to balance clarity and efficiency; the established environmental situation awareness baseline integrates multi-dimensional data to dynamically adapt to the airspace situation; based on the baseline, boundary and whitelist filtering is performed to effectively eliminate interference and focus on unknown targets, and the situation can be marked to provide environmental basis for threat determination; the generated structured target data set integrates core information and attributes to facilitate efficient subsequent processing.

[0080] S103. Based on the target data set, the similarity between each pixel in the AI ​​image and the target feature template is calculated using the cosine similarity formula, and a heat map is generated based on the similarity.

[0081] In one specific implementation, step S103 includes the following steps:

[0082] Step 1031: Generate the target feature vector of the target feature template based on the target data set.

[0083] Specifically, step 1031 can be achieved through the following process:

[0084] Step b1: Identify basic feature parameters from the target dataset.

[0085] The basic characteristic parameters include: the spatial position, flight status parameters, morphological parameters, and infrared radiation parameters of the unknown target.

[0086] For example, basic feature parameters are extracted from the dataset: three-dimensional spatial coordinates (20000m, 12000m, 600m), flight status parameters: speed 320km / h, heading 31°, morphological parameters: length 3m, width 1.5m, and fuselage outline coordinate sequence [(19998m, 12000m, 600m), (20002m, 12000m, 600m), (20000m, 12001.5m, 600m), (20000m, 11998.5m, 600m)], and infrared radiation parameters: mid-wave infrared radiation intensity 1300W / m², long-wave infrared radiation intensity 1083W / m².

[0087] Step b2: Convert the basic feature parameters into a standardized input format, input the standardized basic feature parameters into the situation-enhanced feature analysis model, and extract local texture detail features through the convolutional neural network in the situation-enhanced feature analysis model.

[0088] Among them, the situation-enhancing feature analysis model refers to a model that integrates convolutional neural networks and attention mechanisms, which is used to extract and enhance deep features from basic features. This embodiment does not specifically limit the specific structure of the model.

[0089] For example, after standardizing these parameters according to a preset format, they are input into a situational enhancement feature analysis model. The 3×3 convolution kernel in the model slides through the target image with a stride of 1 to extract local texture details such as fuselage surface texture and tail edge.

[0090] Step b3: Strengthen the correlation between local texture detail features through the attention mechanism in the situational enhancement feature analysis model, and generate depth features based on the strengthened correlation between local texture detail features.

[0091] Among them, the multiple deep features refer to high-order abstract features extracted from the basic features of the target, which are used to accurately describe the essential attributes of the target;

[0092] Furthermore, these multiple depth features can include morphological correlation features and infrared radiation correlation features; wherein, the morphological correlation features reflect the target's external shape and structure, and these features can include the average curvature value of the contour of the local region where the pixel is located, as well as the edge gradient direction angle calculated by the Sobel operator and quantized into a specified principal direction encoding; the infrared radiation correlation features can reflect the target's thermal radiation characteristics, and these features can include the absolute radiation intensity value of the pixel and the ratio of the radiation intensity of the pixel in the mid-wave infrared and long-wave infrared characteristic bands.

[0093] In step b3, the structure of the situation-enhanced feature analysis model is based on the core architecture of "input layer - convolutional feature extraction layer - attention enhancement layer - feature association generation layer". The input layer receives the standardized basic feature parameters of the target. The convolutional feature extraction layer is composed of a convolutional neural network with multiple layers of convolutional kernels of different sizes. It captures the local texture details of the target image through sliding calculation. The attention enhancement layer is located after the convolutional layer. It filters the local texture details through a dynamic weight allocation algorithm, assigning higher weights to features with high correlation to target recognition, thereby strengthening the correlation between these features. The feature association generation layer, based on the enhanced feature correlation, combines algorithms such as curve fitting and band separation to integrate the local features into deep features that include morphological correlation features and infrared radiation correlation features. Finally, the deep features are output for subsequent similarity calculation.

[0094] For example, the model's attention mechanism assigns a weight of 0.8 to the tail fin edge texture features to strengthen the association, and calculates the average curvature value of the contour using a curve fitting algorithm, as shown in the formula. Where k is the average curvature value of the contour, y' is the first derivative of the contour curve, and y'' is the second derivative. Substituting these values ​​into the fuselage contour coordinate sequence, we obtain the mean value of y' is 0.1 and the mean value of y'' is 0.0021. Substituting these values ​​into the formula, we get... The value is 0.002, rounded to two decimal places. The edge gradient direction angle is calculated using the Sobel operator and quantized to obtain a 3-bit binary code 101, corresponding to the value 5. In the infrared radiation correlation feature, the absolute radiation intensity value can be the average of medium wave and long wave, such as (1300+1083) / 2=1191.5W / m², and the radiation intensity ratio is 1300 / 1083≈1.2, which finally generates the depth feature.

[0095] Step b4: After concatenating the depth features into a four-dimensional comprehensive feature vector, perform Z-score normalization to obtain the normalized feature vector of the pixel.

[0096] Z-score standardization refers to a method of converting data into a standard distribution with a mean of 0 and a standard deviation of 1, which is used to eliminate the influence of feature units.

[0097] Step b3 can combine the average curvature value of the contour, the edge gradient direction angle, and the ratio of the radiation intensity of the pixel in the mid-wave infrared and long-wave infrared characteristic bands into a four-dimensional comprehensive feature vector.

[0098] Step b5: Based on the current airspace environment data and early warning information, dynamically select the target feature template with the highest relevance value from the preset feature library, call the reference feature vector of the target feature template, and obtain a target feature vector by normalizing the feature statistical distribution of the reference feature vector in the preset feature library through Mahalanobis distance.

[0099] Among them, the preset feature library refers to the database that stores various target feature templates. The target feature templates include the baseline feature vector and feature statistical distribution information. Mahalanobis distance normalization refers to the method of calculating and standardizing the distance by combining the data covariance matrix, which is used to optimize the adaptability of the baseline feature vector.

[0100] Step 1032: Compare the standardized feature vector of each pixel with the target feature vector. In the process of calculating the similarity value using the cosine similarity formula, a diagonal weighted matrix is ​​introduced. The weight allocation strategy of the diagonal weighted matrix is ​​used to adapt to environmental changes and obtain the similarity.

[0101] The cosine similarity formula can be found in relevant technologies and will not be elaborated here. The diagonal weighted matrix is ​​a matrix whose diagonal elements are weight values. The construction of the diagonal weighted matrix is ​​based on the dimension of the standardized feature vector. The construction method needs to dynamically determine the weight of the diagonal elements in combination with the current spatial environment data.

[0102] Its weight allocation strategy can be dynamically adjusted according to the airspace environment to highlight the weight of features with strong environmental adaptability. For example, if the visibility is low and the lighting is insufficient, the weight of infrared radiation correlation features corresponding to diagonal elements is increased, while the weight of morphological correlation features is decreased. If the wind speed is high and the target attitude is unstable, the weights are adjusted in the opposite direction, while ensuring that the sum of the weights of all diagonal elements is 1. This strategy makes the cosine similarity calculation focus more on features with strong environmental adaptability, thereby improving the accuracy of the similarity results.

[0103] Step 1033: Linearly map the similarity to a preset grayscale range to obtain the corresponding grayscale value. Combine the grayscale values ​​of all pixels into an initial grayscale image. Perform convolution filtering on the initial grayscale image using a Gaussian kernel with a specified standard deviation to obtain a thermal image.

[0104] Among them, thermal image refers to an image that maps similarity to grayscale values ​​and is filtered. This image can intuitively present the degree of matching between the target and the preset features through the grayscale depth. Gaussian kernel is a smoothing kernel used for convolution filtering. The filtering intensity is controlled by specifying the standard deviation and is used to eliminate noise in the initial grayscale image.

[0105] For example, a similarity of 1 is mapped to a grayscale value of 255, and other pixels are mapped according to their corresponding similarities to form an initial grayscale image. A 5×5 Gaussian kernel with a standard deviation of 1.5 is used to convolve and filter the initial grayscale image to obtain a thermal image, in which the T001 target area is a bright white at level 255, the grayscale value within a 10m radius is 180-220, and the grayscale value of the background area is below 50.

[0106] The S103 scheme described above provides comprehensive data for feature processing through basic feature parameters; it captures local details through standardization and convolutional networks, and generates deep features that reflect morphology and infrared core attributes based on an attention mechanism; Z-score standardization ensures vector comparability; combining environmental filtering and normalization of target templates enhances the targeting of the comparison; and the introduction of a dynamic diagonal weighted matrix to calculate cosine similarity makes the results adaptable to the environment. Therefore, the final generated thermal image can intuitively present the feature matching region.

[0107] S104. Based on the spatial location and morphological parameters of the unknown target, locate the region to be analyzed containing the unknown target in the thermal image, introduce a multi-scale feature generation algorithm, generate multi-scale feature vectors by dividing the depth features of the region to be analyzed into different proportions, and perform hierarchical comparison between the multi-scale feature vectors and existing vector templates in a preset multi-type target template library to obtain the feature matching degree.

[0108] Among them, hierarchical comparison refers to the process of calculating the similarity between vectors of different proportions in multi-scale feature vectors and template vectors of corresponding types and scales in the template library.

[0109] Multi-scale feature generation algorithms refer to algorithms that convert the depth features of the region to be analyzed into feature vectors according to different proportions in a multi-scale scaling system through interpolation, scaling and other techniques. The core purpose is to improve the comprehensiveness and accuracy of subsequent feature matching through multi-dimensional feature coverage.

[0110] As a specific implementation method, such as Figure 2 As shown, step S104 includes the following steps:

[0111] Step 1041: Determine the multi-scale scaling system, which includes: compressed scale, original scale, and expanded scale.

[0112] The compression scale ratio is used to enhance the comparison of the overall contour features of the unknown target, the original scale ratio is used to preserve the original details of the depth features of the region to be analyzed, and the expansion scale ratio is used to expand the feature dimensions to enhance the blurred edge regions of the unknown target. For example, the multi-scale ratio system is a compression scale ratio of 0.8, an original scale ratio of 1.0, and an expansion scale ratio of 1.2.

[0113] Step 1042: Based on the multi-scale feature generation algorithm and the multi-scale ratio system, generate multi-scale feature vectors from the depth features of the region to be analyzed according to different ratios.

[0114] Step 1043: Call the preset multi-type target template library. The preset multi-type target template library stores feature vector templates according to target type and motion state, and establishes a template index table.

[0115] Among them, the preset multi-type target template library is a database that stores feature vector templates according to target type and motion state, and at the same time, a template index table is established to quickly locate the corresponding template.

[0116] Step 1044: Compare the feature vectors corresponding to the original scale ratio in the multi-scale feature vectors with the first matching degree of the reference template in the template index table, the feature vectors corresponding to the compressed scale ratio with the second matching degree of the compressed template in the template index table, and the feature vectors corresponding to the expanded scale ratio with the third matching degree of the expanded template in the template index table.

[0117] It should be understood that the above three matching degrees can all be calculated using the cosine similarity formula, and this embodiment does not limit this.

[0118] Step 1045: Merge the first matching degree, the second matching degree, and the third matching degree to obtain the feature matching degree.

[0119] It should be noted that this embodiment does not specifically limit the method of fusion.

[0120] Steps 1041 to 1045 above construct a multi-scale proportional system to cover the feature requirements of the overall outline, details, and edge regions of the target, avoiding the omission of key features at a single scale; use interpolation algorithms to generate multi-scale feature vectors to ensure the completeness and accuracy of data at each scale; call the classification template library and use indexes to quickly locate and improve comparison efficiency; hierarchical comparison enables precise matching of features at each scale with corresponding templates, and the weights are dynamically adjusted in combination with the environment to obtain a comprehensive feature matching degree, avoiding the one-sidedness of single-scale results. Therefore, this process improves the comprehensiveness and reliability of matching, provides a scientific basis for subsequent threat determination, and enhances the system's adaptability to diverse targets.

[0121] S105. Based on feature matching degree, determine whether the unknown target is a target to be intercepted. If the unknown target is a target to be intercepted, predict the predicted flight path of the target to be intercepted within a preset time period in the future.

[0122] Specifically, S105 may include the following processes:

[0123] Step 1051: Combine the feature matching degree, the scale ratios in the multi-scale ratio system, and the flight state-related features in the depth features to form the state space of the reinforcement learning algorithm, and define the action space.

[0124] In this context, the state space of a reinforcement learning algorithm refers to the set of environmental information perceived and used by the agent for decision-making; flight state-related features refer to the core parameters describing the motion characteristics of an unknown target, such as the real-time velocity, heading angle, and acceleration of the unknown target; and the action space refers to the set of adjustment actions that the agent can execute in reinforcement learning, used to provide the agent with operational directions for optimizing target matching effects. This space includes the following actions:

[0125] Numerical adjustments to the scale ratios in a multi-scale scaling system, such as adjusting the scale ratio within a range of 0.7-0.9 with an adjustment step of ±0.05, adjusting the scale ratio within a range of 0.9-1.1 with an adjustment step of ±0.05, and adjusting the scale ratio within a range of 1.1-1.3 with an adjustment step of ±0.05.

[0126] Actions for adding or deleting feature dimensions used for feature comparison, such as adding two high-contribution dimensions, "tail fin infrared radiation intensity" and "fuselage contour smoothness", and deleting one redundant dimension, "background infrared noise".

[0127] Step 1052: Use the difference between the new feature matching degree obtained after dynamic adjustment and the feature matching degree before adjustment as the core reward item, determine the additional reward item based on the flight state associated features, and use the weighted sum of the core reward item and the additional reward item as the reward.

[0128] The additional reward item refers to the auxiliary reward index determined based on the correlation characteristics of the flight state of the unknown target. It is used to consider the impact of the action on the adaptability of the target's motion state. The more stable the target's motion and the more it fits the requirements of the protected airspace control, the higher the value of the additional reward item. This embodiment does not make specific limitations on the calculation method of the reward item, and can be set accordingly according to the actual situation.

[0129] Step 1053: The agent selects and executes an action based on the current state. The environment provides feedback on the new feature matching degree and corresponding reward obtained after the action is executed. The agent learns the optimal policy immediate reward by maximizing the accumulated reward, so as to dynamically adjust the feature matching degree.

[0130] Among them, environmental feedback refers to the result of the new feature matching degree and corresponding reward returned by the system after the agent performs an action; cumulative reward refers to the total reward obtained by the agent in multiple actions, which is used to guide the agent to learn the long-term optimal strategy rather than short-term gains; optimal strategy refers to the action selection rule learned by the agent by maximizing the cumulative reward, which can guide the system to dynamically adjust the feature matching parameters to achieve the optimal matching effect.

[0131] Step 1054: Based on the new feature matching degree obtained after dynamic adjustment, determine whether the unknown target is a target to be intercepted.

[0132] In practical applications, in low-altitude airspace safety protection scenarios, after determining the state space and action space of the unknown target T001, the reward item is first determined. For example, before adjustment, the feature matching degree of T001 is 0.993. The agent initially selects the combination of "expanding the scale ratio +0.05" and "adding the tail fin infrared radiation intensity dimension". After adjustment, the new feature matching degree is obtained by recalculating through the hierarchical comparison process. The core reward item is 0.998-0.993=0.005.

[0133] Analyzing the correlation characteristics of flight status, the reasonable speed range for high-speed small UAVs is preset to 280-350 km / h, the heading angle fluctuation threshold to ±2°, and the reasonable acceleration range to 1000-1800 km / h². Weights of 0.4, 0.3, and 0.3 are assigned to these three parameters, respectively. A percentage-based scoring system is used: 90 points for a speed of 320 km / h, 95 points for a heading angle fluctuation of ±1° within 5 seconds, and 92 points for acceleration of 1440 km / h². The additional bonus is calculated as 90×0.4 + 95×0.3 + 92×0.3 = 92.1, which is then normalized to the 0-0.01 range using the following normalization formula: ,in, This is an additional reward item after normalization. The original score is 0.00921.

[0134] Next, the overall reward is calculated: the current airspace visibility is 4.5km, which is of medium complexity. The core reward item is assigned a weight of 0.6, and the additional reward item a weight of 0.4. The overall reward formula is then used. Calculate, where R is the total reward. The weight of the core reward items, As the core reward item, As for the weight of the additional reward items, As the additional reward term after normalization, substituting it into the equation, we get R = 0.6 × 0.005 + 0.4 × 0.00921 = 0.006684. The agent uses the Q-learning algorithm to select the action combination with the highest Q value, i.e., the initial Q value is 0.006. After executing the action, the scale ratio is adjusted to 1.25, and a new dimension of tail fin infrared radiation intensity is added, i.e., the value is 1320 W / m².

[0135] Finally, environmental feedback and policy learning were performed. The system provided a new feature matching degree of 0.998 and a comprehensive reward of 0.006684. The agent updated the Q value using a temporal difference formula. After 10 iterations, the agent learned the optimal policy, and the feature matching degree of T001 stabilized above 0.997.

[0136] The embodiments of this application can comprehensively evaluate the value of actions through a comprehensive reward mechanism, avoiding decision-making bias caused by a single indicator; the agent selects actions based on the Q-learning algorithm and updates the strategy through temporal difference, ensuring that the learning process is efficient and in line with the needs of actual scenarios; the optimal strategy formed after multiple iterations can dynamically adjust the feature matching parameters to adapt to changes in the target motion state and fluctuations in the spatial environment, so that the feature matching degree remains at a stable and efficient level.

[0137] In one specific implementation, step S105, predicting the predicted flight path of the target to be intercepted within a preset future time period, includes:

[0138] Step 1055: Call the historical flight track data and corresponding flight status parameters of the target to be intercepted from the target data set to generate a historical flight track image.

[0139] Among them, historical trajectory data refers to the continuous spatial coordinate sequence of the target to be intercepted over a period of time stored in the target data set, which is used to reflect the target's past flight path; flight status parameters refer to the core indicators describing the target's motion characteristics, including the target's real-time speed, heading angle, acceleration and other data.

[0140] Step 1055: Use an image feature fitting algorithm to perform polynomial fitting on the trajectory curve in the historical flight trajectory image to extract the direction change feature points and acceleration-related pixel change patterns in the trajectory curve, and obtain the fitting result.

[0141] Among them, the image feature fitting algorithm refers to the algorithm that approximates the curves in the track image through mathematical modeling, and is used to extract the key features of the track;

[0142] The specific type of image feature fitting algorithm can be the least squares polynomial fitting algorithm. This algorithm constructs a polynomial function model, uses the pixel coordinates (x, y) of the original trajectory curve as sample data, minimizes the sum of squared residuals from the sample points to the fitted curve, and solves for the optimal coefficients of the polynomial, so that the fitted curve fits the original trajectory to the greatest extent. At the same time, based on the fitted polynomial, the first derivative of the curve can be calculated by numerical differentiation algorithms, such as forward difference and central difference, to locate the feature points of directional change. Combined with the correspondence between the trajectory pixel interval and the acceleration parameter, the law of acceleration-related pixel change is summarized, and finally the extraction and fitting of trajectory features are completed.

[0143] Among them, the direction change feature point refers to the coordinate point in the track curve where the heading changes significantly, marking the turning point of the target's flight direction; the acceleration-related pixel change law refers to the correlation pattern between the pixel distribution in the track image and the target's acceleration change, and the acceleration change will be reflected by the density and curvature of the track curve.

[0144] Step 1056: Combining the motion pattern stability in the depth features and the current flight state parameters, predict the predicted flight trajectory within a future preset time period based on the fitting results.

[0145] In the above scheme, dynamic mode stability refers to the index in depth characteristics that reflects whether the target's flight state is regular, which is quantified by acceleration fluctuation amplitude and heading angle change frequency; predicted flight trajectory refers to the spatial coordinate sequence that the target may pass through in a preset time period of seconds in the future, based on historical trajectory and current state, providing a basis for the scheduling of interception equipment.

[0146] In practical applications, in low-altitude airspace security protection scenarios, the following steps are performed on T001, which is identified as a high-threat target to be intercepted: First, retrieve the three-dimensional coordinate data and corresponding flight status parameters for each second within the past 60 seconds from the target data set of T001. For example, at the 50th second (20000m, 12000m, 600m), ..., at the 60th second (20762m, 12413m, 600m), the speed is 320km / h, the heading angle is 31°, and the acceleration is 1440km / h². Project the three-dimensional coordinates onto the xy plane, with the z-axis fixed at 600m. Connect the coordinate points with a red solid line, mark the speed value of each point, and generate a historical flight track image.

[0147] Next, a third-order polynomial fitting algorithm is used to fit the trajectory curve, and the fitting formula is given by... Where y is the longitudinal coordinate of the trajectory in meters, x is the abscissa of the trajectory in meters, and a, b, c, and d are fitting coefficients. Five typical coordinate points were selected: (20000, 12000), (20228.6, 12124), (20457.2, 12248), (20685.8, 12372), and (20762, 12413).

[0148] The equations were solved using the least squares method to obtain a = b= Given c=0.47 and d=2600, the fitting formula is: Find the first derivative. We discovered abrupt changes in the derivative at x=20300m and x=20600m, which were identified as characteristic points of directional change. Analyzing the track pixels, we found that for every 228.6m increase in x, y increases by 41.3m, with a constant pixel interval. We concluded that the pixel interval remains unchanged when the acceleration is stable.

[0149] Finally, the stability index of the T001 motion mode was extracted, namely, an acceleration fluctuation amplitude of 2%, which was considered stable. The current speed of 320 km / h was converted to m / s, corresponding to 320×1000 / 3600≈88.89 m / s. The x-direction velocity component was 88.89×cos31°≈76.2 m / s, and the y-direction velocity component was 88.89×sin31°≈45.8 m / s. Pre-setting the next 10 seconds, at the 61st second, x=20762+76.2≈20838.2 m. Substituting these values ​​into the fitting formula yielded... Given z=600m, coordinates are 20838.2m, 12458.8m, 600m; and so on, the predicted coordinates for each second in the next 10 seconds are calculated to form a predicted flight path. This path is used to subsequently schedule three interceptor aircraft. The interceptor arrives at the predicted coordinates of T001 at 65 seconds: 20762+76.2×5≈21143m, 12413+45.8×5≈12642m, and interception is carried out at 600m. The temporary unknown target identifier T001 is used throughout the interception process to ensure accurate correlation.

[0150] This application's embodiments visualize historical flight track data and flight state parameters, generating historical flight track images that provide an intuitive basis for track analysis and avoid the abstractness of pure data. A multinomial fitting algorithm is used to extract the correlation between the direction change feature points and acceleration of the track, accurately capturing the core features of the target track and providing a reliable mathematical model for prediction. Combining motion mode stability with current state correction results ensures the accuracy of future second-level track predictions, especially achieving efficient extrapolation when the target motion is stable, and reducing errors through feature point correction when the motion is unstable.

[0151] S106. Based on the predicted flight path and the threat level of the target to be intercepted, allocate corresponding interception equipment and control the interception equipment to move and communicate in real time between the equipment to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

[0152] Among them, the threat level refers to the classification of the degree of danger of a target based on the target attributes in the deep features. It is usually divided into three levels: high, medium and low, which determines the scheduling priority of the interception equipment.

[0153] As a specific implementation, step S106, based on the predicted flight trajectory and the threat level of the target to be intercepted, allocates corresponding interception equipment and controls the interception equipment to move and communicate in real time between the equipment to form a 3D interception network, so that the target to be intercepted enters the interception range of the 3D interception network, including:

[0154] Step 1061: Determine the threat level of the target to be intercepted based on the target attributes in the deep features. According to the number of targets to be intercepted and their threat levels, schedule interception devices from the interception device library according to a preset multiple relationship. Among them, targets with high threat levels are given priority to be assigned interception devices with high-speed response capabilities.

[0155] Among them, the interception device library refers to a resource library that stores various interception devices, including attributes such as device model, response speed, and interception range;

[0156] The preset multiplier relationship is a dynamic numerical system based on the threat level hierarchy of the targets to be intercepted. The core principle is that "the higher the threat level, the greater the redundancy of the interception equipment". The base multiplier for a single high-threat target is 1:3, the base multiplier for a single medium-threat target is 1:2, and the base multiplier for a single low-threat target is 1:1. If there are multiple targets, the total scheduling quantity is determined according to "multiplier of a single target × number of targets". When high-threat and medium-low-threat targets coexist, the multiplier ratio of high-threat targets is given priority.

[0157] Considering the situation of a single high-threat target T001 in the airspace scenario, its threat level was determined to be high. Therefore, strictly following the preset ratio of 1:3, three interceptors with high-speed response capabilities were dispatched from the interceptor equipment pool. This not only met the multiple interception needs of the high-threat target, but also avoided excessive consumption of equipment resources.

[0158] For example, the attribute information of T001 is extracted from the depth features, including flight speed of 320km / h, infrared radiation intensity of 1250W / m², and fuselage length of 3m. Using a weighted scoring method, i.e., speed weight 0.4, infrared radiation intensity weight 0.3, and size weight 0.3, the score is calculated as 320 / 400×40+1250 / 1500×30+3 / 5×30=32+25+18=75 points. Since the high threat level threshold is 70 points, T001 is determined to be a high threat target. According to the preset ratio of 1:3, that is, 1 high threat target corresponds to 3 interception devices, 3 high-speed response interceptors are dispatched from the interception device library, with a response time ≤2 seconds and a maximum flight speed of 450km / h.

[0159] Step 1062: Divide the protected area into multiple sub-areas with fixed diameters. Each sub-area is assigned at least one interceptor device. Send a scheduling instruction to the interceptor device via frequency hopping encrypted communication technology, which includes the predicted flight path, sub-area boundary, temporary identifier of the target to be intercepted, and interception priority. This will enable the interceptor device to take off according to the scheduling instruction and dynamically adjust its own trajectory through real-time inter-device communication to form a 3D interception network.

[0160] Among them, frequency hopping encrypted communication technology refers to the technology of ensuring the security of dispatch command transmission by randomly switching communication frequencies and encrypting data; sub-region refers to a small area of ​​airspace that is divided into protection areas according to a fixed diameter, used to clarify the area of ​​responsibility of interception equipment; 3D interception network refers to the interception range network formed by multiple interception equipment in three-dimensional airspace by dynamically adjusting their trajectories, covering the predicted trajectory of the target.

[0161] For example, the airspace is divided into three sub-regions with a fixed diameter of 5km. The boundaries of the first sub-region are x∈[20500m, 21000m], y∈[12400m, 12900m], and z∈[500m, 700m]. The boundaries of the second sub-region are x∈[21000m, 21500m], y∈[12900m, 13400m], and z∈[500m, 700m]. The boundaries of the third sub-region are x∈[21500m, 22000m], y∈[13400m, 13900m], and z∈[500m, 700m]. Three interceptors, C1, C2, and C3, are assigned to each sub-region.

[0162] Frequency hopping encrypted communication technology is used to send scheduling instructions to three interceptor aircraft. The instructions include the predicted flight path of T001 for the next 10 seconds, the corresponding sub-area boundary, the temporary identifier T001, and the interception priority 1. After receiving the instructions, the three interceptor aircraft take off and share position information through real-time inter-aircraft communication. C1 adjusts to (20700m, 12400m, 600m), C2 to (21200m, 12900m, 600m), and C3 to (21700m, 13400m, 600m), forming a 3D interception network.

[0163] Step 1063: Bind the interception device and the target to be intercepted, and receive the flight data and target tracking image returned by the interception device to form a data association pair.

[0164] For example, C1 is bound to T001 to receive the flight data transmitted back by C1, namely (20700m, 12400m, 600m), flight speed 380km / h, heading angle 31°, and tracking image of T001; similarly, C2 and T001 are bound to each other to receive the corresponding flight data and tracking image, forming three sets of data association pairs: C1-T001, C2-T001, and C3-T001.

[0165] Step 1064: Based on the data association, dynamically adjust the movement parameters of the intercepting device according to the relative positional relationship between the intercepting device and the target to be intercepted.

[0166] Among them, the data association pair refers to the corresponding data set containing the device's flight data and the target's tracking image after the interception device is bound to the target; the movement parameters refer to the parameters that control the movement of the interception device, including flight speed, heading angle, and altitude adjustment rate, which are used to correct the device's trajectory to adapt to the target's movement.

[0167] For example, based on the data association pair C1-T001, the real-time relative position of C1 and T001 is calculated using a three-dimensional distance formula. Where d is the three-dimensional distance, x1, y1, z1 are the real-time coordinates of C1, such as 20700m, 12400m, 600m, and x2, y2, z2 are the real-time coordinates of T001, such as 20762m, 12413m, 600m. Substituting into the formula, we get d=63.4m. Since the reasonable distance range is 100m-150m, it is determined that the distance is too close, and the movement parameters of C1 are dynamically corrected. The flight speed is reduced from 380km / h to 350km / h, and the heading angle is adjusted from 31° to 34° to ensure that C1 and T001 maintain a reasonable relative distance.

[0168] Step 1065: When the target to be intercepted enters the interception range of the 3D interception network, a lock command is sent to the interception device to indicate that the tracking within the interception range has been completed.

[0169] For example, the actual flight position of T001 is monitored in real time. When T001 flies to 20900m, 12500m, and 600m, the position enters the interception range of C2. The interception radius of C2 is 2km. The sensors on C2 detect that the signal strength of T001 has reached the interception threshold. A lock command is sent to C2. The command includes the current coordinates of T001 (20900m, 12500m, 600m) and the lock priority (1). After receiving the command, C2 activates the target lock function and completes the tracking within the interception range. This lock state will be used for subsequent intercept weapon launch steps to ensure that T001 is continuously and accurately locked.

[0170] This application's embodiments achieve rational resource allocation and ensure interception capabilities by scheduling interception equipment that matches the target threat level. Subsequently, a three-dimensional interception network covering the predicted target trajectory is constructed through encrypted communication and inter-machine collaboration to improve the interception success rate. Based on this, a data association pair is established between the interception equipment and the target to ensure accurate tracking and avoid confusion during multi-target collaboration. Next, the system dynamically corrects the movement parameters of the interception equipment to continuously adapt to the target's real-time movement, maintaining the effective tracking posture of the interception network. Finally, when the target enters the interception range, the system locks on, providing a reliable prerequisite for subsequent interception.

[0171] Figure 3This is a schematic diagram illustrating a specific implementation of an AI image recognition and high-speed tracking collaborative system provided in this application. (Refer to...) Figure 3 The system may include:

[0172] The acquisition module 31 is used to guide the photoelectric turntable to take pictures after determining that the radar has detected an unknown target based on the radar echo image, so as to acquire an AI image containing at least one unknown target in the airspace.

[0173] The elimination module 32 is used to remove data of existing targets in the preset whitelist from the AI ​​image through the environmental situation awareness baseline corresponding to the airspace, so as to obtain a set of target data carrying situation energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located.

[0174] The calculation module 33 is used to calculate the similarity between each pixel in the AI ​​image and the target feature template using the cosine similarity formula based on the target data set, and to generate a heat map based on the similarity.

[0175] The positioning and comparison module 34 is used to locate the region to be analyzed containing the unknown target in the thermal image based on the spatial location and morphological parameters of the unknown target. A multi-scale feature generation algorithm is introduced to generate multi-scale feature vectors by taking the depth features of the region to be analyzed at different proportions. The multi-scale feature vectors are then compared hierarchically with existing vector templates in a preset multi-type target template library to obtain the feature matching degree.

[0176] The determination module 35 is used to determine whether the unknown target is a target to be intercepted based on the feature matching degree, and if the unknown target is a target to be intercepted, to predict the predicted flight path of the target to be intercepted within a preset time period in the future;

[0177] The allocation control module 36 is used to allocate corresponding interception equipment based on the predicted flight trajectory and the threat level of the target to be intercepted, and to control the interception equipment to move and communicate in real time between the equipment to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

[0178] The AI ​​image recognition and high-speed tracking collaborative system of this application is used to implement the aforementioned AI image recognition and high-speed tracking collaborative method. Therefore, the specific implementation of the AI ​​image recognition and high-speed tracking collaborative system can be found in the embodiment section of the AI ​​image recognition and high-speed tracking collaborative method above. The specific implementation can be referred to the description of the corresponding embodiment, and will not be repeated here.

[0179] like Figure 4As shown, this application also provides an electronic device, including: a memory 41 for storing a computer program; and a processor 42 for executing the computer program to implement the steps of any of the above-described AI image recognition and high-speed tracking collaborative methods.

[0180] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described AI image recognition and high-speed tracking collaborative methods.

[0181] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0182] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the AI ​​image recognition and high-speed tracking collaborative method.

[0183] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0184] The above provides a detailed description of the AI ​​image recognition and high-speed tracking collaborative method and system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A collaborative method for AI image recognition and high-speed tracking, characterized in that, include: After determining that the radar has detected an unknown target based on the radar echo image, the photoelectric turntable is guided to take pictures to obtain an AI image containing at least one unknown target in the airspace. By using the environmental situation awareness baseline corresponding to the airspace, data of existing targets in the preset whitelist are removed from the AI ​​image to obtain a target data set carrying situational energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located. This includes: acquiring digital map data of the protected area and importing it into the regional deployment architecture of the airspace; acquiring environmental data of the airspace in real time by accessing meteorological sensors or preset meteorological data interfaces; fusing the digital map data, the regional deployment architecture, and the environmental data of the airspace to obtain a three-dimensional airspace basic model; calculating the performance attenuation coefficient of different detection methods under the current environment using a preset empirical model based on the environmental data; dynamically offsetting and correcting the airspace boundary based on the preset fixed airspace boundary in the three-dimensional airspace basic model according to real-time wind speed and direction; and combining the temporary whitelist rules dynamically generated based on the environmental data, the performance attenuation coefficient, and the corrected airspace boundary to form the environmental situation awareness baseline. Based on the target dataset, the similarity between each pixel in the AI ​​image and the target feature template is calculated using the cosine similarity formula. Based on the similarity, a heat map is generated. Based on the spatial location and morphological parameters of the unknown target, the region to be analyzed containing the unknown target is located in the thermal image. A multi-scale feature generation algorithm is introduced to generate multi-scale feature vectors by taking the depth features of the region to be analyzed at different proportions. The multi-scale feature vectors are then compared hierarchically with existing vector templates in a preset multi-type target template library to obtain the feature matching degree. Based on the feature matching degree, it is determined whether the unknown target is a target to be intercepted. If the unknown target is a target to be intercepted, the predicted flight path of the target to be intercepted is predicted within a future preset time period. Based on the predicted flight path and the threat level of the target to be intercepted, corresponding interception equipment is allocated, and the interception equipment is controlled to move and communicate with each other in real time to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

2. The method according to claim 1, characterized in that, The step involves using the environmental situational awareness baseline corresponding to the airspace to remove data of existing targets from the AI ​​image that are already on a preset whitelist, thereby obtaining a target data set carrying situational energy markers, including: The brightness data of the airspace environment is obtained. If the brightness data is lower than the preset brightness threshold, night brightness enhancement technology is introduced to enhance the brightness of the AI ​​image. The AI ​​image with enhanced brightness is then adaptively adjusted in resolution to obtain the AI ​​image with adjusted resolution. Based on the regional deployment architecture of the airspace and the environment in which the airspace is located, an environmental situation awareness baseline corresponding to the airspace is adaptively established. Based on the environmental situation awareness baseline, the airspace boundary and whitelist of the protected area are determined. Image data exceeding the airspace boundary are removed to obtain the first optimized AI image. Data of existing targets in the preset whitelist are removed from the first optimized AI image to obtain the second optimized AI image. Situation energy markers are determined based on the second optimized AI image and the environmental situation awareness baseline. The data in the AI ​​image after secondary optimization is structured, and the structured data and the tagging information are combined to obtain the target data set. The tagging information includes: situational energy tag, ground detection equipment identifier, interception equipment status and unknown target temporary identifier.

3. The method according to claim 1, characterized in that, The introduced multi-scale feature generation algorithm generates multi-scale feature vectors from the depth features of the region to be analyzed according to different proportions. These multi-scale feature vectors are then compared hierarchically with existing vector templates in a pre-set multi-type target template library. The feature matching degree is obtained, including: A multi-scale scaling system is determined, which includes: a compressed scale ratio, an original scale ratio, and an expanded scale ratio; Based on the multi-scale feature generation algorithm and the multi-scale ratio system, the depth features of the region to be analyzed are generated into multi-scale feature vectors according to different ratios. A preset multi-type target template library is invoked. The preset multi-type target template library stores feature vector templates according to target type and motion state, and establishes a template index table. The feature vectors corresponding to the original scale ratio in the multi-scale feature vectors are compared with the first matching degree of the reference template in the template index table, the feature vectors corresponding to the compressed scale ratio are compared with the second matching degree of the compressed template in the template index table, and the feature vectors corresponding to the expanded scale ratio are compared with the third matching degree of the expanded template in the template index table. The first matching degree, the second matching degree, and the third matching degree are combined to obtain the feature matching degree.

4. The method according to claim 1, characterized in that, The process of calculating the similarity between each pixel in the AI ​​image and the target feature template using the cosine similarity formula based on the target dataset, and generating a heat map based on the similarity, includes: Generate a target feature vector based on the target data set; The standardized feature vector of each pixel is compared with the target feature vector. In the process of calculating the similarity value using the cosine similarity formula, a diagonal weighted matrix is ​​introduced. The weight allocation strategy of the diagonal weighted matrix is ​​used to adapt to environmental changes and obtain the similarity. The similarity is linearly mapped to a preset grayscale range to obtain the corresponding grayscale value. The grayscale values ​​of all pixels are combined to form an initial grayscale image. The initial grayscale image is then convolved and filtered using a Gaussian kernel with a specified standard deviation to obtain a thermal image.

5. The method according to claim 1, characterized in that, The prediction of the target's flight path within a preset future time period includes: The historical flight track data and corresponding flight status parameters of the target to be intercepted are retrieved from the target data set to generate a historical flight track image; An image feature fitting algorithm is used to perform polynomial fitting on the trajectory curve in the historical flight trajectory image to extract the direction change feature points and acceleration-related pixel change patterns in the trajectory curve, and the fitting result is obtained. By combining the motion pattern stability in the depth features and the current flight status parameters, the predicted flight trajectory within a preset time period is predicted based on the fitting results.

6. The method according to claim 1, characterized in that, Based on the predicted flight path and the threat level of the target to be intercepted, corresponding interception equipment is allocated, and the interception equipment is controlled to move and communicate in real time between the equipment to form a 3D interception network, so that the target to be intercepted enters the interception range of the 3D interception network, including: The threat level of the target to be intercepted is determined based on the target attributes in the deep features. Based on the number of targets to be intercepted and their threat levels, interception devices are scheduled from the interception device library according to a preset multiple relationship. The protected area is divided into multiple sub-areas with fixed diameters. Each sub-area is assigned at least one interceptor device. A scheduling instruction containing the predicted flight path, sub-area boundary, temporary identifier of the target to be intercepted, and interception priority is sent to the interceptor device through frequency hopping encrypted communication technology. This enables the interceptor device to take off according to the scheduling instruction and dynamically adjust its own trajectory through real-time inter-aircraft communication to form a 3D interception network. The interception device and the target to be intercepted are bound together, and the interception device sends back its own flight data and target tracking images to form a data association pair. Based on data correlation, the relative positional relationship between the interception device and the target to be intercepted is determined, and the movement parameters of the interception device are dynamically corrected. When the target to be intercepted enters the interception range of the 3D interception network, a lock command is sent to the interception device, indicating that the tracking within the interception range has been completed.

7. An AI image recognition and high-speed tracking collaborative system, characterized in that, For performing the AI ​​image recognition and high-speed tracking collaborative method as described in any one of claims 1 to 6, comprising: The acquisition module is used to guide the photoelectric turntable to take pictures after determining that the radar has detected an unknown target based on the radar echo image, so as to acquire an AI image containing at least one unknown target in the airspace. The elimination module is used to remove data of existing targets in the preset whitelist from the AI ​​image through the environmental situation awareness baseline corresponding to the airspace, so as to obtain a set of target data carrying situational energy markers. The environmental situation awareness baseline is adaptively established based on the regional deployment architecture of the airspace and the environment in which the airspace is located. The calculation module is used to calculate the similarity between each pixel in the AI ​​image and the target feature template using the cosine similarity formula based on the target dataset, and to generate a heat map based on the similarity. The positioning and comparison module is used to locate the region to be analyzed containing the unknown target in the thermal image based on the spatial location and morphological parameters of the unknown target. A multi-scale feature generation algorithm is introduced to generate multi-scale feature vectors by taking the depth features of the region to be analyzed at different proportions. The multi-scale feature vectors are then compared hierarchically with existing vector templates in a preset multi-type target template library to obtain the feature matching degree. The determination module is used to determine whether the unknown target is a target to be intercepted based on the feature matching degree, and if the unknown target is a target to be intercepted, to predict the predicted flight path of the target to be intercepted within a future preset time period; The allocation control module is used to allocate corresponding interception equipment based on the predicted flight trajectory and the threat level of the target to be intercepted, and to control the interception equipment to move and communicate with each other in real time to form a 3D interception network so that the target to be intercepted enters the interception range of the 3D interception network.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the AI ​​image recognition and high-speed tracking collaborative method as described in any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the AI ​​image recognition and high-speed tracking collaborative method as described in any one of claims 1 to 6.

Images