Unmanned aerial vehicle confrontation action classification method based on SlowFast-graph Fourier transform

Abstract

The invention provides an unmanned aerial vehicle confrontation action classification method based on SlowFast-graph Fourier transform, and the method comprises the steps: obtaining the training graph data of an unmanned aerial vehicle, carrying out the preprocessing of the training graph data, dividing the graph data into high-frame-rate graph data and low-frame-rate graph data, thereby giving full play to the advantages of different-frame-rate data in feature capture, and achieving the recognition of the confrontation actions of the unmanned aerial vehicle. The perception and decision-making capability of the unmanned aerial vehicle to a complex environment is improved; respectively inputting two data sequences with fast and slow differences into a SlowFast-graph Fourier transform network, and fusing features generated by a fast path into a slow path by adopting a one-way transverse connection mode so as to realize effective fusion and interaction of the features; and inputting the fused features into an action classification network module to realize accurate classification operation of the actions of the unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of multi-source data fusion, and more specifically to a UAV adversarial action classification method based on SlowFast-graph Fourier transform. Background Art

[0002] In military and security settings, drones must engage in confrontations with enemy drones or defense systems. Through reinforcement learning, drones can learn optimal strategies in these confrontational environments to evade threats or deliver effective strikes. Drone confrontations often occur in highly dynamic and uncertain environments, such as cities, forests, or battlefields. Obstacles, weather conditions, and enemy behavior in these environments can change at any time, complicating drone perception, decision-making, and execution. Drones must integrate data from multiple sensors (such as cameras, radar, and infrared sensors) to achieve comprehensive environmental awareness. Drone confrontation maneuver classification involves categorizing drone maneuvers into several categories based on the different tasks and objectives in the confrontation scenario. These categories can include basic flight maneuvers, maneuvers, and special maneuvers. The goal of this classification is to provide drones with a clear action framework, enabling them to quickly select and execute optimal strategies in complex environments. During these confrontational games, drones must process large amounts of data in real time and make rapid decisions. However, existing data processing algorithms often struggle to meet these stringent real-time requirements. Therefore, efficient and accurate data fusion remains a significant technical challenge, leading to decision delays and performance degradation. Summary of the Invention

[0003] In order to solve the problem of insufficient real-time processing and fusion of massive data by UAVs in complex environments, this paper proposes a UAV adversarial action classification method based on SlowFast-graph Fourier transform.

[0004] In order to achieve the above technical effects, the technical solutions of the present invention are as follows:

[0005] A method for classifying UAV adversarial actions based on SlowFast-graph Fourier transform includes the following steps:

[0006] Acquire UAV training graph data, and convert the graph data into a low-frame-rate, low-temporal-resolution graph data sequence and a high-frame-rate, high-temporal-resolution graph data sequence through a sampling operation;

[0007] Constructing a SlowFast-Graph Fourier Transform (SF-GFT) network, wherein the SlowFast-Graph Fourier Transform network includes a Slow-GFT transformation network module, a Fast-GFT transformation network module, and an action classification network module;

[0008] Inputting the low-frame-rate, low-time-resolution graph data sequence into the Slow-GFT transformation network module to obtain slow-path features; inputting the high-frame-rate, high-time-resolution graph data sequence into the Fast-GFT transformation network module to obtain fast-path features; and superimposing the fast-path features onto the slow-path features to obtain fused features.

[0009] The fused features are input into the action classification network module to classify the actions of the drone.

[0010] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0011] The present invention obtains and preprocesses the UAV training graph data, divides the graph data into high- and low-frame-rate graph data, and fully utilizes the advantages of data with different frame rates in feature capture, thereby improving the UAV's perception and decision-making capabilities of complex environments; the two data sequences with different speeds are respectively input into the SlowFast-graph Fourier transform network, and a unidirectional horizontal connection is adopted to integrate the features generated by the fast path into the slow path, thereby realizing effective feature fusion and interaction; the fused features are input into the action classification network module to realize accurate classification of UAV actions. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The flowchart of the UAV adversarial action classification method based on SlowFast-Graph Fourier Transform is shown in an embodiment of the present invention.

[0013] Figure 2 This is a diagram of the SlowFast-graph Fourier transform network structure shown in an embodiment of the present invention.

[0014] Figure 3 This is a flowchart of the drone action classification workflow shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0016] The terms used in this invention are for the purpose of describing specific embodiments only and are not intended to limit the invention. The singular forms "a," "the," and "the" used in this invention and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0017] It should be understood that although the terms "first," "second," "third," etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if" as used herein may be interpreted as "when," "when," or "in response to determining."

[0018] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1

[0020] This embodiment proposes a UAV adversarial action classification method based on SlowFast-graph Fourier transform, such as Figure 1 The flowchart of this method is shown in FIG.

[0021] Acquire UAV training graph data, and convert the graph data into a low-frame-rate, low-temporal-resolution graph data sequence and a high-frame-rate, high-temporal-resolution graph data sequence through a sampling operation;

[0022] Constructing a SlowFast-graph Fourier transform network, wherein the SlowFast-graph Fourier transform network includes a Slow-GFT transform network module, a Fast-GFT transform network module and an action classification network module;

[0023] Inputting the low-frame-rate, low-time-resolution graph data sequence into the Slow-GFT transformation network module to obtain slow-path features; inputting the high-frame-rate, high-time-resolution graph data sequence into the Fast-GFT transformation network module to obtain fast-path features; and superimposing the fast-path features onto the slow-path features to obtain fused features.

[0024] The fused features are input into the action classification network module to classify the actions of the drone.

[0025] In this embodiment, the drone training graph data is obtained and preprocessed, and the graph data is divided into high and low frame rate graph data, so as to give full play to the advantages of different frame rate data in feature capture, thereby improving the drone's perception and decision-making capabilities of complex environments; the two data sequences with different speeds are respectively input into the SlowFast-graph Fourier transform network, and a unidirectional horizontal connection method is used to integrate the features generated by the fast path into the slow path, so as to achieve effective fusion and interaction of features; the fused features are input into the action classification network module to achieve accurate classification operations of drone actions.

[0026] Example 2

[0027] On the basis of Example 1, the technical effects of this solution are further demonstrated. Specifically:

[0028] Obtain drone training data and convert it into graph data G as training network input.

[0029] Furthermore, during the processing of drone training data, each drone is abstracted as a "point" in three-dimensional space, and a coordinate system is constructed within this three-dimensional space. To more efficiently summarize the various types of data, a drone graph data sequence is created based on the constructed three-dimensional spatial coordinate system. This sequence is constructed in a similar way to a skeleton sequence and contains rich drone state information, specifically the drone's three-dimensional coordinates, pose matrix, and velocity data. Mathematically, drone-related information can be represented by a graph G = (V, E). Here, the set V is a set of nodes, |V| = n, where n represents the number of nodes, and each node in the set corresponds to an entity in the graph; the set E represents edges, reflecting the relationships between nodes.

[0030] In this embodiment, graph data can effectively express data relationships and intuitively present data relevance, helping drones better understand the meaning behind the data and making more reasonable decisions. This processing and integration approach not only facilitates subsequent data processing and analysis, but also presents the data in a more organized and relevant form, providing strong support for drones' decision-making and execution in complex missions.

[0031] By performing a sampling operation, the input image data is converted into two image data sequences with different characteristics: one with a low frame rate and low temporal resolution, and the other with a high frame rate and high temporal resolution. Separating the data into high and low frame rate data is designed to fully leverage the advantages of different frame rates in capturing features, thereby improving the drone's perception and decision-making capabilities in complex environments.

[0032] Furthermore, in the SlowFast-Graph Fourier Transform-based drone adversarial motion classification method, the slow and fast paths perform different yet complementary feature extraction tasks, jointly enabling accurate situational awareness and decision-making for drone swarms. The slow path primarily extracts slowly varying features, such as spatial details, from low-frame-rate, low-temporal-resolution data sequences. During data sampling in the slow path, the sampling time step is set to τ, meaning a sample is taken every τ seconds. Let the data length sampled by the slow path be T, then the corresponding raw data length is T × τ. The fast path focuses on capturing rapidly varying features, such as the temporal motion of the drone swarm, from high-frame-rate, high-temporal-resolution data sequences. The parameter α is introduced to represent the frame rate ratio between the fast and slow paths, satisfying α > 1. The sampled data length of the fast path is αT. Both paths process the same raw data, differing in that the data sampling density of the fast path is α times greater than that of the slow path. Regarding channel design, the fast and slow paths share similar design principles, with the channel capacity of the slow path set to C. To measure the channel difference between the fast path and the slow path, we introduce the parameter β to represent the channel ratio between the two. Based on this setting, the channel capacity of the fast path is lower than that of the slow path, specifically βC.

[0033] In an optional embodiment, the Fast-GFT transformation network includes a high-resolution graph data input layer, a first graph Fourier transform layer, a first frequency domain filter layer, a first inverse graph Fourier transform layer, a first activation function layer and a first global pooling layer connected in sequence; the Slow-GFT transformation network includes a low-resolution graph data input layer, a second graph Fourier transform layer, a second frequency domain filter layer, a second inverse graph Fourier transform layer, a second activation function layer and a second global pooling layer connected in sequence; the action classification network module includes a fully connected layer, a Softmax classifier layer and an output layer connected in sequence. The SlowFast-graph Fourier transform network structure diagram is shown in FIG. Figure 2 shown.

[0034] In an optional embodiment, the step of inputting the high frame rate and high time resolution graph data sequence into the Fast-GFT transformation network module to obtain fast path features comprises:

[0035] Inputting the high frame rate and high time resolution image data sequence into the first image Fourier transform layer and converting it into a frequency domain signal;

[0036] Inputting the frequency domain signal into the first frequency domain filter layer to obtain a filtered frequency domain signal;

[0037] Converting the filtered frequency domain signal back to a spatial domain signal via the first inverse Fourier transform layer;

[0038] Extracting nonlinear features in the spatial domain signal using the first activation function layer;

[0039] The first global pooling layer compresses the spatial dimension of the feature map into a single value and aggregates global information to obtain fast path features;

[0040] The step of inputting the low frame rate and low time resolution graph data sequence into the Slow-GFT transformation network module to obtain a slow path feature specifically includes:

[0041] Inputting the low frame rate and low time resolution image data sequence into the second image Fourier transform layer and converting it into a frequency domain signal;

[0042] Inputting the frequency domain signal into the second frequency domain filter layer to obtain a filtered frequency domain signal;

[0043] Converting the filtered frequency domain signal back to a spatial domain signal via the second inverse Fourier transform layer;

[0044] Extracting nonlinear features in the spatial domain signal using the second activation function layer;

[0045] The second global pooling layer compresses the feature map spatial dimension into a single value and aggregates global information to obtain slow path features.

[0046] In an optional embodiment, the step of superimposing the fast path feature onto the slow path feature to obtain the fused feature adopts a time-to-channel approach, and the specific steps include:

[0047] The feature size of the fast path is reshaped and transposed to make the feature size of the fast path consistent with that of the slow path; the features of the transformed fast path are superimposed on the features of the slow path to achieve the fusion of the fast and slow path features.

[0048] In this example, these two different data sequences, each with different speeds, are fed into a SlowFast-graph Fourier transform network. Within this network architecture, unidirectional horizontal connections are used to integrate features generated by the fast path into the slow path, achieving effective feature fusion and interaction. When processing these two types of data, this network helps extract spectral features from the graph data and mine its deeper insights, providing strong support for subsequent drone action classification and decision-making.

[0049] Furthermore, since the fast and slow pathways differ significantly in the temporal dimension, to effectively fuse their features, we employ a lateral concatenation transformation technique to align their feature dimensions. Horizontal concatenation is a widely used technique for fusing features of varying spatial resolutions and semantic levels. The feature size corresponding to the slow pathway can be expressed as {C, T, V}, while the feature size of the fast pathway is {1 / α×C, α×T, V}. Here, C represents the channel size, T is the data length, and V is the node feature vector. To fuse the features of the fast pathway into the slow pathway, we employ a time-to-channel approach. Specifically, the feature size of the fast pathway is reshaped and transposed to {1 / α×1 / α×C, T, V}, thereby aligning it with the feature size {C, T, V} of the slow pathway. Finally, the transformed features are superimposed on the slow pathway features, achieving an effective fusion of the features of the fast and slow pathways.

[0050] In an optional embodiment, the image data sequence includes the three-dimensional coordinates, posture matrix and speed data of the drone.

[0051] In an optional embodiment, the specific steps of processing the graph data by the first graph Fourier transform layer and the second graph Fourier transform layer include:

[0052] Construct the Laplacian matrix L of the graph. The Laplacian matrix is ​​usually defined as L = DA, where D is the degree matrix and A is the adjacency matrix. In graph signals, the Laplacian matrix is ​​used to describe the signal difference between the central node and its neighboring nodes;

[0053] The Laplace matrix is ​​symmetrically normalized, and the expression of the symmetrically normalized Laplace matrix is:

[0054] L sym =ID -1 / 2 AD -1 / 2

[0055] Among them, A represents the adjacency matrix of the graph, and D represents the degree matrix (diagonal matrix, D ii =∑ j A ij ), I represents the identity matrix;

[0056] Perform eigenvalue decomposition on the Laplace matrix L to obtain the eigenvalues ​​and corresponding eigenvectors; the expression is:

[0057] L=UΛU T

[0058] Among them, U represents the eigenvector matrix, Λ represents the eigenvalue diagonal matrix, the eigenvalue represents the spectrum information of the graph, and the eigenvector represents the different spectral components of the graph;

[0059] The signal of the graph data sequence is converted from the vertex domain to the spectrum domain. The graph Fourier transform expression of the signal x of the graph data sequence is:

[0060]

[0061] in, represents the signal of the graph data sequence after graph Fourier transform, and U represents the eigenvector matrix.

[0062] Furthermore, the frequency domain filter layer filters the signal of the graph data sequence after the graph Fourier transform. In the spectral domain, the signal can be filtered by the filter g(Λ): Then the filtered signal is obtained by inverse transformation: x filtered =Ug(Λ)U T x.

[0063] Furthermore, in the inverse Fourier transform layer, the signal is restored from the spectrum domain to the vertex domain through inverse transformation; its expression is:

[0064] In an optional embodiment, the activation function expressions in the first activation function layer and the second activation function layer are:

[0065] ReLU(x)=max(0,x)

[0066] Where x represents the input signal. When the input x≥0, the output is x; when x<0, the output is 0.

[0067] In this embodiment, ReLU solves the complex pattern recognition problem through nonlinear activation, efficient computation, and sparsity.

[0068] In an optional embodiment, the global pooling operation expressions in the first global pooling layer and the second global pooling layer are:

[0069]

[0070] Among them, H represents the height of the feature map, W represents the width of the feature map, h, w represent the position coordinates, F h，w Represents the (h, w) position element in the channel.

[0071] In an optional embodiment, the fully connected layer performs a linear transformation on the features of the previous output and inputs the transformed features into the SoftMax classifier layer. The calculation expression in the fully connected layer is:

[0072] y=σ(W g+b)

[0073] Where g∈R n Represents the input vector, n represents the dimension, W∈Rm×n Represents the weight matrix, dimension represents m×n, b∈R m represents the bias vector, m represents the dimension, σ(·) represents the activation function, y∈R m Represents the output vector, and m represents the dimension.

[0074] In an optional embodiment, the input vector g is multiplied by the weight matrix W to obtain a linear transformation result Wg, which is added with the bias vector b to obtain Wg+b, and a nonlinear transformation is performed through the activation function σ(·) to obtain the final output y.

[0075] Furthermore, the output of the fully connected layer is connected to a Softmax classifier. The Softmax function is mainly used for the probability output of multi-classification problems. It is used to map the integrated features to specific category labels, thereby realizing the classification of actions.

[0076] In an optional embodiment, the Softmax classifier layer normalizes the numerical values ​​and converts them into probability distributions of each category of actions, and inputs the probability distributions of each category of actions into the output layer to obtain action classification results; the calculation expression of the Softmax classifier layer is:

[0077]

[0078] Where y represents an n-dimensional probability distribution vector, z represents an n-dimensional real number vector, and e represents the base of the natural logarithm.

[0079] Furthermore, the output value of the Softmax function ranges between (0, 1), which makes it very suitable for representing probability.

[0080] In this embodiment, the fully connected layer performs a linear transformation on the feature vectors output by the previous layer, integrating the features across all dimensions. The output is then fed into the SoftMax function module. SoftMax normalizes the values ​​and converts them into probability distributions for each action category, ensuring that the sum of the probabilities is 1, thus revealing the likelihood of different action categories. This processed data is then passed to the output layer, which outputs the action classification result—the model's prediction of the action category for the input graph data.

[0081] In an optional embodiment, the drone obtains corresponding classified action instructions according to the action classification results, optimizes the action during execution, and outputs the optimized action.

[0082] In an optional embodiment, the optimizing the action execution process includes using a Kalman filter algorithm to update the action execution process in real time.

[0083] Furthermore, the drone leverages this classification method to accurately identify various actions and execute corresponding operations accordingly. During the action execution process, a Kalman filter algorithm is introduced to optimize it. Based on a dynamic model of the system state and effectively processing measurement data, this algorithm continuously updates and predicts the system state, enabling real-time and precise tracking of the drone's actions. This significantly improves the stability and reliability of the drone's actions in complex environments.

[0084] Kalman Filter is an algorithm that uses the linear system state equation to optimally estimate the system state through system input and output observation data.

[0085] The first step is the prediction step:

[0086] State prediction: Predict the state at the current moment based on the state estimate at the previous moment and the state transfer matrix of the system.

[0087] The formula is: k =F k a k-1 +B k u k +w k

[0088] Among them, a k represents the state of the system at time k, F k Represents the state transition matrix, describing how the state evolves from time k-1 to time k, B k Represents the control input matrix, describing the control input u k Impact on state, w k represents process noise, assuming a Gaussian distribution with a mean of 0 and a covariance of Q k .

[0089] Measurement model: The observation equation formula is: N k =H k a k +v k

[0090] Among them, N k represents the observed value of the system at time k, H k Represents the observation matrix, describing how the state is mapped to the observation value, v k represents the observation noise, which is assumed to be Gaussian distributed with a mean of 0 and a covariance of R k .

[0091] The second step is the Kalman filter calculation steps:

[0092] Step 1: Prediction. In the prediction step, the system model is used to predict the state and covariance at the current moment.

[0093] Prediction status:

[0094] in, represents the predicted state value at time k, Represents the estimated value of the state at time k-1.

[0095] Forecast covariance:

[0096] in, represents the predicted state covariance value at time k, P k-1 Represents the estimated value of the state covariance at time k-1.

[0097] Step 2: Update.

[0098] In the update step, the observations are used to correct the predictions and obtain a more accurate state estimate.

[0099] Calculate the Kalman gain; its expression is:

[0100]

[0101] Among them, K k Represents the Kalman gain, which is used to weigh the confidence of the predicted value and the observed value.

[0102] Update the state estimate; its expression is:

[0103]

[0104] in, represents the updated state estimate at time k, Represents the observed residual value.

[0105] Update the covariance estimate; its expression is:

[0106]

[0107] P k represents the updated covariance value at time k, and I represents the identity matrix.

[0108] In this embodiment, the problems of the existing technology in the field of drone reinforcement learning, such as the difficulty in achieving efficient and accurate data fusion, the inability to meet the requirements of real-time processing and fusion of massive data in complex environments, and the high computational complexity of related algorithms are solved. The situational awareness capability, decision-making accuracy, and the precision and complexity of the fusion data algorithm of the drone are comprehensively improved. The action execution process is optimized to continuously update and predict the system status, so that the drone's action execution can be tracked in real time and accurately.

[0109] Example 3

[0110] Specifically, based on Example 1 and Example 2, the solution is described in conjunction with specific embodiments. The drone action classification workflow diagram of this embodiment is as follows: Figure 3 shown.

[0111] In a complex battlefield environment, our reconnaissance aircraft are operating five unmanned reconnaissance aircraft. These drones must perceive their surroundings in real time, identify enemy drone movements, and employ countermeasures based on these identifications. The enemy has dispatched five drones, preparing to raid our positions. These drones may employ rapid movements, sudden turns, or weapon attacks. Therefore, our reconnaissance aircraft require efficient data processing and motion classification methods to respond and implement appropriate maneuvers, such as evasive maneuvers, formation flight encirclement, and weapon counterattacks.

[0112] Step 1: Collect training data for drones in battlefield environments, including their 3D coordinates, pose matrix, and velocity data. This data is converted into a graph called G = (V, E), where each drone is represented as a node in the graph, V is the set of nodes, |V| = n, where n = 5 represents the number of nodes, and E represents edges. Edges between nodes represent relationships between drones (e.g., distance, relative velocity, etc.).

[0113] In step 2, the system preprocesses the drone image data. The system samples the input image data and divides it into two paths. The low-frame-rate, low-temporal-resolution image data path is suitable for processing relatively stable, slowly changing information, such as the layout of buildings in a city, and is sampled every 5 seconds. The high-frame-rate, high-temporal-resolution image data path is used to capture rapidly changing information, such as the rapid maneuvers of enemy drones, and is sampled every 0.5 seconds.

[0114] Exemplarily, α represents the frame rate ratio between the fast path and the slow path, where α=10.

[0115] In step 3, the graph Fourier transform network processes the graph data. Low-frame-rate graph data enters the Slow-GTF transform network, which analyzes the long-term impact of building layout on flight paths and potential threats. High-frame-rate graph data enters the Fast-GTF transform network, which rapidly identifies the immediate movements and threats of enemy drones. These two networks are laterally connected to share key information, such as the general movement trends of enemy drones.

[0116] Step 4: The fully connected neural network performs action classification. The fused features are input into the fully connected neural network (FC) layer to classify the drone's actions. Drone actions are rich and varied, and the most common ones can be classified into basic flight actions, maneuvers, special actions, etc. The following is a detailed introduction:

[0117] Basic flight maneuvers: take-off, hovering, forward, backward, ascent, descent, left turn, right turn.

[0118] Maneuvers: flying left, flying right, diving, flying upwards, rolling left, rolling right, jumping, circling.

[0119] Special moves: triangle formation, V formation, weapon attack, evasion and dodging, 360-degree flip.

[0120] In step 5, the Kalman filter algorithm optimizes the action. Based on the classification results, the drone adopts the corresponding countermeasure strategy. For example, if the enemy drone is identified as "diving rapidly," an evasive action Ai is output. The Kalman filter algorithm optimizes the action execution process, updates and predicts the system state in real time, and outputs the drone's final output action Ai+1, ensuring the accuracy and stability of the drone's actions.

[0121] In this embodiment, the situational awareness capability and decision-making accuracy of the UAV are significantly improved, while the computational complexity is reduced while ensuring fusion accuracy, providing strong support for the efficient operation of the UAV in complex confrontation environments and promoting the further development of UAV technology in the military and security fields.

[0122] Each embodiment of the present invention is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiment. The device embodiment described above is merely exemplary. The modules described as separate components may or may not be physically separated. When implementing the scheme of the present invention, the functions of each module can be implemented in the same one or more software and / or hardware. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the scheme of this embodiment.

[0123] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A UAV adversarial action classification method based on SlowFast-Graph Fourier Transform, characterized by: The following steps are involved: Acquire UAV training graph data, and convert the graph data into a low-frame-rate, low-temporal-resolution graph data sequence and a high-frame-rate, high-temporal-resolution graph data sequence through a sampling operation; Constructing a SlowFast-graph Fourier transform network, wherein the SlowFast-graph Fourier transform network includes a Slow-GFT transform network module, a Fast-GFT transform network module and an action classification network module; Inputting the low-frame-rate, low-time-resolution graph data sequence into the Slow-GFT transformation network module to obtain slow-path features; inputting the high-frame-rate, high-time-resolution graph data sequence into the Fast-GFT transformation network module to obtain fast-path features; and superimposing the fast-path features onto the slow-path features to obtain fused features. The fusion features are input into the action classification network module to classify the actions of the drone.

2. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 1, characterized in that: The Fast-GFT transformation network includes a high-resolution graph data input layer, a first graph Fourier transform layer, a first frequency domain filter layer, a first inverse graph Fourier transform layer, a first activation function layer and a first global pooling layer connected in sequence; the Slow-GFT transformation network includes a low-resolution graph data input layer, a second graph Fourier transform layer, a second frequency domain filter layer, a second inverse graph Fourier transform layer, a second activation function layer and a second global pooling layer connected in sequence; the action classification network module includes a fully connected layer, a Softmax classifier layer and an output layer connected in sequence.

3. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 1, characterized in that: The step of inputting the high frame rate and high time resolution graph data sequence into the Fast-GFT transformation network module to obtain fast path features specifically includes: Inputting the high frame rate and high time resolution image data sequence into the first image Fourier transform layer and converting it into a frequency domain signal; Inputting the frequency domain signal into the first frequency domain filter layer to obtain a filtered frequency domain signal; Converting the filtered frequency domain signal back to a spatial domain signal via the first inverse Fourier transform layer; Extracting nonlinear features in the spatial domain signal using the first activation function layer; The first global pooling layer compresses the spatial dimension of the feature map into a single value and aggregates global information to obtain fast path features; The step of inputting the low frame rate and low time resolution graph data sequence into the Slow-GFT transformation network module to obtain a slow path feature specifically includes: Inputting the low frame rate and low time resolution image data sequence into the second image Fourier transform layer and converting it into a frequency domain signal; Inputting the frequency domain signal into the second frequency domain filter layer to obtain a filtered frequency domain signal; Converting the filtered frequency domain signal back to a spatial domain signal via the second inverse Fourier transform layer; Extracting nonlinear features in the spatial domain signal using the second activation function layer; The second global pooling layer compresses the feature map spatial dimension into a single value and aggregates global information to obtain slow path features.

4. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 1, characterized in that: The fast path feature is superimposed on the slow path feature to obtain the fused feature in a time-to-channel manner, and the specific steps include: The feature size of the fast path is reshaped and transposed to make the feature size of the fast path consistent with that of the slow path; the features of the transformed fast path are superimposed on the features of the slow path to achieve the fusion of the fast and slow path features.

5. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 3, characterized in that: The specific steps of processing the graph data by the first graph Fourier transform layer and the second graph Fourier transform layer include: Construct the Laplacian matrix L of the graph. The Laplacian matrix is ​​usually defined as L = DA, where D is the degree matrix and A is the adjacency matrix. In graph signals, the Laplacian matrix is ​​used to describe the signal difference between the central node and the neighboring nodes; The Laplace matrix is ​​symmetrically normalized, and the expression of the symmetrically normalized Laplace matrix is: L sym =I-D -1 / 2 AD -1 / 2 Where A represents the adjacency matrix of the graph, D represents the degree matrix, and I represents the identity matrix; Perform eigenvalue decomposition on the Laplace matrix L to obtain the eigenvalues ​​and corresponding eigenvectors; the expression is: L=UΛU T Among them, U represents the eigenvector matrix, Λ represents the eigenvalue diagonal matrix, the eigenvalue represents the spectrum information of the graph, and the eigenvector represents the different spectral components of the graph; The signal of the graph data sequence is converted from the vertex domain to the spectrum domain. The graph Fourier transform expression of the signal x of the graph data sequence is: in, represents the signal of the graph data sequence after graph Fourier transform, and U represents the eigenvector matrix.

6. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 3, characterized in that: The activation function expressions in the first activation function layer and the second activation function layer are: ReLU(x)=max(0,x) Where x represents the input signal. When the input x≥0, the output is x; when x<0, the output is 0.

7. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 3, characterized in that: The global pooling operation expressions in the first global pooling layer and the second global pooling layer are: Among them, H represents the height of the feature map, W represents the width of the feature map, h, w represent the position coordinates, F h，w Represents the (h, w) position element in the channel.

8. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to any one of claims 1 to 7, characterized in that: The fully connected layer performs a linear transformation on the features of the previous output and inputs the transformed features into the SoftMax classifier layer. The calculation expression of the fully connected layer is: y=σ(W g+b) Where g∈R n Represents the input vector, n represents the dimension, W∈R m×n Represents the weight matrix, dimension represents m×n, b∈R m represents the bias vector, m represents the dimension, σ(·) represents the activation function, y∈R m represents the output vector, and m represents the dimension; The Softmax classifier layer normalizes the values ​​and converts them into the probability distribution of each category of action. The probability distribution of each category of action is input into the output layer to obtain the action classification result. The calculation expression of the Softmax classifier layer is: Where y represents an n-dimensional probability distribution vector, z represents an n-dimensional real number vector, and e represents the base of the natural logarithm.

9. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to any one of claims 1 to 8, characterized in that: The UAV obtains the corresponding classified action instructions according to the action classification results, optimizes the action execution process, and outputs the optimized action.

10. The method for classifying UAV adversarial actions based on SlowFast-Graph Fourier Transform according to claim 9, characterized in that: The optimization during the action execution process includes using a Kalman filter algorithm to update the action execution process in real time.

Images