An unmanned aerial vehicle intelligent defense and attack control system

Abstract

The application belongs to the technical field of unmanned aerial vehicle attack and defense, and discloses an intelligent defense and attack control system for unmanned aerial vehicles; the method comprises the following steps: acquiring environment sensing data of the unmanned aerial vehicle, wherein the environment sensing data comprises state data of the unmanned aerial vehicle itself and image data of the surrounding environment; pre-processing the image data of the surrounding environment to obtain model input image data; performing target detection on the model input image data by using a variant Poolformer algorithm to acquire comprehensive information of the target object; solving an optimal attack and defense strategy by using a perception integrated attack algorithm according to the state data of the unmanned aerial vehicle itself and the comprehensive information of the target object; sending the optimal attack and defense strategy to an unmanned aerial vehicle execution terminal to control the unmanned aerial vehicle to execute the optimal attack and defense strategy; the optimal attack and defense strategy can be dynamically generated according to the real-time change of the environment situation, and the survival ability and striking efficiency of the unmanned aerial vehicle in the hostile environment are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle attack and defense technology, and more particularly, relates to an intelligent defense and attack control system for unmanned aerial vehicles. BACKGROUND

[0002] The patent with the application publication number CN116820126A discloses a deep reinforcement learning system and method based on multi-unmanned aerial vehicle path planning. A virtual environment model of a multi-unmanned aerial vehicle path planning scene is built, and a deep reinforcement learning algorithm is used to train an intelligent agent unmanned aerial vehicle in the model. The state information collected during the flight of the unmanned aerial vehicle is input into the neural network of the deep reinforcement learning to train the model. During the training process, an adversarial attack algorithm is used to interfere with the Markov decision sequence received by the intelligent agent. A state and reward dual defense framework is constructed to complete the defense against the attack disturbance. The disturbed Markov sequence is input into the defense framework for data recovery. The recovered data sequence is placed in an experience pool for the unmanned aerial vehicle to make subsequent action decisions, and the optimal flight path of the unmanned aerial vehicle is obtained through multiple rounds of model training. The trained reinforcement learning model is subjected to adversarial attack, and the model that passes the defense test is deployed to an unmanned aerial vehicle edge computing platform.

[0003] However, the existing unmanned aerial vehicle defense and attack control system still has many deficiencies in the actual environment. In a complex environment, the target recognition and tracking capability is weak, especially in harsh weather, complex terrain, or electromagnetic interference conditions, which easily leads to target loss or misidentification, seriously affecting the effect. In terms of threat assessment, a simple distance threshold or a fixed weight assessment method is often used, which cannot accurately reflect the real threat level of the target, leading to decision-making errors when facing multiple target threats. In terms of attack and defense strategy formulation, the traditional system mostly uses preset fixed strategies, lacks the ability to quickly respond to changes in the battlefield situation, and cannot adjust the attack and defense strategy in a timely manner according to the dynamic changes of the situation of the enemy and ourselves, which easily leads to task failure or loss of unmanned aerial vehicles in a high-intensity confrontation environment. At the same time, the existing system cannot effectively balance multiple key factors during task execution, often resulting in overly aggressive or overly conservative decision-making. In addition, when facing new threats or unknown situations, the combat effectiveness cannot be continuously improved.

[0004] In view of this, the present application provides an intelligent defense and attack control system for unmanned aerial vehicles to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: an intelligent defense and attack control system for unmanned aerial vehicles, comprising: a data acquisition module for acquiring environment perception data of the unmanned aerial vehicle, the environment perception data including state data of the unmanned aerial vehicle itself and image data of the surrounding environment;

[0006] a data processing module, configured to pre-process image data of a surrounding environment to obtain model input image data;

[0007] a target detection module, configured to perform target detection on the model input image data by using a variant Poolformer algorithm to obtain comprehensive information of a target object;

[0008] a strategy fitting module, configured to solve an optimal attack and defense strategy by using a perception-integrated attack algorithm according to state data of the unmanned aerial vehicle and the comprehensive information of the target object, and send the optimal attack and defense strategy to an unmanned aerial vehicle execution terminal to control the unmanned aerial vehicle to execute the optimal attack and defense strategy; the modules are connected through wired and / or wireless modes.

[0009] Further, the state data of the unmanned aerial vehicle includes position data, motion state, energy state, weapon state, health management state and environment state; the environment state includes temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information.

[0010] The obstacle information includes three-dimensional spatial position coordinates of the obstacle, obstacle size and motion trajectory of a dynamic obstacle.

[0011] Further, the motion trajectory of the dynamic obstacle is obtained in the following manner:

[0012] The dynamic obstacle is regarded as a moving particle, and the unmanned aerial vehicle and other obstacles generate virtual gravitational field and repulsive field for the particle, and the particle moves under the action of the resultant force of the gravitational field and the repulsive field.

[0013] The unmanned aerial vehicle is defined as a gravitational source; the size of the gravitational field is inversely proportional to the distance between the unmanned aerial vehicle and the dynamic obstacle; the direction of the gravitational field points to the position of the unmanned aerial vehicle;

[0014] Other obstacles are defined as repulsive sources, and the size of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle; the direction of the repulsive field points away from the repulsive source;

[0015] For each time step, the resultant force acting on the particle is calculated, which is equal to the sum of the gravitational field and the sum of the repulsive field; the resultant force is substituted into the Newton motion equation to obtain the acceleration a, and the numerical integration is performed to obtain the velocity and displacement of the particle; a preset time range is set, and the changes of the velocity and displacement of the particle within the time range are used to construct a three-dimensional position sequence of the unmanned aerial vehicle, which is the motion trajectory of the dynamic obstacle.

[0016] Further, the pre-processing manner of the image data of the surrounding environment includes:

[0017] The original image data is subjected to attention fusion equalization processing to obtain preliminary input image data; the preliminary input image data is subjected to image cropping to obtain a key region image; pixel values in the key region image are scaled to a range of 0-1 or -1 to 1 to obtain model input image data;

[0018] The attention fusion equalization processing manner includes:

[0019] An attention structure based on a deep convolutional neural network is defined; the input of the attention structure is original image data, and the output is an attention weight map of the same size as the original image data; each pixel value in the attention weight map represents the importance degree of the corresponding position;

[0020] An image dataset containing images under various scenes and conditions is collected, and the images contain target objects; the images in the image dataset are manually annotated, and the positions and categories of the target objects are marked on the images using a bounding box;

[0021] The attention structure is integrated into a classification network or a target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure multiplies the output attention weight map with the original image data as a mask to obtain a feature map;

[0022] The original image data is subjected to evolutionary histogram equalization to obtain a globally enhanced image; the feature map and the globally enhanced image are multiplied to obtain preliminary input image data.

[0023] Further, the evolutionary histogram equalization manner includes:

[0024] The original image data is divided into a plurality of overlapping small block regions, and the size of each small block region is pre-set; a certain overlapping region is set between adjacent small blocks; each small block region is subjected to histogram equalization processing to obtain a reconstructed small block region;

[0025] For each reconstructed small block region, the normalized coordinates of the center point in the original image data are calculated; the center point coordinates are recorded; according to the center point coordinates, the Euclidean distance d of the reconstructed small block region to the center of the original image data is calculated; for each reconstructed small block region, the number of overlapping pixels with other reconstructed small block regions is counted; for each reconstructed small block region, a gray co-occurrence matrix is calculated; based on the comprehensive features in the gray co-occurrence matrix; the comprehensive features include contrast DA, energy DH, homogeneity DF, and entropy DS;

[0026] Contrast DA = ∑ I ∑ J (I-J) γ ×P(I,J) δ; wherein, I and J are the row and column indexes of the gray co-occurrence matrix; P(I, J) is an element in the gray co-occurrence matrix, representing the frequency of pixel pairs between the gray values I and J; γ is an exponential factor controlling the degree of influence of the gray difference on the contrast; and δ is an exponential factor controlling the degree of influence of the frequency of pixel pairs on the contrast;

[0027] Homogeneity wherein, a and b are parameters controlling the shape of the function; and the entropy DS = ∑ I ∑ J P(I, J) x log(P(I, J));

[0028] The integrated features are normalized and weightedly fused to obtain an integrated texture complexity index corresponding to the reconstructed small block region; for each reconstructed small block region, an edge detection operator is applied to extract the number of edge pixels and the cumulative gradient amplitude, and weightedly fused to obtain an edge intensity index of the reconstructed small block region; the integrated texture complexity index and the edge intensity index are fused into an integrated feature index C;

[0029] The position weight of the reconstructed small block region h is defined as wherein, σ is a parameter controlling the weight decay rate; ω1 is an influence parameter; ω2 is an exponential factor controlling the distance, used to control the degree of influence of the distance on the weight;

[0030] The overlap weight of the reconstructed small block region h is defined as wherein, a1 is a scaling coefficient; No is the number of pixels overlapped by the reconstructed small block region h and other reconstructed small block regions, Nb is the total number of pixels of the reconstructed small block region h; a2 is an overlap exponential factor; γ2 is a control parameter; the position weight and the overlap weight of the reconstructed small block region h are multiplied to obtain an integrated weight coefficient corresponding to the reconstructed small block region;

[0031] For the overlapped regions in the reconstructed small block region, the pixel values therein are weightedly averaged according to the integrated weight coefficient corresponding to the reconstructed small block region to obtain fused pixel values; for the non-overlapped regions in the reconstructed small block region, the corresponding pixel values are directly adopted to obtain a complete globally enhanced image.

[0032] Further, the image cropping of the preliminary input image data comprises:

[0033] Based on the dynamic obstacle, the view volume of each time step in the time range is determined according to the motion trajectory, for each time step t on the motion trajectory, taking the position of the unmanned aerial vehicle as the origin, a camera coordinate system is established, and the three coordinate axis directions of the camera coordinate system are determined according to the attitude angle of the unmanned aerial vehicle; according to the field of view angle, the coordinates of the eight vertices of the camera frustum in the camera coordinate system are calculated; the vertex coordinates in the camera coordinate system are converted to the world coordinate system by using the position and attitude angle of the unmanned aerial vehicle, to obtain the coordinates of the eight vertices in the world coordinate system, and a three-dimensional frustum or pyramid is constructed by using the coordinates of the eight vertices in the world coordinate system; the three-dimensional frustum or pyramid is the field of view range of the unmanned aerial vehicle at the time step t; then a series of time-varying three-dimensional view volumes V(t) are obtained;

[0034] For each time step t, the corresponding three-dimensional view volume V(t) is projected onto a two-dimensional image plane to obtain a series of two-dimensional polygonal regions P(t);

[0035] A mask image M with the same size as the global enhanced image is initialized, and all pixel values in the mask image M are initially 0; for each time step t, the pixel values in the corresponding two-dimensional polygonal region P(t) are set to 1 in the mask image M; the global enhanced image is multiplied pixel by pixel with the mask image M to obtain a key region image.

[0036] Further, the manner of performing target detection on the model input image data comprises:

[0037] A variant Poolformer model is constructed, and the basic structure of the variant Poolformer model comprises an encoder and a decoder; the encoder is defined to comprise a main feature branch and a parallel Fourier transform branch; the Fourier transform branch performs two-dimensional discrete Fourier transform on the model input image data, converts it from the spatial domain to the frequency domain, and performs low-pass filtering and high-pass filtering on the model input image data in the frequency domain;

[0038] A Conv1x1 convolution layer is inserted between the Fourier transform branch and the main feature branch; the weight W1 corresponding to the Conv1x1 layer of the main feature branch is initialized to 1, and the weight W2 corresponding to the Conv1x1 layer of the Fourier transform branch is initialized to 0;

[0039] Initially, the weight of the Conv1x1 layer corresponding to the Fourier transform branch is 0, and during the training process, the gradients of W1 and W2 are calculated using the back propagation algorithm, and W1 and W2 are updated using the optimizer; the weight of the Conv1x1 layer corresponding to the Fourier transform branch gradually changes, so that the Fourier transform branch has an impact on the output;

[0040] After the Conv1x1 layer, the features of the backbone feature branch and the Fourier transform branch are fused to obtain a fused feature tensor, and the fused feature tensor is sent to a decoder, and the decoder applies a series of convolution layers and a prediction head to the fused feature tensor; the prediction head outputs a target detection result, and the target detection result includes a target bounding box and a class probability of a target object;

[0041] Comprehensive information of each target object is extracted from the target detection result, and the comprehensive information includes position information, speed information, and a threat level; the position information is a coordinate position and a target bounding box size of the target object; a motion speed is calculated by tracking a displacement of the target object in the continuous frame model input image data, and the motion speed is the speed information;

[0042] According to the class probability and the speed information of the target object, a potential threat degree of the target object to the unmanned aerial vehicle is evaluated, and a corresponding threat level value is given, that is, the threat level.

[0043] Further, the manner of solving the optimal attack and defense strategy by using the perception integrated attack algorithm includes:

[0044] A series of attack and defense behaviors that the unmanned aerial vehicle can take are defined to constitute a decision space; each attack and defense behavior is parameterized by one or several continuous variables;

[0045] According to the type of the target object, a kinematic model of the target object is defined as a constant speed model, a constant acceleration model, or a maneuvering target model; parameters of the kinematic model of the target object include a position and a speed, and initial values of the parameters are obtained from the target detection result; a Kalman filtering algorithm is used to update the parameters of the kinematic model online; a kinematic model of the unmanned aerial vehicle is defined as a 6-degree-of-freedom rigid body motion model;

[0046] State data of the unmanned aerial vehicle itself and the comprehensive information of the target object are combined into a state vector; control inputs of the unmanned aerial vehicle are defined to include a thrust and a control surface deflection; the control inputs of the unmanned aerial vehicle are combined into a control vector;

[0047] Then, a state transition equation is defined as: x(T+1)=f(x(T),u(T)), wherein x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, and u(T) represents the control vector of the unmanned aerial vehicle at time T; f() is a nonlinear function obtained from the kinematic models of the target object and the unmanned aerial vehicle;

[0048] The state transition equation is used as a constraint to define a reward function; a formula of the reward function is:

[0049] R(s,a)=v1×Rki(s,a)+v2×Rsu(s,a)+v3×Rdi(s,a)+v4×Rti(s,a);

[0050] wherein, v1, v2, v3 and v4 are weight balance coefficients of corresponding items; R(s, a) is the benefit of the UAV taking the control vector a under the state vector s; Rki(s, a) is the elimination threat function; Rsu(s, a) is the self survival function; Rdi(s, a) is the distance penalty function; Rti(s, a) is the time penalty function;

[0051] All state vectors constitute a state space, and the state space, decision space, state transition equation and reward function are constructed into a reinforcement learning problem model; and the optimal attack and defense strategy is solved based on the reinforcement learning problem model.

[0052] Further, the elimination threat function Rki(s, a) = wki x th(s, a); wherein, wki is the weight coefficient of eliminating threat; th(s, a) represents the threat level of the target object to the UAV taking the control vector a under the state vector s;

[0053] Self survival function wherein, wda is the weight coefficient of survival penalty; dam(s, a) is the damage value of the UAV taking the control vector a under the state vector s; hea(s) is the health state of the UAV under the state vector s; τ1 and τ2 are adjustable parameters;

[0054] Distance penalty function wherein, dis(s, a) is the distance between the UAV taking the control vector a under the state vector s and the target object; wea is the effective range distance of the weapon of the UAV; σ1 is a rate adjustment parameter, and δ1 is a degree adjustment parameter;

[0055] Time penalty function wherein, tmax is the maximum time step length that the UAV can execute, i.e. the time limit of the task; t1 is the time step length that the UAV has executed; is a time penalty exponential parameter.

[0056] Further, the manner of solving the optimal attack and defense strategy based on the reinforcement learning problem model comprises:

[0057] A decision strategy network is constructed based on the reinforcement learning problem model, and the decision strategy network is composed of an Actor network and a Critic network, both of which adopt a deep neural network structure; the input of the Actor network is a state vector, and the output is a control vector; the input of the Critic network is a state vector and a control vector, and the output is a benefit; the network structure of the Actor network and the Critic network is a full connection network or a convolution network; a deep reinforcement learning algorithm is adopted to optimize the decision strategy network, so as to obtain a trained decision strategy network;

[0058] Acquire the current state data of the unmanned aerial vehicle, for each target object, the current state data of the unmanned aerial vehicle and the comprehensive information of the target object constitute a real-time state vector, take the real-time state vector as input, output a series of continuous behavior vectors through the trained decision strategy network, decode to obtain the attack and defense behavior and its parameters as the preliminary attack and defense strategy;

[0059] The continuous behavior vector constitutes a discrete behavior sequence;Convolution is performed on the discrete behavior sequence and the Gaussian kernel to obtain a smooth behavior sequence;The smooth behavior sequence is decoded to obtain the optimal attack and defense strategy.

[0060] The technical effects and advantages of the unmanned aerial vehicle intelligent defense and attack control system of the present application are as follows:

[0061] The present application can obtain environmental situation information in all directions and multiple dimensions, not only improves the understanding ability of complex environment, but also realizes effective identification of hidden targets, greatly reduces the probability of missed detection and false detection;In the target evaluation aspect, through comprehensive analysis of the motion characteristics, threat level and attack intention of the target, a more scientific and accurate threat evaluation system is established, which provides a reliable basis for subsequent decision-making;In the attack and defense decision-making aspect, the limitation of traditional fixed strategy is broken through, the optimal attack and defense strategy can be dynamically generated according to the real-time change of environmental situation, the survival ability and strike efficiency of unmanned aerial vehicle in the hostile environment are significantly improved;It can independently weigh attack opportunity, threat avoidance and energy consumption and other key factors to ensure efficient completion of the task;In addition, it also has the ability of autonomous learning and experience accumulation, which can continuously optimize and improve the decision-making model and continuously improve the combat effectiveness;In practical application, it significantly reduces the burden of operators and improves the cooperative combat capability of unmanned aerial vehicles, which lays an important foundation for building intelligent and networked unmanned aerial vehicle system. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a schematic diagram of an unmanned aerial vehicle intelligent defense and attack control system of the present application;

[0063] Figure 2 It is a schematic diagram of a security authentication system based on a secure computer. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] Embodiment 1

[0066] Please refer to Figure 1 The unmanned aerial vehicle intelligent defense and attack control system described in this embodiment comprises:

[0067] The data acquisition module is configured to acquire environment perception data of the unmanned aerial vehicle, wherein the environment perception data comprises state data of the unmanned aerial vehicle itself and image data of the surrounding environment.

[0068] The data processing module is configured to pre-process the image data of the surrounding environment to obtain model input image data.

[0069] The target detection module is configured to use a variant Poolformer algorithm to perform target detection on the model input image data to obtain comprehensive information of the target object.

[0070] The strategy fitting module is configured to use a perception-integrated attack algorithm to solve an optimal attack and defense strategy according to the state data of the unmanned aerial vehicle itself and the comprehensive information of the target object, and send the optimal attack and defense strategy to the unmanned aerial vehicle execution terminal to control the unmanned aerial vehicle to execute the optimal attack and defense strategy. The modules are connected through wired and / or wireless means to realize data transmission between the modules.

[0071] The state data of the unmanned aerial vehicle itself comprises position data, motion state, energy state, weapon state, health management state and environment state. The position data is the three-dimensional spatial position coordinates (longitude, latitude and height) of the unmanned aerial vehicle, the motion state comprises the velocity vector (horizontal velocity and vertical velocity), the attitude angle (yaw angle, pitch angle and roll angle) and the acceleration vector of the unmanned aerial vehicle, and the energy state comprises the remaining amount of fuel / battery, the power output of the engine / motor and the energy consumption rate.

[0072] The weapon state is the remaining amount and state of various weapons. For each weapon, a state index is defined to represent its working state. For example, a discrete value or a continuous value is used for quantification, 0 represents complete failure, 1 represents slight failure, and 2 represents normal working; or a real number between 0 and 1, 0 represents complete failure, 1 represents perfect, and the intermediate value represents different degrees of failure or aging.

[0073] The health management state is the health state (also quantified by discrete values or continuous values) of various subsystems (airframe, avionics, control, etc.), and the environment state comprises temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information. The obstacle information comprises the three-dimensional spatial position coordinates of the obstacle, the size of the obstacle and the motion trajectory of the dynamic obstacle. These state data are acquired and integrated by various sensors of the unmanned aerial vehicle itself.

[0074] The motion trajectory of the dynamic obstacle can be acquired in the following ways:

[0075] The dynamic obstacle is regarded as a moving particle, and the unmanned aerial vehicle and other obstacles generate virtual gravitational field and repulsive field for the particle, and the particle moves under the action of the resultant force of the gravitational field and the repulsive field; the unmanned aerial vehicle is defined as a gravitational source; the size of the gravitational field is inversely proportional to the distance between the unmanned aerial vehicle and the dynamic obstacle; the direction of the gravitational field points to the position of the unmanned aerial vehicle; other obstacles (fixed or dynamic) are defined as repulsive sources, and the size of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle, and the direction of the repulsive field points away from the repulsive source; for each time step, the resultant force acting on the particle is calculated, which is equal to the sum of the gravitational field plus the sum of the repulsive field; the resultant force is substituted into the Newton motion equation to obtain the acceleration a, and the numerical integration is performed to obtain the velocity and displacement of the particle, and the change of the velocity and displacement of the particle in a preset time range is used to construct a three-dimensional position sequence of the unmanned aerial vehicle, which is used to represent the motion trajectory of the dynamic obstacle.

[0076] The pre-processing manner of the image data of the surrounding environment comprises:

[0077] The original image data is subjected to attention fusion and equalization processing to obtain preliminary input image data; the preliminary input image data is subjected to image cropping to obtain a key area image; the pixel values in the key area image are scaled to the range of 0-1 or -1 to 1 (using the minimum-maximum normalization or mean-variance normalization method), to obtain model input image data.

[0078] The attention fusion and equalization processing manner comprises:

[0079] An attention structure based on a deep convolutional neural network is defined, which is used to learn the importance weight of different regions in the image; the input of the attention structure is the original image data, and the output is an attention weight map with the same size as the original image data; each pixel value in the attention weight map represents the importance degree of the corresponding position, and the greater the value, the more important it is.

[0080] An image dataset comprising images under various scenes and conditions is collected, and the images contain target objects of interest (such as people, vehicles, aircraft, etc.); the images in the image dataset are manually annotated, and the positions and categories of the target objects are marked on the images using a bounding box; the attention structure is integrated into a classification network or a target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure outputs an attention weight map, which is multiplied by the original image data as a mask to obtain a feature map; the original image data is subjected to evolutionary histogram equalization to obtain a globally enhanced image; the feature map and the globally enhanced image are multiplied to obtain preliminary input image data.

[0081] The training process of the comprehensive network model comprises:

[0082] The loss function is defined, the comprehensive network structure is trained by using the image data set, the gradient of the attention structure is calculated according to the value of the loss function of the comprehensive network model, and the parameters of the attention structure are updated, so that the attention structure automatically focuses on the region where the target object is located after training.

[0083] The evolutionary histogram equalization method comprises the following steps:

[0084] The original image data is divided into a plurality of overlapping small block regions, usually in the form of a sliding window; the size of each small block region is pre-set, such as 64x64 or 128x128 pixels; a certain overlapping region is set between adjacent small blocks to ensure the smoothness of the boundary processing.

[0085] Each small block region is subjected to histogram equalization processing to obtain a reconstructed small block region; specifically, each small block region is converted into a corresponding gray image, and for each gray image, the number of pixels of each gray level (0-255) is counted and a gray histogram is constructed; the gray histogram reflects the distribution of pixels of different gray levels in the image.

[0086] Based on the gray histogram, the cumulative distribution function value CDF(i) of each gray level i is calculated;

[0087] Wherein, n(0) represents the number of pixels of gray level 0, n(1) represents the number of pixels of gray level 1, and so on, n(i) is the number of pixels of gray level i; N is the total number of pixels in the gray image; for each gray level i, its new mapping gray level j=(L-1)xCDF(i) is calculated according to its cumulative distribution function value CDF(i); wherein L is the total number of gray levels, usually 256; the gray level i of each pixel in the small block region is mapped to the new mapping gray level j; a remapped small block region is obtained; denoted as a reconstructed small block region.

[0088] For each reconstructed small block region, the normalized coordinates of the center point in the original image data are calculated; denoted as the center point coordinates; according to the center point coordinates, the Euclidean distance d of the reconstructed small block region to the center of the original image data is calculated; for each reconstructed small block region, the number of pixels overlapping with other reconstructed small block regions is counted; for each reconstructed small block region, its gray co-occurrence matrix (which reflects the spatial distribution characteristics of pixels with different gray values in the image) is calculated; based on the comprehensive features in the gray co-occurrence matrix; the comprehensive features include contrast DA, energy DH (which is the square of the mean value, reflecting the uniformity of the gray distribution in the image), homogeneity DF and entropy DS.

[0089] Contrast DA = ∑ I ∑ J(I-J) γ x P(I, J) δ ; wherein, I and J are the row and column indexes of the gray co-occurrence matrix, corresponding to different gray values; P(I, J) is an element in the gray co-occurrence matrix, representing the frequency of pixel pairs between gray values I and J; γ is an exponential factor controlling the degree of influence of gray difference on contrast; δ is an exponential factor controlling the degree of influence of pixel pair frequency on contrast.

[0090] Homogeneity wherein, a and b are parameters controlling the shape of the function; the entropy DS = ∑ I ∑ J P(I, J) x log(P(I, J));

[0091] The integrated features are normalized and weighted to obtain an integrated texture complexity index corresponding to the reconstructed small block region; for each reconstructed small block region, an edge detection operator (such as Sobel, Canny, etc.) is applied to extract the number of edge pixels and the cumulative gradient amplitude, and weighted fusion is performed to obtain an edge intensity index of the reconstructed small block region; the integrated texture complexity index and the edge intensity index are fused into an integrated feature index C; specifically, a simple weighted sum or a nonlinear fusion function such as multilayer perception (MLP) is used.

[0092] Define the position weight of the reconstructed small block region h wherein, σ is a parameter controlling the decay rate of the weight; ω1 is an influence parameter adjusting the degree of influence of the integrated feature index C on the weight; ω2 is an exponential factor controlling the distance, used to control the degree of influence of the distance on the weight.

[0093] Define the overlap weight of the reconstructed small block region h wherein, a1 is a scaling coefficient, used to adjust the magnitude of the overlap weight; No is the number of overlapping pixels of the reconstructed small block region h and other reconstructed small block regions, Nb is the total number of pixels of the reconstructed small block region h; a2 is an overlap exponential factor, used to control the degree of influence of the number of overlapping pixels on the weight; γ2 is a control parameter, used to adjust the degree of influence of the integrated feature index on the weight.

[0094] Multiply the position weight and the overlap weight of the reconstructed small block region h to obtain an integrated weight coefficient corresponding to the reconstructed small block region; for the overlapping region in the reconstructed small block region, the pixel values therein are weighted and averaged according to the integrated weight coefficient of the corresponding reconstructed small block region to obtain the fused pixel values; for the non-overlapping region in the reconstructed small block region, the corresponding pixel values are directly used to obtain a complete globally enhanced image.

[0095] The image cropping method for the preliminary input image data includes:

[0096] Based on the dynamic obstacle, the view volume of each time step in the time range is determined according to the motion trajectory, for each time step t on the motion trajectory, according to the position and attitude angle of the unmanned aerial vehicle at the time step t, the corresponding view range of the unmanned aerial vehicle is calculated, and a series of time-varying three-dimensional view volumes V(t) are obtained; specifically, the camera coordinate system is established with the position of the unmanned aerial vehicle as the origin, the directions of the three coordinate axes of the camera coordinate system are determined according to the attitude angle of the unmanned aerial vehicle; according to the field of view angle (calculated by the internal parameters of the camera), the coordinates of the eight vertices of the camera frustum in the camera coordinate system are calculated; the vertex coordinates in the camera coordinate system are converted to the world coordinate system by using the position and attitude angle of the unmanned aerial vehicle, and the coordinates of the eight vertices in the world coordinate system are obtained, and a three-dimensional frustum or pyramid is constructed by using the coordinates of the eight vertices in the world coordinate system; the three-dimensional frustum or pyramid is the view range of the unmanned aerial vehicle at the time step t; and then a series of time-varying three-dimensional view volumes V(t) are obtained.

[0097] For each time step t, the corresponding three-dimensional view volume V(t) is projected onto a two-dimensional image plane to obtain a series of two-dimensional polygonal regions P(t), representing the image regions that can be observed by the image acquisition equipment of the unmanned aerial vehicle at each time step; initialize a mask image M with the same size as the global enhanced image, and all pixel values in the mask image M are initially 0; for each time step t, the pixel values in the corresponding two-dimensional polygonal region P(t) are set to 1 in the mask image M, and the final mask image M contains the image regions that can be observed by the unmanned aerial vehicle in the entire time range; the global enhanced image is multiplied pixel by pixel with the mask image M to obtain the key region image.

[0098] The manner of target detection on the model input image data includes:

[0099] A variant Poolformer model is constructed, and the basic structure of the variant Poolformer model includes an encoder and a decoder; the encoder is defined to include a main feature branch and a parallel Fourier transform branch; the main feature branch extracts features from the model input image data using a multi-head self-attention mechanism, and the decoder aggregates global context information using a pooling attention mechanism.

[0100] The Fourier transform branch performs two-dimensional discrete Fourier transform (2D-DFT) on the model input image data, converts it from the spatial domain to the frequency domain, and performs low-pass filtering and high-pass filtering on the model input image data in the frequency domain; a Conv1x1 convolution layer is inserted between the Fourier transform branch and the main feature branch; the weight W1 corresponding to the Conv1x1 layer of the main feature branch is initialized to 1, and the weight W2 corresponding to the Conv1x1 layer of the Fourier transform branch is initialized to 0.

[0101] Initially, the Fourier transform branch does not contribute to the output because the weights of the Conv1x1 layer corresponding to the Fourier transform branch are zeros; during the training process, the gradients of W1 and W2 are calculated using the backpropagation algorithm, and W1 and W2 are updated using an optimizer (such as SGD, Adam, etc.); the weights of the Conv1x1 layer corresponding to the Fourier transform branch gradually change, so that the Fourier transform branch has an impact on the output.

[0102] After passing through the Conv1x1 layer, the features of the backbone feature branch and the Fourier transform branch are fused to obtain a fused feature tensor, and the fused feature tensor is input into the decoder; the decoder applies a series of convolutional layers and prediction heads to the fused feature tensor; the prediction head outputs the target detection result, which includes the target bounding box and class probability of the target object.

[0103] The training process is as follows:

[0104] Collect and prepare the training dataset, which contains images under various scenes and conditions, and label the target objects (people, vehicles, aircraft, etc.) in the images to obtain the bounding boxes and class labels of the target objects; public datasets such as COCO, VOC, etc. can be used, or custom datasets can be used; divide the training dataset into a training set and a validation set; initialize other trainable parameters (such as convolution kernel weights, batch normalization parameters, etc.) of the Poolformer model variant, which can use Xavier or Kaiming initialization; define the loss function, which includes two parts: classification loss (such as cross-entropy loss) and regression loss (such as smooth L1 loss), which are used to supervise the object detection task; use the mini-batch stochastic gradient descent method for training; in each training iteration, a batch of image data and corresponding labels are sampled from the training dataset; input the image data into the Poolformer model variant to obtain the predicted bounding box and class probability; calculate the value of the loss function; use the backpropagation algorithm to calculate the gradients of the trainable parameters (including the Conv1x1 layer weights); use the optimizer (such as SGD, Adam, etc.) to update the trainable parameters according to the gradients; periodically evaluate the model performance on the validation set, such as the mAP (mean average precision) and other indicators; adjust the hyperparameters, such as the learning rate and the regularization strength, to obtain better performance according to the model performance on the validation set; save the Poolformer model variant with the best model performance.

[0105] Comprehensive information of each target object is extracted from the target detection result, and the comprehensive information includes position information, speed information, and threat level; the position information is the coordinate position (the center coordinate of the target bounding box) and the size of the target bounding box of the target object; the motion speed is calculated by tracking the displacement of the target object in the continuous frame model input image data, that is, the speed information; according to the category probability and the speed information of the target object, the potential threat degree of the target object to the unmanned aerial vehicle is evaluated, and a corresponding threat level value is given, that is, the threat level; specifically, for different categories of target objects, the potential threat degree value of the target objects to the unmanned aerial vehicle is defined in advance; for example, the threat degree of a pedestrian is low, and the threat degree of equipment is high; the threat of a stationary target object is small, and the threat of a high-speed moving target object is large; according to the speed information of the target object, the potential threat degree value is adjusted; and the threat level is obtained.

[0106] The way of solving the optimal attack and defense strategy by using the perception integrated attack algorithm includes:

[0107] A series of attack and defense behaviors that the unmanned aerial vehicle can take are defined to constitute a decision space; for example, acceleration, deceleration, change of heading, launching of weapons, evasion, etc.; each attack and defense behavior is parameterized by one or several continuous variables, such as acceleration, heading angle, etc.; according to the type of the target object, the kinematic model of the target object is defined as a constant speed model, a constant acceleration model, or a maneuvering target model; the parameters of the kinematic model of the target object include position and speed, and the initial values of the parameters are obtained from the target detection result; the Kalman filtering algorithm is used to update the parameters of the kinematic model online; and the kinematic model of the unmanned aerial vehicle is defined as a 6-DOF rigid body motion model.

[0108] The state data of the unmanned aerial vehicle itself and the comprehensive information of the target object are combined into a state vector; the control input of the unmanned aerial vehicle is defined to include thrust and control surface deflection; the control input of the unmanned aerial vehicle is combined into a control vector; and the state transition equation is defined as: x(T+1) = f(x(T), u(T)), wherein x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, u(T) represents the control vector (attack and defense behavior) of the unmanned aerial vehicle at time T; f() is a nonlinear function obtained from the kinematic model of the target object and the unmanned aerial vehicle; specifically, according to the 6-DOF rigid body motion model, the differential equations between the state data of the unmanned aerial vehicle itself and the rate of change thereof are listed, for example, the rate of change of the position is equal to the corresponding speed; the rate of change of the speed is equal to the function of the corresponding acceleration and the angular velocity; the rate of change of the attitude angle is equal to the nonlinear function of the corresponding angular velocity; the thrust determines the acceleration, and the control surface deflection determines the angular acceleration; the control input is substituted into the function of the corresponding acceleration and the angular velocity, that is, the control input is coupled with the state vector; the differential equations of the position and the speed of the target object are established according to the kinematic model of the target object; for example, for the constant speed model, the rate of change of the position of the target is equal to the speed of the target object; the differential equations are discretized to obtain the difference form, and all the discretized differential equations are combined, that is, f().

[0109] The state transition equation is used as a constraint to define a reward function, and the benefits of the control vector (attack and defense behavior) are quantified; the formula of the reward function is:

[0110] R(s, a) = v1 x Rki(s, a) + v2 x Rsu(s, a) + v3 x Rdi(s, a) + v4 x Rti(s, a);

[0111] Wherein v1, v2, v3 and v4 are weight balance coefficients of the corresponding items; R(s, a) is the benefit of the control vector a of the unmanned aerial vehicle under the state vector s; Rki(s, a) is a threat elimination function; Rsu(s, a) is a self survival function; Rdi(s, a) is a distance penalty function; Rti(s, a) is a time penalty function.

[0112] The threat elimination function Rki(s, a) = wki x th(s, a); wherein wki is a weight coefficient of threat elimination; th(s, a) represents the threat level of the target object to the unmanned aerial vehicle when the control vector a is taken under the state vector s; the threat level is a discrete value (such as 1-5 levels) or a continuous value (such as between 0 and 1).

[0113] The self survival function wherein wda is a weight coefficient of survival penalty, used to adjust the penalty size; dam(s, a) is the damage value suffered by the UAV when taking the control vector a under the state vector s; hea(s) is the health status of the UAV under the state vector s, which is a comprehensive value of the remaining fuel and structural integrity; τ1 and τ2 are adjustable parameters.

[0114] Distance penalty function wherein dis(s, a) is the distance between the UAV and the target object when taking the control vector a under the state vector s; wea is the effective firing range distance of the weapon of the UAV; σ1 is a rate adjustment parameter, which controls the increasing rate of the distance; δ1 is a degree adjustment parameter, which controls the increasing degree of the distance.

[0115] Time penalty function wherein tmax is the maximum time step length that can be executed by the UAV, i.e., the time limit of the task; t1 is the time step length that has been executed by the UAV; is a time penalty index parameter; as the executed time approaches the time limit, the time penalty will rapidly increase, controlling the increasing rate.

[0116] All state vectors constitute a state space, and the state space, decision space, state transition equation and reward function are constructed into a reinforcement learning problem model; a decision strategy network is constructed based on the reinforcement learning problem model, the decision strategy network is composed of an Actor network and a Critic network, and a deep neural network structure is usually used; the input of the Actor network is the state vector, and the output is the control vector; the input of the Critic network is the state vector and the control vector, and the output is the return; the network structure of the Actor network and the Critic network is a fully connected network, a convolutional network or other types, which needs to be designed according to the representation form of the state and behavior; a deep reinforcement learning algorithm (such as deep Q network (DQN), policy gradient, etc.) is used to optimize the decision strategy network, and a trained decision strategy network is obtained.

[0117] The current state data of the UAV is obtained, for each target object, the current state data of the UAV and the comprehensive information of the target object are combined to form a real-time state vector, the real-time state vector is taken as the input, a series of continuous behavior vectors are output through the trained decision strategy network, and the attack and defense behaviors and their parameters are decoded to obtain the preliminary attack and defense strategy. Specifically, the real-time state vector is input into the trained Actor network, the Actor network will output the corresponding optimal behavior according to the real-time state vector, and the optimal behavior is decoded to obtain the attack and defense behaviors and their parameters.

[0118] The continuous behavior vector constitutes a discrete behavior sequence, which can have a sharp change, for example, the change of acceleration or heading angle between adjacent time steps is too large, which can cause the UAV to be unable to actually execute or execute inefficiently; the discrete behavior sequence is convolved with a Gaussian kernel to obtain a smoothed behavior sequence; the smoothed behavior sequence is decoded to obtain an optimal attack and defense strategy.

[0119] It should be noted that the output of the Actor network is usually a continuous behavior vector, which needs to be decoded into specific attack and defense behaviors and their parameters, for example, different components of the behavior vector are decoded into acceleration, heading angle rate of change, weapon launch parameters, etc.

[0120] In this embodiment, the environmental situation information can be obtained in all directions and multiple dimensions, which not only improves the understanding ability of complex environment, but also realizes effective identification of hidden targets, greatly reduces the probability of missed detection and false detection; in target evaluation, a more scientific and accurate threat evaluation system is established by comprehensively analyzing the motion characteristics, threat level and attack intention of the target, which provides a reliable basis for subsequent decision-making; in attack and defense decision-making, the limitations of traditional fixed strategies are broken through, and the optimal attack and defense strategy can be dynamically generated according to the real-time changes of the environmental situation, which significantly improves the survival ability and strike efficiency of the UAV in the hostile environment; it can autonomously weigh the attack opportunity, threat avoidance and energy consumption and other key factors to ensure efficient completion of the task; in addition, it also has the ability of autonomous learning and experience accumulation, which can continuously optimize and improve the decision-making model and continuously improve the combat effectiveness; in practical application, it significantly reduces the burden of operators and improves the cooperative combat capability of UAVs, laying an important foundation for building intelligent and networked UAV systems.

[0121] Embodiment 2

[0122] Please refer to Figure 2 The embodiment not described in detail is described in the description of embodiment 1, and the UAV intelligent defense and attack control method comprises:

[0123] Step 1, obtaining the environmental perception data of the UAV, the environmental perception data comprising the state data of the UAV itself and the image data of the surrounding environment;

[0124] Step 2, preprocessing the image data of the surrounding environment to obtain model input image data;

[0125] Step 3, using the variant Poolformer algorithm to detect the target object in the model input image data to obtain the comprehensive information of the target object;

[0126] Step 4, according to the state data of the unmanned aerial vehicle itself and the comprehensive information of the target object, an optimal attack and defense strategy is solved by using a perception integrated attack algorithm; the optimal attack and defense strategy is sent to the unmanned aerial vehicle execution terminal to control the unmanned aerial vehicle to execute the optimal attack and defense strategy.

[0127] Embodiment 3

[0128] The embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the operation mode of the unmanned aerial vehicle intelligent defense and attack control method provided by the computer program.

[0129] Since the electronic device introduced in the embodiment is the electronic device used to implement the unmanned aerial vehicle intelligent defense and attack control method in the embodiment, the specific implementation of the electronic device and various changes thereof can be understood by those skilled in the art based on the unmanned aerial vehicle intelligent defense and attack control method introduced in the embodiment, so the implementation of the electronic device in the method in the embodiment will not be introduced in detail. As long as the electronic device used to implement the unmanned aerial vehicle intelligent defense and attack control method in the embodiment is implemented by those skilled in the art, it belongs to the scope of protection of the present application.

[0130] The above formulas are dimensionless values, and the formulas are obtained by collecting a large amount of data to simulate the most recent real situation, and the preset parameters and threshold values in the formulas are set by those skilled in the art according to the actual situation.

[0131] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application is within the protection scope of the present application. It should be noted that, for ordinary technical users in the technical field, some improvements and decorations without departing from the principle of the present application are also considered as the protection scope of the present application.

Claims

1. An unmanned aerial vehicle intelligent defense and attack control system, characterized in that, The method comprises the following steps: a data acquisition module is configured to acquire environmental perception data of the UAV, wherein the environmental perception data comprises state data of the UAV itself and image data of the surrounding environment; a data processing module is configured to preprocess the image data of the surrounding environment to obtain model input image data; a target detection module is configured to use a variant Poolformer algorithm to perform target detection on the model input image data to obtain comprehensive information of the target object; the target detection on the model input image data comprises the following steps: a variant Poolformer model is constructed, wherein the basic structure of the variant Poolformer model comprises an encoder and a decoder; the encoder is defined to comprise a main feature branch and a parallel Fourier transform branch; the Fourier transform branch performs two-dimensional discrete Fourier transform on the model input image data, converts the model input image data from a spatial domain to a frequency domain, and performs low-pass filtering and high-pass filtering on the model input image data in the frequency domain; a Conv1x1 convolution layer is inserted between the Fourier transform branch and the main feature branch; the weight W1 of the Conv1x1 layer corresponding to the main feature branch is initialized as 1, and the weight W2 of the Conv1x1 layer corresponding to the Fourier transform branch is initialized as 0; initially, the weight of the Conv1x1 layer corresponding to the Fourier transform branch is 0; during the training process, the gradients of W1 and W2 are calculated using a back propagation algorithm, and W1 and W2 are updated using an optimizer; the weight of the Conv1x1 layer corresponding to the Fourier transform branch gradually changes, so that the Fourier transform branch has an impact on the output; after passing through the Conv1x1 layer, the features of the main feature branch and the Fourier transform branch are fused to obtain a fused feature tensor; the fused feature tensor is input into the decoder; the decoder applies a series of convolution layers and a prediction head to the fused feature tensor; the prediction head outputs a target detection result, which comprises a target bounding box of the target object and a class probability; comprehensive information of each target object is extracted from the target detection result, wherein the comprehensive information comprises position information, speed information and a threat level; the position information is the coordinate position and the size of the target bounding box of the target object; the speed information is calculated by tracking the displacement of the target object in consecutive frame model input image data; a threat level value is assigned to the target object according to the class probability and the speed information, which is the threat level of the target object to the UAV; a strategy fitting module is configured to use a perception-integrated attack algorithm to solve an optimal attack and defense strategy according to the state data of the UAV itself and the comprehensive information of the target object; the optimal attack and defense strategy is sent to a UAV execution terminal to control the UAV to execute the optimal attack and defense strategy; the modules are connected through wired and / or wireless means. 2.The unmanned aerial vehicle intelligent defense and attack control system of claim 1, wherein, the state data of the UAV itself comprises position data, motion state, energy state, weapon state, health management state and environment state; the environment state comprises temperature, air pressure, wind speed, visibility, electromagnetic interference intensity and obstacle information. The obstacle information includes three-dimensional spatial position coordinates of the obstacle, obstacle size, and a motion trajectory of the dynamic obstacle. 3.The unmanned aerial vehicle intelligent defense and attack control system of claim 2, wherein, The motion trajectory of the dynamic obstacle is obtained in the following manner: The dynamic obstacle is regarded as a moving particle, and the unmanned aerial vehicle and other obstacles generate virtual gravitational fields and repulsive fields for the particle, and the particle moves under the action of the resultant force of the gravitational fields and the repulsive fields; The unmanned aerial vehicle is defined as a gravitational source, the size of the gravitational field is inversely proportional to the distance between the unmanned aerial vehicle and the dynamic obstacle, and the direction of the gravitational field points to the position of the unmanned aerial vehicle; Other obstacles are defined as repulsive sources, the size of the repulsive field is inversely proportional to the distance between the repulsive source and the corresponding particle, and the direction of the repulsive field points away from the repulsive source; For each time step, the resultant force acting on the particle is calculated, the resultant force is equal to the sum of the gravitational fields plus the sum of the repulsive fields, the resultant force is substituted into the Newtonian motion equation, the acceleration a is solved, and the particle velocity and displacement are obtained by numerical integration, a preset time range is set, and the change of the particle velocity and displacement in the time range is used to construct a three-dimensional position sequence of the unmanned aerial vehicle, that is, the motion trajectory of the dynamic obstacle.

4. The unmanned aerial vehicle intelligent defense and attack control system of claim 3, wherein, The image data of the surrounding environment is preprocessed in the following manner: The original image data is subjected to attention fusion and equalization processing to obtain preliminary input image data, the preliminary input image data is subjected to image cropping to obtain a key area image, and the pixel values in the key area image are scaled to the range of 0-1 or -1 to 1 to obtain model input image data; The attention fusion and equalization processing is performed in the following manner: An attention structure based on a deep convolutional neural network is defined, the input of the attention structure is the original image data, and the output is an attention weight map with the same size as the original image data; each pixel value in the attention weight map represents the importance degree of the corresponding position; An image dataset containing images under various scenes and conditions is collected, the images contain target objects; the images in the image dataset are manually annotated, and the positions and categories of the target objects are marked on the images using a bounding box; The attention structure is integrated into a classification network or a target detection network to obtain a comprehensive network model; in the feature extraction stage of the classification network or the target detection network, the attention structure multiplies the output attention weight map with the original image data as a mask to obtain a feature map; The original image data is subjected to evolutionary histogram equalization to obtain a globally enhanced image; the feature map and the globally enhanced image are multiplied to obtain preliminary input image data.

5. The unmanned aerial vehicle intelligent defense and attack control system of claim 4, wherein, The evolutionary histogram equalization is performed in the following manner: The original image data is divided into a plurality of overlapping small block regions, the size of each small block region is pre-set, and a certain overlap region is provided between adjacent small blocks; each small block region is subjected to histogram equalization processing to obtain a reconstructed small block region; For each reconstructed patch region, calculate the normalized coordinates of the center point in the original image data; denoted as the center point coordinates; according to the center point coordinates, calculate the Euclidean distance d of the reconstructed patch region to the center of the original image data; for each reconstructed patch region, count the number of pixels overlapping with other reconstructed patch regions; for each reconstructed patch region, calculate its gray co-occurrence matrix; calculate the comprehensive feature based on the gray co-occurrence matrix; the comprehensive feature includes contrast DA, energy DH, homogeneity DF and entropy DS; Contrast DA =∑ I ∑ J (I-J) γ ×P(I,J) δ ; where I and J are the row and column indices of the gray co-occurrence matrix; P(I,J) is an element in the gray co-occurrence matrix, representing the frequency of occurrence of pixel pairs between gray values I and J; γ is an exponential factor controlling the degree of influence of the gray difference on the contrast; δ is an exponential factor controlling the degree of influence of the frequency of occurrence of pixel pairs on the contrast. Homogeneity where a and b are parameters that control the shape of the function; the entropy DS = ∑ I ∑ J P(I,J) x log(P(I,J)); The comprehensive features are normalized and weighted fused to obtain the comprehensive texture complexity index corresponding to the reconstructed patch region; for each reconstructed patch region, the edge detection operator is applied to extract the number of edge pixels and the cumulative gradient amplitude, and weighted fusion is performed to obtain the edge intensity index of the reconstructed patch region; the comprehensive texture complexity index and the edge intensity index are fused into a comprehensive feature index C; Defining the position weight of the reconfigured patch region h where σ is a parameter that controls the speed of weight decay; ω1 is an influence parameter; and ω2 is an exponential factor that controls the distance, used to control the degree of influence of distance on the weight. Defining the overlap weight of the reconstructed tile region h wherein a1 is a scaling factor; No is the number of pixels of the reconstructed tile region h overlapping with other reconstructed tile regions, Nb is the total number of pixels of the reconstructed tile region h; a2 is an overlap exponent factor; γ2 is a control parameter; multiplying the position weight of the reconstructed tile region h and the overlap weight to obtain the comprehensive weight coefficient of the corresponding reconstructed tile region; For the overlapping region in the reconstructed patch region, the pixel values therein are weighted and averaged according to the comprehensive weight coefficient of the corresponding reconstructed patch region to obtain the fused pixel value; for the non-overlapping region in the reconstructed patch region, the corresponding pixel value is directly used to obtain a complete globally enhanced image.

6. The unmanned aerial vehicle intelligent defense and attack control system of claim 5, wherein, The image cropping of the preliminary input image data includes: Based on the motion trajectory of the dynamic obstacle, the field of view volume of each time step in the time range is determined, for each time step t on the motion trajectory, the camera coordinate system is established with the position of the unmanned aerial vehicle as the origin, the directions of the three coordinate axes of the camera coordinate system are determined according to the attitude angle of the unmanned aerial vehicle; according to the field of view angle, the coordinates of the eight vertices of the camera frustum in the camera coordinate system are calculated; the vertex coordinates in the camera coordinate system are converted to the world coordinate system by using the position and attitude angle of the unmanned aerial vehicle, to obtain the coordinates of the eight vertices in the world coordinate system, and a three-dimensional frustum or pyramid is constructed by using the coordinates of the eight vertices in the world coordinate system; the three-dimensional frustum or pyramid is the field of view range of the unmanned aerial vehicle at time step t; then a series of time-varying three-dimensional field of view volumes V(t) are obtained; For each time step t, the corresponding three-dimensional field of view volume V(t) is projected onto a two-dimensional image plane to obtain a series of two-dimensional polygonal regions P(t); Initialize a mask image M with the same size as the globally enhanced image, and all pixel values in the mask image M are initially 0; for each time step t, the pixel values in the corresponding two-dimensional polygonal region P(t) are set to 1 in the mask image M; the globally enhanced image and the mask image M are multiplied pixel by pixel to obtain a key region image.

7. The unmanned aerial vehicle intelligent defense and attack control system of claim 6, wherein, The way of solving the optimal attack and defense strategy by using the perception integrated attack algorithm includes: A series of attack and defense behaviors that the unmanned aerial vehicle can take are defined to form a decision space; each attack and defense behavior is parameterized by one or several continuous variables; According to the type of the target object, the kinematic model of the target object is defined as a constant velocity model, a constant acceleration model or a maneuvering target model; parameters of the kinematic model of the target object include position and velocity, and initial values of the parameters are obtained from the target detection result; the parameters of the kinematic model are updated online by using a Kalman filtering algorithm; the kinematic model of the unmanned aerial vehicle is defined as a 6-degree-of-freedom rigid body motion model; The state data of the unmanned aerial vehicle itself and the comprehensive information of the target object are combined into a state vector; the control input of the unmanned aerial vehicle is defined to include thrust and control surface deflection; the control input of the unmanned aerial vehicle is combined into a control vector; Then, a state transition equation is defined as: x(T+1)=f(x(T),u(T)), wherein x(T+1) is the state vector at time T+1; x(T) is the state vector at time T, and u(T) represents the control vector of the unmanned aerial vehicle at time T; f() is a nonlinear function obtained from the kinematic models of the target object and the unmanned aerial vehicle; The state transition equation is used as a constraint to define a reward function; the formula of the reward function is: R(s,a)=v1×Rki(s,a)+v2×Rsu(s,a)+v3×Rdi(s,a)+v4×Rti(s,a); wherein v1, v2, v3 and v4 are weight balance coefficients of corresponding items; R(s,a) is the benefit of the control vector a of the unmanned aerial vehicle under the state vector s; Rki(s,a) is a threat elimination function; Rsu(s,a) is a self survival function; Rdi(s,a) is a distance penalty function; and Rti(s,a) is a time penalty function; All state vectors form a state space, and the state space, a decision space, the state transition equation and the reward function are constructed into a reinforcement learning problem model; an optimal attack and defense strategy is obtained based on the reinforcement learning problem model.

8. The unmanned aerial vehicle intelligent defense and attack control system of claim 7, wherein, The threat elimination function Rki(s,a)=wki×th(s,a); wherein wki is a weight coefficient of threat elimination; and th(s,a) represents a threat level of the target object to the unmanned aerial vehicle under the control vector a of the unmanned aerial vehicle under the state vector s; Survival function where wda is the weight coefficient of survival penalty; dam(s, a) is the damage value suffered by the UAV when the control vector a is taken under the state vector s; hea(s) is the health status of the UAV under the state vector s; τ1 and τ2 are adjustable parameters; Distance penalty function where dis(s, a) is the distance between the UAV and the target object when the state vector is s and the control vector a is taken by the UAV; wea is the effective weapon range distance of the UAV; σ1 is a rate adjustment parameter, and δ1 is a degree adjustment parameter. Time penalty function where tmax is the maximum time step that the UAV can execute, i.e., the time limit of the task; ti is the time step that the UAV has already executed; is a time penalty index parameter.

9. The unmanned aerial vehicle intelligent defense and attack control system of claim 8, wherein, The manner of obtaining the optimal attack and defense strategy based on the reinforcement learning problem model comprises: A decision strategy network is constructed based on the reinforcement learning problem model, the decision strategy network is composed of an Actor network and a Critic network, and both adopt a deep neural network structure; the input of the Actor network is the state vector, and the output is the control vector; the input of the Critic network is the state vector and the control vector, and the output is the benefit; the network structure of the Actor network and the Critic network is a fully connected network or a convolutional network; a deep reinforcement learning algorithm is used to optimize the decision strategy network, so as to obtain a trained decision strategy network; The current state data of the unmanned aerial vehicle is obtained, for each target object, the current state data of the unmanned aerial vehicle and the comprehensive information of the target object are combined into a real-time state vector, the real-time state vector is used as input, a series of continuous behavior vectors are output through the trained decision strategy network, attack and defense behaviors and parameters are decoded to obtain a preliminary attack and defense strategy. The continuous behavior vector constitutes a discrete behavior sequence; the discrete behavior sequence is convolved with a Gaussian kernel to obtain a smoothed behavior sequence; and the smoothed behavior sequence is decoded to obtain an optimal attack-defense strategy.

Images