A method for cooperative perception of a mobile target based on edge intelligent aerial unmanned system

Abstract

The application discloses a kind of based on edge intelligent air unmanned system's mobile target cooperative perception method, it is related to unmanned system cooperative perception technical field.The present application includes four steps of heterogeneous sensor data acquisition and sliding window local cache, track-data-global number cooperative association, federal learning distributed Gaussian process model cooperative optimization, track prediction and uncertainty quantification fusion, information gain driven intelligent auction task allocation, and the intelligent closed loop of perception-uncertainty evaluation-resource scheduling is jointly constructed.The present application uses Gaussian process non-parametric modeling to get rid of fixed motion model dependence, reduces communication overhead by lightweight interaction and asynchronous relaxation consistency, realizes the optimal configuration of perception resource by dynamic auction mechanism, significantly improves high-maneuvering target tracking accuracy and robustness, can realize wide-area cooperative situation awareness stably in bandwidth limited, heterogeneous platform, strong interference scene, and is suitable for low-altitude supervision, regional security, search and rescue and the like scene.

Description

Technical Field

[0001] This invention relates to collaborative perception technology for unmanned systems, and more particularly to a collaborative perception method for maneuvering targets based on an edge-intelligent aerial unmanned system. Background Technology

[0002] Unmanned system collaborative perception refers to multiple distributed observation platforms connected via wireless networks sharing information and collaborating on computation. The aim is to construct a unified, accurate, and real-time wide-area situational awareness picture. Its core objective is to overcome the limitations of single platforms in terms of field of view, continuous monitoring capabilities, and reliability, thereby achieving stable tracking of non-cooperative, highly maneuverable targets. This technology is crucial in applications such as regional surveillance, air defense early warning, and low-altitude emergency management.

[0003] However, building such a system faces several interconnected fundamental technical challenges. First, adaptive perception and prediction of complex maneuvering targets is the primary challenge. Traditional tracking algorithms rely on pre-defined motion models, such as typical uniform speed or uniform acceleration models. When the target performs nonlinear maneuvers such as sharp turns or speed changes, model mismatch can lead to decreased tracking accuracy or even target loss. While improved methods based on multi-model switching exist, their generalization ability to unknown maneuvering patterns is limited, and the maintenance and optimization of model sets are complex, resulting in long response times. Therefore, there is an urgent need for an edge data-driven intelligent target trajectory perception framework that reduces reliance on fixed models.

[0004] Secondly, achieving efficient collaboration under limited communication and heterogeneous platform conditions is a key bottleneck in the implementation of this method. In actual deployments, communication bandwidth between platforms is limited, and links may be unstable. Existing collaborative methods often require frequent exchange of raw observation data or intermediate states, which usually leads to huge communication overhead and is difficult to implement. At the same time, there are differences in sensor types, data formats, and processing capabilities among platforms. Asynchronous data and inconsistent formats can introduce fusion errors, impair the performance of collaborative sensing, and in severe cases, cause target loss. Therefore, it is necessary to design a collaborative mechanism with low communication bandwidth requirements and compatibility with heterogeneous data characteristics, so that each platform can reduce its dependence on continuous, high-quality communication, effectively fuse multi-source information, and improve the consistency of overall sensing.

[0005] Furthermore, the system's resource allocation and task scheduling lack adaptive coupling with the perception process. In multi-target tracking scenarios, how to dynamically allocate limited airborne resources to different observation platforms based on the target's real-time state directly affects the efficiency and accuracy of collaborative perception. Traditional scheduling methods are usually based on fixed priority rules or simple geometric relationships, failing to fully consider the real-time uncertainty of target motion and the performance differences between different platforms when performing the same task. This leads to low utilization of limited airborne resources and a decline in overall system performance.

[0006] In summary, existing technical solutions still have significant shortcomings in addressing complex maneuvering target modeling, efficient collaboration with low communication overhead, and adaptive resource scheduling for dynamic targets. They lack a holistic solution that can systematically integrate these aspects and achieve closed-loop optimization. Therefore, there is an urgent need to develop a novel cross-platform collaborative perception method capable of achieving adaptive, highly efficient edge-intelligent collaborative tracking capabilities under the dual challenges of communication constraints and target maneuverability. This invention is proposed precisely to address this need.

[0007] A review of existing literature revealed the most similar implementation scheme as follows: Chinese patent application number CN202310559945.6, entitled "A Dynamic Cooperative Sensing and Localization Method for Space-Based Navigation Enhanced Ad hoc Networks," describes a method that establishes dynamic cooperative relationships between nodes based on evolutionary game theory to achieve autonomous allocation and adjustment of sensing tasks; subsequently, a cooperative localization state space model is constructed, and node positions are jointly estimated and optimized using a recursive filtering algorithm. However, this scheme is based on traditional filtering methods and does not address the adaptive modeling and cooperative learning problems of complex maneuvering targets, lacking intelligent resource scheduling capabilities for sensing uncertainties. Another similar scheme is Chinese patent application number CN202511225538.7, entitled "A Space-Based Cooperative Sensing and Communication Method for Low-Altitude Airspace Management," which describes a method that establishes a space-based satellite network, dynamically schedules multi-satellite cooperative detection through a central node, and centrally merges data after on-board preprocessing to generate a low-altitude situation map for threat assessment and distribution. However, this patent focuses on the centralized scheduling and rapid space-ground collaborative processing of space-based systems to improve the monitoring coverage and response speed of low-altitude targets. It does not involve adaptive trajectory modeling and learning mechanisms for highly maneuverable targets, nor does it have the intelligent decision-making capability to dynamically optimize resource allocation based on real-time perception uncertainty. The Chinese patent application number is CN202410293109.2, entitled: A method, device, electronic device and storage medium for three-dimensional collaborative perception based on space, air and ground. Its specific approach is to establish a cost-benefit model and resource constraints, and use a genetic algorithm to jointly optimize the scheduling and task allocation of multiple types of perception resources in space, air and ground to achieve three-dimensional collaborative observation planning. However, this proposed method focuses on system-level resource planning and static task assignment, but does not address adaptive trajectory modeling, distributed collaborative learning, and real-time dynamic decision-making based on uncertainty in target tracking. Therefore, it cannot solve the problems of continuous and accurate tracking and real-time resource optimization for highly maneuverable targets. Chinese patent application number CN202510153793.9, entitled "Dynamic Target Intelligent Tracking Method and System Based on Air-Ground Unmanned Collaborative System," specifically describes a method that uses extended Kalman filtering to fuse the kinematic states of air and ground platforms, combined with environmental maps and path planning, to achieve collaborative state estimation and tracking control of dynamic targets. However, this approach still relies on traditional filtering algorithms and explicit path planning, lacks adaptive learning capabilities for complex target maneuvering patterns, and lacks cross-platform distributed model collaboration and uncertainty-based intelligent resource scheduling mechanisms. Therefore, it has limitations in handling highly nonlinear motion and multi-platform autonomous collaborative optimization. Summary of the Invention

[0008] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to overcome the limitations of resource scheduling.

[0009] To achieve the above objectives, the present invention provides a method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system, characterized in that the method includes the following steps: Step 1: Each platform collects observation data through heterogeneous sensors and performs local linearization processing to maintain a local database for the sliding time window; Step 2: Employ a cross-platform collaborative association mechanism to complete local track-data matching and global track-number unified matching; Step 3: Based on federated learning and the alternating multiplier method, the hyperparameters of the Gaussian process model on each platform are optimized in a distributed and collaborative manner to achieve global model consistency; Step 4: Use the optimized Gaussian process model to predict the target trajectory and quantify uncertainty, and complete distributed information fusion; Step 5: Construct an intelligent auction mechanism based on prediction uncertainty and information gain, and perform dynamic perception task allocation and resource scheduling; The above steps form a closed loop of perception-assessment-scheduling, enabling collaborative tracking of highly maneuverable targets.

[0010] Furthermore, in step 1, the heterogeneous sensors on each platform operate independently according to their own mechanisms, asynchronously generating observation streams containing raw measurements and local timestamps. The data formats, accuracy, and update rates are different for each platform. The collected raw data undergoes local linearization to obtain linearized observation variance. ,in , The location of the target. For the locally linearized matrix For local linearization points, To observe the true value, The observation noise has a mean of 0 and a covariance of R. The platform opens a sliding window data cache of fixed time length, continuously storing data from its most recent few periods to maintain the database within the sliding time window. The database contains sampling timestamps and corresponding linearized measurements. Old data is removed in real time to control memory usage and ensure that the system always tracks and makes decisions based on the latest and most relevant local information.

[0011] Furthermore, step 2 also includes: Step 2.1: For each existing target track maintained by this platform, call its trajectory prediction model to obtain its expected value and covariance at the current moment. This state is then projected onto the sensor measurement space through the observation equation to obtain the expected value of the observation. Centered on, information covariance Establish an association gate for the shape of an ellipse, where R is the local linearization matrix, and R is the covariance matrix of the observation noise; Step 2.2: Calculate the Mahalanobis distance from each of the new data obtained by this platform in the current period to the center of each associated gate. Given an association threshold If the Mahalanobis distance is greater than this threshold, i.e. ,in If the threshold is not met, the association match between the observation and the track is rejected; otherwise, the association match is entered, resulting in a candidate set of observations for each target. A cost matrix is ​​constructed based on this candidate observation set. Where N is the number of target tracks, M is the number of observations, and the elements in the matrix are... This is the Mahalanobis distance between the two matching elements; if a match is rejected, the element is ∞. Step 2.3: Based on this, construct an augmented cost matrix. If the number of targets is greater than the number of observations, add virtual observations; otherwise, add virtual targets. The cost is a maximum value. Construct the cost matrix into an augmented cost square matrix. Where K = max(N, M), the Hungarian algorithm is then used to solve the following optimization problem. ,

[0012] Where X is the matching matrix of 0-1 matrix elements, To broaden the cost matrix The element in the i-th row and j-th column, The element in the i-th row and j-th column of matrix X is used to obtain the target track-observation matching pair; Step 2.4: Successfully associated points are used to update the database of the corresponding track; for new observations, if new observations continue to appear within a certain time window, they are identified as new targets, otherwise the observations are removed; for unmatched targets, the original data is used for prediction without updating the dataset, and if no associated existing track is obtained within a certain time period, the track is terminated. Step 2.5: Each platform outputs a set of local track information within a sliding window, including: track ID, timestamp, track status, and track covariance, and uploads it to the central platform; the central platform receives track information sent by all neighboring platforms and maintains a set of global track information. The center matches the received trajectories with the global trajectory; specifically, for each trajectory i from platform q, the Mahalanobis distance between the global trajectory and the trajectory of each platform is calculated one by one. ,in , ) represents the expectation and covariance of the global trajectory. , ) represents the expectation and covariance of trajectory i; then a gating mechanism is used, if the Mahalanobis distance is greater than this threshold, i.e. ,in If the threshold is not met, the association matching of this track pair is rejected; thus, a trajectory matching cost matrix is ​​formed. Where H is the number of global tracks, and the elements in the matrix are... This is the Mahalanobis distance between the two matching elements; if a match is rejected, the element is ∞. Step 2.6: Based on this, perform virtual track completion on the trajectory matching cost matrix to obtain the augmented trajectory matching cost matrix. Furthermore, the Hungarian algorithm is used to solve the problem. ,

[0013] in A matching matrix of 0-1 matrix elements. To augment the trajectory matching cost matrix The element in the i-th row and j-th column, For matrix The element in the i-th row and j-th column is used to obtain the target track-track matching pair; After obtaining the optimal matching sequence, the central node labels the matching sub-platform track sequences according to the global track sequence and returns the label-track data packet to the sub-platform, thereby achieving the unification of the numbering, track and observation data of each platform.

[0014] Furthermore, in step 3, based on the matched number-track-observation data, distributed model parameter collaborative optimization based on federated learning is adopted. Without sharing the original data, the collective data is used to improve the cognitive model quality of each platform for target track maneuvering, thereby fundamentally improving the prediction ability for complex maneuvering targets.

[0015] Furthermore, step 3 also includes: Step 3.1: Each platform independently builds and maintains a Gaussian process model for each target it tracks, based on its local dataset. , where x(t) is the target trajectory, which is described as a function that varies with time. This is a priori mean function, usually set to zero or a simple trend. The covariance function defines the correlation between states at different time points. Under this model, the state of the target at any time is not only determined by its neighboring times, but also by the global constraints of the observed samples in the entire time series. Step 3.2: Each platform uses its maintained Gaussian process model and the matched dataset. Construct hyperparameters for the prior Gaussian process Maximum likelihood function , in Let be the vector formed by the linearized measurements of platform q. For locally linearized diagonal matrices, Let glym matrix be the kernel function with respect to the sampling timestamp. The maximum likelihood function represents a constant; based on this maximum likelihood function and the local dataset, each platform solves for the optimal hyperparameters by maximizing the maximum likelihood function using methods such as gradient descent. As model parameters when platform collaboration is not considered; Step 3.3: In order to achieve consensus among the Gaussian process models of various platforms and integrate global information, a federated learning framework is adopted. Under the premise of protecting the privacy of local data on each platform, a set of globally consistent hyperparameters is found to maximize the joint likelihood of the local models of all platforms. Step 3.4: Achieve a global solution using the alternating multiplier method, with each platform maintaining a local set of hyperparameters. With Lagrange multipliers .

[0016] Furthermore, step 3.4 also includes: Step 3.4.1: Each platform q publishes the local hyperparameters obtained under the current iteration round k to the central node. With Lagrange multipliers The central node collects all updated local hyperparameters and Lagrange multipliers from all platforms, and then performs global aggregation to obtain global variables.

[0017] in The number of parameters received. >0 is the preset penalty function factor; this global variable integrates the local information of all platforms to form a new global consensus parameter, and then broadcasts the global consensus parameter to each platform; when performing global aggregation, it is not required to receive all platform parameters. For parameters that lack updates, the original parameter information is used for global aggregation. Step 3.4.2: Global variables broadcast by the receiving center on each platform and compared it with the local Lagrange multipliers To fix the issue, when local datasets or global variables are updated on each platform, the following optimization problem is solved locally to update the local hyperparameters:

[0018] in .) represents the optimization variable when the objective function is minimized. The first term in the formula represents the likelihood optimization objective, and the second and third terms represent the degree of global consensus. This update drives the local parameters to move closer to the global consensus while fitting their own data. This step can be performed in parallel with the global variable update. When the platform does not receive the global variable, it still uses the global variable from the previous moment for local updates. Step 3.4.3: Each platform updates its own Lagrange multipliers based on the latest local and global parameters to strengthen consistency constraints. ( ) This formula is essentially an update of the dual variable of the consistency constraint, achieved through continuous iterative updates. The system can gradually reduce the deviation between the parameters of each platform and the global parameters, thereby improving the convergence and consistency of the federated learning process.

[0019] Step 3.4.4 The federated learning process ends when the maximum number of iterations is reached.

[0020] Furthermore, step 4 also includes: Step 4.1: Using the optimized unified model and the associated dataset, given a future time t... The predicted waypoints Each platform constructs a joint distribution of linearized data and points to be predicted based on the prior Gaussian process distribution; Step 4.2: Each platform, based on its constructed prior Gaussian process model, performs prediction on the track points at each time step, using the matched dataset. You can get The posterior Gaussian distribution has an expectation and covariance of , respectively. , , in This represents the covariance between data points. = Represents the data and the track points to be predicted The covariance is obtained through a prior Gaussian process. For locally linearized diagonal matrices, The vector formed by the linearized measurements of platform q; both can be used as the platform q with respect to waypoints. The predicted value and the measurement of prediction uncertainty are as follows: the smaller the posterior covariance, the higher the confidence of the platform q in predicting the waypoint based on the current matching data; conversely, the larger the posterior covariance, the greater the uncertainty of the prediction result at that moment. Step 4.3: To obtain globally consistent perception results and achieve complementarity among platforms in terms of perception data and capabilities, a distributed information fusion mechanism is adopted to integrate the tracking results of each platform. Each platform uploads its trajectory prediction results to the central platform, which then calculates the results according to information weights. Adaptive fusion of covariance and variance:

[0021] Separate center platform for track points Predicted values ​​and measures of prediction uncertainty; inverse of the covariance matrix The covariance of platform q represents the amount of predictive information. A smaller covariance indicates a more reliable prediction result and a greater impact on the fusion result; conversely, a larger covariance indicates higher prediction uncertainty and a correspondingly lower contribution. Through this adaptive weighted fusion mechanism, the central platform can fully leverage the complementarity of different platforms in terms of observation range, sensing accuracy, and data quality to obtain globally consistent and more reliable trajectory prediction results. The fused result... and These serve as the central platform's global prediction value and global prediction uncertainty measure for waypoints, respectively, and are broadcast to each sub-platform to provide a unified perception basis for subsequent collaborative decision-making. Furthermore, through information fusion, each platform generates trajectory prediction results with uncertainty quantification. Step 5, based on these prediction results, adopts an intelligent auction mechanism to generate a set of executable and efficient task deployment instructions, thereby achieving the global optimal or near-optimal configuration of sensing resources in a dynamic environment.

[0022] Furthermore, step 5 also includes: Step 5.1: Based on the trajectory prediction results with uncertainty quantification, each platform performs a self-evaluation of each trajectory perception and calculates the benefits. Unlike the traditional benefit-cost framework that requires the separate construction of complex benefit and cost functions, this method does not introduce a complex task benefit model separately. Instead, it directly uses the reduction in prediction uncertainty as the perception benefit and combines it with the energy consumption cost required for the platform to perform the sampling action for a comprehensive evaluation. ) in ), The degree to which the observations collected by platform q at the next sampling time contribute to the reduction of uncertainty is characterized. For The sampling space centered on platform q can be geographically dispersed when the platform q has high mobility, and tend to be concentrated otherwise. This represents the energy cost of platform q in sensing target trajectory i; therefore, this evaluation metric can measure the maximum ability of platform q to improve tracking performance from the current state, and can avoid the problem of simply pursuing prediction accuracy while ignoring platform resource consumption, thus achieving an adaptive balance between sensing benefits and energy consumption constraints. Step 5.2: After evaluating the value of each local track tracking, each platform begins to execute task bidding. Each platform considers its own energy consumption limitations, comparing the energy consumption of its allocated target tracks with its maximum energy capacity to determine the list of executable track tracking. To reduce communication burden, each platform only compiles its top K highest bids to form data packets. Send it to the central node; Step 5.3: After receiving bids from all platforms, the central node sorts all bid pairs to obtain a global bid list. Based on this global bidding list, the central node platform evaluates the priority of target task i by considering factors such as the platform's tracking revenue for the target, the platform's relative scarcity of tasks, and the number of platforms that have already assigned the task:

[0023] in This represents the number of platforms that have been allocated to target i. This indicates the number of platforms on which tracking target i can be executed. and It is a decreasing function of the dependent variable; this priority function works by combining three parts: To reduce the priority of further allocation for targets that have already acquired a significant amount of platform resources, Used to increase the resource scarcity weight for targets with fewer executable platforms. This is used to characterize the optimal tracking capability that the current platform set can provide for the target. Through the above priority evaluation, the central node can prioritize the processing of target tasks with high tracking benefits, insufficient resource coverage, and scarce executable platforms, thereby avoiding uneven allocation of platform resources and improving the overall efficiency of multi-platform collaborative tracking task allocation. Step 5.4: The central node scans the platform-tracking task evaluation list sequentially using a greedy strategy. If target i has not yet been assigned and the task load of platform q has not exceeded the limit, target i is assigned to platform q; otherwise, the bid is skipped and the next item is continued. After the scan is completed, the platform-tracking allocation result is obtained. The central node sends it to each sub-platform. After receiving the result, the sub-platform reorganizes the next bid list until each platform no longer submits a new bid, thus completing the task allocation. Intelligent task allocation is executed repeatedly at a fixed frequency to adapt to changes in target movement, platform attitude, and dynamic changes in communication topology. When a platform detects a task execution failure, it can immediately report to the central node, which will trigger a temporary re-attempt to achieve rapid adaptive task reallocation. This centralized auction mechanism can be seamlessly coupled with the trajectory prediction module in the above steps. The platform directly uses the posterior covariance obtained from the prediction results to calculate the bid, achieving low computational complexity in the perception-decision closed loop. The central node only needs to sort and greedily allocate the limited bids, and the communication overhead is small. Each platform only uploads the top K highest bids and does not need to upload the original measurement or complete state information. Overall, this scheme significantly reduces the implementation complexity while maintaining theoretical rationality, and achieves efficient task coordination and dynamic perception resource scheduling across platforms.

[0024] Furthermore, steps 1-5 constitute a self-driven intelligent closed loop. The allocation decision in step 5 guides the data collection action in step 1. The new data generated, after being correlated in step 2, is used to drive the further co-evolution of the model in step 3. The evolved model then generates more accurate predictions in step 4, thereby leading to better resource allocation. This cycle repeats continuously, enabling the entire system to achieve adaptive coordination of perception, cognition, decision-making, and execution in complex dynamic environments, continuously maintaining optimal global situational awareness capabilities.

[0025] The present invention has the following technical effects: (1) This invention can improve the tracking success rate and accuracy of the system for highly maneuverable and unknown maneuverable targets by an order of magnitude. The system can adaptively learn the target's motion pattern and provide reliable prediction and confidence assessment within a probabilistic framework, greatly reducing tracking loss caused by model mismatch. Especially when the target performs unconventional tactical maneuvers, the system can still maintain stable trajectory estimation, thereby reducing the system's dependence on complex multi-model library design, parameter tuning, and switching logic, and reducing the complexity and cost of algorithm development and maintenance. At the same time, higher tracking accuracy and resource utilization mean that the same task can be completed with fewer platform or sensor resources, or more complex tasks can be performed with the same resources, significantly improving cost-effectiveness.

[0026] (2) This invention significantly improves the accuracy and robustness of the cooperative sensing system in real-world unstable communication environments. It enables the system to maintain a globally asymptotically consistent and high-quality situational awareness even under real-world conditions of intermittent link disruptions and heterogeneous observations across platforms. This significantly reduces the risk of a precipitous drop or collapse in system performance due to communication problems, making the entire system architecture more flat and robust. This directly enhances the sustainability of the target tracking system in complex electromagnetic environments, allowing it to provide a usable and consistent field perception map even under conditions of strong interference, partial platform damage, or separation from formation. Furthermore, this mechanism reduces the extreme requirements on communication infrastructure, making it possible to build efficient cooperative networks using existing or low-cost communication equipment, resulting in significant equipment adaptability and cost advantages.

[0027] (3) This invention significantly improves the overall average tracking accuracy and resource utilization efficiency of the system in multi-target, multi-platform tracking scenarios. The system can automatically and in real time direct sensing resources to the most critical and unclear task in the situation, realizing adaptive optimization of resource allocation, and keeping the system agile and efficient at all times. This allows limited hardware resources to play their maximum sensing role, achieving better performance under the same resource conditions, and has significant economic benefits. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of an edge intelligent trajectory collaborative perception scenario according to a preferred embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the overall implementation of a cross-platform track collaborative perception method based on edge intelligence, according to a preferred embodiment of the present invention. Figure 3 This is a schematic diagram of a cross-platform track collaborative perception method based on edge intelligence, according to a preferred embodiment of the present invention. Figure 4 This is a schematic diagram of the track-data-number matching (step 2) process according to a preferred embodiment of the present invention; Figure 5 This is a flowchart illustrating the collaborative training (step 3) of a Gaussian process model based on federated learning, according to a preferred embodiment of the present invention. Figure 6 This is a flowchart illustrating the perception task allocation (step 5) based on an intelligent auction mechanism, according to a preferred embodiment of the present invention. Detailed Implementation

[0029] refer to Figures 1-6 This invention proposes a cross-platform trajectory collaborative perception method based on edge intelligence, comprising the following steps: Step 1: The heterogeneous sensors on each platform operate independently according to their own mechanisms, asynchronously generating observation streams containing raw measurements and local timestamps. The data format, accuracy, and update rate may vary. Local linearization is performed on the collected raw data to obtain the linearized observation variance. ,in , The location of the target. For locally linearized matrices, For local linearization points, To observe the true value, With a mean of 0 and a covariance of R To reduce observation noise, the platform establishes a fixed-time sliding window data cache, continuously loading data from its most recent periods to maintain the database within the sliding window. The database contains sampling timestamps and corresponding linearized measurements. Old data is removed in real time to control memory usage, ensuring that the system always tracks and makes decisions based on the latest and most relevant local information. This design ensures the timeliness of data processing and stabilizes the computational load, providing a data foundation for subsequent target tracking.

[0030] Step 2: Based on the obtained sliding window data, a cross-platform collaborative association mechanism is adopted to further bind the messy sensor data with the correct target identity, realizing the association and matching of data and tracks. A schematic diagram of the specific process is shown below. Figure 4 As shown, the details are as follows: Step 2.1: For each existing target track maintained by this platform, call its trajectory prediction model to obtain its expected value and covariance at the current moment. This state is then projected onto the sensor measurement space through the observation equation to obtain the expected value of the observation. Centered on, information covariance Establish an association gate for the shape of an ellipse, where For locally linearized matrices, R Let be the covariance matrix of the observation noise.

[0031] Step 2.2: Calculate the Mahalanobis distance from each of the new data obtained by this platform in the current period to the center of each associated gate. Given an association threshold If the Mahalanobis distance is greater than this threshold, i.e. ,in If the threshold is not met, the association match between the observation and the track is rejected; otherwise, the association match is entered, resulting in a candidate set of observations for each target. A cost matrix is ​​constructed based on this candidate observation set. Where N is the number of target tracks, M is the number of observations, and the elements in the matrix are... This is the Mahalanobis distance between the two elements; if a match is rejected, the element is ∞.

[0032] Step 2.3: Based on this, construct an augmented cost matrix. If the number of targets is greater than the number of observations, add virtual observations; otherwise, add virtual targets. The cost is a maximum value. Construct the cost matrix into an augmented cost square matrix. Where K = max(N, M), the Hungarian algorithm is then used to solve the following optimization problem. ,

[0033] Where X is the matching matrix of 0-1 matrix elements, To broaden the cost matrix The element in the i-th row and j-th column, Let X be the element in the i-th row and j-th column of matrix X, and then the target track-observation matching pair is obtained.

[0034] Step 2.4: Successfully associated points are used to update the database of the corresponding tracks. For new observations, if new observations continue to appear within a certain time window, they are identified as new targets; otherwise, the observation is removed. For unmatched targets, the original data is continuously used for prediction, and their dataset is not updated. If no associated existing track is found within the specified time, the track is terminated.

[0035] Step 2.5: Each platform outputs a set of local track information within a sliding window, including: track ID, timestamp, track status, and track covariance, and uploads it to the central platform. The central platform receives track information sent by all neighboring platforms and maintains a set of global track information. The center matches the received trajectories with the global trajectory. Specifically, for each trajectory i from platform q, the Mahalanobis distance between the global trajectory and the trajectory of each platform is calculated one by one. ,in , ) represents the expectation and covariance of the global trajectory. , () represents the expectation and covariance of trajectory i. A gating mechanism is then employed; if the Mahalanobis distance is greater than a certain threshold, i.e. ,in If the threshold is not met, the association matching of this track pair is rejected. This results in the formation of a trajectory matching cost matrix. Where H is the number of global tracks, and the elements in the matrix are... This is the Mahalanobis distance between the two elements; if a match is rejected, the element is ∞.

[0036] Step 2.6: Based on this, perform virtual track completion on the trajectory matching cost matrix to obtain the augmented trajectory matching cost matrix. Furthermore, the Hungarian algorithm is used to solve the problem. ,

[0037] in A matching matrix of 0-1 matrix elements. To augment the trajectory matching cost matrix The element in the i-th row and j-th column, For matrix The element in the i-th row and j-th column is used to obtain the target track-track matching pair. After the central node obtains the optimal matching sequence, it labels the matched sub-platform track sequences according to the global track sequence and returns the label-track data packet to the sub-platform, thereby achieving the unification of the numbering-track-observation data of each platform.

[0038] Step 3: Based on the matched number-track-observation data, a distributed model parameter collaborative optimization based on federated learning is adopted. Without sharing the original data, the collective data is used to improve the cognitive model quality of each platform for target track maneuvers, thereby fundamentally improving the prediction capability for complex maneuvering targets. A schematic diagram of the specific process is shown below. Figure 5 As shown, the details are as follows: Step 3.1: Each platform independently builds and maintains a Gaussian process model for each target it tracks, based on its local dataset. , where x(t) is the target trajectory, which is described as a function that varies with time. This is a priori mean function, usually set to zero or a simple trend. The covariance function defines the correlation between states at different time points. For example, a squared exponential kernel can be used to characterize smoothly changing target maneuvers, where the hyperparameters of the kernel function are... This model explicitly describes the shape of the prior function space, determining the effectiveness of posterior predictions for target tracking. Under this model, the state of the target at any given time is determined not only by its neighboring times but also by global constraints imposed by the observed samples throughout the entire time series.

[0039] Step 3.2: Each platform uses its maintained Gaussian process model and the matched dataset to complete the process. Construct hyperparameters for the prior Gaussian process Maximum likelihood function , in Let be the vector formed by the linearized measurements of platform q. For locally linearized diagonal matrices, Let glym matrix be the kernel function with respect to the sampling timestamp. This represents a constant. Based on this maximum likelihood function and the local dataset, each platform solves for the optimal hyperparameters by maximizing the maximum likelihood function using methods such as gradient descent. This serves as a model parameter when platform collaboration is not considered.

[0040] Step 3.3: To achieve consensus among Gaussian process models across platforms and integrate global information, a federated learning framework is adopted. Under the premise of protecting the privacy of local data on each platform, a consistent set of hyperparameters is found to maximize the joint likelihood of the local models on all platforms.

[0041] Step 3.4: Achieve a global solution using the alternating multiplier method, with each platform maintaining a local set of hyperparameters. With Lagrange multipliers The following alternating iterative steps are executed in parallel: Step 3.4.1: Each platform q publishes its local hyperparameters obtained under the current iteration round k to the central node. With Lagrange multipliers The central node collects all updated local hyperparameters and Lagrange multipliers from all platforms, and then performs global aggregation to obtain global variables. . in The number of parameters received. >0 represents the preset penalty function factor. This global variable integrates local information from all platforms to form a new global consensus parameter, which is then broadcast to each platform. During global aggregation, it is not required to receive parameters from all platforms; for parameters that lack updates, the original parameter information is used for global aggregation.

[0042] Step 3.4.2: Global variables broadcast by the receiving center on each platform and compared it with the local Lagrange multipliers The settings are fixed. When local datasets or global variables are updated on each platform, the following optimization problem is solved locally to update the local hyperparameters: , in The period (.) represents the optimization variable when the objective function reaches its minimum. The first term in the formula represents the likelihood optimization objective, and the second and third terms represent the degree of global consensus. This update drives the local parameters to converge towards the global consensus while fitting their own data. This step can be performed in parallel with the global variable update. When the platform does not receive the global variables, it still uses the global variables from the previous time step for local updates.

[0043] Step 3.4.3: Each platform updates its own Lagrange multipliers based on the latest local and global parameters to strengthen consistency constraints. ( ), This formula is essentially an update of the dual variable of the consistency constraint, achieved through continuous iterative updates. The system can gradually reduce the deviation between the parameters of each platform and the global parameters, thereby improving the convergence and consistency of the federated learning process.

[0044] Step 3.4.4: The federated learning process ends when the maximum number of iterations is reached.

[0045] Through the federated learning strategy based on the alternating multiplier method described above, multiple platforms can achieve consistent optimization of the hyperparameters of the Gaussian process model without sharing the original observation data, thereby ensuring the consistency, accuracy, and real-time performance of trajectory prediction across platforms. This strategy balances privacy protection, scalability, and computational efficiency in edge computing environments, providing a foundation for subsequent prediction using the Gaussian process model.

[0046] Step 4: After collaborative optimization, each platform obtains a highly consistent Gaussian process model. This step performs forward-looking trajectory prediction and rigorously quantifies the reliability of the prediction results, as detailed below: Step 4.1: Using the optimized unified model and the associated dataset, given a future time... The predicted waypoints Each platform constructs a joint distribution of linearized data and points to be predicted based on the prior Gaussian process distribution; Step 4.2: Each platform, based on its constructed prior Gaussian process model, performs prediction on the track points at each time step, using the matched dataset. You can get The posterior Gaussian distribution has an expectation and covariance of , respectively. , , in This represents the covariance between data points. = Represents the data and the track points to be predicted The covariance is obtained through a prior Gaussian process. For locally linearized diagonal matrices, The vector formed by the linearized measurements of platform q; both can be used as the platform q with respect to waypoints. The predicted value and the measurement of prediction uncertainty are as follows: the smaller the posterior covariance, the higher the confidence of the platform q in predicting the waypoint based on the current matching data; conversely, the larger the posterior covariance, the greater the uncertainty of the prediction result at that moment. Step 4.3: To obtain globally consistent perception results and achieve complementarity among platforms in terms of perception data and capabilities, a distributed information fusion mechanism is adopted to integrate the tracking results of each platform. Each platform uploads its trajectory prediction results to the central platform, which then categorizes them according to information weights. Adaptive fusion of covariance and variance: , , Separate center platform for track points The predicted value and the measure of prediction uncertainty. The inverse of the covariance matrix. The covariance of platform q represents the amount of predictive information. A smaller covariance indicates a more reliable prediction result and a greater impact on the fusion result; conversely, a larger covariance indicates higher prediction uncertainty and a correspondingly lower contribution. Through this adaptive weighted fusion mechanism, the central platform can fully leverage the complementarity of different platforms in terms of observation range, sensing accuracy, and data quality to obtain globally consistent and more reliable trajectory prediction results. The fused result... and These values ​​serve as the central platform's global prediction value and global prediction uncertainty measure for waypoints, respectively, and are broadcast to each sub-platform to provide a unified perception basis for subsequent collaborative decision-making.

[0047] Step 5: After information fusion, each platform generates trajectory prediction results with uncertainty quantification. Based on these prediction results, this step adopts an intelligent auction mechanism to generate a set of executable and efficient task deployment instructions, achieving globally optimal or near-optimal allocation of sensing resources in a dynamic environment, as detailed below: Step 5.1: Based on the trajectory prediction results with uncertainty quantification, each platform performs a self-evaluation of each trajectory perception and calculates the benefits. Unlike the traditional benefit-cost framework that requires the separate construction of complex benefit and cost functions, this method does not introduce a complex task benefit model separately. Instead, it directly uses the reduction in prediction uncertainty as the perception benefit and combines it with the energy consumption cost required for the platform to perform the sampling action for a comprehensive evaluation. ) in, ), The degree to which the observations collected by platform q at the next sampling time contribute to the reduction of uncertainty is characterized. For The sampling space centered on platform q can be geographically dispersed when the platform q has high mobility, and tend to be concentrated otherwise. This represents the energy cost of platform q sensing the target trajectory i. Therefore, this evaluation metric can measure the maximum ability of platform q to improve tracking performance from the current state, avoiding the problem of simply pursuing prediction accuracy while ignoring platform resource consumption, and achieving an adaptive balance between sensing benefits and energy consumption constraints.

[0048] Step 5.2: After evaluating the value of each local track tracking operation, each platform begins executing task bids. Each platform considers its own energy consumption limitations, comparing the energy consumption of its allocated target tracks with its maximum energy capacity to determine the list of executable track tracks. To reduce communication overhead, each platform only compiles its top K bids into data packets. It is then sent to the central node.

[0049] Step 5.3: After receiving bids from all platforms, the central node sorts all bid pairs to obtain a global bid list. Based on this global bidding list, the central node platform evaluates the priority of target task i by considering factors such as the platform's tracking revenue for the target, the platform's relative scarcity of tasks, and the number of platforms that have already assigned the task:

[0050] in This represents the number of platforms that have been allocated to target i. This indicates the number of platforms on which tracking target i can be executed. and As a decreasing function of the dependent variable, this priority function works in combination with three parts: To reduce the priority of further allocation for targets that have already acquired a significant amount of platform resources, Used to increase the resource scarcity weight for targets with fewer executable platforms. This is used to characterize the optimal tracking capability that the current platform set can provide for the target. Through the above priority evaluation, the central node can prioritize the processing of target tasks with high tracking benefits, insufficient resource coverage, and scarce executable platforms, thereby avoiding uneven allocation of platform resources and improving the overall efficiency of multi-platform collaborative tracking task allocation.

[0051] Step 5.4: The central node scans the platform-tracking task evaluation list sequentially using a greedy strategy. If target i has not yet been assigned and the task load of platform q has not exceeded the limit, target i is assigned to platform q; otherwise, the bid is skipped and the next item is moved on. After the scan is completed, the platform-tracking allocation result is obtained. The central node then distributes this result to each sub-platform. Upon receiving the result, the sub-platform reorganizes its bid list for the next round of bidding until no platform submits a new bid, thus completing the task allocation.

[0052] Intelligent task allocation is executed repeatedly at a fixed frequency to adapt to changes in target movement, platform attitude, and dynamic changes in communication topology. When a platform detects a task execution failure, it can immediately report it to the central node, which triggers a temporary replay to achieve rapid adaptive task reallocation.

[0053] This centralized auction mechanism can be seamlessly coupled with the trajectory prediction module described above. The platform directly uses the posterior covariance obtained from the prediction results to calculate the bid, achieving low computational complexity in the perception-decision closed loop. The central node only needs to sort and greedily allocate a limited number of bids, with minimal communication overhead. Each platform only uploads the top K highest bids, without needing to upload original measurements or complete state information. Overall, this scheme significantly reduces implementation complexity while maintaining theoretical rationality, achieving efficient cross-platform task coordination and dynamic perception resource scheduling.

[0054] Steps 1 through 5 constitute a self-driven intelligent closed loop. The allocation decision in step 5 guides the data collection actions in step 1. The new data generated, after being correlated in step 2, is used to drive the further co-evolution of the model in step 3. The evolved model then produces more accurate predictions in step 4, thereby leading to better resource allocation. This cycle repeats continuously, enabling the entire system to achieve adaptive coordination of perception, cognition, decision-making, and execution in complex and dynamic environments, continuously maintaining optimal global situational awareness capabilities.

Claims

1. A method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system, characterized in that, The method includes the following steps: Step 1: Each platform collects observation data through heterogeneous sensors and performs local linearization processing to maintain a local database for the sliding time window; Step 2: Employ a cross-platform collaborative association mechanism to complete local track-data matching and global track-number unified matching; Step 3: Based on federated learning and the alternating multiplier method, the hyperparameters of the Gaussian process model on each platform are optimized in a distributed and collaborative manner to achieve global model consistency; Step 4: Use the optimized Gaussian process model to predict the target trajectory and quantify uncertainty, and complete distributed information fusion; Step 5: Construct an intelligent auction mechanism based on prediction uncertainty and information gain, and perform dynamic perception task allocation and resource scheduling; The above steps form a closed loop of perception-assessment-scheduling, enabling collaborative tracking of highly maneuverable targets.

2. The method for cooperative perception of maneuvering targets based on edge intelligent unmanned aerial systems as described in claim 1, characterized in that, In step 1, the heterogeneous sensors on each platform operate independently according to their own mechanisms, asynchronously generating observation streams containing raw measurements and local timestamps. The data formats, accuracy, and update rates are different. Local linearization is performed on the collected raw data to obtain the linearized observation variance. ,in , The location of the target. For the locally linearized matrix For local linearization points, To observe the true value, The observation noise has a mean of 0 and a covariance of R. The platform opens a sliding window data cache of fixed time length, continuously storing data from its most recent few periods to maintain the database within the sliding time window. The database contains sampling timestamps and corresponding linearized measurements. Old data is removed in real time to control memory usage and ensure that the system always tracks and makes decisions based on the latest and most relevant local information.

3. The collaborative perception method for maneuvering targets based on an edge-intelligent aerial unmanned system as described in claim 2, characterized in that, Step 2 also includes: Step 2.1: For each existing target track maintained by this platform, call its trajectory prediction model to obtain its expected value and covariance at the current moment. This state is then projected onto the sensor measurement space through the observation equation to obtain the expected value of the observation. Centered on, information covariance Establish an association gate for the shape of an ellipse, where R is the local linearization matrix, and R is the covariance matrix of the observation noise; Step 2.2: Calculate the Mahalanobis distance from each of the new data obtained by this platform in the current period to the center of each associated gate. Given an association threshold If the Mahalanobis distance is greater than this threshold, i.e. ,in If the threshold is not met, the association match between the observation and the track is rejected; otherwise, the association match is entered, resulting in a candidate set of observations for each target. A cost matrix is ​​constructed based on this candidate observation set. Where N is the number of target tracks, M is the number of observations, and the elements in the matrix are... This is the Mahalanobis distance between the two matching elements; if a match is rejected, the element is ∞. Step 2.3: Based on this, construct an augmented cost matrix. If the number of targets is greater than the number of observations, add virtual observations; otherwise, add virtual targets. The cost is a maximum value. Construct the cost matrix into an augmented cost square matrix. Where K = max(N, M), the Hungarian algorithm is then used to solve the following optimization problem. , Where X is the matching matrix of 0-1 matrix elements, To broaden the cost matrix The element in the i-th row and j-th column, The element in the i-th row and j-th column of matrix X is used to obtain the target track-observation matching pair; Step 2.4: Successfully associated points are used to update the database of the corresponding track; for new observations, if new observations continue to appear within a certain time window, they are identified as new targets, otherwise the observations are removed; for unmatched targets, the original data is used for prediction without updating the dataset, and if no associated existing track is obtained within a certain time period, the track is terminated. Step 2.5: Each platform outputs a set of local track information within a sliding window, including: track ID, timestamp, track status, and track covariance, and uploads it to the central platform; the central platform receives track information sent by all neighboring platforms and maintains a set of global track information. The center matches the received trajectories with the global trajectory; specifically, for each trajectory i from platform q, the Mahalanobis distance between the global trajectory and the trajectory of each platform is calculated one by one. ,in , ) represents the expectation and covariance of the global trajectory. , ) represents the expectation and covariance of trajectory i; then a gating mechanism is used, if the Mahalanobis distance is greater than this threshold, i.e. ,in If the threshold is not met, the association matching of this track pair is rejected; thus, a trajectory matching cost matrix is ​​formed. Where H is the number of global tracks, and the elements in the matrix are... This is the Mahalanobis distance between the two matching elements; if a match is rejected, the element is ∞. Step 2.6: Based on this, perform virtual track completion on the trajectory matching cost matrix to obtain the augmented trajectory matching cost matrix. Furthermore, the Hungarian algorithm is used to solve the problem. , in A matching matrix of 0-1 matrix elements. To augment the trajectory matching cost matrix The element in the i-th row and j-th column, For matrix The element in the i-th row and j-th column is used to obtain the target track-track matching pair; After obtaining the optimal matching sequence, the central node labels the matching sub-platform track sequences according to the global track sequence and returns the label-track data packet to the sub-platform, thereby achieving the unification of the numbering, track and observation data of each platform.

4. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 3, characterized in that, In step 3, based on the matched number-track-observation data, distributed model parameter collaborative optimization based on federated learning is adopted. Without sharing the original data, the collective data is used to improve the cognitive model quality of each platform for target track maneuvering, thereby fundamentally improving the prediction ability for complex maneuvering targets.

5. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 4, characterized in that, Step 3 also includes: Step 3.1: Each platform independently builds and maintains a Gaussian process model for each target it tracks, based on its local dataset. , where x(t) is the target trajectory, which is described as a function that varies with time. This is a priori mean function, usually set to zero or a simple trend. The covariance function defines the correlation between states at different time points. Under this model, the state of the target at any time is not only determined by its neighboring times, but also by the global constraints of the observed samples in the entire time series. Step 3.2: Each platform uses its maintained Gaussian process model and the matched dataset. Construct hyperparameters for the prior Gaussian process Maximum likelihood function , in Let be the vector formed by the linearized measurements of platform q. For locally linearized diagonal matrices, Let glym matrix be the kernel function with respect to the sampling timestamp. The maximum likelihood function represents a constant; based on this maximum likelihood function and the local dataset, each platform solves for the optimal hyperparameters by maximizing the maximum likelihood function using methods such as gradient descent. As model parameters when platform collaboration is not considered; Step 3.3: In order to achieve consensus among the Gaussian process models of various platforms and integrate global information, a federated learning framework is adopted. Under the premise of protecting the privacy of local data on each platform, a set of globally consistent hyperparameters is found to maximize the joint likelihood of the local models of all platforms. Step 3.4: Use the alternating multiplier method to find the globally optimal hyperparameters. Each platform maintains a local set of hyperparameters. With Lagrange multipliers .

6. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 5, characterized in that, Step 3.4 also includes: Step 3.4.1: Each platform q publishes the local hyperparameters obtained under the current iteration round k to the central node. With Lagrange multipliers The central node collects all updated local hyperparameters and Lagrange multipliers from all platforms, and then performs global aggregation to obtain global variables. in The number of parameters received. >0 is the preset penalty function factor; this global variable integrates the local information of all platforms to form a new global consensus parameter, and then broadcasts the global consensus parameter to each platform; when performing global aggregation, it is not required to receive all platform parameters. For parameters that lack updates, the original parameter information is used for global aggregation. Step 3.4.2: Global variables broadcast by the receiving center on each platform and compared it with the local Lagrange multipliers To fix the issue, when local datasets or global variables are updated on each platform, the following optimization problem is solved locally to update the local hyperparameters: in .) represents the optimization variable when the objective function is minimized. The first term in the formula represents the likelihood optimization objective, and the second and third terms represent the degree of global consensus. This update drives the local parameters to move closer to the global consensus while fitting their own data. This step can be performed in parallel with the global variable update. When the platform does not receive the global variable, it still uses the global variable from the previous moment for local updates. Step 3.4.3: Each platform updates its own Lagrange multipliers based on the latest local and global parameters to strengthen consistency constraints. ( ) This formula is essentially an update of the dual variable of the consistency constraint, achieved through continuous iterative updates. The system can gradually reduce the deviation between the parameters of each platform and the global parameters, thereby improving the convergence and consistency of the federated learning process. Step 3.4.4 The federated learning process ends when the maximum number of iterations is reached.

7. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 6, characterized in that, Step 4 also includes: Step 4.1: Using the optimized unified model and the associated dataset, given a future time t... The predicted waypoints Each platform constructs a joint distribution of linearized data and points to be predicted based on the prior Gaussian process distribution; Step 4.2: Each platform, based on its constructed prior Gaussian process model, performs prediction on the track points at each time step, using the matched dataset. You can get The posterior Gaussian distribution has an expectation and covariance of , respectively. , , in This represents the covariance between data points. = Represents the data and the track points to be predicted The covariance is obtained through a prior Gaussian process. For locally linearized diagonal matrices, The vector formed by the linearized measurements of platform q; both can be used as the platform q with respect to waypoints. The predicted value and the measurement of prediction uncertainty are as follows: the smaller the posterior covariance, the higher the confidence of the platform q in predicting the waypoint based on the current matching data; conversely, the larger the posterior covariance, the greater the uncertainty of the prediction result at that moment. Step 4.3: To obtain globally consistent perception results and achieve complementarity among platforms in terms of perception data and capabilities, a distributed information fusion mechanism is adopted to integrate the tracking results of each platform. Each platform uploads its trajectory prediction results to the central platform, which then calculates the results according to information weights. Adaptive fusion of covariance and variance: Separate center platform for track points Predicted values ​​and measures of prediction uncertainty; inverse of the covariance matrix The covariance of platform q represents the amount of predictive information. A smaller covariance indicates a more reliable prediction result and a greater impact on the fusion result; a larger covariance indicates higher prediction uncertainty and a correspondingly lower contribution. Through this adaptive weighted fusion mechanism, the central platform can fully utilize the complementarity of different platforms in terms of observation range, sensing accuracy, and data quality to obtain globally consistent and more reliable trajectory prediction results. The fused result... and These values ​​serve as the central platform's global prediction value and global prediction uncertainty measure for waypoints, respectively, and are broadcast to each sub-platform to provide a unified perception basis for subsequent collaborative decision-making.

8. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 7, characterized in that, After information fusion, each platform generates trajectory prediction results with uncertainty quantification. Step 5, based on the prediction results, adopts an intelligent auction mechanism to generate a set of executable and efficient task deployment instructions, so as to achieve the global optimal or near-optimal configuration of sensing resources in a dynamic environment.

9. The method for cooperative perception of maneuvering targets based on an edge-intelligent unmanned aerial system as described in claim 8, characterized in that, Step 5 further includes: Step 5.1: Based on the trajectory prediction results with uncertainty quantification, each platform performs a self-evaluation of each trajectory perception and calculates the benefits. Unlike the traditional benefit-cost framework that requires the separate construction of complex benefit and cost functions, this method does not introduce a complex task benefit model separately. Instead, it directly uses the reduction in prediction uncertainty as the perception benefit and combines it with the energy consumption cost required for the platform to perform the sampling action for a comprehensive evaluation. ) in ), The degree to which the observations collected by platform q at the next sampling time contribute to the reduction of uncertainty is characterized. For The sampling space centered on platform q can be geographically dispersed when the platform q has high mobility, and tend to be concentrated otherwise. This represents the energy cost of platform q in sensing target trajectory i; therefore, this evaluation metric can measure the maximum ability of platform q to improve tracking performance from the current state, and can avoid the problem of simply pursuing prediction accuracy while ignoring platform resource consumption, thus achieving an adaptive balance between sensing benefits and energy consumption constraints. Step 5.2: After evaluating the value of each local track tracking, each platform begins to execute task bidding. Each platform considers its own energy consumption limitations, comparing the energy consumption of its allocated target tracks with its maximum energy capacity to determine the list of executable track tracking. To reduce communication burden, each platform only compiles its top K highest bids to form data packets. Send it to the central node; Step 5.3: After receiving bids from all platforms, the central node sorts all bid pairs to obtain a global bid list. Based on this global bidding list, the central node platform evaluates the priority of target task i by considering factors such as the platform's tracking revenue for the target, the platform's relative scarcity of tasks, and the number of platforms that have already assigned the task: in This represents the number of platforms that have been allocated to target i. This indicates the number of platforms on which tracking target i can be executed. and It is a decreasing function of the dependent variable; this priority function works by combining three parts: To reduce the priority of further allocation for targets that have already acquired a significant amount of platform resources, Used to increase the resource scarcity weight for targets with fewer executable platforms. This is used to characterize the optimal tracking capability that the current platform set can provide for the target. Through the above priority evaluation, the central node can prioritize the processing of target tasks with high tracking benefits, insufficient resource coverage, and scarce executable platforms, thereby avoiding uneven allocation of platform resources and improving the overall efficiency of multi-platform collaborative tracking task allocation. Step 5.4: The central node scans the platform-tracking task evaluation list sequentially using a greedy strategy. If target i has not yet been assigned and the task load of platform q has not exceeded the limit, target i is assigned to platform q; otherwise, the bid is skipped and the next item is continued. After the scan is completed, the platform-tracking allocation result is obtained. The central node sends it to each sub-platform. After receiving the result, the sub-platform reorganizes the next bid list until each platform no longer submits a new bid, thus completing the task allocation. Intelligent task allocation is executed repeatedly at a fixed frequency to adapt to changes in target movement, platform attitude, and dynamic changes in communication topology. When a platform detects a task execution failure, it can immediately report to the central node, which will trigger a temporary re-attempt to achieve rapid adaptive task reallocation. This centralized auction mechanism can be seamlessly coupled with the trajectory prediction module in the above steps. The platform directly uses the posterior covariance obtained from the prediction results to calculate the bid, achieving low computational complexity in the perception-decision closed loop. The central node only needs to sort and greedily allocate the limited bids, and the communication overhead is small. Each platform only uploads the top K highest bids and does not need to upload the original measurement or complete state information. Overall, this scheme significantly reduces the implementation complexity while maintaining theoretical rationality, and achieves efficient task coordination and dynamic perception resource scheduling across platforms.

10. The method for cooperative perception of maneuvering targets based on an edge-intelligent aerial unmanned system as described in claim 9, characterized in that, Steps 1-5 constitute a self-driven intelligent closed loop. The allocation decision in step 5 guides the data collection action in step 1. The new data generated is correlated in step 2 and used to drive the further co-evolution of the model in step 3. The evolved model then generates more accurate predictions in step 4, thereby leading to better resource allocation. This cycle repeats, and the entire system can achieve adaptive coordination of perception, cognition, decision-making and execution in complex dynamic environments, continuously maintaining the best global situational awareness capability.

Images