Ship target identification system and method based on UEAI behavior tree

Abstract

The invention discloses a ship target identification method and system based on a UEAI behavior tree. The method comprises the steps that a simulation system scene is built, scene resources are prepared, and a scene modeling tool of an unreal engine is used for building the scene resources into an infrared scene according to requirements; presetting a criterion, and constructing a target feature database; running the simulation system scene, and capturing a picture under a preset view angle; processing various targets in the picture in real time by using a UEAI behavior tree, and carrying out target feature calculation and logic judgment; and executing the criterion, extracting the shape feature and the radiance feature of the target, comparing the shape feature and the radiance feature with the target feature database to obtain a target feature result, and outputting a target recognition result of the current frame. The system comprises a target generation module, a background generation module and a target identification module. According to the method, external input information can be replaced, target identification is prevented from being influenced by severe weather or complex sea conditions, expansion and maintenance are facilitated, meanwhile, the real-time target detection efficiency is improved, and the rendering frame rate is increased.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar target recognition, and relates to a ship target recognition system and method based on a UEAI behavior tree. Background Art

[0002] The primary task of modern guidance systems is to accurately identify targets from a large number of decoys and complex environmental backgrounds during orbital flight. Therefore, target recognition is the core and also the difficulty of the guidance task. Since the cost of directly conducting physical experiments is too high, semi-physical experiments are carried out in advance, specific working conditions are simulated in a simulation scenario, and then the anti-missile function of the target or the target detection and recognition algorithm function of the detector is tested using the simulation scenario. This is also one of the common solutions currently.

[0003] In terms of guidance digital simulation, Luo Xishuang et al. simulated and solved the entire process of the guidance control system simulation platform and the missile attacking the target, obtained target information, missile information, and the perspective information of the missile tracking the target. Then, the simulation system performed infrared imaging on the target based on the information obtained above, and then used image processing algorithms to process the image to obtain target recognition information. Finally, the infrared simulation platform transmitted the target recognition information to the guidance control system simulation platform.

[0004] In terms of target detection and recognition, Shang Wenli et al. improved a new method for processing data training sets and modified some mechanisms in the YOLOv5s network to address the problems of low recognition accuracy and serious missed detections in the target recognition and detection method based on deep learning. They proposed a method for recognizing maritime ship targets based on SAR images and the YOLOv5s network. After dividing the ship dataset according to a ratio, they performed flipping operations on the training set images and labels, modified the backbone module in the YOLOv5s network, introduced the CBAM attention mechanism into it, and modified the target box regression loss function to EIOU, improving the recognition accuracy of the deep learning method and reducing the missed detection rate of the model.

[0005] During the three-dimensional scene simulation process, most of the current target detection and recognition algorithms are designed based on deep learning. They perform predictive analysis on the image sequence captured and transmitted back by the detector based on the trained model to achieve the function of detecting and recognizing the targets in the image. The entire target detection and recognition process is relatively long, and the requirements for hardware are high, resulting in poor real-time performance of the detector for target recognition and detection, and seriously affecting the rendering efficiency of the three-dimensional scene simulation program. Summary of the Invention

[0006] The present invention aims to solve the technical problems of poor real-time performance of target recognition and detection by detectors and low rendering efficiency in the prior art. The present invention provides a ship target recognition method and system based on a UEAI behavior tree, and the technical solution adopted is as follows:

[0007] A ship target recognition method based on a UEAI behavior tree, comprising the steps of:

[0008] S1. Build a simulation system scene;

[0009] S2. Preset criteria and construct a target feature database;

[0010] S3. Run the simulation system scene and capture the images from a preset perspective;

[0011] S4. Use the UEAI behavior tree to process various targets in the image in real time, and perform target feature calculation and logical judgment;

[0012] S5. Execute the criteria, compare with the target feature database to obtain a target feature result, and output the target recognition result of the current frame.

[0013] In an embodiment of the present invention, the step S1 includes:

[0014] S11. Prepare scene resources, where the scene resources include ship models, temperature field textures, atmospheric texture resources, and various interferences;

[0015] S12. Use the scene modeling tool of the Unreal Engine to build the scene resources into an infrared scene according to requirements.

[0016] In an embodiment of the present invention, the step S12 includes:

[0017] S121. Generate targets using the scene resources, where the targets include target ships, virtual seekers, and shipborne interferences, and the target ships, the virtual seekers, and the shipborne interferences move according to the running trajectories;

[0018] S122. Generate the sea surface background and the sky background in the scene using the scene resources, and simulate the transmission effect of infrared radiation in the atmosphere in the scene;

[0019] S123. The movement of each element in the scene is realized by simulating the running trajectories, and the movement of the target ship, the movement of the virtual seeker, and the function of the target ship releasing ship interferences when the virtual seeker approaches the target ship are realized by using blueprint functions.

[0020] In an embodiment of the present invention, the step S2 includes:

[0021] S21. Create a blackboard and an AI behavior tree in the Unreal Engine. Add relevant parameters to the blackboard to determine whether a target has been captured and store all the target tags currently captured. Build relevant branch logic in the AI behavior tree to perform ballistic correction after capturing the target or continue the scanning process after failing to capture the target. The criteria include the shape characteristics of the target and the radiant luminance characteristics of the target.

[0022] S22. Obtain images of common ship targets in the daytime environment through simulation. According to different shooting angles and shooting times, use the blueprint function constructed according to the calculation methods of each target characteristic in the criteria to calculate the shape characteristics and radiant luminance statistical characteristics of the target at different missile incoming angles. Store them according to the ship model, incoming angle, incoming time, corresponding shape characteristics, and radiant luminance characteristic structure, and build a target characteristic database.

[0023] In one embodiment of the present invention, the step S4 includes:

[0024] S41. The target ship and the interference need to be placed in the scene after creating blueprints.

[0025] S42. The virtual seeker is created using the Pawn type and placed in the scene.

[0026] S43. Create a blueprint for an AI controller under the Actor class, add an AI perception component, and configure the AI sense as AI vision perception.

[0027] S44. Determine whether a target has been captured and store all the target tags currently captured, and perform ballistic correction after capturing the target or continue the scanning process after failing to capture the target.

[0028] In one embodiment of the present invention, the step S5 includes:

[0029] S51. Extract the shape characteristics of the target. The shape characteristics include the aspect ratio, compactness, and circularity of the target, and compare them with the shape characteristic parameters in the target characteristic database to obtain the percentage of the comparison difference of the shape characteristics.

[0030] S52. Extract the radiant luminance characteristics of the infrared imaging target, and compare them with the radiant luminance characteristic parameters in the target characteristic database to obtain the percentage of the comparison difference of the luminance characteristics.

[0031] S53. Among the percentage of the comparison difference of the shape characteristics and the percentage of the difference of the luminance characteristics, select the first 5 groups of percentage of the difference. According to the target type, add the percentage of the difference of the target characteristics of the same type of target to obtain the difference degree weight between the currently captured target and the pre-calculated target in the target characteristic database. Select the target with the smallest difference degree weight as the target recognition result for output.

[0032] In one embodiment of the present invention, the step S51 includes:

[0033] S511. Calculate the aspect ratio of the target. The aspect ratio L of the target is defined as the ratio of the major axis α to the minor axis β of the minimum circumscribed ellipse of the target image. During the simulation process, given that the horizontal FOV of the detector is θ, the vertical FOV is φ, the distance between the detector and the ship is d, according to the horizontal resolution N and vertical resolution M of the simulation scene, and the width W of the target in the simulation screen photo , the angular span Δα of the target in the horizontal direction in the simulation screen is obtained as:

[0034]

[0035] Based on the horizontal angular span Δα of the target and the distance d between the target and the detector, the actual length L of the target real is:

[0036]

[0037] The vertical angular span Δβ of the target is:

[0038]

[0039] The actual height H of the target real is:

[0040]

[0041] Then the aspect ratio L of the target is:

[0042] L = L real / H real ; (5)

[0043] S512. Calculate the compactness of the target. The compactness C is defined as the ratio of the perimeter P of the minimum circumscribed circle of the target to the pixel area A occupied by the target, expressed as:

[0044] C = 4πA / P 2 (6)

[0045] The length L of the target in the screen photp is used as the diameter of the circumscribed circle, and the pixel area of the target is obtained using the built-in blueprint nodes of the Unreal Engine, and the compactness of the target is obtained as:

[0046]

[0047] S513. Calculate the circularity Cir of the target. Based on the pixel area A of the target and the diameter L of the minimum circumscribed circle of the target photo , the calculation formula for the circularity of the target is obtained as:

[0048]

[0049] S514. Compare the calculated aspect ratio, compactness, and circularity of the target with the parameters in the target feature database respectively to obtain the percentage difference. Store the percentage difference in ascending order in the graph as a key-value pair with the common target type and the percentage difference.

[0050] In one embodiment of the present invention, the step S52 includes:

[0051] The brightness features of the infrared imaging target include brightness statistical features and invariant moment features. Denote the quantized brightness level of the infrared image as B, and q(b) as the probability density function of the brightness value b, expressed as:

[0052] q(b) = N p (b) / M p (9)

[0053] In formula (7), N p (b) is the number of pixels with the brightness value b, and M p is the total number of image pixels. Then the brightness statistical features include:

[0054] Brightness mean: μ b = Σb·q(b) (10)

[0055] Brightness variance:

[0056] Energy: E b = -Σ[q(b)] 2 (12)

[0057] Entropy: S b = -Σq(b)·log[q(b)] (13)

[0058] When calculating the target shape features, obtain the pixel area occupied by the target, extract the radiation brightness value of each pixel in the pixel area occupied by the target, calculate the corresponding brightness mean, brightness variance, energy, and entropy values, and store the calculation results in an array;

[0059] Compare the brightness statistical features calculated for the captured target with the pre-calculated brightness statistical feature database. Select the corresponding radiation brightness statistical feature values of each ship under this working condition as reference values according to the missile incoming angle and the time. Calculate the percentage difference between the measured value and the reference value. For example, if the obtained target brightness mean is E α , and the pre-calculated brightness means of common targets are E α1 , E α2 , Eα3 , ……;

[0060] Then the percentage difference between the brightness mean of the unknown target captured in real time and that of the common targets in the database is

[0061] Store the percentage difference in ascending order, with the common target type and the percentage difference as a pair of key-value pairs in a graph;

[0062] The calculation methods of the percentage differences in brightness variance, energy, and entropy are the same as that of the brightness mean percentage difference. After calculation, they are also stored in the graph in ascending order.

[0063] A ship target recognition system based on the UEAI behavior tree, comprising:

[0064] A target generation module, which is used to generate the intrinsic radiation of the target, manage the ships, interferences, and virtual seekers in the scene, load the ships, interferences, and virtual seekers during the scene operation, and make the ships, interferences, and virtual seekers appear in the scene and radiate intrinsic infrared radiation;

[0065] A background generation module, which is used to construct the sky background, ocean background, and simulate the atmospheric transmission effect in the scene, and manage the radiation of the sky background, the radiation of the sea surface background, and the influence of the atmospheric transmission effect on all infrared radiations in the scene;

[0066] A target recognition module, which performs target recognition by capturing the image in the current field of view, captures the total infrared radiation generated by the coupling of the target generation module and the background generation module in the form of a picture, and is used for target recognition judgment.

[0067] In an embodiment of the present invention, the target generation module includes a resource replacement module for target replacement, a motion control module for controlling the target motion process, a temperature texture customization module for target temperature texture replacement, and a trajectory customization module for customizing and modifying the trajectory;

[0068] The background generation module includes a sea surface background customization module for simulating sea surface effects in different regions, a sky background customization module for customizing and modifying the sky background, and an atmospheric texture customization module for atmospheric texture replacement;

[0069] The target recognition module includes an AI behavior tree visual perception module for real-time processing of various targets in the captured picture, an AI behavior tree logic execution module for calculating target features and performing logical judgments, and a target recognition result output module for outputting the target recognition result.

[0070] Advantages of the present invention:

[0071] 1. The ship target recognition method of the present invention combines the behavior tree and AI technology. By modifying the implementation algorithm of AI vision perception in the AI behavior tree, a ship target detection and recognition method with good real-time performance based on this database is designed. This method can replace the external input information, avoiding the possible influence of bad weather or complex sea conditions on target recognition, and is convenient for expansion and maintenance. Compared with other methods that require threshold segmentation and contour detection operations first, this method can directly screen out candidate targets, calculate the target features of the candidate targets, and quickly compare the calculation results with the feature database, improving the efficiency of real-time target detection. Based on the rendering optimization of the Unreal Engine itself for complex scenes, the rendering frame rate of this system is improved, providing strong support for maritime tasks.

[0072] 2. The ship target recognition system of the present invention is based on the AI behavior tree module in the Unreal Engine. By configuring the AI vision perception parameters in the AI behavior tree module, it aims to establish a modular ship target full-link infrared radiation characteristic modeling and simulation system, and build a refined infrared target feature - imaging feature database for ship targets based on this system. This database is convenient for expansion and maintenance. For new ship targets, only the original targets in the scene need to be replaced and the infrared target features under various working conditions need to be calculated to add new data to the database, facilitating subsequent target recognition simulation for this type of ship. Brief Description of the Drawings

[0073] Figure 1 is a flowchart of a ship target recognition method based on the UEAI behavior tree provided by an embodiment of the present invention;

[0074] Figure 2 is a schematic diagram of an anti-ship missile approaching provided by an embodiment of the present invention;

[0075] Figure 3 is a schematic diagram of the execution logic of the AI behavior tree provided by an embodiment of the present invention;

[0076] Figure 4 is a schematic diagram of target type weight calculation provided by an embodiment of the present invention;

[0077] Figure 5 is a schematic diagram of the internal data structure of the feature database provided by an embodiment of the present invention;

[0078] Figure 6 is a schematic diagram of the structure of a ship target recognition system based on the UEAI behavior tree provided by an embodiment of the present invention. Detailed Embodiments

[0079] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0080] The ship target recognition method based on the UEAI (Unreal Engine AI) behavior tree is a target recognition solution that combines the behavior tree and artificial intelligence (AI) technologies. The behavior tree is a tree-like structure used to control the decision-making process and is commonly used in fields such as game AI and robot control. UEAI is a unified agent framework designed to integrate perception, decision-making, and execution into one system. In the ship target recognition scenario, the behavior tree can be used to manage the target recognition process and, combined with technologies such as deep learning and computer vision, achieve accurate recognition and classification of ship targets.

[0081] The present invention provides a ship target recognition method based on the UEAI behavior tree. Referring to the attached Figure 1 , the ship target recognition method based on the UEAI behavior tree includes the steps:

[0082] S1. Build a simulation system scenario;

[0083] S2. Preset criteria and construct a target feature database;

[0084] S3. Run the simulation system scenario and capture the images at the preset viewing angle;

[0085] S4. Use the UEAI behavior tree to process various targets in the image in real time, perform calculations and logical judgments on the target features;

[0086] S5. Execute the criteria, compare with the target feature database to obtain the target feature result, and output the target recognition result of the current frame.

[0087] In an embodiment of the present invention, step S1 includes:

[0088] S11. Prepare the scene resources, which include ship models, temperature field textures, atmospheric texture resources, and various interferences.

[0089] S121. Generate targets using the scene resources. The targets include target ships, virtual seekers, and shipborne interferences. The target ships, virtual seekers, and shipborne interferences move according to the running trajectories; the schematic diagram of anti-ship missile approaching is referred to the attached Figure 2 , and the running modes of each target should generally follow this trajectory.

[0090] S122. Generate the sea surface background and sky background in the scene using the scene resources, and simulate the transmission effect of infrared radiation in the atmosphere in the scene.

[0091] S123. The movement simulation running trajectories of the elements in the scene are realized by using blueprint functions to achieve the movement of the target ship, the movement of the virtual seeker, and the function of the target ship releasing shipborne interferences when the virtual seeker approaches the target ship.

[0092] Blueprint functions are visual scripting tools in UE used to implement game logic. They achieve functions by creating node graphs without writing code. In this function, the movements of the target ship and the virtual seeker are realized by using the trajectory point CSV file read when the scene is loaded. According to the trajectory points and the spline component in UE, a trajectory spline is constructed in the scene. Then, through the timeline function in the blueprint function, a floating-point value from 0 to 1 is output within the specified time t. After that, within the time t, linear interpolation is performed on the spline based on the floating-point value in the range of 0 to 1. The result of the interpolation output is a three-dimensional coordinate point on the spline. As time goes from 0 to t and the floating-point value changes from 0 to 1, this three-dimensional coordinate point gradually moves from the starting point to the ending point of the spline. By updating the floating-point value through the timeline and setting the result of the linear interpolation on the corresponding spline to the target or the virtual seeker, the along-track movement of the target and the seeker can be achieved.

[0093] During the movement of the virtual seeker, the distance between the ship and the virtual seeker is calculated at each sampling time interval. At the same time, in the blueprint responsible for controlling the ship, it is judged in real time whether this distance is greater than or equal to a preset value. When the distance is less than or equal to the preset distance, the blueprint will execute the function of releasing interference. Otherwise, it will do nothing and enter the next sampling interval.

[0094] The function of releasing interference utilizes the projectile component in the UE target. The projection direction of the interference is calculated through the missile approaching angle. After projection, when the preset explosion time is reached, the interference is generated.

[0095] The trajectory of the ship's movement is constructed through the trajectory customization module. At the stage of loading resources during the scene operation, the CSV table at the target path will be read to construct the corresponding orbit, and the blueprint function will implement the related functions of the target's along-track movement. The CSV table is a file that stores tabular data in plain text form separated by commas. When read by the program, it is regarded as a character sequence separated by commas, which is convenient for data processing.

[0096] The function of releasing interference is triggered by means of distance detection. When the distance between the virtual seeker and the ship reaches the preset value, the corresponding interference release event will be executed. After the interference is released, it will be executed according to the movement mode of this interference until the preset interference duration is reached, and then the destruction event of the interference will be executed to avoid excessive resources in the scene affecting the running frame rate and simulation effect.

[0097] S12. Use the scene modeling tool of the Unreal Engine to build the scene resources into an infrared scene according to the requirements.

[0098] Import the scene resources into the Content Browser of UE, and then create a new level as the carrier of the scene. Add the sea surface blueprint in the Oceanology plugin to the scene. Drag the blueprint into the scene to generate a sea surface background in the scene. Add a post - processing volume component to the scene, and use the atmospheric transmittance texture and path radiance texture in the scene resources to simulate the atmospheric effect. Use the sky background blueprint of UE. After dragging it into the scene, change the material of the sky background, and load the sky background texture in the scene resources into its material, so that the sky background becomes an infrared sky. Then drag the model from the Content Browser into the scene, and adjust the position, rotation and scale to make the ship stop on the sea surface. Finally, open the blueprints of the ship and the virtual seeker, and implement the motion control and interference release functions in the blueprints.

[0099] In one embodiment of the present invention, step S2 includes:

[0100] S21. Create a blackboard and an AI behavior tree in the Unreal Engine. Add relevant parameters to the blackboard to judge whether a target is captured and store all the target tags currently captured. Build relevant branch logics in the AI behavior tree to execute the ballistic correction after capturing the target or the continuous scanning process after not capturing the target. The criteria include the shape characteristics and radiation brightness characteristics of the target.

[0101] S22. Obtain the images of common ship targets in the daytime environment through simulation. According to different shooting angles and shooting times, calculate the shape characteristics and radiation brightness statistical characteristics of the target at different missile incoming angles according to the blueprint functions constructed by the calculation methods of each target characteristic in the criteria. Store them according to the ship model, incoming angle, incoming time, corresponding shape characteristics and radiation brightness characteristic structure, and build a target characteristic database.

[0102] In one embodiment of the present invention, step S4 includes:

[0103] S41. The target ship and the interference need to be placed in the scene after creating blueprints. First, create a blueprint class of the target ship in the Content Browser of the Unreal Engine, and implement the functions of target movement, detecting the distance between the missile and the ship, and interference release in the blueprint according to the implementation process in S123. Then, create a blueprint class of the interference, add a projectile component to the blueprint, set the projectile parameters and the explosion timing, and add a blueprint event to trigger the projectile component. Finally, instantiate these blueprints and drag them from the Content Browser into the scene, and the corresponding target can be generated in the scene to complete the creation of the target and the interference in the scene;

[0104] S42. The virtual seeker is created using the Pawn type and placed in the scene. In the Content Browser, you can select the type that the created blueprint inherits. To create a virtual seeker, you need to select the blueprint inheritance class. Select the Pawn type in the blueprint type to create it. After creation, add a SceneCapture2D component to its blueprint. This component can capture the images within the camera's field of view in real time after being enabled and can customize the detector parameters. Then, follow the implementation process in S123 to implement the function of the virtual seeker moving along the trajectory. Finally, drag the virtual seeker from the Content Browser into the scene. After the scene runs, the virtual seeker will move along the trajectory;

[0105] S43. Create a blueprint for an AI controller under the Actor class, add an AI perception component and configure the AI sense as AI vision perception. Create a new blueprint in the Content Browser. When selecting the inherited parent class, there is a subclass named Controller (controller) branch in the Actor class directory below. You can find the AIController (AI controller) class under this branch and select this class as the parent class to create the blueprint. Add an AIPerception (AI perception component) to the newly created blueprint. Click on this component to configure it in the Details panel on the right. There is a parameter named "Sense Configuration" in the Details panel of the AI perception component. Select AI vision perception in its selection list to complete the creation process of the AI controller. Finally, open the blueprint of the virtual seeker and change the AI controller class of its Pawn to the AI controller created in this step in the Details panel on its right;

[0106] S44. Determine whether the target is captured and store all the target tags captured currently, and perform the ballistic correction after capturing the target or continue the scanning process after not capturing the target. The execution logic of the AI behavior tree is as shown in the appendix Figure 3 shown. First, the detector scans. Whether the target is captured. If "yes", execute the criterion to determine the type of the captured target; if "no", the detector continues to scan. Then, determine whether the target is a ship. If "yes", output the ship target type; if "no", return.

[0107] The criteria in the configuration process include the shape features of the target and the radiation brightness features of the target.

[0108] In an embodiment of the present invention, step S5 includes:

[0109] S51. When executing the criterion, first extract the shape features of the target. The shape features include the aspect ratio, compactness, and circularity of the target, and compare them with the shape feature parameters in the target feature database to obtain the percentage of the comparison difference of the shape features.

[0110] S511. Calculate the aspect ratio of the target. The aspect ratio L of the target is defined as the ratio of the major axis α to the minor axis β of the minimum circumscribed ellipse of the target image. During the simulation, given that the horizontal FOV of the detector is θ, the vertical FOV is φ, the distance between the detector and the ship is d, according to the horizontal resolution N and vertical resolution M of the simulation scenario, and the width W of the target in the simulation image photo , the angular span Δα of the target in the horizontal direction in the simulation image is obtained as follows:

[0111]

[0112] Based on the horizontal angular span Δα of the target and the distance d between the target and the detector, the actual length L of the target real is as follows:

[0113]

[0114] The vertical angular span Δβ of the target is:

[0115]

[0116] The actual height H of the target real is as follows:

[0117]

[0118] Then the aspect ratio L of the target is:

[0119] L = L real / H real ; (5)

[0120] S512. Calculate the compactness of the target. The compactness C is defined as the ratio of the perimeter P of the minimum circumscribed circle of the target to the pixel area A occupied by the target, expressed as:

[0121] C = 4πA / P 2 (6)

[0122] The length L of the target in the image photo is used as the diameter of the circumscribed circle. The pixel area of the target is obtained using the built-in blueprint nodes of the Unreal Engine, and the compactness of the target is obtained as:

[0123]

[0124] S513. Calculate the circularity Cir of the target. Based on the pixel area A of the target and the diameter L of the minimum circumscribed circle of the target photo , the calculation formula for the circularity of the target is obtained as:

[0125]

[0126] S514. Compare the calculated aspect ratio, compactness, and circularity of the target with the parameters in the target feature database respectively to obtain the percentage difference. Arrange the percentage differences in ascending order and store them in a graph as key-value pairs with common target types and percentage differences.

[0127] After obtaining the calculation results of the three shape features, compare them with the shape feature parameter database of common ships such as Ford-class aircraft carriers, Burke-class destroyers, Ticonderoga-class cruisers, etc. Select the corresponding shape feature values of each ship under this working condition according to the missile incoming angle and the time.

[0128] According to the percentage difference between the calculated value and the existing data. For example, if the aspect ratio of the obtained target is L α , and the pre-calculated aspect ratios of common targets are L α1 , L α2 , L α3 , …… etc. Then the percentage difference between the aspect ratio of the unknown target captured in real time and the aspect ratio of the common target in the database is: Finally, arrange these results in ascending order and store them in a graph (Map) as key-value pairs with common target types and percentage differences. The calculation method of the percentage difference in the comparison of compactness and circularity is the same as that of the aspect ratio, and they are stored in a graph respectively after calculation.

[0129] Store the hit probabilities of the targets in descending order in an array, and then continue to calculate the radiation brightness characteristics. The brightness characteristics of infrared imaging targets mainly include point characteristics, brightness statistical characteristics, and invariant moments, etc. In this system, the brightness statistical characteristics and invariant moment characteristics are mainly used.

[0130] S52. Extract the radiation brightness characteristics of the infrared imaging target, compare them with the radiation brightness characteristic parameters in the target feature database, and obtain the percentage difference in the brightness characteristics;

[0131] The brightness characteristics of the infrared imaging target include brightness statistical characteristics and invariant moment characteristics. Denote the quantized brightness level of the infrared image as B, and q(b) as the probability density function of the brightness value b, which is expressed as:

[0132] q(b) = N p (b) / M p (9)

[0133] In formula (7), N p (b) is the number of pixels with the brightness value b, and M p is the total number of image pixels. Then the brightness statistical characteristics include:

[0134] Brightness mean: μ b = Σb·q(b) (10)

[0135] Luminance variance:

[0136] Energy: E b = -Σ[q(b)] 2 (12)

[0137] Entropy: S b = -Σq(b)·log[q(b)] (13)

[0138] When calculating the target shape features, the pixel area occupied by the target is obtained, the radiance value of each pixel in the pixel area occupied by the target is extracted, the corresponding luminance mean, luminance variance, energy, and entropy values are calculated, and the calculation results are stored in an array.

[0139] When preparing resources, based on the temperature field textures of common ships, we calculate the luminance statistical features of these ships in advance and store them for reference. Compare the luminance statistical features calculated from the captured target with the luminance statistical feature database that has been calculated in advance. Select the corresponding radiance statistical feature values of each ship under this working condition as reference values according to the missile incoming angle and the time of occurrence, and calculate the percentage difference between the measured value and the reference value. For example, the obtained target luminance mean is E α , and the calculated luminance mean of the common target is E α1 , E α2 , E α3 , …….

[0140] Then the percentage difference between the luminance mean of the unknown target captured in real time and the luminance mean of the common target in the database is

[0141] Store the percentage difference in ascending order, with the common target type and the percentage difference as a key-value pair in a graph.

[0142] The calculation method of the percentage difference of luminance variance, energy, and entropy is the same as that of the percentage difference of luminance mean. After calculation, they are also stored in the graph in ascending order.

[0143] S53. Among the percentage differences of shape feature comparison and the percentage differences of luminance feature, select the top 5 groups of percentage differences. According to the target type, add up the percentage differences of the target features of the same type of target. The specific process is as shown in the appendix Figure 4 as follows.

[0144] The sum of the percentage differences of each target feature of Ship 1 is:

[0145]

[0146] where P 船irepresents the difference between the currently captured target and ship i, E 长宽比i represents the percentage difference between the previously calculated captured target and the calculated aspect ratio characteristics of the common ship i, E 紧凑度i 、E 圆形度i 、……, and so on.

[0147] According to formula (14), the difference degree weight between the currently captured target and the pre-calculated target in the target feature database can be obtained, and the target with the smallest difference degree weight is selected as the target recognition result for output.

[0148] The construction of the database for storing various characteristic values of common ship targets mentioned in the criterion calculation process is carried out on the basis of completing the infrared scene construction.

[0149] To build the database, we need to obtain various images of multiple common ship targets in the simulation environment through simulation. According to different shooting angles and different shooting times, and according to the blueprint function constructed by the calculation method of each target feature in the criterion, calculate the shape characteristics and radiation brightness statistical characteristics of the target under different missile incoming angles, and store them according to the ship model, incoming angle, incoming time, corresponding shape characteristics and radiation brightness characteristic structure, as shown in the appendix Figure 5 as shown.

[0150] After the feature database is constructed, it is stored in the program using a structure array, loaded during the scene operation. When performing parameter comparison, first match the time, then perform a nearest match for the incoming angle, and then obtain the ship type and its shape brightness statistical feature data group that meet the matching result according to the matching result for calculation and comparison.

[0151] The detection and recognition method of spatial infrared targets based on the AI behavior tree in the Unreal Engine of the present invention can analyze the images captured by the virtual detector in real-time in virtual simulation experiments, and based on the provided recognition basis, accurately recognize the targets in the scene from the complex environmental background and the scene with interference.

[0152] The present invention also provides a ship target recognition system based on the UEAI behavior tree. Referring to the appendix Figure 6 , the ship target recognition system based on the UEAI behavior tree includes: a target generation module, a background generation module, and a target recognition module.

[0153] The target generation module is used to generate the intrinsic radiation of the target, manage elements such as ships, jammers, and virtual seekers in the scene, load the above elements during the scene operation, and make each element appear in the scene and radiate intrinsic infrared radiation; the background generation module is used to construct the sky background, ocean background and simulate the atmospheric transmission effect in the scene, and manage the radiation of the sky background, the radiation of the sea surface background and the influence of the atmospheric transmission effect on all infrared radiations in the scene; the target recognition module captures the image in the current field of view for target recognition. The value of each pixel point in the image is the radiation brightness of the target or the background, and captures the total infrared radiation generated by the coupling of the target generation module and the background generation module in the form of a picture for target recognition judgment.

[0154] The target generation module manages the intrinsic infrared radiation of each element, and the background generation module manages the infrared radiation influence brought by the environment. After the two are coupled, the detector can receive realistic infrared radiation.

[0155] In summary, the target generation module and the background generation module are the basis of the system. The target generation module is used to create the target, control the movement and intrinsic radiation of the target. The background generation module is used to create the environmental background, generate the sky background radiation, the sea surface background radiation and simulate the atmospheric transmission effect. After the interaction between the background generation module and the target generation module, the total infrared radiation of the target is generated. The target recognition module is the core of the system function. It captures the infrared image generated by the coupling of the target generation module and the background generation module through the virtual seeker, obtains the target through its AI behavior tree perception module, judges the type of the target through the AI behavior tree logic execution module, and finally outputs the recognition result through the target recognition result output module.

[0156] In an embodiment of the present invention, the target generation module includes a resource replacement module, a motion control module, a temperature texture customization module and a trajectory customization module; the background generation module includes a sea surface background customization module, a sky background customization module and an atmospheric texture customization module; the target recognition module includes an AI behavior tree visual perception module, an AI behavior tree logic execution module, and a target recognition result output module.

[0157] The resource replacement module and the motion control module are combined into a model customization module. In this system, the models are divided into target models and jammer models. Among them, the target model is generally controlled by blueprint functions, and the target can be replaced by using blueprint functions. Use the prepared ship model resources, pass its path to the model customization module, and the function will read the model resources of the target path and use it to replace the original model.

[0158] The interference model is modeled using the Niagara particle system in the Unreal Engine according to the release process of common shipborne interferences and the physical effects of interferences under an infrared camera. The particle system is used to simulate the whole process of interference from explosion to dissipation and some specific physical characteristics of the interference, such as the radiant luminance and transmittance of the interference. The replacement of the interference model can also be carried out through blueprint functions, and different interference effects can be achieved by selecting different Niagara systems.

[0159] The motion control module is a more target - type implementation of various motion modes. By presetting the target type, relevant motion functions are implemented to control the motion process of the target. For example, the motion control of the interference is implemented in the Unreal Engine to enable it to be automatically released when the virtual seeker approaches the ship within a certain distance and follow the actual interference motion mode.

[0160] Temperature texture customization module: Blueprint functions in the Unreal Engine can control the material parameters of the model. By implementing relevant blueprint functions, the texture parameters of the model can be controlled. With the prepared texture resources, the path of the texture resources is used as the input of the function, and the target model is passed into the function by reference. The function replaces the texture in the material of the target model in the form of material parameters, and finally the replacement of the target temperature texture is completed after the function finishes.

[0161] Trajectory customization module: The trajectories in the scene are generally generated by splines. The sampling point data used by the splines, that is, the track data, is obtained by externally reading a CSV file. By implementing blueprint functions, when the scene is loaded, the CSV table is read from the target path to obtain its sampling point data. The sampling point information includes the world coordinates (X, Y, Z) and rotation (Pitch, Yaw, Roll) of the object sampling frame, and a track composed of sampling points is generated in the scene. Finally, the track is bound to the target model. The custom modification of the track can be achieved by changing the sampling point data in the table.

[0162] Ocean background customization module: The ocean background uses a dynamic ocean component generated by the Oceanology plugin. After modifying the parameters of roughness and average emissivity in the material, and based on the 24 - hour average temperature of the ocean in a certain area obtained from reference materials, the ocean is infrared - processed. Parameters such as the temperature parameter, roughness, and average emissivity of the ocean can be customized and modified to achieve the effect of simulating the sea surface in different regions.

[0163] The parameter input in the atmosphere texture custom module is to use Modtran software to calculate the multi-band atmospheric radiation data under different cut heights and different sunlight, and generate a multi-band atmospheric texture based on this, which includes a set of atmospheric transmittance and atmospheric path radiation in a certain height range and a certain zenith angle range at different distances. The above atmospheric texture resources are stored in two texture arrays in the Unreal Engine, and then the distance is used as the index of the texture array in the post-processing material, and the height and zenith angle are used as the UV of the texture to obtain the atmospheric transmittance and atmospheric path radiation under specific UV coordinates, thereby realizing the simulation of the atmospheric background. Modtran software uses different calculation parameters to achieve the replacement of atmospheric textures.

[0164] The sky background customization module can modify the parameters on the interface to modify the weather type and the intensity of the corresponding weather parameters, thereby realizing customized modification of the sky background.

[0165] The AI ​​behavior tree visual perception module is used to process various targets in the captured images in real time. The AI ​​behavior tree logic execution module is used to calculate target features and perform logical judgments. The target recognition result output module is used to output target recognition results.

[0166] The present invention forms a modular ship target full-link refined imaging simulation system, in which external input information can be replaced, avoiding the impact of bad weather or complex sea conditions on target identification, and is easy to expand and maintain. At the same time, the present invention uses the AI ​​behavior tree module of the Unreal Engine to improve the efficiency of real-time target detection, increase the rendering frame rate, and provide strong support for maritime missions.

[0167] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A ship target recognition method based on UEAI behavior tree, characterized in that: Includes steps: S1. Build simulation system scenario; S2, preset criteria and build target feature database; S3, running the simulation system scene to capture the image at a preset viewing angle; S4. Use the UEAI behavior tree to process various targets in the picture in real time, and perform target feature calculation and logical judgment; S5. Execute the criterion, compare with the target feature database to obtain the target feature result, and output the target recognition result of the current frame.

2. According to the method for ship target recognition based on UEAI behavior tree in claim 1, it is characterized in that: The step S1 comprises: S11. Prepare scene resources, including ship models, temperature field textures, atmospheric texture resources, and various interferences; S12. Use the scene modeling tool of Unreal Engine to build the scene resources into an infrared scene according to requirements.

3. According to claim 2, a ship target recognition method based on UEAI behavior tree is characterized in that: The step S12 comprises: S121, generating a target using the scene resources, wherein the target includes a target ship, a virtual seeker, and a shipborne interference, and the target ship, the virtual seeker, and the shipborne interference run according to a running trajectory; S122, using the scene resources to generate a sea surface background and a sky background in the scene, and simulating the transmission effect of infrared radiation in the atmosphere in the scene; S123. The movement simulation of each element in the scene is realized by the running trajectory, and the movement of the target ship, the movement of the virtual seeker, and the function of the target ship releasing ship interference when the virtual seeker approaches the target ship are realized by using blueprint functions.

4. According to the method of ship target recognition based on UEAI behavior tree in claim 3, it is characterized in that: The step S2 comprises: S21. Create a blackboard and an AI behavior tree in the Unreal Engine, add relevant parameters in the blackboard for determining whether a target is captured and storing all currently captured target labels, and construct relevant branch logic in the AI ​​behavior tree for executing trajectory correction after capturing a target or continuing scanning after failing to capture a target, wherein the judgment criteria include the shape characteristics of the target and the radiation brightness characteristics of the target; S22. Images of common ship targets in daytime environments are obtained through simulation. According to the shooting angles and shooting times, the shape characteristics and radiation brightness statistical characteristics of the targets under different missile attack angles are calculated according to the blueprint function constructed by the calculation method of each target feature in the criterion. The target feature database is built by storing the corresponding shape characteristics and radiation brightness characteristic structures according to the ship model, attack angle, attack time, and corresponding shape characteristics.

5. According to claim 4, a ship target recognition method based on UEAI behavior tree is characterized in that: The step S4 comprises: S41, the target ship and the interference need to be placed in the scene after creating a blueprint; S42, the virtual seeker is created using the Pawn type and placed in the scene; S43. Create an AI controller blueprint under the Actor class, add the AI ​​perception component and configure the AI ​​sense to AI vision perception; S44, determining whether the target is captured and storing all target tags currently captured, executing trajectory correction after the target is captured or continuing the scanning process after the target is not captured.

6. According to the method of ship target recognition based on UEAI behavior tree in claim 5, it is characterized in that: The step S5 comprises: S51, extracting shape features of the target, the shape features including aspect ratio, compactness and circularity of the target, and comparing them with shape feature parameters in the target feature database to obtain a comparison difference percentage of the shape features; S52, extracting the radiation brightness characteristics of the infrared imaging target, and comparing them with the radiation brightness characteristic parameters in the target characteristic database to obtain the contrast difference percentage of the brightness characteristics; S53. Select the first five groups of difference percentages from the contrast difference percentages of the shape features and the difference percentages of the brightness features, add the difference percentages of the target features of the same type of targets according to the target type, obtain the difference weight between the current captured target and the pre-calculated target in the target feature database, and select the target with the smallest difference weight as the target recognition result output.

7. The ship target recognition method based on UEAI behavior tree according to claim 6 is characterized in that: The step S51 comprises: S511. Calculate the aspect ratio of the target. The aspect ratio L of the target is defined as the ratio of the major axis α to the minor axis β of the minimum circumscribed ellipse of the target image. In the simulation process, it is known that the horizontal FOV of the detector is θ, the vertical FOV is φ, and the distance between the detector and the ship is d. According to the horizontal resolution N and vertical resolution M of the simulation scene, the width W of the target in the simulation picture is photo , the horizontal angle span Δα of the target in the simulation screen is obtained as: According to the horizontal angle span Δα of the target and the distance d between the target and the detector, the actual length L of the target is real for: The vertical angle span Δβ of the target is: The actual height of the target H real for: Then the aspect ratio L of the target is: L=L real / H real ; (5) S512, calculate the compactness of the target, where the compactness C is defined as the ratio of the circumference P of the minimum circumscribed circle of the target to the pixel area A occupied by the target, expressed as: C=4πA / P 2 (6) The length of the target on the screen is L photp As the diameter of the circumscribed circle, the pixel area of ​​the target is obtained using the Unreal Engine's built-in blueprint node, and the target's compactness is: S513, calculate the circularity Cir of the target, based on the pixel area A of the target and the diameter L of the minimum circumscribed circle of the target photo , the calculation formula for the target circularity is: S514. Compare the calculated aspect ratio, compactness and circularity of the target with the parameters in the target feature database to obtain the difference percentage, and store the difference percentage in a graph in ascending order with common target type and difference percentage as a key-value pair.

8. According to claim 6, a ship target recognition method based on UEAI behavior tree is characterized in that: The step S52 comprises: The brightness characteristics of infrared imaging targets include brightness statistical characteristics and invariant moment characteristics. The brightness level of the quantized infrared image is denoted as B, and q(b) is the probability density function of the brightness value b, which can be expressed as: q(b)=N p (b) / M p (9) In formula (7), N p (b) is the number of pixels with brightness value b, M p is the total number of image pixels, then the brightness statistical features include: Mean brightness: μ b =Σb·q(b)(10) Brightness variance: Energy: E b =-Σ[q(b)] 2 (12) Entropy: S b =-Σq(b)·log[q(b)](13) When calculating the target shape features, the pixel area occupied by the target is obtained, the radiation brightness value of each pixel in the pixel area occupied by the target is extracted, the corresponding brightness mean, brightness variance, energy and entropy value are calculated, and the calculation results are stored in an array; The brightness statistical characteristics calculated by the captured target are compared with the brightness statistical characteristics database calculated by the realization. According to the missile attack angle and the time, the corresponding radiation brightness statistical characteristic value of each ship under this working condition is selected as the reference value, and the difference percentage between the measured value and the reference value is calculated. For example, the average brightness of the target is E α , the average brightness of the common target is E α1 , E α2 , E α3 ,……; The difference percentage between the brightness mean of the unknown target captured in real time and the common targets in the database is The difference percentages are stored in a graph in ascending order, with common target types and difference percentages as a pair of key-value pairs; The calculation method of the difference percentage of brightness variance, energy, and entropy is the same as the calculation method of the difference percentage of brightness mean. After the calculation is completed, they are also stored in the graph in ascending order.

9. A ship target recognition system based on UEAI behavior tree, characterized in that: include: The target generation module is used to generate the intrinsic radiation of the target, manage the ships, jammers, and virtual seekers in the scene, load the ships, jammers, and virtual seekers when the scene is running, and make the ships, jammers, and virtual seekers appear in the scene and radiate intrinsic infrared radiation; Background generation module, used to construct sky background, ocean background and simulate atmospheric transmission effect in the scene, manage the radiation of sky background, sea background and the influence of atmospheric transmission effect on all infrared radiation in the scene; The target recognition module recognizes the target by capturing the image in the current field of view, and captures the total infrared radiation generated by the coupling of the target generation module and the background generation module in the form of a picture for target recognition judgment.

10. A ship target recognition system based on UEAI behavior tree according to claim 9, characterized in that: The target generation module includes a resource replacement module for target replacement, a motion control module for controlling the target motion process, a temperature texture customization module for target temperature texture replacement, and a trajectory customization module for trajectory customization modification; The background generation module includes a sea surface background customization module for simulating sea surface effects in different areas, a sky background customization module for customizing and modifying the sky background, and an atmosphere texture customization module for replacing atmosphere texture; The target recognition module includes an AI behavior tree visual perception module for real-time processing of various targets in the captured image, an AI behavior tree logic execution module for calculating target features and performing logical judgment, and a target recognition result output module for outputting target recognition results.

Images