A method for active resource allocation of a mar system based on pose prediction under a 6g network framework

Abstract

The application provides a 6G network framework-based MAR system active resource allocation method based on pose prediction, relates to the technical field of 6G mobile communication, and first extracts spatial environment structure dependence features of user historical poses by using a graph attention network, extracts time dimension dynamic features by combining a gated recurrent unit, constructs a space-time pose prediction model to output predicted pose information in the next time slot, then establishes a nonlinear mapping between the predicted pose and the resource demand quantity of a calculation task by using a Gaussian process regression, outputs a task quantity prediction value and uncertainty estimation, and finally adopts a double-delay deep deterministic policy gradient-parameterized deep Q network algorithm to make decisions on a mixed action space with the optimization target of minimizing end-to-end time delay, and outputs resource allocation actions in the current time slot; the application can realize fine resource scheduling with task time sequence constraint perception, effectively reduce end-to-end time delay, and improve user experience quality.

Description

Technical Field

[0001] This invention relates to the field of 6G mobile communication technology, and more particularly to an active resource allocation method for a MAR system based on pose prediction under a 6G network framework. Background Technology

[0002] Mobile Augmented Reality (MAR) systems are comprehensive application systems that deeply integrate next-generation information technologies such as computer vision, 5G / 6G communication, and Mobile Edge Computing (MEC). They aim to overlay computer-generated virtual information onto the real environment observed by the user in real time and accurately. The smooth and realistic experience of MAR relies on the synergy of key technologies: tracking and registration, and virtual-real fusion. Tracking and registration refers to the continuous measurement of the position and orientation of physical entities and their alignment within a coordinate system, ensuring spatial consistency between virtual objects and physical entities. Virtual-real fusion considers the impact of the real environment on virtual objects; for example, virtual objects need to maintain consistent lighting characteristics with the real environment in real time. The efficient collaboration of these two key technologies strongly depends on the precise and dynamic allocation of resources by 6G networks and edge computing platforms.

[0003] In recent years, research on wireless resource management strategies for AR applications has received widespread attention. ZHANG L et al. explored the task offloading and placement problem of AR overlay rendering in multi-user mobile augmented reality systems, aiming to maximize user experience quality and minimize service costs. CHENG S et al. jointly optimized video configuration selection, bandwidth allocation, and computation offloading decisions to address the computationally intensive and latency-sensitive MAR video transmission problem. MAO S et al. designed a resource management framework for augmented reality-enabled in-vehicle metaverse to balance vehicle energy consumption and operator revenue. Facing the challenges of real-time task response capabilities and multi-agent resource competition in MAR systems, CHENG S et al. proposed a central control mechanism to jointly optimize bandwidth allocation and video configuration to minimize AoAI. However, these studies abstract the computational tasks of AR applications into independent units, failing to fully consider the inherent temporal constraints within the tasks. Figure 1 As shown, the three key technologies often involve multiple complex and strictly time-constrained computational tasks.

[0004] Furthermore, current mainstream AR development engines schedule different computational tasks in a pipeline manner. For example, ARKit and ARCore use a tracking and registration method based on Visual Simultaneous Localization and Mapping (vSLAM). Their workflow involves the camera capturing images, the vSLAM module outputting poses, and finally, the rendering module performing virtual fusion. This sequential processing means that when rendering content, the pose information used by the system no longer represents the user's current perspective, directly impacting the rendering effect and resource utilization efficiency in the subsequent virtual-real fusion process. To address this issue, firstly, the camera pose information needs to shift from passive optimization to active prediction. This requires not only accurately calculating where the camera is now, but also learning to predict where the camera is about to go. Then, based on the predicted camera pose information, the system needs to proactively allocate resources for key tasks such as tracking registration and virtual-real fusion.

[0005] Common pose prediction algorithms fall into two categories: filter-based and deep learning-based. SONG X et al. designed a "minimum hysteresis" filter to reduce jitter caused by noise in pose data, while simultaneously reducing hysteresis latency through backtracking updates and compensation strategies. NWOBODO O et al. employed a framework fusing Kalman filters and particle filters to predict camera pose, dynamically adjusting the contribution weights of the Kalman and particle filters based on sensor data, motion state, and computational resource capacity, achieving accurate and smooth appearance of virtual objects in the user's field of view. Filter-based pose prediction methods are typically based on simple linear or uniformly accelerated motion models; however, they suffer from difficulties in modeling complex nonlinear motion and slow response to sudden environmental changes in complex motion scenarios. In recent years, deep learning-based pose prediction methods have offered new possibilities for extracting motion pattern styles from training data. ABUHASHISH F et al. used an LSTM network to process inertial measurement unit data, generating the motion trajectory of an AR terminal and predicting the field of view, thereby reducing the energy consumption requirements of mobile and resource-constrained devices in real-time scenarios. To reduce the impact of pose estimation lag on the AR experience, CHENH et al. proposed a motion uncertainty coding and decoding network to address the misalignment between virtual objects and the physical world caused by directly using the current pose for rendering. This network uses LSTM to quantify the uncertainty of the input data and predict the uncertainty of the output motion.

[0006] While existing deep learning-based pose prediction schemes have made significant progress in improving the real-time performance of MAR systems, these methods primarily focus on learning motion patterns from time series. However, changes in camera pose are essentially reflections of its motion trajectory in the environmental space. Furthermore, treating different parts of the input sequence equally or using simple weighting fails to capture historical states or environmental features most relevant to the current motion. This capability is crucial in dynamic or cluttered scenes, as it requires ignoring irrelevant or distracting visual information and focusing more on stable environmental structures. There is an urgent need for a novel pose prediction framework that can simultaneously and effectively capture both spatial environmental structure dependencies and temporal dynamic features. Summary of the Invention

[0007] This invention proposes an active resource allocation method for MAR systems based on pose prediction under a 6G network framework. It aims to solve the problems in existing mobile augmented reality systems where resource allocation does not fully consider task timing constraints and pose prediction lacks spatial environment structure modeling, resulting in long end-to-end latency and poor virtual-real fusion experience.

[0008] This invention provides an active resource allocation method for a MAR system based on pose prediction under a 6G network framework, characterized by the following steps:

[0009] Step S1: Construct a spatiotemporal pose prediction model based on graph attention network and gated recurrent unit. Use the graph attention network to extract spatial environment structure dependency features from the historical pose information of mobile augmented reality users. Use the gated recurrent unit to extract the temporal dimension dynamic features from the historical pose information. Then, output the predicted pose information of the next time slot through the decoder.

[0010] Step S2: Construct a task quantity prediction proxy model based on Gaussian process regression, input the predicted pose information into the task quantity prediction proxy model, establish a nonlinear mapping relationship between the predicted pose information and the computational task resource requirements, and output the predicted value of the computational task resource requirements for the next time slot and its uncertainty estimate.

[0011] Step S3: With minimizing end-to-end latency as the optimization objective, the predicted resource requirements of the computing tasks are used as prior information. A dual-latency deep deterministic strategy gradient-parameterized deep Q-network algorithm is employed to make decisions on the hybrid action space, output the resource allocation action for the current time slot, and perform resource allocation on the computing tasks of the mobile augmented reality system according to the resource allocation action. The hybrid action space includes task offloading decisions, wireless resource block allocation decisions, the computing resource allocation ratio for tracking and registration tasks, and the computing resource allocation ratio for virtual-real fusion tasks.

[0012] Furthermore, the specific method of step S1 includes:

[0013] The historical pose information of mobile augmented reality (AR) users is represented as follows:

[0014]

[0015] in, Represents coordinates, Indicates azimuth;

[0016] The length of the observation sequence is Then the pose observation sequence can be represented as:

[0017] The predicted pose is represented as ;

[0018] Historical pose information is first processed by an MLP layer to extract features, which are represented as follows:

[0019]

[0020] in, and It is an MLP weighting and biasing. It is the ReLU activation function;

[0021] Next, the spatial features are processed through the Graph Attention Network (GAT). for

[0022]

[0023] in, It is the attention mechanism function. and It is the relevance weight; , and These are the softmax function, the LeakyReLU activation function, and the sigmoid function, respectively.

[0024] It will be used as input to the gated recurrent unit (GRU) for the following operations:

[0025]

[0026] in, It's an update gate. It's a door reset. It is a characteristic of the GRU output in the previous time slot. It is a hidden feature. It is an activation function; , , , , and These are the weights;

[0027] Finally, the GRU output features The input is fed into the decoder GRU, and the output is the predicted pose information. for:

[0028]

[0029] in, For s-linear transformation layer;

[0030] The position prediction loss function and the angle loss function in the user pose information are as follows:

[0031]

[0032] Therefore, the loss function for pose prediction is:

[0033] .

[0034] Furthermore, the specific method of step S2 includes:

[0035] Assuming the resource requirements of the virtual fusion task It is the user pose vector The function follows a Gaussian process distribution:

[0036]

[0037]

[0038] in, It is a mean function; It is the covariance function;

[0039] For the pose input of the training set and corresponding resource demand observations Considering observation noise Predict pose Corresponding task volume prediction value Compared with observed values The joint prior distribution is expressed as:

[0040]

[0041] in, It is the covariance matrix of the training set. It is the cross-covariance matrix between the training set and the test points;

[0042] According to Bayes' theorem, given the observation data , and test input The posterior distribution of the predicted values ​​is also a Gaussian distribution, following the following distribution:

[0043]

[0044] Among them, the predicted mean With prediction variance They are respectively:

[0045]

[0046]

[0047] Among them, the predicted mean This will be used as an estimate of the system's resource demand at the next moment, while the prediction variance... This quantifies the uncertainty of the model.

[0048] Furthermore, the method is applied to the following task architecture: the computational tasks of the mobile augmented reality system are modeled as a directed acyclic graph; the directed acyclic graph includes parallel execution of tracking and registration task branches and pose prediction and virtual-real fusion task branches, and is equipped with virtual entry nodes and virtual exit nodes; the tracking and registration task branches include sub-tasks of vSLAM map update, pose estimation and tracking, and scene reconstruction executed according to dependencies; the pose prediction and virtual-real fusion task branches include sub-tasks of pose prediction, lighting estimation, realistic rendering, and occlusion processing executed according to dependencies.

[0049] Furthermore, the specific method for minimizing end-to-end latency as the optimization objective in step S3 includes:

[0050] The end-to-end latency of the pose prediction and virtual-real fusion task is defined as the total completion time obtained by summing the communication latency and computation latency of each subtask in the branch according to the task dependency relationship. Among them, the communication latency is calculated based on the Shannon formula according to the channel bandwidth, channel gain, transmit power and noise. The local computation latency is obtained by dividing the computation task amount by the proportion of local computation resources allocated. The edge computation latency is obtained by dividing the computation task amount by the proportion of computation resources allocated to the edge server. The earliest start time of the dependent subtask is the maximum value of the sum of the earliest completion time of all direct predecessor tasks and the transmission latency. The earliest completion time is the earliest start time plus the computation latency.

[0051] The optimization objective is set to minimize the end-to-end latency of the pose prediction and virtual-real fusion tasks, with the completion latency of the tracking registration task being less than the duration of one system time slot as a constraint.

[0052] Furthermore, in step S3, the specific method for using the dual-delay deep deterministic strategy gradient-parameterized deep Q-network algorithm to make decisions on the hybrid action space and output the resource allocation action for the current time slot includes:

[0053] S3-1: Construct a state space, which consists of observable information from all AR terminals within the current time slot, represented as follows: ; This represents observable information; the observable information includes: channel gain. Historical pose information Track the number of registrations Virtual-real integration task volume ;

[0054] S3-2: Construct action space A, which includes: unloading decision. Resource block allocation decision Track the proportion of registered computing resources and the ratio of virtual and real computing resources ;

[0055] S3-3: The target Q value is calculated using a dual Q network mechanism, that is, two independent Critic networks are used to calculate the Q value of the action respectively, and the smaller of the two is taken as the target Q value to alleviate the problem of overestimation of action value.

[0056] The target network parameters of the two Critics are used and If so, then the value of the target network is represented as:

[0057]

[0058] S3-4: A delayed update mechanism is adopted, setting the update frequency of the Actor network to be lower than that of the Critic network, and updating the Actor network once after the Critic network updates a preset number of times.

[0059] S3-5: A target action smoothing mechanism is adopted, in which truncated Gaussian noise is added to the output of the target policy network to generate smooth target actions, wherein the amplitude of the truncated Gaussian noise is controlled by preset parameters.

[0060] S3-6: Based on the state space, action space and the above mechanism, with the goal of minimizing the end-to-end latency of pose prediction and virtual-real fusion tasks, output the resource allocation action for the current time slot.

[0061] Furthermore, in step S3, the Actor network and Critic network in the dual-delay deep deterministic policy gradient-parameterized deep Q-network algorithm both adopt a three-layer fully connected structure, with 512, 256 and 128 hidden layer neurons respectively, and use the ReLU activation function; the discount factor is 0.98.

[0062] Furthermore, before step 2, the method includes: using mutual information to analyze the nonlinear relationship between the pose vector and the resource demand, using Moran's I index to determine the clustering effect of the resource demand in the pose space, and when a clustering effect exists, using the task quantity prediction proxy model based on Gaussian process regression.

[0063] Compared with the prior art, the present invention has the following advantages:

[0064] 1. Improve pose prediction accuracy: By constructing a spatiotemporal prediction model that combines graph attention network and gated recurrent unit, and extracting the spatial environment structure dependence and temporal dynamic features of user motion, this method can more accurately predict the pose information of the next time slot compared with existing methods, providing a reliable basis for subsequent resource allocation.

[0065] 2. Achieve accurate prediction of task load: By using Gaussian process regression to establish a nonlinear mapping between pose and resource requirements, not only can a point estimate of the task load be given, but the uncertainty of the prediction can also be quantified. This allows the system to reserve resource margins in high uncertainty scenarios and to aggressively allocate resources in low uncertainty scenarios, effectively balancing computational efficiency and system stability.

[0066] 3. Reduce end-to-end latency: The dual-latency deep deterministic strategy gradient-parameterized deep Q-network algorithm is adopted to jointly optimize the hybrid action space (discrete unloading / resource block decision and continuous computing resource ratio). The optimization objective is to minimize the end-to-end latency of pose prediction and virtual-real fusion tasks, which significantly reduces the ghosting and lag of virtual content and improves the quality of user experience.

[0067] 4. Refined resource scheduling: This invention fully considers the temporal constraints between AR application tasks and introduces a directed acyclic graph into the resource allocation model to realize task-related perception of computational offloading and resource allocation. This overcomes the shortcomings of existing methods that treat tasks as independent units and improves resource utilization efficiency in multi-user and multi-edge server scenarios.

[0068] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0069] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0070] Figure 1 A flowchart for computational tasks in AR applications;

[0071] Figure 2 This is a sequence diagram for AR application tasks;

[0072] Figure 3 Directed acyclic graph for AR application computation tasks;

[0073] Figure 4 Here is a diagram of the GAT-GRU pose prediction model architecture;

[0074] Figure 5 This is a diagram of the TD3-PD3QN framework.

[0075] Figure 6 The graph shows the convergence curves of the loss during model training and validation.

[0076] Figure 7 A complex nonlinear analysis diagram of pose and resource requirements;

[0077] Figure 8 A comparison chart of GPR model predicted values ​​and actual values ​​and their confidence intervals;

[0078] Figure 9 This is a graph showing the prediction performance of the GPR model on random samples. Detailed Implementation

[0079] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0080] This invention proposes a pose prediction model based on Graph Attention Network (GAT) and Gated Recurrent Unit (GRU). The model utilizes GAT to extract spatial features of dynamic changes in AR user pose, focusing on the most relevant pose information, and employs GRU to extract long-term temporal dependencies and short-term changes in the pose trajectory.

[0081] This invention constructs a probabilistic surrogate model based on Gaussian Process Regression (GPR) with information entropy to achieve a nonlinear mapping between predicted pose and resource demand, obtain accurate estimates of computational workload and quantify uncertainty, and provide accurate prior information for proactive resource decision-making.

[0082] This invention designs a joint optimization mechanism for communication and computing resources based on task timing constraints. With minimizing end-to-end latency as the optimization objective, it utilizes the Twin Delayed Deep Deterministic Policy Gradient-Parameterized Deep Dueling Double Q-Network (TD3-PD3QN) algorithm to solve the latency minimization problem with mixed actions.

[0083] This invention considers an edge network architecture for MAR systems consisting of MEC servers and AR terminals, wherein the set of MEC servers is represented as follows: The AR terminal set is represented as Let m and n represent the indices of the MEC server and AR terminal, respectively, and M and N represent the number of MEC servers and AR terminals, respectively. The computing resource capacity of MEC server m is expressed as... The computing resource capacity of AR terminal n is expressed as: The system's time slot set is defined as follows: The length of a single time slot is defined as The MEC server assists the AR terminal in completing AR application computing tasks. An AR terminal device can select at most one MEC server for task offloading within a time slot. To adapt to the dynamic changes in uplink and downlink traffic between the MEC server and the AR terminal, this invention adopts a hybrid access mechanism of dynamic Time Division Duplex (TDD) and Orthogonal Frequency-Division Multiple Access (OFDMA) at the physical layer transmission scheme. This scheme enables the network to dynamically adjust the transmission direction (uplink or downlink) of wireless resources and perform multi-user scheduling at a smaller time granularity (time slot) based on the actual service flow generated by the AR application. The wireless resource capacity is... The radio resource blocks are divided into B uniform blocks, each containing a specific number of subcarriers within a time slot. The set of radio resource blocks for the MEC server is represented as follows: .

[0084] like Figure 2 Therefore, in this invention, the AR application begins image acquisition at time slot t, and then performs vSLAM map updates, pose estimation and tracking, and scene reconstruction based on the image at the current time. Simultaneously, pose prediction is performed at time slot t, and illumination estimation, realistic rendering, and occlusion processing are performed based on the predicted pose information.

[0085] AR application representation TR represents the tracking and registration task, and VA represents the pose prediction and virtual-real fusion task. To represent the dependencies between the subtasks of the AR application, this invention abstracts them into a directed acyclic graph, such as... Figure 3 As shown. Since tracking registration, pose prediction, and virtual-real fusion are executed in parallel, they compete for computing and communication resources. Therefore, to address the problem of unclear scheduling logic caused by multiple initial subtasks, this invention sets up virtual entry and exit nodes, denoted as S and E.

[0086] The graph representation of tracking registration is as follows . This indicates the set of registered subtasks to be tracked, and the node represents that. This indicates tracking the registration of subtasks, where when When vSLAM map is updated, Time represents pose estimation and tracking, when Time represents scene reconstruction. The set of edges is represented as , Subtasks Kazuko's Mission Dependencies between subtasks Must be in subtask Execution can only begin after the previous one has finished. ,otherwise, I represents the number of registered subtasks to track. The basic properties of registered subtasks to track can be represented as a quadruple. If the AR terminal uninstalls and registers with the MEC server m, then... ,otherwise If transmitting subtasks At that time, the communication between MEC server m and AR terminal n occupied resource block b, then ,otherwise, . Subtasks Transmit to subtask The amount of data. Indicates completion of subtask The amount of computing resources required at that time.

[0087] Pose prediction and virtual-real fusion are represented as . This represents the set of subtasks for pose prediction and virtual-real fusion. This represents the pose prediction or virtual-real fusion subtask, where when =1 indicates pose prediction, when =2 indicates illumination estimation, when =3 indicates realism, when =4 indicates occlusion. The set of edges is represented as... , Subtasks Kazuko's Mission The dependencies between them, where J represents the number of virtual and physical rendering subtasks. If the subtasks... Must be in subtask Execution can only begin after the previous one has finished. ,otherwise, The basic properties of the virtual-real fusion subtask can be represented as a quadruple. AR terminal offload pose prediction or virtual-real fusion subtask To MEC server m, then ,otherwise, If transmitting subtasks At that time, the communication between MEC server m and AR terminal n occupied resource block b, then ,otherwise, . Subtasks Transmit to subtask The amount of data. Indicates completion of subtask The amount of computing resources required at that time.

[0088] In MAR systems, the user's Quality of Experience (QoE) is highly sensitive to latency. End-to-end latency is a key factor affecting immersion and can be mainly categorized into two parts: communication latency and computation latency.

[0089] The communication latency between MEC server m and AR terminal n is:

[0090] (1)

[0091] Among them, when In this case, formula (1) is used to calculate the tracking registration subtask. Communication delay; when At that time, formula (1) is the sub-task of material forecasting and virtual-real integration. Communication delay. It is the transmission rate between MEC server m and AR terminal n, which can be expressed using Shannon's formula as follows: ,in, This represents the bandwidth of channel b. It is channel gain. It's noise. This represents the transmission power. When AR terminal n sends data to MEC server m, the transmission power of AR terminal n is... When MEC server m sends data to AR terminal n, the transmit power of MEC server m is... .

[0092] If the computation task is executed on the AR terminal, its local execution time is:

[0093] (2)

[0094] Among them, tracking registration subtask Local computing latency The proportion of computing resources allocated to it by the AR terminal It is positively correlated, virtual fusion subtask Local computing latency The proportion of computing resources allocated to it by the AR terminal It is positively correlated.

[0095] If the computation task is executed on the MEC server, its MEC execution time is:

[0096] (3)

[0097] Among them, tracking registration subtask MEC computation latency The proportion of computing resources allocated to it by the MEC server It is positively correlated, virtual fusion subtask MEC computation latency The proportion of computing resources allocated to it by the MEC server It is positively correlated.

[0098] This invention defines the earliest start time of a dependent task as... The earliest completion time is defined as Subtasks or of It is ready only after all prerequisite tasks are completed and the calculation results are transmitted. The calculation formula is as follows:

[0099] (4)

[0100] in, Subtasks or The direct precursor mission. It is an XOR operation.

[0101] Subtask or of It can be represented as

[0102] (5)

[0103] Since the MEC server is periodically selected for computational task offloading in each time slot, this invention ensures that the tracking registration result is ultimately returned to the AR terminal device. Furthermore, the virtual-real fusion result needs to be displayed on the AR terminal device, therefore it also needs to be returned to the AR terminal device. Therefore, the total tracking registration latency is:

[0104] (6)

[0105] The end-to-end latency for pose prediction and virtual-real fusion is:

[0106] (7)

[0107] This invention aims to minimize the end-to-end latency of pose prediction and virtual-real fusion, as this process directly determines the AR video effect seen by AR users on their terminal devices. Minimizing this latency effectively reduces the ghosting and lag of virtual content, a key optimization direction for improving the immersion, smoothness, and realism of the AR experience. Simultaneously, a hard constraint is that the tracking registration latency must be less than one time slot. The fundamental reason is that tracking registration is the foundation of all subsequent AR processing; it must be completed before the next image acquisition (i.e., the start of time slot t+1) to ensure real-time updates of the world model. If this process times out, it will directly cause virtual objects to "drift" relative to the real world or even lose tracking, causing the entire AR experience to collapse.

[0108] Based on this, the following objective function is constructed:

[0109] (8)

[0110] in, and These represent the available computing resources for the AR terminal and the MEC server, respectively.

[0111] To improve the user QoE of MAR systems, this invention proposes a pose-prediction based proactive resource allocation (PP-PRA) strategy. This strategy predicts the camera pose at the next moment, enabling computational resource offloading and forward allocation of communication radio resource blocks. First, a pose prediction framework based on a Graph Attention Network (GAT) and a Gated Recurrent Unit (GRU) learns the camera trajectory from spatial environmental structure and temporal dynamic features, improving prediction accuracy. Second, a Gaussian regression-based task estimation algorithm is used to establish a nonlinear mapping between pose and task data volume, and the task data volume is predicted. Finally, a deep reinforcement learning algorithm dynamically optimizes computational offloading and radio resource block scheduling decisions based on the predicted task data volume.

[0112] (1) Pose prediction based on GAT-GRU;

[0113] like Figure 4 As shown, the pose prediction framework constructed in this invention consists of two parts: an encoder and a decoder. The encoder, composed of a Gaussian Artificial Atlas (GAT) and a Giant Root Array (GRU), is used to extract the spatial and temporal features of the AR user's historical pose information. The decoder, implemented using a GRU, is used to output future pose information.

[0114] AR user's historical pose information is represented as ,in Represents coordinates, Indicates the azimuth angle. The observation sequence length is... Then the pose observation sequence can be represented as The predicted pose is represented as .

[0115] Historical pose information is first processed by an MLP layer to extract features, which are represented as follows:

[0116] (9)

[0117] in, and It is an MLP weighting and biasing. It is the ReLU activation function.

[0118] Next, spatial features are analyzed using GAT. for:

[0119] (10)

[0120] in, It is the attention mechanism function. and It is the relevance weight. , and These are the softmax function, the LeakyReLu activation function, and the sigmoid function, respectively.

[0121] It will be used as input to the GRU and the following operations will be performed:

[0122] (11)

[0123] in, It's an update gate. It's a door reset. It is a characteristic of the GRU output in the previous time slot. It is a hidden feature. It is an activation function. , , , , and These are the weights.

[0124] Finally, the GRU output features The input is fed into the decoder GRU, and the output is the predicted pose information. for:

[0125] (12)

[0126] s-linear transform layer. The position prediction loss function and angle loss function in the user pose information are respectively...

[0127] (13)

[0128] (14)

[0129] Therefore, the loss function for pose prediction is:

[0130] (15)

[0131] (2) Task estimation algorithm based on Gaussian regression;

[0132] In MAR scenarios, there is often a complex nonlinear mapping relationship between the pose data of AR devices and the resource requirements corresponding to each time slot. Due to the high uncertainty of the underlying physical mapping mechanism, this invention employs mutual information based on information entropy to capture pose data. The impact of periodicity or locational abrupt changes on resources. Let... For pose vector, For resource requirements, the definition of mutual information can effectively capture large fluctuations in resource requirements caused by pose abrupt changes. Any component of the pose vector is related to the resource requirement. Mutual information between them is defined as:

[0133] (16)

[0134] in, It is a joint probability distribution. and It is a marginal probability distribution.

[0135] Anisotropy and Spatial Autocorrelation: Moran's I index is used to analyze whether resource demand exhibits a clustering effect in pose space. If A higher value indicates that similar poses tend to generate similar resource requirements, which provides a theoretical basis for using Gaussian processes. A partial dependency plot (PDP) is used to display the average response curve of the target variable as a function of a single pose component, in order to uncover the underlying physical constraint logic.

[0136] Based on the foregoing analysis, there is a significant angle-dominant relationship and complex nonlinear characteristics between pose and resource requirements, making it difficult for traditional linear regression models to fit such high-dimensional surfaces. Therefore, this invention employs Gaussian Regression Model (GPR) to construct a non-parametric surrogate model. GPR achieves high-precision approximation without pre-setting specific physical equations and provides an estimate of the uncertainty (confidence interval) of the prediction results, enhancing the robustness of resource scheduling.

[0137] Assuming the resource requirements of the virtual fusion task It is the user pose vector The function follows a Gaussian process distribution:

[0138] (17)

[0139] (18)

[0140] in, It is a mean function; It is the covariance function (kernel function), used to measure the similarity between different poses.

[0141] For the pose input of the training set and corresponding resource demand observations Considering observation noise Predict pose Corresponding task volume prediction value Compared with observed values The joint prior distribution is expressed as:

[0142] (19)

[0143] in, It is the covariance matrix of the training set. This is the cross-covariance matrix between the training set and the test set. According to Bayes' theorem, given the observed data... , and test input The posterior distribution of the predicted values ​​is also a Gaussian distribution, following the following distribution:

[0144] (20)

[0145] Among them, the predicted mean With prediction variance They are respectively:

[0146] (twenty one)

[0147] (twenty two)

[0148] Among them, the predicted mean This will be used as an estimate of the system's resource demand at the next moment, while the prediction variance... This quantifies the uncertainty of the model.

[0149] (3) Resource allocation algorithm based on TD3-PD3QN;

[0150] To achieve the end-to-end delay minimization defined by the objective function (8), this invention needs to solve a typical hybrid action space optimization problem. The main characteristic of this problem is the heterogeneity of the decision variables. The task unloading decision O and the communication resource block allocation decision U are discrete binary variables, while the terminal and edge server computing resource allocation decisions are high-dimensional continuous variables. Traditional deep reinforcement learning algorithms have obvious limitations in dealing with this type of problem. For example, the parameterized DQN algorithm uses an Actor network to output the continuous parameters corresponding to each discrete action, and then inputs the discrete action and continuous parameters together into the Q network to evaluate the value. This method is prone to overestimation of action value. Another common approach is action space relaxation based on continuous control algorithms, such as the DDPG algorithm or the TD3 algorithm. The principle is to relax the discrete decision variables into continuous variables, and the Actor network directly outputs continuous action values. Then, during the execution phase, the values ​​are mapped back to discrete binary values ​​through a step function or threshold determination. This method destroys the continuity of the gradient, making it difficult for the Actor network to backpropagate effectively. Therefore, this invention adopts the TD3-PD3QN algorithm, which improves the TD3 algorithm framework by introducing a double Q network, delaying policy updates, and smoothing action targets, thereby alleviating the training oscillations caused by gradient mutations at the discrete-continuous interface in traditional algorithms.

[0151] state space Composed of observations from all AR devices, denoted as . This represents observable information, including channel gain. Historical pose information

[0152] Track the number of registrations Virtual-real integration task volume .

[0153] Action space Includes: Uninstallation decision Resource block allocation decision Track the proportion of registered computing resources and the ratio of virtual and real computing resources .

[0154] The reward function is expressed as .

[0155] Figure 5 This is the TD3-PD3QN algorithm framework. The specific steps are as follows:

[0156] (1) The agent obtains the system state of the current time slot. This state integrates the user pose information for the next time slot predicted by the GAT-GRU model, the channel state, and the resource demand distribution output by the task load estimation agent model.

[0157] (2) The Actor network determines the current state. Direct output of mixed motion This action includes discrete components (such as the choice of offloading tasks to the edge cloud or processing them locally) and continuous components (such as the specific bandwidth or computation frequency ratio allocated to each task).

[0158] (3) After the system executes the allocation action, the MAR environment provides an immediate reward. Guided by minimizing end-to-end latency, transition to the next state. .

[0159] (4) Transfer the trajectory tuple The samples are stored in the experience replay pool, and then a small batch of samples is randomly selected for network training.

[0160] (5) The target Q value is calculated using two independent target Critic networks. Overestimation is suppressed by truncating the double Q learning mechanism, and the Actor network parameters are optimized by using a delayed strategy update technique.

[0161] Dual-Q network: TD3 uses two independent Critic networks to calculate the Q-value of the action, and takes the smaller one as the target Q-value. The target network parameters of the two Critic networks are... and If so, then the value of the target network is represented as:

[0162] (twenty three)

[0163] The core idea is that Q-learning and its variants often suffer from overestimation bias. Choosing a smaller Q-value during Q-updates can effectively reduce this problem. Through the mutual constraints of the two Q-networks, the network is more inclined to select actions that perform well on both networks, rather than having the policy biased towards suboptimal solutions due to errors in one network. During TD3 training, the two Q-networks are updated independently to reduce their correlation and avoid error accumulation due to gradient propagation. Their loss functions are as follows:

[0164] (twenty four)

[0165] (25)

[0166] Delayed Policy Update: TD3 introduces a delayed policy update mechanism based on DDPG, meaning the update frequency of the Actor network is lower than that of the Critic network, typically set to update the Actor only once after the Critic network updates twice. The theoretical basis for this strategy is that the Critic network has a large estimation error in the early stages of training. If the Actor network updates frequently, it may optimize along unstable gradient directions, leading to deviations from the optimal policy space. Delayed policy updates allow the Critic network to reach a more stable state after multiple updates, reducing the unstable learning impact caused by large Critic errors in the early stages of training. Furthermore, TD3 also adopts the soft update mechanism of DDPG, allowing the target network parameters to converge to the main network parameters with smaller steps, thereby further enhancing training stability. The update formula is:

[0167] (26)

[0168] (27)

[0169] (28)

[0170] in, and These are the soft update system data for the actor network and the critic network, respectively.

[0171] Target Action Smoothing: In DDPG, the target value is calculated directly based on the output of the target policy network. However, this direct use of the policy network output can lead to unstable estimations by the policy overfitting the Q-network. To mitigate this problem, TD3 adds slight random Gaussian noise to the output of the target policy network and calculates the expected value of the action in the training data to approximate the average action. The action smoothing formula is:

[0172] (29)

[0173] in, It is a truncated Gaussian noise that satisfies:

[0174] (30)

[0175] In the formula To mitigate the noise amplitude, this design reduces the estimation error of the policy network by overfitting the network, while enhancing the policy's exploratory nature. When updating the Critic network, target actions with random perturbations are used, making the Critic estimation smoother within a certain range, thereby improving the policy's generalization ability under different environmental conditions.

[0176] Example

[0177] This invention simulates an application scenario based on a typical MAR system, including the number of MEC servers. The number of ARs served by each server is AR users are evenly distributed across the edge network. The MEC service radius is 50 meters. The transmit power of the MEC server and AR terminal is 46 dBm and 23 dBm, respectively. The noise power is -174 dBm / Hz, and the total channel bandwidth is [not specified]. =30MHz, channel gain is affected by Rayleigh fading. The total computing resources for the MEC server and AR device are respectively... and .

[0178] The GAT network consists of two graph attention layers, each containing four attention heads. The hidden layer has a dimension of 64. It aggregates the features of neighboring nodes through a multi-head attention mechanism and uses the ELU activation function and a Dropout rate of 0.2 for regularization.

[0179] The GRU network is a two-layer stacked gated recurrent unit structure with an input dimension of 256 and a hidden layer dimension of 64. It captures temporal dependencies through reset gate and update gate mechanisms, uses a dropout rate of 0.2 between layers, and finally maps the 64-dimensional features to a 3-dimensional output through a fully connected layer.

[0180] TD3-PD3QN is composed of an Actor-Critic architecture, which includes an Actor network for parameterized actions and two Critic networks with adversarial double-Q network structures. The Actor network employs a three-layer fully connected structure, with 512, 256, and 128 neurons per layer, using the ReLU activation function in each layer, and ultimately outputting continuous action parameters through the Sigmoid function. Each Critic network consists of three shared layers that process the joint input of state and action parameters, and a two-stream branch structure that separately calculates state value and action advantage. Discount factor =0.98.

[0181] The GAT-GRU spatiotemporal pose prediction framework is trained on the TUM RGB-D dataset (https: / / cvg.cit.tum.de / data / datasets / rgbd-dataset). This invention uses four models as benchmark models for comparison:

[0182] GAT-LSTM: GAT and LSTM are used as encoders to extract spatiotemporal features of pose information, and LSTM is used as a decoder.

[0183] GCN-GRU: GCN is used for spatial feature extraction, while GRU extracts features in the time dimension.

[0184] Social-LSTM: Introduces a "social pooling" layer to aggregate the hidden states of neighboring users.

[0185] St-GAT (Spatio-Temporal GAT) is an advanced spatio-temporal prediction model that combines LSTM with GAT to capture the spatio-temporal interaction behavior of pedestrians. As one of the mainstream benchmarks in academia, the introduction of STGAT helps to verify the specific competitiveness of the GAT-GRU architecture optimized for MAR scenes in this invention on pose prediction tasks.

[0186] Table 1 shows a comparison of the quantitative metrics of each model on the standard test set. Experimental results show that the GAT-GRU designed in this invention significantly outperforms GCN-GRU, Social-GRU, and St-GAT in pose prediction accuracy, and its inference execution efficiency is nearly twice that of GAT-LSTM. Compared with GAT-LSTM, the accuracy difference is small, but GAT-GRU achieves a better balance between accuracy and efficiency, meeting the stringent end-to-end latency requirements of the MAR system.

[0187] Table 1 Comparison of prediction performance of different models

[0188]

[0189] Figure 6 The diagram illustrates the loss convergence curves of the pose prediction model during training, used to verify the performance of the proposed GAT-GRU model. The left graph shows the change in model loss on the training set with each iteration. The curve rapidly decreases from a high level and then stabilizes, indicating that GAT-GRU can quickly capture the spatiotemporal features of historical poses, demonstrating good fitting speed on the training data. The right graph shows the change in model loss on the validation set. The GAT-GRU curve (blue) also maintains a low and stable value on the validation set without any rebound, proving that the model not only learns the training patterns but also possesses excellent generalization performance, accurately predicting unseen user movement trajectories and effectively avoiding overfitting.

[0190] First, the complexity of the mapping relationship between pose and resource requirements is verified using the ETH-UCY dataset InhwanBae / ETH-UCY-Trajectory-Visualizer. Figure 7 (a) represents the relationship between location x and resource demand, (b) represents the relationship between location y and resource demand, and (c) represents the angle. The relationship with resource requirements. These three figures demonstrate that resource requirements exhibit a distinctly nonlinear manifold distribution in pose space. High-load samples are not randomly distributed but rather highly clustered along specific coordinate trajectories. The heteroscedasticity and periodic abrupt changes exhibited by the local data intuitively reveal the strong correlation between AR rendering load and terminal pose. Figure 8 The regression fitting performance of the GPR surrogate model was evaluated. Experimental results show that the test sample points exhibit significant aggregation characteristics, closely surrounding the ideal reference line, and the vast majority of observed means lie within the 95% confidence interval envelope. This indicates that the model can achieve accurate mapping and does not show significant systematic bias in any sample. Further observation of the outliers marked in red in the figure shows that although the predicted values ​​slightly exceed the statistical interval, they are still located in the diagonal neighborhood and do not show significant trend deviation. This verifies that the model has sensitive response characteristics and good statistical robustness when dealing with sudden high load demands in AR rendering tasks. Figure 9 The analysis shows a comparison between the predicted results and the actual values ​​for 10 random test samples. The red predicted mean closely matches the black actual observed value. All actual values ​​fall precisely within the range given by the model. Within the interval, and mostly within a narrower interval. This dynamic feedback of prediction variance provides a reliability assessment for resource scheduling in AR systems. When the prediction variance is small, the system can adopt an aggressive resource allocation strategy; otherwise, a safety margin is reserved, thus effectively balancing computational efficiency and system stability.

[0191] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for active resource allocation in a MAR system based on pose prediction under a 6G network framework, characterized in that, The method includes the following steps: Step S1: Construct a spatiotemporal pose prediction model based on graph attention network and gated recurrent unit. Use the graph attention network to extract spatial environment structure dependency features from the historical pose information of mobile augmented reality users. Use the gated recurrent unit to extract the temporal dimension dynamic features from the historical pose information. Then, output the predicted pose information of the next time slot through the decoder. Step S2: Construct a task quantity prediction proxy model based on Gaussian process regression, input the predicted pose information into the task quantity prediction proxy model, establish a nonlinear mapping relationship between the predicted pose information and the computational task resource requirements, and output the predicted value of the computational task resource requirements for the next time slot and its uncertainty estimate. Step S3: With minimizing end-to-end latency as the optimization objective, the predicted resource requirements of the computing tasks are used as prior information. A dual-latency deep deterministic strategy gradient-parameterized deep Q-network algorithm is employed to make decisions on the hybrid action space, output the resource allocation action for the current time slot, and perform resource allocation on the computing tasks of the mobile augmented reality system according to the resource allocation action. The hybrid action space includes task offloading decisions, wireless resource block allocation decisions, the computing resource allocation ratio for tracking and registration tasks, and the computing resource allocation ratio for virtual-real fusion tasks.

2. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, The specific method of step S1 includes: The historical pose information of mobile augmented reality (AR) users is represented as follows: in, Represents coordinates, Indicates azimuth; The length of the observation sequence is Then the pose observation sequence can be represented as: The predicted pose is represented as ; Historical pose information is first processed by an MLP layer to extract features, which are represented as follows: in, and It is an MLP weighting and biasing. It is the ReLU activation function; Next, spatial features are processed through the Graph Attention Network (GAT). for in, It is the attention mechanism function. and It is the relevance weight; , and These are the softmax function, the LeakyReLU activation function, and the sigmoid function, respectively. It is the unnormalized attention coefficient, representing the time... Neighbor nodes Features of the central node The degree of importance; The unnormalized attention coefficients of neighboring nodes; Represents the weight vector of the attention mechanism; Normalized attention weights represent the values ​​at time steps. ,node When aggregating local information, it is assigned to neighboring nodes. The actual proportion of attention; Indicates the first Feature transformation weight matrix for each attention head; It will be used as input to the gated recurrent unit (GRU) for the following operations: in, It's an update gate. It's a door reset. It is a characteristic of the GRU output in the previous time slot. It is a hidden feature. It is an activation function; , and These are the input weights from the input space feature sequence to the update gate, reset gate, and candidate hidden state, respectively. , and These are the loop weights from the previous hidden state to the update gate, reset gate, and candidate hidden state, respectively. Finally, the GRU output features The input is fed into the decoder GRU, and the output is the predicted pose information. for: in, For s-linear transformation layer; , and These represent the predicted spatial location (x-coordinate and y-coordinate) and predicted azimuth angle in the next time slot, respectively. Location prediction loss function based on user pose information The angle loss functions are respectively : Where MSE(·) represents the mean square error operator, This represents the set of predicted location coordinate vectors. Let K represent the set of true (ground truth) location coordinate vectors, and let K represent the total number of samples. and These represent the predicted and true x-coordinates of the nth sample in a two-dimensional spatial coordinate system, respectively. and : These represent the predicted and actual ordinates of the nth sample in a two-dimensional coordinate system, respectively; This represents the loss value in azimuth prediction; This represents the predicted azimuth angle of the i-th sample output by the model; This represents the true azimuth angle of the i-th sample; and These represent the sine and cosine functions, respectively, used to transform the difference between the predicted angle and the true angle to the trigonometric function space on the unit circle; This represents the arctangent function in the four quadrants, used to calculate the position of the quadrant from the sine and cosine values. The shortest angular distance within the interval; This means that the square of the calculated minimum angular displacement error is used to construct the optimization objective in the mean square sense; Therefore, the loss function for pose prediction for: in, The weighting coefficients represent the proportion of the contribution of position loss and angle loss to the total loss function.

3. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, The specific methods for step S2 include: Assuming the resource requirements of the virtual fusion task It is the user pose vector The function follows a Gaussian process distribution: x in, It is a mean function; It is the covariance function; Represented by pose vector The resource demand is a stochastic process function with independent variables, i.e., a proxy model for the task quantity to be estimated. Represent a Gaussian process; and These represent two different input pose feature vectors; For the pose input of the training set and corresponding resource demand observations Considering observation noise Predict pose Corresponding task volume prediction value Compared with observed values The joint prior distribution is expressed as: in, It is the covariance matrix of the training set. It is the cross-covariance matrix between the training set and the test points; According to Bayes' theorem, given the observation data , and test input The posterior distribution of the predicted values ​​is also a Gaussian distribution, following the following distribution: Among them, the predicted mean With prediction variance They are respectively: Among them, the predicted mean This will be used as an estimate of the system's resource demand at the next moment, while the prediction variance... This quantifies the uncertainty of the model; The variance representing the independent Gaussian white noise in the observed data; This represents the identity matrix, which has the same dimension as the number of samples.

4. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, The method is applied to the following task architecture: the computational tasks of the mobile augmented reality system are modeled as a directed acyclic graph; the directed acyclic graph includes parallel execution of tracking and registration task branches and pose prediction and virtual-real fusion task branches, and is equipped with virtual entry nodes and virtual exit nodes; the tracking and registration task branches include sub-tasks of vSLAM map update, pose estimation and tracking, and scene reconstruction executed according to dependencies; the pose prediction and virtual-real fusion task branches include sub-tasks of pose prediction, lighting estimation, realistic rendering, and occlusion processing executed according to dependencies.

5. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 4, characterized in that, The specific methods for minimizing end-to-end latency as the optimization objective in step S3 include: The end-to-end latency of the pose prediction and virtual-real fusion task is defined as the total completion time obtained by summing the communication latency and computation latency of each subtask in the branch according to the task dependency relationship. Among them, the communication latency is calculated based on the Shannon formula according to the channel bandwidth, channel gain, transmit power and noise. The local computation latency is obtained by dividing the computation task amount by the proportion of local computation resources allocated. The edge computation latency is obtained by dividing the computation task amount by the proportion of computation resources allocated to the edge server. The earliest start time of the dependent subtask is the maximum value of the sum of the earliest completion time of all direct predecessor tasks and the transmission latency. The earliest completion time is the earliest start time plus the computation latency. The optimization objective is set to minimize the end-to-end latency of the pose prediction and virtual-real fusion tasks, with the completion latency of the tracking registration task being less than the duration of one system time slot as a constraint.

6. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, Step S3 employs a dual-delay deep deterministic policy gradient-parameterized deep Q-network algorithm to make decisions on the hybrid action space and output the resource allocation action for the current time slot. The specific method includes: S3-1: Construct a state space, which consists of observable information from all AR terminals within the current time slot, represented as follows: ; This represents observable information; the observable information includes: channel gain. Historical pose information Track the number of registrations Virtual-real integration task volume ; S3-2: Construct action space A, which includes: unloading decision Resource block allocation decision Track the proportion of registered computing resources and the ratio of virtual and real computing resources ; S3-3: The target Q value is calculated using a dual Q network mechanism, that is, two independent Critic networks are used to calculate the Q value of the action respectively, and the smaller of the two is taken as the target Q value to alleviate the problem of overestimation of action value. The target network parameters of the two Critics are used and If so, then the value of the target network is represented as: in, This represents the target Q value calculated in time slot t; This represents the immediate reward value fed back by the environment after the system performs a resource allocation action in time slot t; This represents the weighting discount factor used to balance current rewards with expected long-term rewards. This indicates the state of the system in the next time slot after performing an action, including information such as predicted pose, channel gain, and task load prediction. This indicates that the target participant network is based on the state. The output predicts the action for the next time slot; and The parameters are respectively and Two independent networks of target critics; This represents the minimization operator, which selects the smaller value from the outputs of the two target Critic networks; Indicates the mixed action space Find the action that maximizes the Q value. S3-4: A delayed update mechanism is adopted, setting the update frequency of the Actor network to be lower than that of the Critic network, and updating the Actor network once after the Critic network updates a preset number of times. S3-5: A target action smoothing mechanism is adopted, in which truncated Gaussian noise is added to the output of the target policy network to generate smooth target actions, wherein the amplitude of the truncated Gaussian noise is controlled by preset parameters. S3-6: Based on the state space, action space and the above mechanism, with the goal of minimizing the end-to-end latency of pose prediction and virtual-real fusion tasks, output the resource allocation action for the current time slot.

7. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, In step S3, the Actor network and Critic network in the dual-delay deep deterministic strategy gradient-parameterized deep Q-network algorithm both adopt a three-layer fully connected structure, with 512, 256 and 128 hidden layer neurons respectively, and use the ReLU activation function; the discount factor is 0.

98.

8. The active resource allocation method for a MAR system based on pose prediction under a 6G network framework according to claim 1, characterized in that, Before step 2, the method further includes: using mutual information to analyze the nonlinear relationship between the pose vector and the resource demand, using Moran's I index to determine the clustering effect of the resource demand in the pose space, and when a clustering effect exists, using the task quantity prediction proxy model based on Gaussian process regression.

Images