An edge cloud cooperative digital human lightweight deployment and scheduling system

Abstract

The application discloses an edge cloud cooperative digital human lightweight deployment and scheduling system, relates to the technical field of edge computing cooperative processing, and comprises a pipeline cutting module, a semantic analysis module, an elastic compression module, a resource sensing module, a task scheduling module and a spectrum feedback module; the pipeline cutting module is used for constructing an operator-level multi-modal pipeline directed acyclic graph according to a digital human interaction request, collecting an operator-level benchmark time consumption matrix, solving an optimal cutting point and outputting an asymmetric cutting scheme; by converting cloud edge scheduling into a graph balance segmentation problem and adopting linear scanning discrete optimization search, pipeline bubbles caused by the traditional symmetric cutting ignoring transmission delay are eliminated; a sliding window variance and dynamic stability attenuation score are introduced to distinguish transient performance and sustainable performance to avoid the cumulative jitter risk of long-time tasks; a bidirectional matching model is constructed to realize supply and demand coupling evaluation, so that the system selects a node which is stable and just matches.

Description

Technical Field

[0001] This invention relates to the field of edge computing collaborative processing technology, and in particular to a lightweight deployment and scheduling system for digital humans in edge cloud collaboration. Background Technology

[0002] Edge computing collaborative processing technology is a computing paradigm that organically combines the centralized and powerful computing power of cloud computing with the distributed local processing capabilities of the edge. By deploying computing, storage, and network resources close to the data source or user's network edge, it enables dynamic task offloading, hierarchical data processing, and elastic resource scheduling between the cloud, edge, and device. It retains the advantages of the cloud in model training, global decision-making, and massive data storage, while leveraging the strengths of edge nodes in low-latency response, bandwidth saving, and privacy protection. This effectively supports new application scenarios that are latency-sensitive and data-intensive, such as real-time rendering of digital humans, autonomous driving, and the industrial internet.

[0003] Current 3D model rendering methods mostly employ global uniform compression or fixed LOD (Level of Detail) strategies, failing to map user interaction intent to the importance distribution in the model space. This results in high semantic relevance areas (such as facial expressions and gestures) not receiving priority rendering, while valuable edge computing power and bandwidth resources are wasted in low relevance areas. Furthermore, existing scheduling algorithms select edge nodes based solely on static peak computing power or average load rate, ignoring the temporal fluctuations in edge node resource supply. For long-duration, high real-time tasks like digital human rendering, achieving instantaneous performance targets does not guarantee sustainable service capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a lightweight deployment and scheduling system for digital humans with edge cloud collaboration. This solves the problem that existing scheduling algorithms only match edge nodes based on static peak computing power or average load rate, ignoring the fluctuation characteristics of edge node resource supply in the time domain. For long-duration, high real-time tasks such as digital human rendering, instantaneous performance does not necessarily mean that sustainable service capabilities are met.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a lightweight deployment and scheduling system for digital humans in edge cloud collaboration, comprising:

[0008] Pipeline segmentation module, semantic parsing module, elastic compression module, resource awareness module, task scheduling module, and spectrum feedback module;

[0009] The pipeline segmentation module is used to construct an operator-level multimodal pipeline directed acyclic graph based on digital human interaction requests, collect operator-level baseline time consumption matrix, solve for the optimal segmentation point, and output an asymmetric segmentation scheme.

[0010] The semantic parsing module is used to identify the intent of digital human interaction requests and generate semantic label vectors, and then transform the semantic label vectors into an initial importance weight map of the three-dimensional model region.

[0011] The elastic compression module is used to obtain the available resource vector of the target edge node in the asymmetric segmentation scheme, solve the optimal compression configuration of the region based on the initial importance weight map of the 3D model region and the available resource vector as the constraint, and assemble a lightweight model instance according to the optimal compression configuration of the region.

[0012] The resource awareness module is used to collect multi-dimensional resource time-series sequences of candidate edge nodes, calculate the sliding window variance, and combine the task load characteristics of the asymmetric partitioning scheme to calculate the comprehensive stability computing power score.

[0013] The task scheduling module is used to perform bidirectional matching of the resource requirements and stability computing power comprehensive scores of the asymmetric partitioning scheme and lightweight model instance to determine the optimal node combination, and control the optimal node combination to execute the digital human rendering task according to the asymmetric partitioning scheme.

[0014] The spectrum feedback module is used to extract the energy proportion of the rendering output frame interval sequence in the human eye sensitive frequency band when the optimal node combination is used to perform the digital human rendering task. The energy proportion is used to correct the initial importance weight map of the 3D model region to obtain the corrected importance weight map. The corrected importance weight map is fed back to the elastic compression module to resolve the optimal compression configuration of the region.

[0015] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for solving the optimal splitting point and outputting the asymmetric splitting scheme are as follows:

[0016] Traverse each connecting edge in the directed acyclic graph of the operator-level multimodal pipeline as a candidate splitting position;

[0017] For each candidate segmentation position, the baseline time of all operators before the candidate segmentation position is accumulated to obtain the expected time on the cloud side, and the baseline time of all operators after the candidate segmentation position is accumulated to obtain the expected time on the edge side.

[0018] Calculate the absolute difference between the expected time on the cloud side and the expected time on the edge side;

[0019] Compare the absolute differences corresponding to all candidate splitting positions, and select the candidate splitting position with the smallest absolute difference as the optimal splitting point;

[0020] Based on the optimal split point, the directed acyclic graph of the operator-level multimodal pipeline is divided into cloud-side sub-pipelines and edge-side sub-pipelines;

[0021] Extract the operator names, inter-operator data transmission protocols, and expected synchronization timestamps from the cloud-side sub-pipeline and the edge-side sub-pipeline, and combine them into an asymmetric partitioning scheme.

[0022] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for converting semantic tag vectors into an initial importance weight map of a three-dimensional model region are as follows:

[0023] Load the pre-trained semantic rendering region mapping knowledge base, input the semantic label vector into the pre-trained semantic rendering region mapping knowledge base for matching and retrieval, and output a list of 3D model mesh partition identifiers associated with the semantic label vector;

[0024] Iterate through all mesh partition identifiers of the 3D model and determine whether the current mesh partition identifier exists in the 3D model mesh partition identifier list;

[0025] If the current mesh partition identifier exists in the 3D model mesh partition identifier list, then assign the current mesh partition identifier a preset high weight value;

[0026] If the current mesh partition identifier does not exist in the 3D model mesh partition identifier list, then assign the current mesh partition identifier a preset low weight value;

[0027] Arrange the weight values ​​corresponding to all grid partition identifiers according to the topology of the 3D model, and generate an initial importance weight map of the 3D model regions.

[0028] As a preferred embodiment of the edge cloud collaborative digital human lightweight deployment and scheduling system described in this invention, the specific steps for solving the optimal compression configuration of the region using the initial importance weight map of the three-dimensional model region as the weighting basis and the available resource vector as the constraint condition are as follows:

[0029] Read the maximum video memory capacity, maximum computing throughput, and maximum bandwidth from the available resource vector;

[0030] For each grid partition in the initial importance weight map of the 3D model region, enumerate all candidate compression levels;

[0031] For each candidate compression level, estimate the corresponding memory usage, computing power consumption, and data transmission volume;

[0032] Remove candidate compression levels that exceed the video memory capacity limit, the computing power consumption limit, or the data transmission limit, and retain the set of feasible compression levels.

[0033] In the set of feasible compression levels, the candidate compression level that maximizes the perceived utility score is selected as the optimal compression configuration for the region based on the nonlinear perception utility formula.

[0034] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for calculating the stability computing power comprehensive score based on the task load characteristics combined with the asymmetric partitioning scheme are as follows:

[0035] Extract the expected execution time of the edge-side sub-pipeline from the asymmetric partitioning scheme as a task load feature;

[0036] Collect multidimensional resource time series of candidate edge nodes and calculate the sliding window variance;

[0037] Obtain the baseline computing power score of candidate edge nodes, and calculate the comprehensive stability computing power score based on the dynamic stability decay formula.

[0038] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for determining the optimal node combination by bidirectional matching of the resource requirements and stability computing power comprehensive scores of the asymmetric partitioning scheme and the lightweight model instance are as follows:

[0039] The total video memory requirement, total computing power requirement, and total bandwidth requirement are obtained by summarizing the optimal compression configuration of the region for the lightweight model instance.

[0040] A node adaptability evaluation model is constructed, with the stability computing power comprehensive score as the numerator and the Euclidean distance between the resource demand and available resource vectors as the denominator.

[0041] Traverse all candidate edge nodes and substitute them into the node fit evaluation model to calculate the fit score;

[0042] Candidate edge nodes with a fit score higher than a preset threshold and whose available resource vectors simultaneously meet the total video memory requirement, total computing power requirement, and total bandwidth requirement are selected to form a set of candidate nodes.

[0043] Select the node with the highest fit score from the candidate node set as the edge-side execution node;

[0044] In the asymmetric partitioning scheme, find the cloud node that matches the cloud-side subpipeline and use it as the cloud-side execution node;

[0045] The optimal node combination is obtained by combining edge-side execution nodes and cloud-side execution nodes.

[0046] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for extracting the energy proportion of the rendering output frame interval sequence in the human eye-sensitive frequency band when executing the digital human rendering task using the optimal node combination are as follows:

[0047] During the execution of the digital human rendering task with the optimal node combination, the timestamps of consecutive frames are recorded at a fixed sampling period.

[0048] Calculate the timestamp difference between adjacent frames to obtain the rendered output frame interval sequence;

[0049] Perform a short-time Fourier transform on the rendered output frame interval sequence to obtain the spectral energy distribution;

[0050] Define a mask function for the frequency band sensitive to human eyes. The mask function for the frequency band sensitive to human eyes takes a value of 1 in the range of 8 Hz to 15 Hz and a value of 0 in the other frequency bands.

[0051] The energy of the sensitive frequency band is obtained by multiplying the spectral energy distribution point by point with the mask function of the human eye sensitive frequency band and then integrating.

[0052] The total fluctuation energy is obtained by integrating the spectral energy distribution across the entire frequency band, and the energy percentage is calculated based on the frequency domain resonance risk formula.

[0053] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps for obtaining the corrected importance weight map by correcting the initial importance weight map of the three-dimensional model region using energy proportion are as follows:

[0054] Determine whether the energy percentage exceeds the preset resonance risk threshold;

[0055] Among them, when the energy ratio exceeds the preset resonance risk threshold, the grid partition with the largest historical jitter amplitude in the current optimal node combination is identified as the high-risk partition.

[0056] The corrected weight value is obtained by multiplying the weight value of the high-risk partition in the initial importance weight map of the 3D model region by a preset attenuation coefficient.

[0057] Keep the weight values ​​of all grid partitions except the high-risk partitions unchanged in the initial importance weight map of the 3D model region, and rearrange the modified weight values ​​of the high-risk partitions and the unmodified weight values ​​of the other grid partitions according to the topology of the 3D model to generate the modified importance weight map.

[0058] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, the specific steps of feeding back the corrected importance weight graph to the elastic compression module to resolve the optimal compression configuration for the region are as follows:

[0059] The elastic compression module receives the corrected importance weight map and replaces the initial importance weight map of the 3D model region;

[0060] Keep the available resource vector unchanged, and re-enumerate all candidate compression levels based on the revised importance weight graph;

[0061] Re-execute the process of screening the feasible compression level set and calculating the nonlinear perceived utility formula;

[0062] A convergence criterion is introduced: the iteration terminates when the difference between the updated optimal regional compression configuration and the previous optimal regional compression configuration is lower than a preset convergence threshold.

[0063] If the difference is not lower than the preset convergence threshold, the spectrum feedback module will be triggered again to collect a new round of rendering output frame interval sequence and repeat the correction process.

[0064] When the convergence criterion is met, the optimal compression configuration of the final region is output and the target edge nodes are triggered to assemble lightweight model instances according to the optimal compression configuration of the final region.

[0065] As a preferred embodiment of the edge cloud collaborative lightweight deployment and scheduling system for digital humans described in this invention, it further includes:

[0066] When the energy percentage does not exceed the preset resonance risk threshold, skip the step of correcting the initial importance weight map of the three-dimensional model region and maintain the current optimal compression configuration of the region unchanged;

[0067] After the digital human rendering task is completed, the energy percentage, the corrected importance weight map, and the updated optimal compression configuration of the region recorded in this task will be stored in the historical execution log.

[0068] When subsequent digital human interaction requests with the same semantic label vector arrive, the historically corrected importance weight map matching the current candidate edge node is retrieved from the historical execution log first as the initial value of the initial importance weight map of the 3D model region, and the first resonance risk response time is shortened.

[0069] The beneficial effects of this invention are as follows: by transforming cloud-edge scheduling into a graph-based balanced partitioning problem and employing linear scanning discrete optimization search, pipeline bubbles caused by neglecting transmission delay in traditional symmetrical partitioning are eliminated; by introducing sliding window variance and dynamic stability decay scores, instantaneous performance and sustainable performance are distinguished to avoid the risk of cumulative jitter in long-term tasks; a bidirectional matching model is constructed to achieve supply and demand coupling evaluation, prompting the system to select nodes that are both stable and perfectly matched; subjective visual experience is objectified into frequency domain signals and a closed-loop correction mechanism is formed, enabling the compression strategy to actively suppress the resonance risk of human eye-sensitive frequency bands, thereby achieving a synergistic improvement in digital human rendering quality, system stability, and user experience consistency under limited heterogeneous resources. Attached Figure Description

[0070] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] Figure 1 This is a schematic diagram of the edge cloud collaborative digital human lightweight deployment and scheduling system in the embodiment. Detailed Implementation

[0072] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0073] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0074] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0075] Reference Figure 1 As one embodiment of the present invention, this embodiment provides a lightweight deployment and scheduling system for edge cloud collaborative digital humans, comprising:

[0076] The module includes a pipeline segmentation module, a semantic parsing module, an elastic compression module, a resource awareness module, a task scheduling module, and a spectrum feedback module.

[0077] The pipeline segmentation module is used to construct an operator-level multimodal directed acyclic graph based on digital human interaction requests, collect operator-level baseline time-consuming matrices, solve for the optimal segmentation point, and output an asymmetric segmentation scheme.

[0078] Furthermore, each connecting edge in the directed acyclic graph of the operator-level multimodal pipeline is traversed as a candidate splitting position. For each candidate splitting position, the baseline time of all operators before the candidate splitting position is accumulated to obtain the expected time on the cloud side, and the baseline time of all operators after the candidate splitting position is accumulated to obtain the expected time on the edge side.

[0079] Calculate the absolute difference between the expected time on the cloud side and the expected time on the edge side; compare the absolute differences corresponding to all candidate split positions, and select the candidate split position with the smallest absolute difference as the optimal split point; divide the directed acyclic graph of the operator-level multimodal pipeline into cloud-side sub-pipelines and edge-side sub-pipelines based on the optimal split point; extract the operator names, inter-operator data transmission protocols, and expected synchronization timestamps contained in the cloud-side sub-pipelines and edge-side sub-pipelines, and combine them into an asymmetric splitting scheme.

[0080] It should be noted that by traversing the connected edges of the pipeline's directed acyclic graph and calculating the absolute difference between the expected time of the cloud and edge sides based on the baseline time, the optimal split point can be selected. This enables precise and balanced distribution of operator-level task load on both ends of the cloud and edge, effectively avoiding rendering latency bottlenecks caused by overload on one side. At the same time, extracting operator names, transmission protocols, and synchronization timestamps to form an asymmetric splitting scheme provides a fine-grained execution basis for subsequent collaborative scheduling of heterogeneous resources, improving the overall throughput efficiency and real-time response of the multimodal pipeline.

[0081] In particular, this partitioning method transforms the scheduling problem of cloud-edge collaboration into a balanced partitioning problem on a graph. It uses an operator-level baseline time matrix to eliminate the differences in execution efficiency of heterogeneous hardware, allowing partitioning decisions to be based entirely on actual processing power rather than theoretical computing power. The asymmetric design itself acknowledges the natural asymmetry between the cloud and the edge in terms of computing density and communication bandwidth, allowing both sides to undertake subtasks of different sizes but aligned in time. This avoids pipeline bubbles caused by traditional symmetric partitioning that ignore transmission latency and synchronization overhead from a physical perspective.

[0082] The process of traversing the connecting edges and calculating the absolute difference in time consumption employs a discrete optimization search strategy based on linear scanning. The motivation for this strategy lies in the fact that the digital human rendering pipeline has strict serial dependency characteristics and cannot be arbitrarily broken down and reassembled like parallel computing. Therefore, it is necessary to find a time balance point while maintaining the topological order. The expression for calculating the absolute difference between the expected time consumption on the cloud side and the expected time consumption on the edge side is as follows:

[0083] ;

[0084] In the formula, This represents the absolute difference in load at the candidate split positions. This represents the cumulative baseline time of all operators preceding the candidate segmentation position. This represents the cumulative baseline time for all operators after the candidate segmentation position.

[0085] The operation of accumulating the baseline time involves constructing a prefix sum array, transforming the subgraph time query, which originally required repeated calculations, into a constant-time array access, thus reducing the computational complexity of the partitioning decision. Selecting the minimum absolute difference as the optimal criterion utilizes the scheduling idea of ​​minimizing the maximum completion time, ensuring that the sub-pipelines on both the cloud and edge sides can complete their respective stages of tasks as simultaneously as possible, thereby reducing the synchronization blocking time caused by one side waiting for the other. Extracting operator names and combining them with the transmission protocol to form an asymmetric partitioning scheme is to establish a physical layer communication contract after logical partitioning, clearly defining the format and timing requirements for cross-node data exchange.

[0086] The semantic parsing module is used to identify the intent of digital human interaction requests and generate semantic label vectors, which are then transformed into an initial importance weight map of the three-dimensional model region.

[0087] Furthermore, a pre-trained semantic rendering region mapping knowledge base is loaded, and semantic label vectors are input into the pre-trained semantic rendering region mapping knowledge base for matching and retrieval, outputting a list of 3D model mesh partition identifiers associated with the semantic label vectors.

[0088] Traverse all mesh partition identifiers of the 3D model and determine whether the current mesh partition identifier exists in the 3D model mesh partition identifier list. If the current mesh partition identifier exists in the 3D model mesh partition identifier list, assign the current mesh partition identifier a preset high weight value. If the current mesh partition identifier does not exist in the 3D model mesh partition identifier list, assign the current mesh partition identifier a preset low weight value. Arrange the weight values ​​corresponding to all mesh partition identifiers according to the topology of the 3D model to generate an initial importance weight map of the 3D model region.

[0089] It should be noted that by loading a pre-trained semantic rendering region mapping knowledge base to transform semantic label vectors into differentiated weight maps of 3D model mesh partitions, user interaction intent can be directly mapped to the importance distribution in the model space. This ensures that highly semantically relevant regions receive priority rendering, while low-relevance regions are automatically downgraded. This significantly improves visual quality at the human eye perception level without increasing overall resource consumption, thus realizing a semantically driven intelligent resource allocation mechanism.

[0090] Specifically, the core of weight generation is to establish cross-modal alignment between semantic and geometric spaces. The pre-trained knowledge base is equivalent to a mapper that translates abstract intentions into concrete grids. It changes the allocation of rendering resources from global uniformity to on-demand focus. By utilizing the differences in human eye sensitivity to different areas of the image, it simulates the selective focus of human attention by adjusting the weight vector without increasing the total rendering budget. Semantic-driven preprocessing provides prior guidance for subsequent elastic compression, avoiding the loss of key semantic information by blindly compressing.

[0091] The step of loading the pre-trained semantic rendering region mapping knowledge base and performing matching retrieval employs a nearest neighbor search algorithm based on vector similarity. This is motivated by the fact that there is no natural mathematical correspondence between the interactive intent described in natural language and the geometric topology of the 3D model; therefore, an intermediate representation layer trained with a large amount of labeled data is necessary to complete the semantic-to-spatial projection. The expression for calculating the matching degree between the semantic label vector and the grid partition identifier in the knowledge base is as follows:

[0092] ;

[0093] In the formula, The cosine similarity between the semantic label vector and the grid partition identifier vector is represented. Represents a semantic label vector. This represents the grid partition identifier vector stored in the knowledge base. This represents the Euclidean norm of a vector.

[0094] The operation of traversing the grid partition identifiers and assigning high and low weight values ​​is the binarization mask generation process. It quantifies continuous semantic relevance into discrete rendering priority labels. Discretization is performed to adapt to the enumeration-based configuration search mechanism in the subsequent compression module and avoid the high computational cost of gradient optimization in continuous space.

[0095] The elastic compression module is used to obtain the available resource vector of the target edge node in the asymmetric segmentation scheme, solve the optimal compression configuration of the region based on the initial importance weight map of the 3D model region and the available resource vector as the constraint, and assemble a lightweight model instance according to the optimal compression configuration of the region.

[0096] Furthermore, the upper limits of video memory capacity, computing power throughput, and bandwidth are read from the available resource vector; for each grid partition in the initial importance weight map of the 3D model region, all candidate compression levels are enumerated; for each candidate compression level, the corresponding video memory usage, computing power consumption, and data transmission volume are estimated; candidate compression levels whose video memory usage exceeds the upper limit of video memory capacity, whose computing power consumption exceeds the upper limit of computing power throughput, or whose data transmission volume exceeds the upper limit of bandwidth are removed, and the set of feasible compression levels is retained; in the set of feasible compression levels, the candidate compression level that maximizes the perceptual utility score is selected as the optimal compression configuration for the region according to the nonlinear perceptual utility formula.

[0097] The formula for nonlinear perception utility is expressed as follows:

[0098] ;

[0099] In the formula, This represents the perceived utility score. The first element in the initial importance weight diagram of the 3D model region The weight values ​​of each grid partition. Represents the semantic sensitivity index. Indicates the first Compression quality metrics for each grid partition at the selected candidate compression level. This indicates a resource stress adjustment factor. Indicates the first Resource consumption of each grid partition at the selected candidate compression level. This represents the upper limit of the corresponding resource in the available resource vector. This indicates the total number of grid partitions.

[0100] It should be noted that by enumerating candidate compression levels under available resource constraints and selecting the optimal configuration based on the nonlinear perceptual utility formula, compression quality and resource overhead can be automatically balanced under multiple constraints of video memory, computing power and bandwidth. This allows high-weight regions to retain higher fidelity while low-weight regions are moderately compressed, avoiding distortion or resource waste in key areas caused by traditional uniform compression. This ensures that lightweight model instances meet hardware boundaries and maximize visual experience at edge nodes.

[0101] In particular, the compressed configuration solution transforms multi-objective optimization into a constrained single-objective utility maximization problem. The nonlinear perceptual utility function reflects the law of diminishing marginal returns, where the benefits of further improving quality in high-weight regions gradually decrease, and the losses from moderately reducing quality in low-weight regions are also controllable. The resource stress adjustment factor enables the system to have environmental adaptability, pursuing higher visual quality when resources are abundant and automatically switching to a conservative strategy when resources are scarce. Dynamic trade-offs make the compression decision no longer a static table lookup, but a continuous optimization that responds to the hardware status in real time, ensuring that the lightweight instance is always on the optimal boundary under the current resource conditions.

[0102] The process of reading the upper limit from the available resource vector and enumerating candidate compression levels employs an exhaustive search strategy with pruning. This is motivated by the fact that the compression levels of digital human models are typically discrete preset levels, and resource consumption between different grid partitions involves complex nonlinear coupling, making traditional convex optimization methods difficult to apply directly. However, by filtering through the triple constraints of memory, computing power, and bandwidth, a large number of infeasible solutions can be quickly eliminated, compressing the search space to a scale that can be computed in real time. The resource feasibility determination expression for candidate compression levels is as follows:

[0103] ;

[0104] In the formula, This represents a Boolean decision indicating whether a candidate compression level is feasible. , , These represent the video memory usage, computing power consumption, and data transfer volume at that compression level, respectively. , , These represent the upper limits of available memory capacity, computing throughput, and bandwidth in the available resource vector, respectively. Represents the logical AND operation.

[0105] The optimal configuration is selected based on the nonlinear perception utility formula. This involves introducing a logarithmic visual quality evaluation function and an exponential resource penalty term. The choice of the logarithmic function conforms to the Weber-Fechner law's description of the characteristics of human eye perception, namely, that marginal returns decrease after the quality is improved to a certain extent. The resource penalty term acts as a Lagrange multiplier, automatically guiding the search direction toward the interior of the resource feasible region by adjusting the penalty intensity without explicitly solving the constraint optimization problem.

[0106] The resource awareness module is used to collect multi-dimensional resource time-series sequences of candidate edge nodes, calculate the sliding window variance, and combine the task load characteristics of the asymmetric partitioning scheme to calculate a comprehensive stability computing power score.

[0107] Furthermore, the expected execution time of the edge sub-pipeline is extracted from the asymmetric partitioning scheme as a task load feature; multi-dimensional resource time series of candidate edge nodes are collected and the sliding window variance is calculated; the benchmark computing power score of candidate edge nodes is obtained, and the comprehensive stability computing power score is calculated based on the dynamic stability decay formula.

[0108] The dynamic stability decay formula is as follows:

[0109] ;

[0110] In the formula, This represents the overall score for stability and computing power. This represents the benchmark computing power score. Indicates the environmental sensitivity coefficient. Represents the variance of the sliding window. This indicates the expected execution time in the task load characteristics. This represents the nonlinear amplification index for the load duration.

[0111] It should be noted that by collecting the sliding window variance of the multidimensional resource time series of candidate nodes and combining it with the expected execution time of the task to calculate the dynamic stability decay score, the sustainable service capability of edge nodes under time-varying load environment can be truly reflected, rather than relying solely on static peak indicators. This effectively avoids the risk of task interruption or frame rate drop caused by resource fluctuations, and provides an environmentally adaptable computing power evaluation basis for high real-time digital human rendering.

[0112] In particular, the key to stability scoring lies in distinguishing between instantaneous performance and sustained performance. The sliding window variance captures the fluctuation characteristics of resource supply in the time domain, rather than the capacity characteristics in the spatial domain. The nonlinear amplification index of load duration reflects the cumulative sensitivity of long tasks to resource jitter. The longer the execution time, the greater the destructive impact of mid-run resource fluctuations. The design breaks through the limitations of traditional scheduling that only considers peak computing power or average load, and explicitly incorporates the uncertainty of the time dimension into the evaluation. This allows the scoring results to predict the probability of a node maintaining stable output in the future, providing a risk screening basis for high real-time rendering tasks.

[0113] The operation of collecting multidimensional resource time-series data and calculating the sliding window variance employs time-series statistical analysis. The resource status of edge nodes exhibits high non-stationarity and abruptness; instantaneous sampling at a single point is highly susceptible to noise interference, leading to misjudgments. The sliding window, through smoothing recent historical data, effectively filters out high-frequency random fluctuations and extracts low-frequency components reflecting the true load trend of nodes. The expression for calculating the sliding window variance of the multidimensional resource time-series data is as follows:

[0114] ;

[0115] In the formula, Represents the variance of the sliding window. Indicates the length of the sliding window. Indicates the first Resource observations at each sampling time. This represents the arithmetic mean of the resource observations within the current sliding window. This indicates the index of the starting sampling time of the current window.

[0116] The dynamic stability decay score, calculated by combining the expected execution time of the task, introduces a time-scale weighting mechanism. It convolves the statistical volatility of resources with the duration of the task. This is because for short-duration tasks, brief performance fluctuations may only cause slight delays, but for long-duration rendering tasks, the same fluctuations may lead to severe frame rate collapse or task timeouts over time. Obtaining a baseline computing power score as the base for decay calculation is to preserve the absolute magnitude of the node's hardware capabilities, preventing the neglect of the node's lower limit of processing power in the process of focusing too much on stability. This ensures that the final score reflects both stability and strength.

[0117] The task scheduling module is used to perform bidirectional matching of the resource requirements and stability computing power comprehensive scores of the asymmetric partitioning scheme and lightweight model instance to determine the optimal node combination, and control the optimal node combination to execute the digital human rendering task according to the asymmetric partitioning scheme.

[0118] Furthermore, based on the regional optimal compression configuration of the lightweight model instance, the total video memory requirement, total computing power requirement, and total bandwidth requirement are obtained; a node adaptability evaluation model is constructed, with the stability computing power comprehensive score as the numerator and the Euclidean distance between the resource requirement and the available resource vector as the denominator.

[0119] Traverse all candidate edge nodes and substitute them into the node fit evaluation model to calculate the fit score; filter candidate edge nodes whose fit scores are higher than the preset threshold and whose available resource vectors simultaneously meet the total memory requirements, total computing power requirements, and total bandwidth requirements, forming a candidate node set; select the node with the highest fit score from the candidate node set as the edge-side execution node; find cloud nodes that match the cloud-side subpipeline in the asymmetric partitioning scheme as cloud-side execution nodes; combine the edge-side execution nodes and the cloud-side execution nodes to obtain the optimal node combination.

[0120] It should be noted that by constructing a node suitability evaluation model with a comprehensive stability computing power score as the numerator and the Euclidean distance between resource requirements and available resources as the denominator for bidirectional matching, the dynamic stability and resource compatibility of nodes can be taken into account simultaneously. This avoids scheduling deviations caused by simply pursuing high computing power or low load, ensuring that the selected edge nodes can stably carry the current task while leaving reasonable margins to cope with sudden disturbances, thereby improving the robustness and task success rate of cloud-edge collaborative scheduling.

[0121] In particular, the bidirectional matching model transforms scheduling decisions from unidirectional adaptation to supply-demand coupled evaluation. It uses stability scores as the numerator to reflect the priority consideration of service quality, and resource Euclidean distance as the denominator to quantify the geometric proximity of task demand and node supply in the multidimensional resource space. The ratio structure has inherent normalization characteristics, avoiding evaluation distortion caused by differences in single-dimensional dimensions. At the same time, it penalizes the two extreme cases of resource surplus and resource shortage, prompting the system to select nodes that are both stable and perfectly matched.

[0122] The step of summarizing the optimal compression configuration of the region to obtain the total resource requirement uses an aggregation and accumulation algorithm. The motivation for this is that the elastic compression module outputs a fine-grained partition-level configuration, while scheduling decisions require a node-level overall load view. By vector-summing the memory, computing power, and bandwidth requirements of each partition, a task requirement vector is constructed that is isomorphic to the node's available resource vector, providing standardized input for subsequent geometric distance calculations. The Euclidean distance calculation expression between the resource requirement and the available resource vector is as follows: ;

[0123] In the formula, This represents the Euclidean distance between the vectors of resource demand and available resources. , , These represent the total video memory requirement, total computing power requirement, and total bandwidth requirement of the lightweight model instance, respectively. , , These represent the available memory capacity, computing throughput, and bandwidth margin in the available resource vector of the candidate edge node, respectively.

[0124] Constructing a node fitness evaluation model and calculating fitness scores involves performing a weighted nearest neighbor search in a multidimensional resource space. The structure design, with stability score as the numerator and Euclidean distance as the denominator, ensures that nodes with higher scores are selected first when distances are the same, and nodes with higher resource matching degree win when stability is similar. The ratio form avoids the problem of difficulty in calibrating weight parameters in multi-objective weighted summation.

[0125] The spectrum feedback module is used to extract the energy proportion of the rendering output frame interval sequence in the human eye sensitive frequency band when the optimal node combination is used to perform the digital human rendering task. The energy proportion is used to correct the initial importance weight map of the 3D model region to obtain the corrected importance weight map. The corrected importance weight map is fed back to the elastic compression module to resolve the optimal compression configuration of the region.

[0126] Furthermore, during the digital human rendering task performed with the optimal node combination, the timestamps of consecutive frames are recorded at a fixed sampling period; the difference between the timestamps of adjacent frames is calculated to obtain the rendering output frame interval sequence; and a short-time Fourier transform is performed on the rendering output frame interval sequence to obtain the spectral energy distribution.

[0127] Define a mask function for the human eye sensitive frequency band. The mask function for the human eye sensitive frequency band takes a value of 1 in the range of 8 Hz to 15 Hz and a value of 0 in the other frequency bands. Multiply the spectral energy distribution with the mask function for the human eye sensitive frequency band point by point and then integrate to obtain the energy of the sensitive frequency band. Integrate the spectral energy distribution across the entire frequency band to obtain the total fluctuation energy, and calculate the energy ratio according to the frequency domain resonance risk formula.

[0128] The formula for the risk of frequency domain resonance is as follows:

[0129] ;

[0130] In the formula, Indicates the percentage of energy. Represents the spectrum energy distribution function. This represents a mask function for frequencies sensitive to the human eye. This indicates the lower limit of the frequency band that the human eye is sensitive to. This indicates the upper limit of the frequency band that the human eye is sensitive to. This indicates the sampling frequency corresponding to a fixed sampling period.

[0131] Furthermore, it is determined whether the energy proportion exceeds the preset resonance risk threshold; when the energy proportion exceeds the preset resonance risk threshold, the grid partition with the largest historical jitter amplitude in the current optimal node combination is identified as the high-risk partition; the weight value of the high-risk partition in the initial importance weight map of the 3D model region is multiplied by the preset attenuation coefficient to obtain the corrected weight value.

[0132] Keep the weight values ​​of all grid partitions except the high-risk partitions unchanged in the initial importance weight map of the 3D model region, and rearrange the modified weight values ​​of the high-risk partitions and the unmodified weight values ​​of the other grid partitions according to the topology of the 3D model to generate the modified importance weight map.

[0133] The elastic compression module receives the corrected importance weight map and replaces the initial importance weight map of the 3D model region; it keeps the available resource vector unchanged and re-enumerates all candidate compression levels based on the corrected importance weight map as the new weighting basis.

[0134] The process of selecting feasible compression levels and calculating nonlinear perceptual utility formulas is re-executed; a convergence criterion is introduced, and the iteration is terminated when the difference between the updated optimal regional compression configuration and the previous optimal regional compression configuration is lower than the preset convergence threshold; if the difference is not lower than the preset convergence threshold, the spectrum feedback module is triggered again to collect a new round of rendering output frame interval sequence and repeat the correction process; when the convergence criterion is met, the final optimal regional compression configuration is output and the target edge nodes are triggered to assemble lightweight model instances according to the final optimal regional compression configuration.

[0135] When the energy percentage does not exceed the preset resonance risk threshold, skip the step of correcting the initial importance weight map of the 3D model region and maintain the current optimal compression configuration of the region unchanged. After the digital human rendering task is completed, store the energy percentage, corrected importance weight map and updated optimal compression configuration of the region recorded in this task into the historical execution log. When subsequent digital human interaction requests with the same semantic label vector arrive, prioritize retrieving the historical corrected importance weight map that matches the current candidate edge node from the historical execution log as the initial value of the initial importance weight map of the 3D model region, and shorten the first resonance risk response time.

[0136] It should be noted that by performing a short-time Fourier transform on the rendered output frame interval sequence and combining it with the mask function of the human eye sensitive frequency band to calculate the energy ratio, the resonance risk of frame rate jitter in the frequency band most easily perceived by the human eye can be accurately quantified. This triggers iterative optimization of the dynamic decay and compression configuration of the high-risk grid partition weights, forming a closed-loop adjustment mechanism from visual physiological feedback to model compression strategy, suppressing subjectively perceptible stuttering and flickering phenomena. At the same time, the historical execution log is used to accelerate the convergence process of subsequent requests with the same semantics, continuously improving the rendering smoothness and user experience consistency under long-term operation.

[0137] In particular, the spectrum feedback closed loop objectifies the subjective visual experience into a computable frequency domain signal. The mask function for frequency bands sensitive to the human eye prevents the system from treating all frame rate jitter equally, focusing only on the frequency components that are most likely to cause discomfort at the physiological level. High-risk partition identification and weight attenuation constitute a negative feedback adjustment path. When resonance risk is detected, the rendering priority of the relevant area is actively reduced to obtain temporal stability. The convergence criterion ensures that the iteration will not fall into infinite oscillation. The reuse of historical logs also transforms online learning into offline knowledge accumulation.

[0138] The short-time Fourier transform (SFT) of the rendered output frame interval sequence is used to obtain the spectral energy distribution. A time-frequency analysis algorithm is employed. Frame rate jitter is not a stationary signal; its frequency components change dynamically over time. Simple time-domain statistics cannot reveal the periodicity of the jitter. The SFT, through windowing and truncation, decomposes the non-stationary signal into a series of locally stationary segments, enabling it to capture transient resonance events within the frequency bands sensitive to the human eye. The SFT calculation expression for the rendered output frame interval sequence is as follows:

[0139] ;

[0140] In the formula, Represents frequency and time Spectral energy distribution at that location This represents the sequence of frame intervals for the rendered output. Indicates that the center is located at Window function, Denotes the complex exponential basis functions. This represents the total length of the frame interval sequence. It represents the imaginary unit.

[0141] Defining a mask function for the frequency band sensitive to the human eye and calculating the energy ratio involves frequency domain filtering and integral normalization. The rectangular window characteristic of the mask function corresponds to the range in visual physiology most sensitive to flicker at a specific frequency. Normalizing the energy of the entire frequency band as the denominator eliminates the interference of the overall frame rate on risk assessment, making the energy ratio a dimensionless indicator that purely reflects the quality of jitter rather than its quantity. Identifying high-risk zones and attenuating the weights establishes a gradient-based negative feedback control loop, which maps the global risk signal at the frequency domain back to the local weight adjustment at the spatial level. By reducing the rendering load in high-risk areas, it actively disrupts the conditions for resonance. The introduction of the convergence criterion is equivalent to a damper in the control system, preventing excessive weight adjustment from causing the image quality to oscillate repeatedly between two extreme states, ensuring that the closed-loop adjustment process monotonically converges to a steady state.

[0142] In summary, this invention eliminates pipeline bubbles caused by neglecting transmission delay in traditional symmetrical partitioning by transforming cloud-edge scheduling into a graph-based balanced partitioning problem and employing linear scanning discrete optimization search. It introduces sliding window variance and dynamic stability decay scoring to distinguish between instantaneous performance and sustainable performance, thus mitigating the risk of accumulated jitter in long-term tasks. A bidirectional matching model is constructed to achieve supply-demand coupling evaluation, prompting the system to select nodes that are both stable and perfectly matched. Subjective visual experience is objectified into frequency domain signals and a closed-loop correction mechanism is formed, enabling the compression strategy to actively suppress the resonance risk in frequency bands sensitive to human vision. Therefore, under limited heterogeneous resources, this invention achieves a synergistic improvement in digital human rendering quality, system stability, and user experience consistency.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A lightweight deployment and scheduling system for edge cloud collaborative digital humans, characterized in that: include: Pipeline segmentation module, semantic parsing module, elastic compression module, resource awareness module, task scheduling module, and spectrum feedback module; The pipeline segmentation module is used to construct an operator-level multimodal pipeline directed acyclic graph based on digital human interaction requests, collect operator-level baseline time consumption matrix, solve for the optimal segmentation point, and output an asymmetric segmentation scheme. The semantic parsing module is used to identify the intent of the digital human's interaction requests and generate semantic label vectors, and then transform the semantic label vectors into an initial importance weight map of the three-dimensional model region. The elastic compression module is used to obtain the available resource vector of the target edge node in the asymmetric segmentation scheme, solve the optimal compression configuration of the region based on the initial importance weight map of the 3D model region and the available resource vector as the constraint, and assemble a lightweight model instance according to the optimal compression configuration of the region. The resource awareness module is used to collect multi-dimensional resource time-series sequences of candidate edge nodes, calculate the sliding window variance, and combine the task load characteristics of the asymmetric partitioning scheme to calculate the comprehensive stability computing power score. The task scheduling module is used to perform bidirectional matching of the resource requirements and stability computing power comprehensive scores of the asymmetric partitioning scheme and lightweight model instance to determine the optimal node combination, and control the optimal node combination to execute the digital human rendering task according to the asymmetric partitioning scheme. The spectrum feedback module is used to extract the energy proportion of the rendering output frame interval sequence in the human eye sensitive frequency band when the optimal node combination is used to perform the digital human rendering task. The energy proportion is used to correct the initial importance weight map of the 3D model region to obtain the corrected importance weight map. The corrected importance weight map is fed back to the elastic compression module to resolve the optimal compression configuration of the region.

2. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 1, characterized in that: The specific steps for finding the optimal split point and outputting the asymmetric splitting scheme are as follows: Traverse each connecting edge in the directed acyclic graph of the operator-level multimodal pipeline as a candidate splitting position; For each candidate segmentation position, the baseline time of all operators before the candidate segmentation position is accumulated to obtain the expected time on the cloud side, and the baseline time of all operators after the candidate segmentation position is accumulated to obtain the expected time on the edge side. Calculate the absolute difference between the expected time on the cloud side and the expected time on the edge side; Compare the absolute differences corresponding to all candidate splitting positions, and select the candidate splitting position with the smallest absolute difference as the optimal splitting point; Based on the optimal split point, the directed acyclic graph of the operator-level multimodal pipeline is divided into cloud-side sub-pipelines and edge-side sub-pipelines; Extract the operator names, inter-operator data transmission protocols, and expected synchronization timestamps from the cloud-side sub-pipeline and the edge-side sub-pipeline, and combine them into an asymmetric partitioning scheme.

3. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 2, characterized in that: The specific steps for converting semantic label vectors into an initial importance weight map of the three-dimensional model region are as follows: Load the pre-trained semantic rendering region mapping knowledge base, input the semantic label vector into the pre-trained semantic rendering region mapping knowledge base for matching and retrieval, and output a list of 3D model mesh partition identifiers associated with the semantic label vector; Iterate through all mesh partition identifiers of the 3D model and determine whether the current mesh partition identifier exists in the 3D model mesh partition identifier list; If the current mesh partition identifier exists in the 3D model mesh partition identifier list, then assign the current mesh partition identifier a preset high weight value; If the current mesh partition identifier does not exist in the 3D model mesh partition identifier list, then assign the current mesh partition identifier a preset low weight value; Arrange the weight values ​​corresponding to all grid partition identifiers according to the topology of the 3D model, and generate an initial importance weight map of the 3D model regions.

4. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 3, characterized in that: The specific steps for solving the optimal compression configuration of a region using the initial importance weight map of the 3D model region as the weighting basis and the available resource vector as the constraint are as follows: Read the maximum video memory capacity, maximum computing throughput, and maximum bandwidth from the available resource vector; For each grid partition in the initial importance weight map of the 3D model region, enumerate all candidate compression levels; For each candidate compression level, estimate the corresponding memory usage, computing power consumption, and data transmission volume; Remove candidate compression levels that exceed the video memory capacity limit, the computing power consumption limit, or the data transmission limit, and retain the set of feasible compression levels. In the set of feasible compression levels, the candidate compression level that maximizes the perceived utility score is selected as the optimal compression configuration for the region based on the nonlinear perception utility formula.

5. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 4, characterized in that: The steps for calculating the comprehensive stability computing power score based on the task load characteristics combined with the asymmetric partitioning scheme are as follows: Extract the expected execution time of the edge-side sub-pipeline from the asymmetric partitioning scheme as a task load feature; Collect multidimensional resource time series of candidate edge nodes and calculate the sliding window variance; Obtain the baseline computing power score of candidate edge nodes, and calculate the comprehensive stability computing power score based on the dynamic stability decay formula.

6. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 5, characterized in that: The specific steps for determining the optimal node combination by bidirectionally matching the resource requirements and stability computing power comprehensive scores of the asymmetric partitioning scheme and the lightweight model instance are as follows: The total video memory requirement, total computing power requirement, and total bandwidth requirement are obtained by summarizing the optimal compression configuration of the region for the lightweight model instance. A node adaptability evaluation model is constructed, with the stability computing power comprehensive score as the numerator and the Euclidean distance between the resource demand and available resource vectors as the denominator. Traverse all candidate edge nodes and substitute them into the node fit evaluation model to calculate the fit score; Candidate edge nodes with a fit score higher than a preset threshold and whose available resource vectors simultaneously meet the total video memory requirement, total computing power requirement, and total bandwidth requirement are selected to form a set of candidate nodes. Select the node with the highest fit score from the candidate node set as the edge-side execution node; In the asymmetric partitioning scheme, find the cloud node that matches the cloud-side subpipeline and use it as the cloud-side execution node; The optimal node combination is obtained by combining edge-side execution nodes and cloud-side execution nodes.

7. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 6, characterized in that: The specific steps for extracting the energy proportion of the rendered output frame interval sequence in the human eye-sensitive frequency band when performing digital human rendering tasks are as follows: During the execution of the digital human rendering task with the optimal node combination, the timestamps of consecutive frames are recorded at a fixed sampling period. Calculate the timestamp difference between adjacent frames to obtain the rendered output frame interval sequence; Perform a short-time Fourier transform on the rendered output frame interval sequence to obtain the spectral energy distribution; Define a mask function for the frequency band sensitive to human eyes. The mask function for the frequency band sensitive to human eyes takes a value of 1 in the range of 8 Hz to 15 Hz and a value of 0 in the other frequency bands. The energy of the sensitive frequency band is obtained by multiplying the spectral energy distribution point by point with the mask function of the human eye sensitive frequency band and then integrating. The total fluctuation energy is obtained by integrating the spectral energy distribution across the entire frequency band, and the energy percentage is calculated based on the frequency domain resonance risk formula.

8. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 7, characterized in that: The specific steps for obtaining the corrected importance weight map by correcting the initial importance weight map of the three-dimensional model region using energy proportion are as follows: Determine whether the energy percentage exceeds the preset resonance risk threshold; Among them, when the energy ratio exceeds the preset resonance risk threshold, the grid partition with the largest historical jitter amplitude in the current optimal node combination is identified as the high-risk partition. The corrected weight value is obtained by multiplying the weight value of the high-risk partition in the initial importance weight map of the 3D model region by a preset attenuation coefficient. Keep the weight values ​​of all grid partitions except the high-risk partitions unchanged in the initial importance weight map of the 3D model region, and rearrange the modified weight values ​​of the high-risk partitions and the unmodified weight values ​​of the other grid partitions according to the topology of the 3D model to generate the modified importance weight map.

9. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 8, characterized in that: The specific steps for feeding the corrected importance weight map back to the elastic compression module to resolve for the optimal compression configuration in the region are as follows: The elastic compression module receives the corrected importance weight map and replaces the initial importance weight map of the 3D model region; Keep the available resource vector unchanged, and re-enumerate all candidate compression levels based on the revised importance weight graph; Re-execute the process of screening the feasible compression level set and calculating the nonlinear perceived utility formula; A convergence criterion is introduced: the iteration terminates when the difference between the updated optimal regional compression configuration and the previous optimal regional compression configuration is lower than a preset convergence threshold. If the difference is not lower than the preset convergence threshold, the spectrum feedback module will be triggered again to collect a new round of rendering output frame interval sequence and repeat the correction process. When the convergence criterion is met, the optimal compression configuration of the final region is output and the target edge nodes are triggered to assemble lightweight model instances according to the optimal compression configuration of the final region.

10. The edge cloud collaborative lightweight deployment and scheduling system for digital humans as described in claim 9, characterized in that: Also includes: When the energy percentage does not exceed the preset resonance risk threshold, skip the step of correcting the initial importance weight map of the three-dimensional model region and maintain the current optimal compression configuration of the region unchanged; After the digital human rendering task is completed, the energy percentage, the corrected importance weight map, and the updated optimal compression configuration of the region recorded in this task will be stored in the historical execution log. When subsequent digital human interaction requests with the same semantic label vector arrive, the historically corrected importance weight map matching the current candidate edge node is retrieved from the historical execution log first as the initial value of the initial importance weight map of the 3D model region, and the first resonance risk response time is shortened.

Images