Unmanned system intention classification and trajectory prediction method, device, equipment and medium

Abstract

The application provides an unmanned system intention classification and trajectory prediction method, device, equipment and medium, relates to the unmanned driving technical field, and the method acquires a plurality of channel signals collected in real time by a preset sensor component and constructs a plurality of channel attribute graph sets, calculates the change rate corresponding to each channel attribute graph information, splices the change rates of all channels to obtain a global hidden potential field, calculates the convergence result of each channel according to the global hidden potential field, projects the fusion result of the convergence result of each channel using a nonlinear projector to obtain a label prediction result, thereby realizing the recognition of a navigation mode and the continuous prediction of a trajectory. The application realizes the continuous time fusion of multi-source heterogeneous sensor data and the smooth trajectory deduction by splicing the multi-channel change rates into a global hidden potential field and performing evolution solving based on second-order partial differentiation, and can be widely applied to unmanned driving systems such as unmanned ships and unmanned planes.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to methods, devices, equipment, and media for classifying intentions and predicting trajectories of unmanned systems. Background Technology

[0002] Currently, for autonomous unmanned systems such as unmanned surface vessels, drones, and unmanned ground vehicles to perform tasks in complex and dynamic environments, accurately perceiving the motion state of surrounding dynamic obstacles, identifying their behavioral intentions, and predicting their future trajectories are core technological prerequisites for achieving safe autonomous navigation and obstacle avoidance. Taking unmanned vessels as an example, in densely trafficked waterways such as ports and waterways, autonomous vessels need to fuse asynchronous data streams from multiple heterogeneous sensors such as radar, AIS, and vision in real time to accurately judge the navigation modes of surrounding vessels (such as heading and speed maintenance, cooperative avoidance, and near-shore berthing) and generate continuous and smooth predicted trajectories that meet physical constraints. Similarly, drones in urban low-altitude areas need to fuse multi-source information such as vision, radar, and communication to identify the motion patterns of surrounding aircraft, and unmanned ground vehicles in mixed traffic flows also face the challenge of recognizing the driving intentions and predicting the trajectories of surrounding vehicles. These scenarios are highly consistent in their technical essence, all requiring the processing of multi-source heterogeneous sensor data, identification of target motion modes, and generation of continuous trajectories that meet dynamic constraints. However, existing technical solutions still have many shortcomings in addressing these problems.

[0003] Current mainstream trajectory prediction and intent recognition methods can be broadly categorized into three types. The first type is based on traditional filtering and state estimation, such as Kalman filtering and its nonlinear extensions. This involves recursively extrapolating from a state-space model of the target's motion, resulting in high computational efficiency. However, the prediction accuracy heavily relies on a pre-defined motion model. When the target undergoes mode switching or strong maneuvers, the fixed model struggles to accurately describe the true dynamics, leading to a sharp increase in error. The second type is based on data-driven deep learning methods, such as Long Short-Term Memory networks and Transformers. These methods learn end-to-end mapping relationships on large-scale trajectory data. While they do not require explicit modeling of motion equations, they often neglect spatial interactions between targets, and the output is a discrete time-point sequence. Trajectory smoothness lacks physical constraints, and generalization ability is limited by the distribution of training data. The third type introduces graph neural networks to model interactions between targets, aggregating neighborhood information through message passing. This partially compensates for the shortcomings of time-series models, but existing graph-based methods still generally face deep-seated technical obstacles.

[0004] First, most existing solutions model trajectory prediction as a recursive or sequence generation over discrete time steps, which is fundamentally architecturally mismatched with the reality that the motion of unmanned systems is essentially a continuous-time-domain physical process. Discretization inevitably introduces approximation errors in the time dimension, losing the fine-grained spatiotemporal variations in the continuous motion process. This mismatch severely limits the model's ability to capture real dynamics, especially when the sensor sampling frequency is low or the target is highly maneuverable. Second, multi-source heterogeneous sensor data often have different sampling frequencies and arrival times, and may be missing due to communication packet loss. Most existing methods require forced time alignment or interpolation preprocessing of the data from each channel, converting it into a unified discrete-time grid before fusion. This not only introduces additional errors but also causes a sharp deterioration in the fusion effect when data is missing, significantly weakening the robustness of the system in practical deployment. Furthermore, in existing solutions, navigation modality recognition and trajectory prediction are typically designed as two relatively independent modules, with classification followed by prediction or parallel processing. There is a lack of deep coupling and bidirectional constraints between semantic-level behavioral cognition and physical-level motion deduction. Errors in modality recognition are directly propagated to trajectory prediction and cannot be corrected. Simultaneously, the physical consistency information in trajectory evolution fails to feed back into modality recognition, limiting overall performance in complex interaction scenarios. Finally, deep learning methods often require stacking numerous network layers or processing long time-series sequences, storing a large number of intermediate activation values ​​during training and inference. Memory overhead increases linearly over time, making it difficult to meet deployment requirements for embedded platforms with limited computing resources and power consumption, such as unmanned surface vessels and drones.

[0005] In summary, existing unmanned system trajectory prediction technologies still have significant shortcomings in areas such as continuous physical dynamic modeling, heterogeneous asynchronous data fusion, deep coupling of behavioral semantics and motion physics, and efficient deployment at the edge.

[0006] In view of the above, this application is hereby submitted. Summary of the Invention

[0007] This invention provides a method, apparatus, device, and medium for unmanned system intent classification and trajectory prediction, which can at least partially improve the above-mentioned problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A method for intention classification and trajectory prediction in unmanned systems is proposed, which uses a pre-trained MCGCN multi-channel graph convolutional network model for processing. The specific processing steps are as follows: Acquire multi-channel signals collected in real time by preset sensor components and construct a multi-channel attribute atlas; Based on the multi-channel attribute map set, the rate of change corresponding to the attribute map information on each channel is calculated, and the rate of change of all channels is spliced ​​to obtain the global hidden potential field. Based on the global hidden potential field, second-order partial differential calculations are performed to obtain the convergence results of each channel. Then, a nonlinear projector is used to project the fused results of the convergence results of each channel to obtain the label prediction results.

[0010] The present invention also provides an unmanned system intent classification and trajectory prediction device, which is used to implement the unmanned system intent classification and trajectory prediction method as described in any one of the above claims, comprising: The data acquisition unit is used to acquire multi-channel signals collected in real time by preset sensor components and construct a multi-channel attribute map. The global hidden potential field calculation unit is used to calculate the rate of change corresponding to the attribute map information on each channel based on the multi-channel attribute map set, and to stitch together the rate of change of all channels to obtain the global hidden potential field. The prediction unit is used to perform second-order partial differential calculations based on the global hidden potential field to obtain the convergence results of each channel, and then uses a nonlinear projector to project the fused results of the convergence results of each channel to obtain the label prediction results.

[0011] The present invention also provides an unmanned system intent classification and trajectory prediction device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the unmanned system intent classification and trajectory prediction method as described above.

[0012] The present invention also provides a readable storage medium storing a computer program that can be executed by a processor of the device in which the storage medium is located, to implement the unmanned system intent classification and trajectory prediction method as described in any of the above claims.

[0013] In summary, this invention provides a method for intention classification and trajectory prediction of unmanned systems, aiming to address the core challenges faced by existing unmanned surface vessels, unmanned aerial vehicles (UAVs), and other unmanned driving systems in complex dynamic environments, such as continuous physical modeling, heterogeneous data fusion, and efficient edge deployment. The core technical path of this solution is as follows: First, multi-channel signals collected in real time by preset sensor components such as radar, AIS, and vision are acquired to construct a multi-channel attribute map set. Each channel attribute map shares a node set and has its own independent adjacency matrix and feature matrix, thus preserving data diversity while providing a complete input dimension for subsequent processing. Based on this, the rate of change corresponding to the information in each channel attribute map is calculated, and the rate of change of all channels is spliced ​​to obtain a global latent potential field, realizing the mapping from discrete sensor observations to a continuous unified state space. Then, second-order partial differential calculations are performed based on the global latent potential field to obtain the convergence results of each channel. The fused result of the convergence results of each channel is projected using a nonlinear projector to obtain the label prediction result, simultaneously completing the identification of the target ship / UAV's navigation mode and the continuous inference of its future trajectory.

[0014] Compared with the prior art, the present invention has the following beneficial effects: (1) By splicing the multi-channel rate of change into a global latent potential field and performing evolutionary solution based on second-order partial differential equation, the deep coupling of semantic behavior recognition and physical motion inference is realized, which significantly improves the trajectory prediction accuracy in complex interactive scenarios; (2) By using continuous integral instead of discrete stepping, the high-order continuity of the output trajectory is ensured from the mathematical bottom layer, so that the prediction results fully comply with the physical constraints of hydrodynamics / flight dynamics of unmanned systems, effectively reducing the mechanical loss of the underlying propulsion / control system; (3) Continuous dynamic modeling enables the system to provide highly reliable prediction compensation by utilizing dynamic inertia in the case of sensor sampling frequency mismatch or instantaneous communication disconnection, without the need for forced time alignment, which greatly enhances the system's anti-asynchronous interference capability and robustness in actual deployment; The present invention can be widely applied to various unmanned driving systems such as unmanned ships, drones and unmanned ground vehicles. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the unmanned system intent classification and trajectory prediction method provided in the first embodiment of the present invention.

[0016] Figure 2 This is a simplified schematic diagram illustrating the test multi-view dataset provided in an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of ACC (%) performance on a text attribute graph and a multi-view text dataset provided in an embodiment of the present invention.

[0018] Figure 4 This is a schematic diagram illustrating the accuracy (%) performance on a multi-feature image dataset provided in an embodiment of the present invention.

[0019] Figure 5 This is a visualization diagram of t-SNE visualization of the comparison model and the proposed method on a handwritten dataset provided in an embodiment of the present invention.

[0020] Figure 6 This is a schematic diagram illustrating the performance of the method provided in this embodiment of the invention on text attribute graphs under different Euler schemes.

[0021] Figure 7 This is a schematic diagram illustrating the sensitivity analysis of α and N on four multi-view datasets in the method provided by this embodiment of the invention.

[0022] Figure 8 This is a schematic diagram illustrating the performance of the method on the Cora and BBC Sports datasets under different training ratios, as provided in this embodiment of the invention.

[0023] Figure 9 This is a schematic diagram of the module of the unmanned system intent classification and trajectory prediction device provided in the second embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] refer to Figure 1 As shown, the first embodiment of the present invention discloses a method for unmanned system intent classification and trajectory prediction, which uses a pre-trained MCGCN multi-channel graph convolutional network model for processing; it can be executed by an unmanned system intent classification and trajectory prediction device (hereinafter referred to as prediction device), specifically, by one or more processors within the prediction device, to implement the following method: S1, acquire multi-channel signals collected in real time by preset sensor components, and construct a multi-channel attribute map; Specifically, step S1 further includes: the multi-channel attribute atlas The expression is: , , This represents the attribute graph information for the v-th channel, where V is the total number of channels. For a set of nodes, This is the adjacency matrix unique to the v-th channel. This is the feature matrix unique to the v-th channel.

[0026] In this embodiment, the multi-channel attribute map set contains attribute map information for multiple channels. Each attribute map information includes a shared node set, a channel-specific adjacency matrix, and a channel-specific feature matrix. Each node in the node set represents a tracked physical entity, such as a ship or a drone. All channels share the same node set, meaning that different sensors are observing the same set of physical objects. Different channels correspond to different types of sensor components. For example, in an unmanned surface vessel (USV) scenario, the radar channel provides the target's position and radial velocity information, the AIS channel provides the ship's static and dynamic navigation data, and the visual channel provides the target's image features. The feature matrix dimensions of each channel can be different to adapt to the original data format of each sensor, demonstrating the inherent compatibility of this method with multi-source heterogeneous data. The adjacency matrix unique to each channel is independently constructed based on the channel's physical characteristics, such as determining the connection weights between nodes based on spatial distance between targets, feature similarity, or other correlation metrics. By constructing a multi-channel attribute atlas, this method preserves the original characteristics of each sensing mode at the data input level, providing a structured input foundation for subsequent multi-channel deep fusion and avoiding the information loss caused by forcibly aligning all sensor data to the same feature space in traditional methods.

[0027] S2, based on the multi-channel attribute map set, calculate the rate of change corresponding to the attribute map information on each channel, and stitch together the rate of change of all channels to obtain the global hidden potential field. Specifically, step S2 further includes: modeling the propagation of attribute graph information on each channel as a continuous dynamic system controlled by ODE, and obtaining the rate of change, the formula of which is: , Let be the continuous hidden state of node i in the v-th channel at time t. Let be the continuous hidden states of node i in the v-th channel at the initial time. For continuous hidden states The partial derivative with respect to time t is expressed as the rate of change. It is a self-evolutionary function. The adjacency matrix unique to the v-th channel The value of the element in the i-th row and j-th column. For paired interactive functions, Let be the continuous hidden state of node j in the v-th channel at time t. For the dedicated encoder of the v-th channel, Let be the original input feature vector of node ii in the v-th channel; By concatenating the rates of change of all channels, the global hidden potential field is obtained, and its formula is: , The global hidden potential field is represented by the continuous coordinates at time t. An abstract scalar or potential function on [a surface]. , representing continuous coordinates in the channel space. To concatenate additivity operators, The continuous hidden states of node i in the first channel at time t Partial derivative with respect to time t, Let i be the continuous hidden state of node i in the Vth channel at time t. Partial derivative with respect to time t, For splicing, Add to group.

[0028] In this embodiment, for each node in each channel, the evolution of its continuous hidden state in the time domain is driven by two parts: the first part is the self-evolution function, which characterizes the autonomous inertial dynamics of the node's own state changing over time, such as the inherent tendency of a ship to maintain its course and speed without external interference; the second part is the neighborhood interaction term, which characterizes the comprehensive influence of surrounding targets on the node's state evolution by calculating and accumulating the pairwise interaction functions of the continuous hidden states of the neighboring nodes corresponding to the non-zero elements in the adjacency matrix of the channel. By modeling the propagation of the attribute graph information of each channel as a continuous dynamic system controlled by ordinary differential equations to obtain the rate of change, this method fundamentally differs from the traditional approach of stacking discrete propagation layers in graph neural networks. This allows the update of the node state to be no longer limited by a fixed number of discrete layers, but to evolve smoothly in the continuous time domain, thereby more accurately capturing the fine-grained spatiotemporal changes of targets such as unmanned ships and drones during continuous motion. The initial value of the continuous hidden state of each node in each channel is obtained by projecting the original input feature vector of that node onto the dedicated encoder of that channel. This encoder maps the original data of different sensor modes to a unified hidden space, enabling the original data from heterogeneous sensors with different physical meanings and dimensions to be processed uniformly in the same representation space. The self-evolutionary function and the paired interaction function are both learnable nonlinear mappings, and their specific forms are determined through optimization during the model pre-training stage. Different channels can share the same set of function forms or set them separately according to the physical characteristics of each channel. By modeling the propagation of graph information of each channel as a continuous dynamic system controlled by ordinary differential equations, this method introduces the constraint of time continuity in the feature extraction stage, so that the feature representation obtained by subsequent fusion naturally carries the smoothness in the time domain, laying a mathematical foundation for the final generation of continuous prediction trajectories that conform to physical laws. When the sensor sampling interval is uneven or instantaneous data loss occurs, this continuous dynamic system can autonomously fill the observation gap using dynamic inertia without the need for additional interpolation or alignment modules, significantly improving the deployment robustness of the system in the actual maritime environment.

[0029] After obtaining the latent state change rates of each node in each channel, the prediction device concatenates the change rates of all channels to obtain the global latent potential field. Specifically, this method introduces the mathematical property of concatenation additivity, connecting the independent latent state change rates of each channel end-to-end to concatenate a unified global representation, which is defined as the global latent potential field. Its core idea is to treat the dynamic information observed by multiple sensor channels as projections of the same underlying physical world in different sensing directions. By applying the concatenation additivity operator during the concatenation process, this method ensures that the result of calculating the change rates of each channel separately before concatenation is mathematically rigorously equivalent to the result of treating all channels as a whole and evolving them uniformly. This provides a theoretical guarantee for transforming the multi-channel fusion problem into a global latent potential field evolution problem in subsequent steps. The construction of the global latent potential field is one of the core innovations of this invention. By concatenating the change rates of all channels into a unified whole, it achieves a leap from "independent calculation of multiple channels" to "global unified representation," enabling subsequent second-order partial differential calculations to simultaneously process heterogeneous information from all channels within a unified mathematical framework. Within this framework, each channel is no longer treated as an isolated sensing source, but rather as a partial derivative of the same global situation field in different coordinate directions. This allows for the full utilization of the inherent correlation and complementarity between channels. This unified representation enables the deep integration of semantic-level behavioral cognition and physical-level motion deduction within the same mathematical framework. Compared to the traditional approach where modality recognition and trajectory prediction are separated, this method effectively avoids the problem of unidirectional error propagation between modules without correction, significantly improving the overall prediction accuracy in complex interactive scenarios.

[0030] S3. Based on the global hidden potential field, second-order partial differential calculations are performed to obtain the convergence results of each channel. Then, a nonlinear projector is used to project the fused results of the convergence results of each channel to obtain the label prediction results.

[0031] Specifically, step S3 further includes: based on the global hidden potential field Second-order partial differential equations are performed to achieve convergence for the v-th channel, and the formula is as follows: , for For time t and position coordinates The second-order partial derivative, For time t and position coordinates An abstract scalar or potential function on [a surface]. Let v be the position coordinates of the v-th channel in the continuous channel coordinate space. for For position coordinates The partial derivatives, For self-evolution operators, For interactive operators; Let the variable After all channels converge By merging, the evolved latent representations are obtained. ; Call the preset non-linear projector The evolved latent representation Mapping to the label space yields the label prediction results. .

[0032] In this embodiment, second-order partial differential equations are performed based on the global latent potential field to achieve independent convergence for each channel. The second-order partial differential equation calculation involves simultaneously differentiating the global latent potential field in both the time and channel coordinate dimensions. Specifically, the first-order partial derivative of the global latent potential field with respect to the channel coordinates is first obtained, and then the partial derivative of that first-order partial derivative with respect to time is obtained, thus yielding the second-order partial differential result. The physical meaning of this calculation process is that the dynamic information observed by each sensor channel is considered as the projection gradient of the same global latent potential field in the corresponding channel coordinate direction. Therefore, the evolution law of each channel is naturally determined by the second-order changes of this global latent potential field in different directions. The second-order partial differential expression includes two components: a self-evolution operator and an interaction operator. The self-evolution operator characterizes the autonomous intrinsic dynamics of the node's own state changing over time within the global latent potential field framework, while the interaction operator characterizes the mutual influence between nodes in the graph topology defined by the adjacency matrices of each channel. By solving this second-order partial differential equation, the system can obtain the convergence results of each channel in the continuous time domain, that is, the steady-state or terminal evolution values ​​of the hidden states of each channel under the condition of satisfying dynamic constraints. This method uses second-order partial differential equations instead of first-order equations to describe the convergence process of each channel. Its advantage lies in the fact that the second-order form naturally includes information on acceleration changes, so that the convergence results not only reflect the magnitude of the state itself, but also include the trend and curvature information of the state changes. This provides a higher-order physical constraint for the subsequent generation of highly smooth predicted trajectories, fundamentally ensuring the continuity of the trajectory in terms of position, velocity, and acceleration, and meeting the stringent requirements of unmanned surface vessel hydrodynamics and unmanned aerial vehicle flight dynamics for trajectory smoothness. Furthermore, treating each channel as a projection of the same global potential field in different directions means that the channels are no longer isolated from each other, but rather form an inherent coupling relationship through the global hidden potential field. This coupling allows the observation information of one channel to be transmitted to the evolution process of other channels through the potential field, thereby achieving true deep fusion of multi-source information and significantly improving the system's fault tolerance and prediction reliability when a single sensor fails or is interfered with.

[0033] After completing the second-order partial differential equation calculation and obtaining the convergence results of each channel, the prediction device fuses the converged hidden states of all channels to obtain the evolved hidden representation. Specifically, the fusion operation involves concatenating the converged hidden state vectors of each channel end-to-end to form a unified global representation vector. The mathematical validity of this concatenation operation is guaranteed by the aforementioned property of concatenation additivity; that is, the result of first performing second-order partial differential convergence on each channel and then concatenating the vectors is mathematically strictly equivalent to the result of treating all channels as a whole and performing a unified evolution. Through this fusion operation, the complementary information of each channel after convergence converges into the same high-dimensional representation space, providing a complete input foundation for subsequent label prediction. The evolved hidden representation contains both the individual dynamic information carried by the independent convergence of each channel and the intrinsic correlation information between channels within the framework of the global hidden potential field, making it a key intermediate result for the deep integration of semantic recognition and physical deduction. Compared to the traditional approach of simply weighting and averaging or splicing the features of each channel before feeding them into the classifier, the fusion representation obtained by the continuous evolution driven by second-order partial derivatives has clear physical meaning and dynamic constraints. Its numerical distribution is more stable and more robust to noise, which can significantly reduce the false recognition rate in subsequent classification tasks.

[0034] Subsequently, a pre-defined nonlinear projector is invoked to map the evolved latent representations into the label space, yielding label prediction results. The nonlinear projector is preferably a multi-layer fully connected network structure, with its input dimension equal to the total dimension of the evolved latent representations and its output dimension equal to the pre-defined total number of navigation modal categories. During the mapping process, the nonlinear projector reduces the dimensionality of the high-dimensional latent representations layer by layer through multi-layer nonlinear transformations, ultimately outputting a vector with a dimension equal to the total number of categories. This vector, after being processed by a normalized exponential function, represents the predicted probability distribution of each navigation modality. The navigation modalities can include, but are not limited to, typical behavior categories such as heading and speed maintenance, cooperative obstacle avoidance, and near-shore berthing. Specific category settings can be flexibly defined according to the actual application scenario; in addition to unmanned surface vessel scenarios, it also includes unmanned aerial vehicle scenarios, autonomous vehicle scenarios, etc.

[0035] Preferably, in this embodiment, the training steps of the MCGCN multi-channel graph convolutional network model specifically include: initializing model parameters. Its formula is: , These are channel-specific parameters within the encoder. These are the classifier parameters; The training data is input into the MCGCN multi-channel graph convolutional network model to generate training results. Based on the training results, the model parameters are updated by minimizing the cross-entropy loss function through gradient descent. The formula for the cross-entropy loss function is as follows: , For the real label matrix, To train the predicted label matrix, Let C be the training sample index set, and C be the number of classes. For the first The sample at the th The true label indicator value for each category, For the first The sample at the th Training prediction label indication values ​​for each category; Repeat the training steps until the model converges or reaches the preset number of iterations to complete the model training.

[0036] Specifically, in this embodiment, the model parameters are first initialized. The parameters of the MCGCN network model mainly include learnable parameters within each channel's dedicated encoder and learnable parameters within the classifier. These two types of parameters together constitute the optimization object of the entire model. Each channel's dedicated encoder is responsible for projecting the original input features of the corresponding channel onto a unified latent space. The quality of its parameters directly determines whether the original sensor data can be effectively extracted into high-level semantic features. The classifier parameters are responsible for mapping the evolved latent representation to the label space. The quality of its parameters directly determines the accuracy of the final modality recognition. At the beginning of training, random initialization or parameter initialization based on an existing pre-trained model can be used to give the model a reasonable starting point for parameters, accelerating the convergence process of subsequent training. Parameter initialization is a routine operation in deep learning model training. The specific initialization strategy can be flexibly selected according to the actual application scenario, such as using normal distribution random sampling or Xavier initialization.

[0037] After parameter initialization, the training data is input into the MCGCN multi-channel graph convolutional network model to generate training prediction results. The training data consists of pre-collected and labeled multi-channel sensor trajectory data, with each training sample containing a multi-channel attribute map and its corresponding real navigation modality label. After the training data is input into the model, the model performs forward computation according to the aforementioned steps S1 to S3: first, the encoders of each channel extract features and concatenate them into a global state vector; then, the evolved latent representation is obtained through continuous-time dynamic system evolution; and finally, the nonlinear projector outputs the training prediction label matrix. Through this forward propagation process, the model maps the multi-channel input of the training data to the predicted probability distribution of the navigation modality to which each sample belongs, completing the end-to-end transformation from raw sensor data to semantic labels. Unlike traditional approaches that require separate training of modality recognition and trajectory prediction modules, this method uses an end-to-end unified training framework, enabling modality recognition and trajectory prediction to share the same underlying feature representation. The two promote each other during training—the trajectory evolution task provides physical constraints for representation learning, while the modality recognition task provides semantic guidance for representation learning, jointly driving the model to learn more robust and discriminative feature representations.

[0038] After obtaining the training prediction results, the model parameters are updated based on these results by minimizing the cross-entropy loss function using the gradient descent algorithm. Specifically, the cross-entropy loss function measures the difference between the probability distribution of the model's predicted output for training samples and the true labels; a smaller value indicates a more accurate prediction. In this loss function, the training sample index set represents all sample nodes with true labels. Typically, in real maritime scenarios, only a small number of samples can obtain accurate labels, which is a typical characteristic of label-scarce scenarios. The number of categories represents the total number of preset navigation modal categories, such as heading and speed maintenance, cooperative obstacle avoidance, and near-shore berthing. The true label indicator value is either 0 or 1; it is 1 when the true category of the sample is the corresponding category, and 0 otherwise. The training prediction label indicator value is the model's predicted probability that the sample belongs to that category, and its value is between 0 and 1. The gradient descent algorithm calculates the partial derivatives of the loss function with respect to each model parameter to obtain the direction and magnitude of parameter updates, adjusting the parameters in the direction that reduces the loss function value. Since this method uses the adjoint state method to calculate the gradient, it does not need to store a large number of intermediate states during the integration process when performing backpropagation. Instead, it derives the gradient in reverse by solving an adjoint differential equation. This design effectively solves the problem of the memory overhead growing linearly with time in long-term trajectory prediction, enabling the complex MCGCN fusion architecture to be trained on small batches of samples and thus successfully deployed on resource-constrained embedded platforms such as unmanned ships and drones.

[0039] Finally, the above training steps are repeated, i.e., the iterative process of "input training data—forward computation to generate predictions—calculate loss—gradient descent to update parameters" is executed repeatedly until the model converges or reaches the preset number of iterations, completing the model training. The criteria for model convergence can be that the loss function value no longer decreases significantly in consecutive iterations, or the prediction accuracy on the validation set no longer improves. During training, the model parameters are fine-tuned according to the gradient of the loss function in each iteration, gradually approaching the optimal parameter combination. Through a sufficient number of iterations, the model's predictive ability gradually improves, eventually converging to a stable state that can accurately identify various navigation modes. It is worth noting that the continuous dynamics framework used in this method also implicitly includes the effect of physical regularization during training—because the evolution process itself is constrained by partial differential equations or ordinary differential equations, the model cannot learn physically infeasible state jumps or trajectory mutations. This is equivalent to introducing a soft physical prior constraint during training, which helps reduce the model's dependence on the amount of labeled data, enabling a model with good generalization ability to be obtained even with a small number of labeled samples. After completing the above training steps, the MCGCN multi-channel graph convolutional network model can be used to execute the unmanned system intent classification and trajectory prediction method described in steps S1 to S3 above, so as to realize the real-time identification of the navigation modes of surrounding targets and the continuous prediction of future trajectories.

[0040] Specifically, in this embodiment, to further illustrate the effectiveness of the method, the method is evaluated on various real-world datasets using a classification task.

[0041] Test dataset setup: To comprehensively evaluate effectiveness and robustness, experiments were conducted on diverse benchmark datasets, covering text attribute graphs, multi-view text data, and visual multi-feature image data. Dataset statistics are as follows: Figure 2 As shown. The training set and test set are divided at 10% and 90% respectively. Among them, Figure 2 Citeseer, Cora, Cornell, Texas, 20Newsgroups, 3Sources, BBCNews, BBCSports, Animals, Caltech7, Handwritten, and SUNRGBD are all datasets.

[0042] Please see Figure 3 , Figure 4 Comparison methods: Comparison with the current best methods, including MMdynamics, LGCNFF, ECMGD, DG-MVL, RCML, FUML, SMMRL, and TUNED.

[0043] Implementation details: MCGCN models multi-view feature propagation based on Laplace diffusion, employing the Euler scheme for discrete continuous dynamics. For each view v, a k-nearest neighbor (kNN) graph is constructed from the input features to obtain the adjacency matrix A. v This leads to the derivation of the normalized graph Laplace L. v All experiments employed the forward Euler scheme (MCGCN-F), with the iterative propagation rule being: H (n+1) =(I αL)H (n) n=0,…,N 1. Where α is the diffusion intensity (controlling the neighborhood propagation intensity), and N is the propagation depth. Default hyperparameters: k=10, N=2, α=3.

[0044] Quantitative results: Figure 3 , Figure 4 The accuracy (ACC) is the result for all multi-view datasets. The best and second-best results are marked in bold and underline, respectively.

[0045] Citation and webpage graph datasets: sparse and noisy, performance depends on stable neighborhood propagation (without excessive smoothing). Methods such as LGCNFF and DG-MVL rely on learnable fusion or feature decoupling, and their performance is weak / unstable when structural information dominates.

[0046] Multi-view text and image datasets: channels are complementary but often subject to noise interference. Uncertainty-aware methods such as MMdynamics, RCML, and FUML mainly operate at the decision level and do not explicitly couple multi-channel interaction with graph propagation.

[0047] The method presented in this paper interprets feature propagation as a numerical solution of a continuous-time dynamic system, which allows for explicit control of propagation intensity and depth, resulting in a more stable representation.

[0048] Qualitative results: Figure 5 The t-SNE visualizations for each model on the Handwritten dataset are presented. The representations generated by our method exhibit clear inter-cluster separation, compact intra-cluster distribution, and strong discriminative ability; in contrast, the compared methods show inter-class overlap or cluster fragmentation, indicating insufficient view alignment or unstable feature propagation. In summary, our method achieves ideal performance on almost all benchmarks, validating the effectiveness and robustness of the unified multi-channel graph convolution framework.

[0049] at last, Figure 6The ablation results on the graph dataset are presented. The performance gap between the diffusion-free (w / oDiffusion) model (removing the PDE-driven propagation and retaining only the view encoder and fusion head) and the complete model indicates that the performance improvement mainly comes from the proposed diffusion-based multi-channel propagation, rather than encoder capacity. Among the diffusion strategies, forward Euler (MCGCN-F) consistently achieves the best performance, while backward Euler (MCGCN-B) and improved Euler (MCGCN-M) perform slightly worse but are still competitive.

[0050] Figure 7 This study investigates the impact of diffusion intensity α on propagation depth N. A larger α enhances neighborhood aggregation within a single step, accelerating information propagation when N is small. In high-dimensional multi-view text datasets, medium to large α combined with shallow or medium propagation consistently achieves excellent performance, indicating effective cross-view information exchange. Graph datasets are more sensitive to parameters; large α or deep propagation leads to over-smoothing due to sparse and rapid connections. For multi-feature image datasets, medium α balances noise suppression and discriminative feature preservation; excessive diffusion causes inter-class interference. Overall, medium diffusion intensity combined with shallow propagation is a robust setting across datasets.

[0051] Figure 8 The performance of our method is demonstrated on two datasets with different training ratios. On the Cora citation network, our method consistently achieves higher accuracy across all training ratios, with a particularly significant advantage in low-label scenarios. On BBCSports, where view semantic alignment is good and labels are relatively abundant, all methods benefit from increased training data, but our method consistently maintains the performance ceiling. This indicates that unified multichannel evolution not only improves robustness under sparse supervision but also preserves discriminative representations when labels are abundant.

[0052] In summary, the core architecture of this invention lies in utilizing a global latent potential field as a unified mathematical framework for all channels. The independent dynamics of each sensor channel are treated as projections of the same potential field in different directions. The convergence process of each channel is described by second-order partial differential equations, enabling intrinsic coupling between channel information during evolution. Observational biases in a single channel can be corrected by transmitting the potential field to other channels, achieving deep and substantial fusion of multi-source heterogeneous sensor data. During the training phase, this invention employs an end-to-end unified training framework, using the adjoint state method to achieve gradient calculation with low memory usage. This allows complex continuous graph dynamics models to be deployed and updated online on resource-constrained embedded platforms. This invention can be widely applied to various unmanned driving systems such as unmanned surface vessels, drones, and unmanned ground vehicles, providing core technical support for autonomous obstacle avoidance and safe navigation in complex dynamic environments.

[0053] Compared with the prior art, the present invention has the following beneficial effects: First, by modeling the propagation of attribute map information of each channel as a continuous dynamic system controlled by ordinary differential equations, and by performing second-order partial differential calculations starting from the global hidden potential field to achieve convergence of each channel, a deep coupling between semantic-level behavioral intent recognition and physical-level motion trajectory deduction is achieved. This solves the fundamental problem in traditional schemes where modality recognition and trajectory prediction are separated, leading to unidirectional error propagation and the inability to correct it. This significantly improves the trajectory prediction accuracy and modality recognition accuracy in complex interactive scenarios.

[0054] Second, the convergence process of each channel is described by second-order partial differential equations in the continuous time domain. The second-order form naturally contains the change information at the acceleration level, so that the convergence result reflects the state magnitude, change trend and curvature information at the same time. This ensures the high-order continuity of the output trajectory at the position, velocity and acceleration levels from the mathematical level, so that the prediction result fully conforms to the physical constraints of the hydrodynamics or flight dynamics of the unmanned system, effectively reducing the mechanical loss and energy consumption of the underlying propulsion system or control surface.

[0055] Third, the framework design of splicing multi-channel change rates into a global hidden potential field gives the system a strong ability to resist asynchronous interference. When the sensor sampling frequency is mismatched or a momentary communication interruption occurs, there is no need for discretization alignment or interpolation preprocessing. The system can autonomously fill the missing observation interval using continuous dynamic inertia and automatically achieve trajectory smoothing correction based on the differential relationship between the old and new states. This greatly enhances the system's deployment robustness in actual maritime environments such as communication packet loss and asynchronous sensor sampling.

[0056] Fourth, by introducing the adjoint state method to calculate the gradient during the training phase, the massive intermediate states in the integration process are avoided, so that the memory overhead does not increase linearly with the increase of the prediction time span. This fundamentally solves the technical bottleneck that complex models in long-term trajectory prediction are difficult to deploy on resource-constrained platforms, enabling the fusion architecture proposed in this invention to achieve efficient online learning and real-time prediction on embedded boards such as unmanned ships and drones.

[0057] Please see Figure 9 A second embodiment of the present invention provides an unmanned system intent classification and trajectory prediction device, which is used to implement the unmanned system intent classification and trajectory prediction method as described in any one of the above claims, including: The data acquisition unit 101 is used to acquire multi-channel signals collected in real time by preset sensor components and construct a multi-channel attribute map. The global hidden potential field calculation unit 102 is used to calculate the rate of change corresponding to the attribute map information on each channel based on the multi-channel attribute map set, and to stitch together the rate of change of all channels to obtain the global hidden potential field. The prediction unit 103 is used to perform second-order partial differential calculations based on the global hidden potential field to obtain the convergence results of each channel, and to project the fusion result of the convergence results of each channel using a nonlinear projector to obtain the label prediction result.

[0058] A third embodiment of the present invention provides an unmanned system intent classification and trajectory prediction device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the unmanned system intent classification and trajectory prediction method as described above.

[0059] The fourth embodiment of the present invention provides a readable storage medium storing a computer program that can be executed by the processor of the device where the storage medium is located to implement the unmanned system intent classification and trajectory prediction method as described in any of the above embodiments.

[0060] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intention classification and trajectory prediction in unmanned systems, characterized in that, The pre-trained MCGCN multi-channel graph convolutional network model is used for processing. The specific processing steps are as follows: Acquire multi-channel signals collected in real time by preset sensor components and construct a multi-channel attribute atlas; Based on the multi-channel attribute map set, the rate of change corresponding to the attribute map information on each channel is calculated, and the rate of change of all channels is spliced ​​to obtain the global hidden potential field. Based on the global hidden potential field, second-order partial differential calculations are performed to obtain the convergence results of each channel. Then, a nonlinear projector is used to project the fused results of the convergence results of each channel to obtain the label prediction results.

2. The method for unmanned system intent classification and trajectory prediction according to claim 1, characterized in that, The multi-channel attribute atlas The expression is: , , This represents the attribute graph information for the v-th channel, where V is the total number of channels. For a set of nodes, This is the adjacency matrix unique to the v-th channel. This is the feature matrix unique to the v-th channel.

3. The unmanned system intent classification and trajectory prediction method according to claim 2, characterized in that, Based on the multi-channel attribute map atlas, the rate of change corresponding to the attribute map information on each channel is calculated, and the rate of change of all channels is concatenated to obtain the global latent potential field, specifically: The propagation of attribute map information on each channel is modeled as a continuous dynamic system controlled by ODE, and the rate of change is obtained, the formula of which is: ; Let be the continuous hidden state of node i in the v-th channel at time t. Let be the continuous hidden states of node i in the v-th channel at the initial time. For continuous hidden states The partial derivative with respect to time t is expressed as the rate of change. It is a self-evolutionary function. The adjacency matrix unique to the v-th channel The value of the element in the i-th row and j-th column. For paired interactive functions, Let be the continuous hidden state of node j in the v-th channel at time t. For the dedicated encoder of the v-th channel, Let be the original input feature vector of node ii in the v-th channel; By concatenating the rates of change of all channels, the global hidden potential field is obtained, and its formula is: , The global hidden potential field is represented by the continuous coordinates at time t. An abstract scalar or potential function on [a surface]. , representing continuous coordinates in the channel space. To concatenate additivity operators, The continuous hidden states of node i in the first channel at time t Partial derivative with respect to time t, Let i be the continuous hidden state of node i in the Vth channel at time t. Partial derivative with respect to time t, For splicing, Add to group.

4. The method for unmanned system intent classification and trajectory prediction according to claim 3, characterized in that, Based on the global latent potential field, second-order partial differential equations are performed to obtain the convergence results of each channel. Then, a nonlinear projector is used to project the fused convergence results of each channel to obtain the label prediction results, specifically: According to the global hidden potential field Second-order partial differential equations are performed to achieve convergence for the v-th channel, and the formula is as follows: , for For time t and position coordinates The second-order partial derivative, For time t and position coordinates An abstract scalar or potential function on [a surface]. Let v be the position coordinates of the v-th channel in the continuous channel coordinate space. for For position coordinates The partial derivatives, For self-evolution operators, For interactive operators; Let the variable After all channels converge By merging, the evolved latent representations are obtained. ; Call the preset non-linear projector The evolved latent representation Mapping to the label space yields the label prediction results. .

5. The method for unmanned system intent classification and trajectory prediction according to claim 1, characterized in that, The specific training steps for the MCGCN multi-channel graph convolutional network model are as follows: Initialize model parameters Its formula is: , These are channel-specific parameters within the encoder. These are the classifier parameters; The training data is input into the MCGCN multi-channel graph convolutional network model to generate training results. Based on the training results, the model parameters are updated by minimizing the cross-entropy loss function through gradient descent. The formula for the cross-entropy loss function is as follows: , For the real label matrix, To train the predicted label matrix, Let C be the training sample index set, and C be the number of classes. For the first The sample at the th The true label indicator value for each category, For the first The sample at the th Training prediction label indication values ​​for each category; Repeat the training steps until the model converges or reaches the preset number of iterations to complete the model training.

6. A device for classifying intent and predicting trajectory in an unmanned system, characterized in that, A method for implementing the unmanned system intent classification and trajectory prediction method as described in any one of claims 1 to 5, comprising: The data acquisition unit is used to acquire multi-channel signals collected in real time by preset sensor components and construct a multi-channel attribute map. The global hidden potential field calculation unit is used to calculate the rate of change corresponding to the attribute map information on each channel based on the multi-channel attribute map set, and to stitch together the rate of change of all channels to obtain the global hidden potential field. The prediction unit is used to perform second-order partial differential calculations based on the global hidden potential field to obtain the convergence results of each channel, and then uses a nonlinear projector to project the fused results of the convergence results of each channel to obtain the label prediction results.

7. A device for classifying intent and predicting trajectory in an unmanned system, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the unmanned system intent classification and trajectory prediction method as described in any one of claims 1 to 5.

8. A readable storage medium, characterized in that, The storage medium contains a computer program that can be executed by a processor of the device in which the storage medium resides, to implement the unmanned system intent classification and trajectory prediction method as described in any one of claims 1 to 5.

Images