Space-based image non-cooperative target intention recognition method and system

Abstract

The invention discloses a space-based image non-cooperative target intention recognition method and system, and the method comprises the steps: constructing a layered intention space, defining a flight task of a non-cooperative target according to the layered intention space, fusing image data with orbit data, simulating the motion track of the non-cooperative target through a kinetic equation corresponding to the flight task, and carrying out the recognition of the non-cooperative target intention. A space-based image data set is acquired, a corresponding simulation scene is set by using an orbit kinetic equation, the scientificity of the scene is increased, the acquired image sequence data not only comprises spatial domain appearance texture information of a spatial non-cooperative target, but also comprises time domain orbit motion feature information, a data basis is provided for an end-to-end intention recognition algorithm, and the recognition accuracy of the end-to-end intention recognition algorithm is improved. The data set is subsequently used for neural network training, and end-to-end autonomous identification from original observation data to a space target intention can be realized. The real-time performance of non-cooperative target intention prediction is improved, so that the intention of a high-maneuvering target is pre-judged, and the actual requirement of high-precision intention recognition in a complex space scene is met.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition and relates to a method and system for recognizing the intention of non-cooperative targets in space-based images. Background Art

[0002] In recent years, the number of resident space objects has surged, posing a serious collision threat to space assets. According to authoritative statistics from the European Space Agency (ESA) Space Debris Office, the number of space debris larger than 10 centimeters in orbit will have risen sharply from 36,500 in September 2023 to 40,500 by March 2025. At the same time, the fierce global space race continues to intensify, with orbital gaming becoming a frequent phenomenon. This has made real-time monitoring and early warning of adversarial satellites a pressing need to safeguard space assets. The core technology for this requirement is the intention recognition of non-cooperative targets.

[0003] Currently, research on intention recognition of non-cooperative targets in space is still in its infancy, with no unified standard paradigm yet established. Current methods for space-based non-cooperative target recognition primarily focus on utilizing the target's orbital information. In practical applications, this orbital information is primarily acquired by space-based sensors, with optical cameras mounted on satellites playing a key role. After collecting raw observation data, these optical cameras must use complex state estimation algorithms to convert it into information such as orbital position, attitude, and its rate of change, or the distance, azimuth, and elevation angles between the target and the observation platform. These serve as input observation data for intention recognition. However, this approach, based on prior estimates, has significant drawbacks. First, error accumulation is inevitable during the state estimation process. Second, the complex and ever-changing space environment, including interference factors such as space radiation and atmospheric perturbations, can negatively impact data accuracy, severely limiting the accuracy of final intention recognition. Furthermore, traditional methods relying on prior estimates of orbital information suffer from poor real-time performance and are unable to meet the dynamic monitoring requirements of rapidly maneuvering non-cooperative targets.

[0004] Optical cameras, as crucial information acquisition devices in space-based systems, capture image sequences that contain rich information about target features and motion. Directly utilizing this image sequence information, bypassing the cumbersome prior estimation process, can not only reduce error propagation but also significantly improve the efficiency and accuracy of intent recognition. For example, when faced with fast-moving, non-cooperative targets, direct recognition based on image sequences can more promptly capture the target's dynamic changes, allowing for early judgment of its intent.

[0005] Correspondingly, in terms of dataset construction, current datasets used for intention recognition of non-cooperative space targets generally consist of relative orbital position and velocity data generated based on a CW dynamics model. While this model can theoretically simulate the target's motion trajectory, it is essentially a mathematical abstraction under ideal orbital conditions, significantly dissimilar to the complex scenes in real space. Raw images contain not only information about the target's appearance and posture, but also the spatial relationship between the target and its surroundings. This visual information is crucial for intention recognition. However, the construction of existing datasets separates image data from orbital data, preventing the rich spatial relationships and motion details contained in the images from being included in model training. This not only limits the training effectiveness of data-driven models such as deep learning, but also severely hinders the development of end-to-end intention recognition technology based on raw images. This prevents the effective mining and utilization of the rich information contained in the images, making it difficult to meet the practical needs of high-precision intention recognition in complex space scenarios. Summary of the Invention

[0006] The purpose of the present invention is to solve the problem in the prior art: the data set is constructed in a way that separates image data from orbital data, and the spatial relationships and motion details contained in the image cannot be effectively mined and utilized, which makes it difficult to meet the actual needs of high-precision intent recognition in complex space scenes. The present invention provides a method and system for non-cooperative target intent recognition in space-based images.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for identifying non-cooperative target intentions in space-based images comprises the following steps:

[0009] Construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission;

[0010] Based on the dynamic equations corresponding to the flight mission, the motion trajectory of non-cooperative targets is simulated, motion data is collected in real time, and a space-based image dataset is obtained;

[0011] Perform intent recognition on space-based image datasets and obtain recognition results.

[0012] A further improvement of the present invention is:

[0013] The flight missions include surveillance missions, reconnaissance missions, spy missions and collision missions.

[0014] When the flight mission is a surveillance mission, the dynamic equation corresponding to the flight mission is:

[0015]

[0016] Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target. It means that at the beginning, there is no relative displacement between the target and the spacecraft in the x direction, and there are initial position differences in the y and z directions. The initial relative velocities in the x, y, and z directions are all 0. The subsequent target will oscillate along the y direction to achieve following.

[0017] When the flight mission is a reconnaissance mission, the dynamic equation corresponding to the flight mission is:

[0018]

[0019] Where p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; ω represents the orbital angular velocity; the formula indicates that there are initial relative positions in the x, y, and z directions, and the velocity components are related to the orbital angular velocity ω and the initial position. The target flies around the spacecraft at the set rate and direction to complete the orbiting action.

[0020] When the flight mission is a peek mission, the dynamic equation corresponding to the flight mission is:

[0021]

[0022] Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; there are initial positions in the x and y directions, but no initial position in the z direction. There is a velocity component in the x direction related to ω and x0, and the target flies over the area where the spacecraft is located along the x direction.

[0023] When the flight mission is a collision mission, the dynamic equation corresponding to the flight mission is:

[0024]

[0025] Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; only the y direction has the initial relative position and velocity component, indicating that the target starts from the initial position in the y direction and approaches the spacecraft along the y direction at a velocity of y0.

[0026] A space-based image non-cooperative target intention recognition system includes the following steps:

[0027] The hierarchical module is used to construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission;

[0028] The dataset module is used to simulate the motion trajectory of non-cooperative targets based on the dynamic equations corresponding to the flight mission, collect motion data in real time, and obtain space-based image datasets;

[0029] The recognition module is used to perform intent recognition on the space-based image dataset and obtain recognition results.

[0030] The flight missions include surveillance missions, reconnaissance missions, spy missions and collision missions.

[0031] A terminal device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods of the present invention when executing the computer program.

[0032] A computer-readable storage medium stores a computer program, wherein the computer program, when executed by a processor, implements the steps of any method described in the present invention.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention discloses a method for identifying the intention of non-cooperative targets using space-based images. The method constructs a hierarchical intention space, defines the flight mission of the non-cooperative target based on the hierarchical intention space, fuses image data with orbital data, and simulates the motion trajectory of the non-cooperative target through the dynamic equation corresponding to the flight mission to obtain a space-based image dataset. The orbital dynamic equation is used to set up a corresponding simulation scene to increase the scientific nature of the scene. The acquired image sequence data contains not only the spatial appearance texture information of the space non-cooperative target, but also the temporal orbital motion feature information, providing a data foundation for the end-to-end intention recognition algorithm. The dataset is subsequently used for neural network training to achieve end-to-end autonomous intention recognition of non-cooperative targets. This improves the real-time prediction of the intention of non-cooperative targets, thereby enabling the prediction of the intention of highly maneuverable targets and meeting the actual needs of high-precision intention recognition in complex space scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 Flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0039] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0040] In the description of the embodiments of the present invention, it should be noted that if the terms "upper," "lower," "horizontal," "inner," etc. appear, the orientation or positional relationship indicated is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the inventive product is typically placed when in use. These terms are merely for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0041] In addition, if the term "horizontal" appears, it does not mean that the component must be absolutely horizontal, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0042] In the description of the embodiments of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0043] The present invention is described in further detail below with reference to the accompanying drawings:

[0044] See also Figure 1 The embodiment of the present invention discloses a method for non-cooperative target intention recognition in space-based images. The intention space defines a classification system for target intentions through a hierarchical model (such as the upper-level task type "surveillance / reconnaissance / voyeurism / collision" and the lower-level motion mode "oscillation follow / circle / flyover / approach"), providing a clear semantic framework for research; the simulation scene is constructed based on the ROS / Gazebo platform, and the relative orbital dynamics (such as the initial position and velocity parameters under the CW equation constraints) and the imaging process of the onboard camera are simulated for each category in the intention space to generate a SIRI dataset containing multiple target morphologies and multiple motion parameters, solving the problem of real-space data acquisition and providing support for algorithm training; the neural network algorithm (such as ImageConvRNN) is guided by the intention space and takes the simulation dataset as input. The CNN extracts image spatial features and the RNN captures the sequence time domain correlation to achieve end-to-end mapping from the original image to the intent label.

[0045] Given the high cost and high risk of space missions, it's difficult to obtain a large amount of real-world on-orbit observational data for constructing a dataset. This embodiment of the present invention leverages advanced simulation technology to build a high-precision simulation platform. First, this embodiment models the intentions of non-cooperative space targets from two perspectives, establishing a hierarchical intent space. Next, based on these intentions, a simulation platform is built and a dataset is prepared. Finally, this embodiment employs several algorithms to verify the validity of the dataset.

[0046] When carrying out missions in space, non-cooperative targets will approach our spacecraft to perform specific exploration tasks. This embodiment of the present invention defines a local horizontal and local vertical coordinate system on the spacecraft. The center of the coordinate system is located at the center of mass of the spacecraft, and the x, y, and z axes are aligned along the orbital radial direction, velocity direction, and normal to the orbital plane, respectively. To achieve the corresponding mission, the target spacecraft's motion must satisfy the orbital dynamics equations.

[0047]

[0048] Among them, x, y, and z represent the relative positions of the target. Represent the speed of the non-cooperative target, r s，ω They represent the rotation radius and angular velocity of the spacecraft respectively.

[0049] In order to obtain the target's intention, the spacecraft will use its onboard camera to continuously shoot the non-cooperative target and obtain a set of image sequences. i At the moment, the observation spacecraft captures an image V with a pixel size of w×h i After time N, the acquired information becomes a w×h×N tensor as the input of the algorithm, that is, the image sequence S of the non-cooperative target V=[V1, V2, ..., V N The task of intent recognition is to find a mapping f(·) that maps the input image sequence to the corresponding intent label i T , i.e. i T =f(S V ).

[0050] The specific steps include:

[0051] Obtain a space-based image dataset, input the space-based image dataset into a neural network for intent recognition, and obtain recognition results;

[0052] The acquisition of the space-based image dataset includes:

[0053] Step 1: Construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission;

[0054] Step 2: Based on the dynamic equations corresponding to the flight mission, simulate the motion trajectory of the non-cooperative target, collect motion data in real time, and obtain a space-based image dataset.

[0055] Specifically, in step 1, the construction of the hierarchical intent space includes the following steps:

[0056] To identify the intent of non-cooperative space targets, embodiments of the present invention define their intent at two levels. The upper-level intent space primarily focuses on the missions of non-cooperative targets, including surveillance, reconnaissance, surveillance, and collision. To achieve the missions corresponding to the upper-level intent space, the target spacecraft's motion must satisfy the orbital dynamics equations.

[0057] Specifically include:

[0058] In the surveillance mission, the non-cooperative target will follow the oscillating relative trajectory, maintain a certain distance, and use the onboard sensors to predict the position and attitude information of our spacecraft.

[0059] For reconnaissance purposes, a non-cooperative target needs to fly around the spacecraft to sense the spacecraft's geometry and payload.

[0060] In the spying intention, the non-cooperative target only needs to detect whether the spacecraft exists, and has lower requirements for the detection of factors such as posture and shape, so its relative trajectory often presents a one-way flyby pattern.

[0061] In a collision intent, the non-cooperative target will rapidly approach the spacecraft and its relative distance will decrease rapidly, showing a closing motion.

[0062] Corresponding to the above tasks, the lower intention space mainly focuses on the movement patterns corresponding to the upper intention space, including: oscillation following, circling flight, flying over and approaching.

[0063] Assume p0 = (x0, y0, z0) T 、 represent the initial relative position and velocity of the target respectively.

[0064] The corresponding initial values ​​are as follows:

[0065] Oscillation following: This indicates that initially there is no relative displacement between the target and the spacecraft in the x-direction, but there are initial position differences in the y- and z-directions; the initial relative velocities in the three directions are all 0. Subsequently, the target will oscillate along the y- and z-directions under the action of dynamics to achieve following.

[0066] Orbital flight: There are initial relative positions in the x, y, and z directions, and the velocity components are related to the orbital angular velocity ω and the initial position. By setting the parameters reasonably, the target can fly around the spacecraft at a specific speed and direction to complete the orbiting action.

[0067] Flyover: There are initial positions in the x and y directions but not in the z direction; there is only a velocity component in the x direction related to ω and x0. The target flies quickly over the area where the spacecraft is located in the x direction at this speed to achieve the flyby intention.

[0068] near: Only the y-direction has the initial relative position and velocity component, which means that the target starts from the initial position in the y-direction and approaches the spacecraft rapidly along the y-direction at a speed of y0, showing an obvious approaching trend.

[0069] During the simulation process, the embodiments of the present invention flexibly adjust the parameters of these formulas to control the relative motion between the spacecraft and the non-cooperative target, thereby completing the simulation of surveillance, reconnaissance, spying, and collision scenarios under different intentions and conditions.

[0070] Specifically, in step 1, to simulate the motion trajectory of a non-cooperative target, this embodiment constructs a simulation scenario, including:

[0071] The embodiment of the present invention utilizes ROS (Robot Operating System) and Gazebo to realize the construction of simulation scenes.

[0072] Among them, ROS is mainly responsible for information management and control of non-cooperative targets, and Gazebo is mainly responsible for three-dimensional scene modeling, physical simulation and image rendering tasks.

[0073] In this embodiment of the present invention, assuming that our spacecraft is stationary in space, we use the ros_control package in ROS to control non-cooperative targets. By publishing motion command topics with different data, we can achieve non-cooperative target motion with different intentions, such as oscillation following, orbiting, flyby, and approach. At the same time, by dynamically adjusting the initial position parameters of the target in ROS, we can systematically generate diverse simulation scenarios, thereby constructing a large-scale, scenario-rich dataset, providing a data foundation for the training and verification of non-cooperative target intention recognition algorithms.

[0074] In the visual perception module, this embodiment of the present invention utilizes the libgazebo o_ros_camera.so plugin, which integrates Gazebo with ROS, to establish a real-time image data transmission channel. This plugin connects the virtual camera sensor to the ROS messaging mechanism, publishing high-resolution image data rendered by Gazebo in real time using the standard ROS topic format. This allows ROS nodes to efficiently acquire and store image sequences, providing continuous data flow support for subsequent visual algorithm development and verification.

[0075] The target spacecraft's base is the core motion chassis structure connecting the wheel rings and the ball-wheel linkage. The left and right wheel rings interact with the two-wheel motion controllers via plug-in mounts, while the ball-wheel linkage primarily provides support. Regarding the drive system, the target spacecraft's wheel linkages, left wheel linkage, and right wheel linkage are defined as left and right joint drives, respectively. Drive control is provided by two-wheel differential controllers preset in the configuration file, enabling accurate modeling of orbital motion. To achieve stable visual monitoring of non-cooperative targets and ensure the camera captures consistently high-quality images, the observation spacecraft employs a visual tracking system. In this system, the camera is attached to a dedicated gimbal, ensuring precise attitude stability in complex space environments. A servo drive system simulation model is configured in Gazebo using ROS function packages. By defining PID parameters in the configuration file, the gimbal can calculate and execute control commands in real time based on the non-cooperative target trajectory topic data subscribed to ROS, enabling comprehensive tracking of the non-cooperative target.

[0076] During the simulation scenario construction process, this embodiment of the present invention leverages the Gazebo simulation platform and adopts a modular design concept, dividing the system into three major components: the spacecraft model, the non-cooperative target model, and the space environment. The 3D models of the spacecraft and non-cooperative targets are primarily described using URDF (Unified Robot Description Format) files. As the standard robot description format in the ROS system, URDF accurately represents the robot's kinematic structure, inertial parameters, and visual geometry. Through customized URDF macros, this embodiment of the present invention enables the modeling of complex structures such as the spacecraft body, solar panels, and robotic arms, ensuring that the dynamic characteristics of the model during motion are highly consistent with the real world. The space environment simulation is primarily accomplished using SDFormat (Simulation Description Format) files. As a scene description format natively supported by Gazebo, SDFormat provides a rich set of environmental modeling elements. By configuring SDFormat files, this embodiment of the present invention accurately simulates environmental factors such as solar radiation and space debris distribution in the space environment. Furthermore, this embodiment of the present invention develops a spacecraft motion plug-in based on real-world orbital mechanics, enabling orbital motion simulation consistent with orbital dynamic equations, ensuring that the spacecraft motion trajectory in the simulation scenario is highly consistent with the actual space environment.

[0077] In order to fully verify the robustness of the intent recognition algorithm in different scenarios, the embodiment of the present invention has constructed a diverse non-cooperative target model library, covering typical space targets such as satellites, manned spacecraft, probes, and space debris. The embodiment of the present invention generates more than 2,000 different simulation scene combinations by adjusting the size parameters, motion parameters, and environmental parameters of the target. In each scenario, Gazebo and ROS systems are used to synchronously collect the target's visual images, kinematic data, and sensor information, ultimately forming a data set containing 6,000 image sequence samples (a total of 12,000 images).

[0078] Furthermore, in this embodiment, the space-based image dataset is input into a neural network for intent recognition, and the recognition results obtained include:

[0079] To explore the effectiveness of the image simulation dataset, the present invention designed a neural network algorithm based on image sequences for verification. The neural network consists of a convolutional layer for spatial feature extraction, a recurrent layer for temporal feature extraction, an attention mechanism layer, and a classification layer. It leverages the advantages of algorithms such as CNN, BIGRU, and the attention mechanism.

[0080] The convolutional layer is responsible for parallel processing of image sequence information and feature extraction.

[0081] In order to reduce data redundancy, the embodiment of the present invention uses maximum pooling and concatenates all feature information after pooling into a tensor, and reduces the dimension through MLP.

[0082] The recurrent layer has local memory capabilities and is responsible for processing the reduced-dimensional data and extracting its time domain features.

[0083] The attention mechanism layer enables the neural network to focus more on the most critical features and increase recognition accuracy.

[0084] The final classification layer uses softmax to assign the most relevant feature labels to the image sequence. Tested, its accuracy reached over 99%, demonstrating the effectiveness of using images for intent recognition and targeting image simulation datasets for intent recognition problems.

[0085] This section designs an ImageConvRNN model for identifying the intent of non-cooperative space targets from image sequences captured by observation spacecraft. The ImageConvRNN network consists of four parts: a CNN layer for spatial image feature extraction, an RNN layer for temporal sequence feature extraction, an optional self-attention layer, and an intent classifier.

[0086] The input to the ImageConvRNN is an image sequence, S_V, consisting of N consecutive images. Each image in the sequence is first processed by a CNN layer to extract basic texture features encoded as feature maps. To reduce the redundancy of high-dimensional feature maps, this embodiment of the present invention adds a max pooling layer after the convolutional layer and flattens the feature maps into vector form for subsequent RNN layer processing.

[0087] The feature vector is then fed into a fully connected layer to extract the coupling relationship between different dimensions and derive a compressed feature form. After extracting the spatial domain features of each image, the temporal domain motion coupling between different time steps in the image sequence is extracted by the RNN layer.

[0088] An optional self-attention layer is also provided to assign weights to high-dimensional features according to their importance.

[0089] Finally, the extracted features are transformed into the same dimension as the number of elements in the intent set through a fully connected layer and then fed into a softmax classifier to obtain the intent recognition result.

[0090] The specific steps include:

[0091] CNN spatial domain image feature extraction:

[0092] The first step in the ImageConvRNN model is to extract spatial texture features for each image in the sequence in parallel using a CNN network. The CNN network consists of convolutional layers, max pooling layers, flattening layers, and fully connected layers. The motivation for using a single convolutional layer is to extract essential texture information from the image. Since the goal of intent recognition is to perceive the target motion pattern characteristics embedded in the image sequence, reducing the number of convolutional layers allows for better focus on low-level structural texture information rather than high-level semantic information, which helps subsequent RNN layers discover relationships between images.

[0093] For the input image sequence S V Each image V in i (i∈I N ), first reshape it to the size of w0×w0 to reduce the computational burden, and then the convolution layer uses a kernel K of size 5×5 n (n∈I K ) and bias b n (n∈I K ) performs a two-dimensional convolution operation as follows:

[0094] Z in =ReLU(conv(V i ,K n )+b n )

[0095] Among them, Z in Represents image V i The rth 2 feature maps, ReLU(·) is a nonlinear activation function, and cnv(·, ·) represents a two-dimensional convolution operation, as shown below:

[0096]

[0097] in,[·] pq represents the p-th row and q-th column element of the output matrix, V(m,n) and K(pm,qn) represent the m-th row and n-th column element of V and the (pm)-th row and (qn)-th column element of K respectively.

[0098] Then, the maximum pooling layer downsamples the feature map to reduce information redundancy and alleviate overfitting. in , the maximum pooling operation is defined as:

[0099]

[0100] in, is the downsampled feature map derived by max pooling, s1 and s2 represent the vertical and horizontal pooling steps, and W1×W2 is the size of the pooling window. In the problem of the embodiment of the present invention, the parameters of max pooling are set to W1=W2=s1=s2=2, so the feature map after max pooling becomes half of its original size.

[0101] Furthermore, each downsampled feature map Flattened to a one-dimensional vector And concatenate all vectors into an overall vector f i , and pass the output dimension d FC The fully connected layer compresses the high-dimensional feature vector f i , so that the subsequent RNN layer can process and alleviate overfitting. In this way, the image V i The derived eigenvector x of i As input to the following RNN layer, it looks like this:

[0102]

[0103] in,

[0104]

[0105] The hyperparameters of the CNN layer include the reshaped image size w0, the number of convolution kernels K, and the output dimension d of the fully connected layer. FC , the optimal choice of hyperparameters will be determined in the experiments in Section V.

[0106] Time domain sequence feature extraction based on RNN:

[0107] Based on the feature vectors extracted for each image by the CNN layer, the next step is to perceive the temporal motion characteristics of non-cooperative targets embedded in the continuous image sequence. While temporal image sequence data is well-matched to the temporal memory capabilities of recurrent RNNs, the vanishing and exploding gradient problems hinder the persistence of long-term information. To overcome the limitations of traditional RNNs, in this embodiment of the present invention, a variant model, BiGRU, is selected as the optimal network structure for spatial non-cooperative target intent recognition.

[0108] The calculation of the reverse network in BiGRU is as follows:

[0109]

[0110] Among them, r t , z t Represent the outputs of the reset gate and update gate respectively, is a candidate hidden state, W r , W z , W h is the one with the corresponding bias vector br , b z , b h The trainable weight matrix of BiGRU is:

[0111] h Bt =h t +h t '

[0112] Among them, h t and h′ t Derived from the calculation formula of the reverse network in BiGRU, the output sequence characteristics of the BiGRU network are:

[0113] H BiGRU =[h B1 ,h B2 ,...,h BN ].

[0114] The hyperparameters of the RNN layer include the image sequence length N and the output dimension d of the RNN unit. RNN , the best choice of hyperparameters will be found in Section V.

[0115] Self-attention based feature weighting:

[0116] The ImageConvRNN model also provides an optional self-attention layer for assigning weights to the high-dimensional output sequence features H of the RNN layer. Self-attention learns the importance of each feature from the input data without relying on additional information, enabling the model to focus on key features. The calculation of the self-attention layer is based on the mapping between the query Q and the key-value pair KV, which can be derived based on the output features H of the RNN layer as follows:

[0117] Q=W Q ·H,(22)

[0118] K=W K ·H,(23)

[0119] V=W V ·H,(24)

[0120] Among them, W Q , W K , W V Represents a learnable weight matrix, and the self-attention output value is as follows:

[0121]

[0122] Among them, d K is the dimension of K, and the softmax function normalizes the correlation between Q and K to the range of [0, 1].

[0123] Intent Recognition:

[0124] The final step of ImageConvRNN is to identify the target’s intent based on the features extracted above. Depending on whether the self-attention mechanism is integrated in the model, there are two cases for the extracted features: H output by the RNN layer and H output by the self-attention layer. a , here, the embodiment of the present invention uses the same symbol H to represent them for simplicity. The output part includes three steps: first, the output dimension is converted to the same number of elements as the intent set through a fully connected layer; then, the probability distribution of different intents is calculated using the softmax function; finally, the intent type with the highest probability is selected as the result. The calculation process of the output part is as follows:

[0125] h o =W·H+b

[0126]

[0127]

[0128] Among them, W and b represent the weight matrix and bias vector of the fully connected layer, h on is the vector h o The nth dimension of represents the probability of the nth intention type, i t is the intention recognition result of the non-cooperative target.

[0129] The data acquisition method disclosed in the embodiment of the present invention can improve the authenticity of the intent recognition scene and generate high-resolution image sequence data. The embodiment of the present invention integrates image data with orbital data. On the one hand, it uses the orbital dynamics equation to set up the corresponding simulation scene, increasing the scientific nature of the scene; on the other hand, it obtains image information that not only contains the position of non-cooperative targets, but also their appearance. It provides a data foundation for end-to-end intent recognition algorithms. Using the data set of the embodiment of the present invention to train neural networks can realize end-to-end neural networks. This improves the real-time prediction of non-cooperative target intentions, thereby enabling the prediction of the intentions of highly maneuverable targets.

[0130] A space-based image non-cooperative target intention recognition system includes the following steps:

[0131] The hierarchical module is used to construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission;

[0132] The dataset module is used to simulate the motion trajectory of non-cooperative targets based on the dynamic equations corresponding to the flight mission, collect motion data in real time, and obtain space-based image datasets;

[0133] The recognition module is used to perform intent recognition on the space-based image dataset and obtain recognition results.

[0134] A schematic diagram of a terminal device provided in one embodiment of the present invention. The terminal device in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of each of the aforementioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in each of the aforementioned device embodiments are implemented.

[0135] The computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to accomplish the present invention.

[0136] The terminal device may be a computing device such as a desktop computer, a notebook computer, a PDA, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0137] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0138] The memory may be used to store the computer programs and / or modules, and the processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory.

[0139] If the module / unit integrated in the terminal device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0140] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for non-cooperative target intention recognition in space-based images, characterized in that: The following steps are involved: Construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission; Based on the dynamic equations corresponding to the flight mission, the motion trajectory of non-cooperative targets is simulated, motion data is collected in real time, and a space-based image dataset is obtained; Perform intent recognition on space-based image datasets and obtain recognition results.

2. The method for non-cooperative target intention recognition in space-based images according to claim 1, characterized in that: The flight missions include surveillance missions, reconnaissance missions, spy missions and collision missions.

3. The method for non-cooperative target intention recognition in space-based images according to claim 2, characterized in that: When the flight mission is a surveillance mission, the dynamic equation corresponding to the flight mission is: Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target. It means that at the beginning, there is no relative displacement between the target and the spacecraft in the x direction, and there are initial position differences in the y and z directions. The initial relative velocities in the x, y, and z directions are all 0. The subsequent target will oscillate along the y direction to achieve following.

4. The method for non-cooperative target intention recognition in space-based images according to claim 2, characterized in that: When the flight mission is a reconnaissance mission, the dynamic equation corresponding to the flight mission is: Where p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; ω represents the orbital angular velocity; the formula indicates that there are initial relative positions in the x, y, and z directions, and the velocity components are related to the orbital angular velocity ω and the initial position. The target flies around the spacecraft at the set rate and direction to complete the orbiting action.

5. The method for non-cooperative target intention recognition in space-based images according to claim 2, characterized in that: When the flight mission is a peek mission, the dynamic equation corresponding to the flight mission is: Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; there are initial positions in the x and y directions, but no initial position in the z direction. There is a velocity component in the x direction related to ω and x0, and the target flies over the area where the spacecraft is located along the x direction.

6. The method for non-cooperative target intention recognition in space-based images according to claim 2, characterized in that: When the flight mission is a collision mission, the dynamic equation corresponding to the flight mission is: Among them, p0 represents the initial relative position of the target and v0 represents the initial relative velocity of the target; only the y direction has the initial relative position and velocity component, indicating that the target starts from the initial position in the y direction and approaches the spacecraft along the y direction at a velocity of y0.

7. A space-based image non-cooperative target intention recognition system, characterized by: The following steps are involved: The hierarchical module is used to construct a hierarchical intention space, define the flight mission of the non-cooperative target based on the hierarchical intention space, and construct the dynamic equation corresponding to the flight mission; The dataset module is used to simulate the motion trajectory of non-cooperative targets based on the dynamic equations corresponding to the flight mission, collect motion data in real time, and obtain space-based image datasets; The recognition module is used to perform intent recognition on the space-based image dataset and obtain recognition results.

8. The space-based image non-cooperative target intention recognition system according to claim 7, characterized in that: The flight missions include surveillance missions, reconnaissance missions, spy missions and collision missions.

9. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images