A digital human motion generation method and generation system
Through feature extraction and coding, hypernetwork optimization and biological inspiration optimization, combined with action scheduling and execution modules, the architecture adjustment, feature extraction and execution efficiency problems of digital human action generation in deep learning technology are solved, achieving more natural and efficient actions generate.
Patent Information
- Application Number
- CN202411165447.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-23
AI Technical Summary
When existing deep learning technologies generate digital human actions, it is difficult to adjust the network architecture, resulting in insufficient applicability and performance of the model, insufficient feature extraction, lack of naturalness and fluency in action generation, and low execution efficiency.
Feature extraction and coding modules are used to extract features such as motion trajectory, joint angle and speed. The hypernetwork optimization module dynamically adjusts the neural network architecture, the biologically inspired optimization module optimizes the action strategy, and the action scheduling and execution modules reasonably plan timing and resource allocation.
It improves the depth and accuracy of feature extraction, the naturalness and fluency of action generation, ensures the efficiency and applicability of the model in different application scenarios, and enhances the user's immersion and interactive experience.
Smart Images

Figure CN118710786B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of deep learning technology, and in particular to a digital human motion generation method and generation system. Background Art
[0002] Deep learning is a branch of artificial intelligence (AI) and belongs to the field of machine learning. It processes and analyzes complex data by simulating the structure and function of the human brain's neural network. Deep learning technology relies on a large amount of training data and powerful computing power to automatically extract features from data, identify patterns, and make predictions. Its core lies in the multi-layer neural network architecture, which can achieve excellent performance in image recognition, speech recognition, natural language processing and other fields by adjusting the connection weights. In deep learning, commonly used models include convolutional neural networks (CNN), recurrent neural networks (RNN) and generative adversarial networks (GAN). Deep learning plays an important role in many applications such as autonomous driving, medical diagnosis, and personalized recommendation systems, and has significantly improved the ability of computers to process unstructured data.
[0003] The digital human motion generation technology mainly involves generating realistic motion sequences for virtual digital humans through deep learning methods. This method uses deep learning algorithms to analyze and learn data from human motions in the real world, thereby generating natural and smooth motions in a virtual environment. This technology can be used in virtual character animation, game development, virtual reality (VR) experience, and human-computer interaction systems to enhance the user's immersion and interactive experience. Its application is not limited to the entertainment industry, but can also be extended to fields such as education, medical care, and military simulation, providing users with a more realistic and personalized virtual experience.
[0004] However, existing deep learning technologies have some shortcomings in generating digital human action sequences. First, traditional deep learning models usually rely on fixed network architectures, which may make it difficult to adjust the number of network layers and neurons when facing data of different complexity and diversity, thereby affecting the applicability and performance of the model. In addition, the depth and accuracy of feature extraction in the existing technology are limited, and it is difficult to fully capture the subtle changes in action data, resulting in the lack of naturalness and fluency in the generated action sequence. For example, a convolutional neural network with a fixed architecture may not be able to fully capture the dynamic relationship between actions when processing complex action data, resulting in the generated actions being not realistic enough. In addition, the existing technology lacks global optimization and dynamic adjustment of action strategies during the action generation process, making it difficult to quickly adapt in uncertain and changing environments. In addition, in terms of timing planning and resource allocation for action execution, existing methods often lack refined management, which can easily lead to low execution efficiency and affect the actual experience of users. The shortcomings of the existing technology in these aspects limit its scalability and practicality in more complex and diverse application scenarios. Summary of the invention
[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a digital human motion generation method and generation system.
[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: A digital human action generation system comprises:
[0007] The feature extraction and encoding module extracts task-related features from the input action data, including the motion trajectory, joint angle, velocity and acceleration of the action. It calculates the key positions of the joints by performing edge detection on the motion trajectory, identifies and calculates the rate of change and trend characteristics by performing morphological transformation analysis on the angle and velocity, combines the features into feature vectors, inputs them into the feature encoder, and encodes the feature vectors to generate feature vector representations.
[0008] The hypernetwork optimization module uses feature vector representation as input, generates main model structure parameters by mapping input feature vectors with parameters of multiple candidate architectures in the hypernetwork, dynamically adjusts the main model structure, calculates the number of main model layers and neurons, compares the performance of the architecture, selects the optimal network architecture, generates dynamic structure parameters, adapts the activation function between input features and network layers, selects appropriate activation functions, generates activation parameters, adjusts network parameters according to the characteristics of the task and data complexity, generates structure parameters and network architecture, and generates network parameter configuration by sharing and adjusting the parameters generated by the hypernetwork;
[0009] The action generation and adjustment module applies the network parameter configuration in the main model to generate an action sequence. It generates an initial action sequence by calculating the linear combination of input features and network parameters and applying a nonlinear activation function. It calculates the adjustment value by comparing the difference between the generated action sequence and the input features, adjusts the generated action sequence, and compares the action sequence paths between multiple modules, selects the path with the minimum difference, and combines the action strategies to obtain the action sequence path.
[0010] The bio-inspired optimization module applies the action sequence path, encodes the action strategy into a gene sequence, generates a population with random initialization, calculates the fitness value of each individual, generates adaptive individuals by performing crossover and mutation operations on the gene sequence, sorts the fitness values, selects adaptive individuals to enter the next generation, generates the optimal individual, calculates the action path selection probability through pheromone initialization, selects the optimal path, and generates the optimal action strategy;
[0011] The action scheduling and execution module uses the optimal action strategy to plan the timing and resource allocation of action execution. It generates an execution plan by judging the time requirements and resource consumption of the action strategy, integrates the dependencies and priorities between action strategies, builds an action execution queue, determines the action execution time and resource allocation by sorting the action execution order, and adjusts the action execution plan by calculating the dependencies and timing constraints of the action execution to obtain the action execution plan.
[0012] As a further solution of the present invention, the feature vector representation includes trajectory information, joint angle data, velocity change rate and acceleration characteristics, the network parameter configuration includes neuron connection weights, hierarchy configuration and activation function type, the action sequence path includes joint motion trajectory, posture adjustment sequence and continuous action chain, the optimal action strategy includes optimized motion steps, path priority and resource allocation scheme, and the action execution plan includes timing scheduling, resource allocation and execution priority.
[0013] As a further solution of the present invention, the feature extraction and encoding module includes:
[0014] The trajectory processing submodule analyzes the motion trajectory data based on the input action data, obtains the three-dimensional coordinate information from the sensor data, performs trajectory analysis, processes the trajectory in segments, calculates the key positions of the joints through edge detection, identifies the key turning points in the trajectory change, extracts the spatial coordinate information, and generates the trajectory position features;
[0015] The angle analysis submodule analyzes the joint angle and speed based on the trajectory position characteristics. It calculates the rotation angle of the joint from the position change, calculates the speed change rate based on the time difference, identifies the angle change trend through morphological transformation processing, and combines the angle change and speed change rate to form the angular velocity feature data describing the joint movement.
[0016] The feature encoding submodule performs feature vector combination processing based on the angular velocity feature data. It vectorizes the angular velocity, position and acceleration features, integrates the features into feature vectors, maintains data consistency through vector normalization, optimizes the data structure by rearranging feature components, encodes them into feature vector matrices, and generates feature vector representations.
[0017] As a further solution of the present invention, the super network optimization module includes:
[0018] The structural parameter generation submodule performs parameter mapping between the input feature vector and the hypernetwork architecture based on the feature vector representation. By judging the response effects of various architectures to the feature vector, the number of main model layers and neuron configuration are generated, and parameter combination tests are performed. By aggregating the mapping results, the configuration with the best response effect is extracted to generate an architecture parameter mapping table.
[0019] The dynamic architecture selection submodule performs performance analysis of candidate network architectures based on the architecture parameter mapping table. By calculating the efficiency and accuracy scores of multiple architectures under the target task, the architecture with the best score is selected as the main model infrastructure. The optimal network layer configuration is determined by analyzing the inter-layer connection relationship in the architecture, and the network layer structure parameters are generated.
[0020] The parameter configuration submodule performs parameter optimization based on the network layer structure parameters. It determines the activation function by analyzing the neuron activation mode of each layer, selects the appropriate weight setting by comparing the performance of multiple weight configurations, applies parameter sharing strategies to improve computing efficiency, obtains the optimal combination through optimization and adjustment, and generates network parameter configuration.
[0021] As a further solution of the present invention, the action generation and adjustment module includes:
[0022] The sequence generation submodule generates action sequences by applying parameters in the main model based on the network parameter configuration, generates neural activation values by calculating the linear combination of input features and network parameters, obtains the initial action sequence by applying activation operations, and generates initial sequence difference data by comparing the initial action sequence with the target action features;
[0023] The sequence adjustment submodule adjusts the action sequence based on the initial sequence difference data. It adjusts the elements of the initial action sequence by calculating the deviation value of each action element in the sequence, and applies the adjustment strategy to correct the deviation in the sequence to obtain the corrected action sequence and generate the adjusted action sequence.
[0024] The path optimization submodule performs sequence path optimization based on the adjusted action sequence. It judges the path smoothness and fluency by comparing the action sequence paths generated by the module, selects the path with the lowest difference as the final output, and selects the optimal path combination through the path integration strategy to obtain the action sequence path.
[0025] As a further solution of the present invention, the bio-inspired optimization module includes:
[0026] The gene encoding submodule encodes the action strategy based on the action sequence path, decomposes the action strategy into gene sequence fragments, initializes the gene library for population generation, and forms a diversified initial population by combining gene fragments to generate the encoding gene library;
[0027] The fitness value sorting submodule calculates the fitness value based on the coding gene library, calculates the matching value by judging the performance of each individual's gene sequence in the task, generates an individual fitness value list, sorts it by fitness value, and selects the individual with the best matching degree for gene transfer to generate a sequence of adapted individuals;
[0028] The path selection submodule selects the optimal path based on the adaptation individual sequence. It calculates the probability of individual action path selection, judges the performance of multiple paths, selects the path with the best priority, and generates the optimal path and the optimal action strategy by combining the selection strategies.
[0029] As a further solution of the present invention, the action scheduling and execution module includes:
[0030] The plan generation submodule plans the action execution sequence and resource allocation based on the optimal action strategy, builds a plan by judging the time requirements and resource constraints of the action strategy, and then generates a preliminary plan by analyzing the execution time window and resource availability;
[0031] The resource allocation submodule allocates resources based on the preliminary plan. It allocates resources by judging the resource requirements of each action and combining the resource usage priority of the execution queue to generate a resource scheduling plan.
[0032] The plan adjustment submodule makes plan adjustments based on the resource scheduling plan. It adjusts the action execution time by analyzing the dependencies and priorities in the action execution plan, optimizes resource allocation, corrects timing constraints, and obtains the action execution plan.
[0033] A method for generating digital human motions, which is performed based on the digital human motion generation system, comprises the following steps:
[0034] S1, based on the input action data, uses edge detection algorithm to extract the motion trajectory features of the action, including the key positions of the joints. At the same time, the change rate and trend characteristics of the joint angle, velocity and acceleration are identified through morphological transformation analysis. The extracted features are combined into feature vectors, input into the feature encoder, and the feature vector representation is generated through the feature encoding algorithm;
[0035] S2, based on the feature vector representation, maps the input feature vector to the parameters of multiple candidate architectures through the hypernetwork mapping algorithm, and uses the dynamic adjustment algorithm to generate the main model structure parameters, including the number of layers, number of neurons, and activation function type of the main model, and then generates dynamic structure parameters, and finally generates network parameter configuration through parameter sharing and adjustment technology;
[0036] S3, based on the network parameter configuration, applies the network parameter configuration in the main model to generate an initial action sequence, generates the initial action sequence through a linear combination algorithm and a nonlinear activation function, then compares the difference between the generated action sequence and the input feature, calculates the adjustment value and adjusts the generated action sequence through a difference minimization algorithm, and finally generates the action sequence path by comparing the action sequence paths between multiple modules and selecting the path with the smallest difference;
[0037] S4, based on the action sequence path, encodes the action strategy into a gene sequence, uses a genetic algorithm to randomly initialize and generate a population, calculates the fitness value of each individual, generates adaptive individuals through crossover and mutation operations, and selects individuals that enter the next generation through fitness value sorting to generate the optimal individual, and finally initializes through pheromone, applies the ant colony algorithm to calculate the action path selection probability, selects the optimal path, and generates the optimal action strategy;
[0038] S5, based on the optimal action strategy, plans the timing and resource allocation of action execution, generates an execution plan through the time requirement judgment algorithm and resource consumption, integrates the dependencies and priorities between action strategies, uses the execution queue construction technology to sort the action execution order, determines the action execution time and resource allocation, and generates an action execution plan by calculating the dependencies and timing constraints of action execution;
[0039] S6, executes actions based on the action execution plan, collects execution data and uses feedback control algorithms to correct deviations in action execution, and continuously adjusts execution strategies through optimized feedback loops to generate optimized execution results.
[0040] Compared with the prior art, the advantages and positive effects of the present invention are:
[0041] In the present invention, by performing fine feature extraction and encoding on the action data, it is more accurate to extract features such as motion trajectory, joint angle, speed and acceleration from the input data. The recognition of the change rate and trend characteristics of the action is also more accurate, which effectively improves the depth and breadth of feature extraction. Secondly, by dynamically optimizing the neural network architecture using the hypernetwork, the network structure can be adjusted for specific tasks and data complexity, ensuring the efficiency and applicability of the model in different application scenarios. Compared with the traditional deep learning model, this method can flexibly select the optimal network architecture and activation function, improving the training efficiency and performance of the model. By combining the selection of the action sequence path with the biological heuristic optimization algorithm, the naturalness and fluency of the action generation can be effectively improved, ensuring that the action of the virtual digital person is more realistic. By encoding the generated action strategy into a gene sequence and optimizing the action path using a genetic algorithm, the automatic optimization of the action generation strategy is realized. The rational planning of the timing of action execution and resource allocation ensures the efficient execution of the action generation system in different application scenarios. This innovative solution significantly improves the user's immersion and interactive experience in virtual character animation production, game development and virtual reality experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0044] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.
[0045] Example:
[0046] See also Figure 1 The present invention provides a technical solution: a digital human motion generation system comprising:
[0047] The feature extraction and encoding module extracts task-related features from the input action data, including the motion trajectory, joint angle, velocity and acceleration of the action. It calculates the key positions of the joints by performing edge detection on the motion trajectory, identifies and calculates the rate of change and trend characteristics by performing morphological transformation analysis on the angle and velocity, combines the features into feature vectors, inputs them into the feature encoder, and encodes the feature vectors to generate feature vector representations.
[0048] The hypernetwork optimization module uses feature vector representation as input, generates main model structure parameters by mapping input feature vectors with parameters of multiple candidate architectures in the hypernetwork, dynamically adjusts the main model structure, calculates the number of main model layers and neurons, compares the performance of the architecture, selects the optimal network architecture, generates dynamic structure parameters, adapts the activation function between input features and network layers, selects appropriate activation functions, generates activation parameters, adjusts network parameters according to the characteristics of the task and data complexity, generates structure parameters and network architecture, and generates network parameter configuration by sharing and adjusting the parameters generated by the hypernetwork;
[0049] The action generation and adjustment module applies the network parameter configuration in the main model to generate an action sequence. It generates an initial action sequence by calculating the linear combination of input features and network parameters and applying a nonlinear activation function. It calculates the adjustment value by comparing the difference between the generated action sequence and the input features, adjusts the generated action sequence, and compares the action sequence paths between multiple modules, selects the path with the minimum difference, and combines the action strategies to obtain the action sequence path.
[0050] The bio-inspired optimization module applies the action sequence path, encodes the action strategy into a gene sequence, generates a population with random initialization, calculates the fitness value of each individual, generates adaptive individuals by performing crossover and mutation operations on the gene sequence, sorts the fitness values, selects adaptive individuals to enter the next generation, generates the optimal individual, calculates the action path selection probability through pheromone initialization, selects the optimal path, and generates the optimal action strategy;
[0051] The action scheduling and execution module uses the optimal action strategy to plan the timing and resource allocation of action execution. It generates an execution plan by judging the time requirements and resource consumption of the action strategy, integrates the dependencies and priorities between action strategies, builds an action execution queue, determines the action execution time and resource allocation by sorting the action execution order, and adjusts the action execution plan by calculating the dependencies and timing constraints of the action execution to obtain the action execution plan.
[0052] The feature vector representation includes trajectory information, joint angle data, velocity change rate and acceleration characteristics, the network parameter configuration includes neuron connection weights, hierarchy configuration and activation function type, the action sequence path includes joint motion trajectory, posture adjustment sequence and continuous action chain, the optimal action strategy includes optimized motion steps, path priority and resource allocation plan, and the action execution plan includes timing scheduling, resource allocation and execution priority.
[0053] In the feature extraction and encoding module, task-related features are extracted from the input motion data. The format of the feature data is time series data, including motion trajectory, joint angle, velocity and acceleration information in the form of multidimensional arrays. First, edge detection technology is used to analyze the motion trajectory. By calculating the gradient and locating the edge points in the trajectory, the key positions of the joint points are calculated, the trajectory data is divided into meaningful motion segments, and the key turning points of the motion are identified. Then, for the joint angle and velocity data, the trend characteristics of the angle change rate and velocity are calculated through morphological transformation analysis. By smoothing and differentiating the angle curve and velocity curve, the characteristics of angular velocity change are extracted to generate angular velocity data describing the motion process. Finally, all the extracted features are combined into a feature vector and input into the feature encoder. The encoder standardizes the feature vector and optimizes the data structure by rearranging the feature components to generate a multidimensional feature vector matrix file, which contains trajectory information, joint angle data, velocity change rate and acceleration features. As the input of the subsequent modules, the integrity and diversity of feature extraction are ensured, providing a rich data foundation for subsequent model training and optimization.
[0054] In the hypernetwork optimization module, the feature vector representation is used as input. First, the feature vector is mapped to the parameters of multiple candidate architectures in the hypernetwork. The parameters include structural information such as the number of neurons, the number of layers, and the activation function. The response of each candidate architecture is evaluated by inputting the feature vector, and the preliminary performance indicators of each architecture are generated. Then, in the dynamic architecture selection process, the efficiency and accuracy scores of each architecture under a specific task are calculated, and the optimal network architecture is determined by comparing the performance scores. In order to optimize the structure, a parameter sharing mechanism and cross-validation are adopted. By analyzing the connection relationship between neurons at different levels, appropriate connection weights are selected, the optimal activation function is applied, the depth and width of the network are dynamically adjusted, and a new network structure configuration is generated. Subsequently, by adapting the activation function between the input features and the network layer, a suitable activation function is selected, activation parameters are generated, and the network parameters are adjusted according to the characteristics of the task and the complexity of the data, and the final network parameter configuration file is generated. The file includes neuron connection weights, hierarchical structure configuration, and activation function type, ensuring the efficiency and adaptability of the model and improving the performance and generalization ability of the model.
[0055] In the action generation and adjustment module, the action sequence is generated in the main model by applying the network parameter configuration. The specific implementation process is to linearly combine the feature vector with the model parameters, generate the initial neural activation value through the feedforward calculation of the neural network layer, and generate the initial action sequence through the nonlinear activation function. The initial action sequence is a posture or motion posture corresponding to a time step. Then, according to the difference between the initial action sequence and the target feature, adjustments are made. By calculating the error vector and deviation value of each action frame, the action sequence is adjusted using the gradient descent method to correct the error. Then, by comparing the action sequence paths generated by multiple modules, the smoothness and continuity of the path are evaluated, and the path with the smallest difference is selected to ensure the continuity of the action. Finally, the path selection is adjusted through the optimization strategy to generate an optimized action sequence path file, which includes joint motion trajectories, posture adjustment sequences, and continuous action chains, providing accurate action generation paths and supporting real-time control of digital human actions.
[0056] In the bio-inspired optimization module, the action strategy is bio-inspired optimized by applying the action sequence path. First, the action strategy is converted into a gene sequence, which contains the key parameters and path selection information for action execution. Based on the sequence, the initial population is generated by random initialization, and the fitness value of each individual is calculated by the fitness function. The fitness value reflects the performance of the individual in a given task environment. Next, new individuals are generated through gene crossover and mutation operations. Then, the population is sorted according to the fitness value, and the individual with the highest fitness value is selected to enter the next generation to ensure that the genes of the best individual can be passed on. The action path selection probability is calculated by pheromone initialization, and the optimized path selection mechanism is adopted. Combined with the optimal combination strategy between individuals, the optimal action strategy file is generated. The file includes optimized motion steps, path priority and resource allocation scheme, which improves the efficiency and flexibility of the action strategy and ensures excellent performance in complex environments.
[0057] In the action scheduling and execution module, the optimal action strategy is used to perform timing planning and resource allocation for action execution. First, according to the motion steps and path priorities in the optimal action strategy file, the time schedule for action execution is planned, a preliminary execution plan is constructed, and the execution order and required resources of each action are analyzed. Subsequently, by evaluating the time requirements and resource consumption of each action, available resources are allocated to optimize resource utilization efficiency. Combining the dependencies and execution priorities between actions, an action execution queue is constructed to ensure the order and continuity of execution. Next, the execution plan is adjusted to correct the timing constraints in the plan, adjust the action execution time and resource allocation, and ensure the consistency of the plan. Finally, an action execution plan file is generated, which includes timing arrangement, resource allocation, and execution priority, providing a detailed scheduling plan for the operation of the system to ensure stability and efficiency during the execution process.
[0058] In this embodiment, the feature extraction and encoding module includes:
[0059] The trajectory processing submodule analyzes the motion trajectory data based on the input action data, obtains the three-dimensional coordinate information from the sensor data, performs trajectory analysis, processes the trajectory in segments, calculates the key positions of the joints through edge detection, identifies the key turning points in the trajectory change, extracts the spatial coordinate information, and generates the trajectory position features;
[0060] The angle analysis submodule analyzes the joint angle and speed based on the trajectory position characteristics. It calculates the rotation angle of the joint from the position change, calculates the speed change rate based on the time difference, identifies the angle change trend through morphological transformation processing, and combines the angle change and speed change rate to form the angular velocity feature data describing the joint movement.
[0061] The feature encoding submodule performs feature vector combination processing based on the angular velocity feature data. It vectorizes the angular velocity, position and acceleration features, integrates the features into feature vectors, maintains data consistency through vector normalization, optimizes the data structure by rearranging feature components, encodes them into feature vector matrices, and generates feature vector representations.
[0062] In the trajectory processing submodule, detailed information of the motion trajectory is obtained by inputting action data. The specific input data format is a time series data set containing timestamps and position information (x, y, z coordinates). First, the three-dimensional coordinate data collected by the sensor is preliminarily parsed to obtain the continuous position changes of the object or joint in space. Through time series analysis, the trajectory data is segmented to identify and mark the key change points in the motion segment. The change point is manifested as a change in speed, direction or acceleration in the motion path. For each segment, the edge detection technology is used to process the trajectory information and calculate the key position of the joint point in space. The specific operation is to find the point with significant change as the key point by calculating the gradient information of the coordinates. Then, the three-dimensional space coordinates are extracted, and the relative position and motion characteristics of the joint point during the motion process are obtained by analyzing the change in the position of the joint point over time. The file finally generated by this submodule is a trajectory position feature data set, which contains the time series coordinate information of all key positions of the joint point during the entire motion process. By extracting the trajectory position features, an accurate spatial reference can be provided for subsequent angle analysis and action recognition, ensuring that the entire action generation system accurately captures and understands the dynamic motion features.
[0063] In the angle analysis submodule, the angle and speed of joint motion are analyzed in detail through the trajectory position features, and the input data format is the key position time series data obtained from the trajectory processing submodule. First, by calculating the change in the position of the joint point between two consecutive moments, the rotation angle of the joint is derived using the spatial geometric relationship. Then, combined with the timestamp information, the velocity change rate, that is, the change rate of the joint position per unit time, is calculated. In this way, the linear motion characteristics and angular rotation characteristics of the joint are combined to form a preliminary angular velocity description. Subsequently, morphological transformation is used to process the angle data, and the noise is removed by signal smoothing technology, and the change trend in motion is identified. Morphological processing can effectively capture subtle angle changes in motion, thereby accurately describing the dynamic characteristics of the joint. Finally, the combination of angle change and velocity change rate generates an angular velocity feature data file, which records in detail the angular velocity characteristics of each joint during motion. The generation of angular velocity feature data provides a basis for subsequent feature encoding, ensuring that the system can accurately simulate and generate motion based on real motion characteristics.
[0064] In the feature encoding submodule, the feature vectors are combined and encoded through the angular velocity feature data, and the input data format is the angular velocity feature data set output by the angle analysis submodule. First, the angular velocity, position and acceleration features are vectorized, and the three feature information is combined into a unified data structure. Then, the combined feature vectors are standardized to ensure the consistency and comparability of the data. Standardization eliminates the dimensional differences of different eigenvalues, making the data relatively uniform on a numerical scale. Subsequently, the feature component rearrangement technology is used to optimize the data structure, and important features are promoted to a more prominent position to improve the recognition and effectiveness of the feature vector. Through the encoding operation, the final processed feature data is converted into a feature vector matrix file, which contains the feature vector representation of each moment and can fully describe the dynamic characteristics of the joint motion. The generation of the feature vector representation provides structured input data for the subsequent hypernetwork optimization, improves the accuracy and robustness of the system in the motion generation process, and ensures that each step is processed based on the optimized and encoded motion features.
[0065] In this embodiment, the super network optimization module includes:
[0066] The structural parameter generation submodule performs parameter mapping between the input feature vector and the hypernetwork architecture based on the feature vector representation. By judging the response effects of various architectures to the feature vector, the number of main model layers and neuron configuration are generated, and parameter combination tests are performed. By aggregating the mapping results, the configuration with the best response effect is extracted to generate an architecture parameter mapping table.
[0067] The dynamic architecture selection submodule performs performance analysis of candidate network architectures based on the architecture parameter mapping table. By calculating the efficiency and accuracy scores of multiple architectures under the target task, the architecture with the best score is selected as the main model infrastructure. The optimal network layer configuration is determined by analyzing the inter-layer connection relationship in the architecture, and the network layer structure parameters are generated.
[0068] The parameter configuration submodule performs parameter optimization based on the network layer structure parameters. It determines the activation function by analyzing the neuron activation mode of each layer, selects the appropriate weight setting by comparing the performance of multiple weight configurations, applies parameter sharing strategies to improve computing efficiency, obtains the optimal combination through optimization and adjustment, and generates network parameter configuration.
[0069] In the structural parameter generation submodule, the input features are mapped to the parameters of the super network architecture through the feature vector representation. The vector space mapping technology is used to process the multi-dimensional feature data. The feature vector format is a multi-dimensional matrix, which contains information such as motion trajectory, joint angle, velocity and acceleration. First, the input feature vector is associated with the parameters of multiple candidate network architectures through linear transformation, and the response effects of different network architectures to the feature vector are calculated, including the reconstruction error of the feature and the output of the activation layer. The mapping relationship between the feature vector and the architecture is realized by matrix multiplication. Each network architecture can generate a set of independent mapping response results, and the results are used to evaluate the responsiveness of the network architecture in various situations. Then, the response of each network is analyzed layer by layer to determine its number of layers and neuron configuration to ensure that it achieves the optimal feature expression ability when processing tasks. In this process, the depth and width of the network layer and the requirements of specific tasks need to be comprehensively considered, and the neuron activation function and hierarchical connection relationship of the architecture configuration are dynamically adjusted. Through comprehensive performance analysis of all candidate network architectures, the response effects of different architectures are integrated by using the aggregation analysis method to generate the optimal structural parameter configuration of the main model. Finally, all the calculated mapping results are summarized into an architecture parameter mapping table, which lists in detail the response of each architecture under different characteristics and its optimized parameter combination. This mapping table provides necessary reference data and configuration basis for the architecture selection and parameter configuration of subsequent submodules. Through efficient mapping, the response speed and accuracy of the hypernetwork can be significantly improved, ensuring the effectiveness of subsequent optimization.
[0070] In the dynamic architecture selection submodule, the performance of the candidate network architecture is analyzed through the architecture parameter mapping table. The table records the performance indicators of different architectures under various feature combinations, including efficiency and accuracy. First, the multi-task performance evaluation method is used to evaluate the performance of each architecture under the target task using the performance score data in the table, considering the computational complexity, resource consumption and output accuracy of each architecture. During the evaluation process, a comprehensive performance analysis is required for each architecture, focusing on the evaluation of the efficiency and accuracy of the architecture. The architecture with the highest score will be used as the basic structure of the main model. Then, further analysis is conducted on the inter-layer connection relationship in the architecture to determine the optimal network layer configuration. The connection matrix is used to describe the connection density and relationship strength between neurons in each layer to ensure that the selected architecture can effectively process input features and adapt to complex task environments. In the selection process, through the comparison and analysis between architectures, the structural configuration with excellent performance in a multi-task environment is selected. The selected architecture needs to have good scalability and adaptability, and can maintain stable output performance under various feature inputs. The network layer structure parameter file finally generated records in detail the number of layers, number of neurons, hierarchical connections and activation functions of the selected architecture, providing a basis for the subsequent optimization of the parameter configuration submodule. Through dynamic architecture selection, the flexibility and performance of the model can be significantly improved, enabling it to adapt to changing task requirements and input feature changes, and improving the overall operating efficiency of the system.
[0071] In the parameter configuration submodule, detailed parameter optimization processing is performed through the network layer structure parameters. The parameter file includes information such as hierarchical structure, neuron connection, and activation function. First, the neuron activation mode of each layer is analyzed. By sampling the input and output data of each layer, the activation distribution and pattern characteristics of the neurons are obtained. During the analysis process, the activation frequency and pattern stability of the neurons are focused on, and the most suitable activation function type is identified to ensure that the neurons can effectively reflect the nonlinear changes of the input characteristics. Subsequently, the performance of multiple weight configurations is compared for each layer structure. By fine-tuning the connection weights of each layer, the impact of different configurations on network performance is observed. The most suitable weight combination is optimized through repeated experiments to obtain the best network performance. Finally, by integrating all optimized parameters and making global adjustments to ensure the overall performance of the network, the final network parameter configuration file is generated. This file records the activation functions, connection weights, and shared parameter strategies of each layer in detail, providing a detailed operating basis for the practical application of the model. Through such optimization and configuration, the network can obtain the maximum task processing capability with the minimum computing resources, ensuring adaptability and accuracy in different task scenarios, and also providing reliable parameter support for subsequent action generation. The result of this process is a streamlined and efficient neural network model configuration that ensures efficiency and stability when performing complex tasks.
[0072] In this embodiment, the action generation and adjustment module includes:
[0073] The sequence generation submodule generates action sequences by applying parameters in the main model based on the network parameter configuration, generates neural activation values by calculating the linear combination of input features and network parameters, obtains the initial action sequence by applying activation operations, and generates initial sequence difference data by comparing the initial action sequence with the target action features;
[0074] The sequence adjustment submodule adjusts the action sequence based on the initial sequence difference data. It adjusts the elements of the initial action sequence by calculating the deviation value of each action element in the sequence, and applies the adjustment strategy to correct the deviation in the sequence to obtain the corrected action sequence and generate the adjusted action sequence.
[0075] The path optimization submodule performs sequence path optimization based on the adjusted action sequence. It judges the path smoothness and fluency by comparing the action sequence paths generated by the module, selects the path with the lowest difference as the final output, and selects the optimal path combination through the path integration strategy to obtain the action sequence path.
[0076] In the sequence generation submodule, the action sequence is generated by applying the network parameter configuration. The input data format includes feature vectors and network parameter configurations. The feature vectors are generated in the previous feature extraction and encoding module, including information about trajectory, joint angles, speed, etc. The network parameter configuration contains information about the neural network structure, such as weights and biases. In the sequence generation process, the input features and network parameters are first linearly combined to calculate the neural activation value. Here, the activation input of each layer is obtained by matrix operation by multiplying the feature vector with the network weight matrix and adding the bias. The activation input is processed by a nonlinear activation function (such as ReLU or tanh) to obtain the output of each layer, which is passed layer by layer until the output layer. The activation value of the output layer represents the initial action sequence, which corresponds to the joint position and posture of the digital human. After obtaining the initial action sequence, it is compared frame by frame with the target action feature, and the error difference at each moment is calculated to generate the initial sequence difference data. The generation and difference analysis of the initial action sequence lays the foundation for the fine adjustment of the action, making the action of the digital human closer to the real human behavior.
[0077] In the sequence adjustment submodule, the action sequence is adjusted by analyzing the initial sequence difference data. The initial sequence difference data provides the error information of each action element. The data is stored in the form of a vector, which identifies the specific deviation between the initial action sequence and the target action feature. The adjustment process starts with calculating the deviation value of each action element, and analyzing the error direction and magnitude of the action frame by frame based on the difference data. Through the deviation value adjustment strategy, the deviation information is applied to the initial action sequence, and the joint position and posture of each frame are corrected according to the deviation vector. Specifically, the gradient of each action is calculated, and the parameters are adjusted to fine-tune the action in the direction of gradient descent. Through multiple iterative corrections, the error range is continuously narrowed, and finally the corrected action sequence is obtained. The adjusted action sequence more accurately reflects the target action characteristics, the error is significantly reduced, and an adjusted action sequence file is generated, which contains complete adjusted action data. This submodule makes the behavior of the digital human more accurate and natural by reducing the deviation of the initial action, providing a stable foundation for subsequent path optimization.
[0078] In the path optimization submodule, the sequence path is optimized by analyzing and comparing the adjusted action sequences. The adjusted action sequences are provided in a time series data format, including the joint positions and states at each time step. In the path optimization stage, the smoothness and fluency of the paths are first evaluated by analyzing multiple action sequence paths generated by the module. For each path, its smoothness score is calculated to evaluate the naturalness of the action transition. Subsequently, the path with the lowest difference is selected as the final output to ensure that the selected path has the best performance in terms of smoothness and target matching. In the path selection process, different path combinations are evaluated and selected through the path integration strategy, the path structure is optimized, and the best path combination scheme is formed. The final generated action sequence path file records the optimized joint motion trajectory, posture adjustment sequence, and continuous action chain. Through path optimization, the system not only improves the accuracy of action generation, but also enhances the practicality and effectiveness of digital humans in practical applications.
[0079] In this embodiment, the bio-inspired optimization module includes:
[0080] The gene encoding submodule encodes the action strategy based on the action sequence path, decomposes the action strategy into gene sequence fragments, initializes the gene library for population generation, and forms a diversified initial population by combining gene fragments to generate the encoding gene library;
[0081] The fitness value sorting submodule calculates the fitness value based on the coding gene library, calculates the matching value by judging the performance of each individual's gene sequence in the task, generates an individual fitness value list, sorts it by fitness value, and selects the individual with the best matching degree for gene transfer to generate a sequence of adapted individuals;
[0082] The path selection submodule selects the optimal path based on the adaptation individual sequence. It calculates the probability of individual action path selection, judges the performance of multiple paths, selects the path with the best priority, and generates the optimal path and the optimal action strategy by combining the selection strategies.
[0083] In the gene encoding submodule, the action strategy is genetically encoded through the action sequence path. First, the input data is the action sequence path, which is stored in the format of a multidimensional array and contains information such as joint positions, posture sequences, and time steps. By decomposing the action strategy into gene sequence fragments, the characteristic parameters of each action, such as movement amplitude, direction, and speed, are extracted, and the parameters are encoded as gene fragments. Gene encoding converts each action feature into a fixed-length gene sequence so that it can participate in genetic operations. Next, the initial population is generated by combining gene fragments to initialize the gene library. During the population generation process, different gene fragments are randomly selected and combined to ensure that the generated initial population is diverse and representative. To ensure diversity, a perturbation factor is introduced in the population generation so that each individual has slight differences in parameter selection, thereby enhancing the adaptability of the population. Finally, the generated encoding gene library is saved in the form of a structured file, which contains the gene sequences of all individuals and their corresponding action features. The encoding gene library provides rich individual information for subsequent fitness value calculation and path optimization, and can enhance the evolutionary potential and adaptability of the population through gene crossover and mutation in subsequent operations.
[0084] In the fitness ranking submodule, the fitness of each individual is calculated through the coding gene library. First, the gene sequence and action characteristics of the individual are extracted from the coding gene library, and the gene sequence is input into the simulation environment to perform action simulation related to the target task. During the action simulation process, the performance evaluation is performed based on the performance of the individual in the task, focusing on indicators such as accuracy, efficiency and stability of the action. Using the indicators, various performance indicators are integrated into a comprehensive fitness value through a linear weighted method, and the calculation results are stored in the fitness value list. Individuals with high fitness values show better adaptability in the task, which means that their gene sequences are more suitable for the current task. Subsequently, according to the calculated fitness values, the individuals in the gene library are sorted, and the top-ranked individuals are selected for gene transfer to ensure that the gene transfer process retains excellent characteristics. After the sorting is completed, the individuals with the highest fitness are selected for crossover and mutation operations to generate the next generation population and generate a sequence file of adapted individuals. This file records the information of the individuals entering the next generation and their fitness values, which serves as the input of the path selection submodule to ensure the directionality and efficiency of the evolutionary process.
[0085] In the path selection submodule, the optimal path selection is performed by adapting the individual sequence. First, based on the adaptation individual sequence, the gene sequence of the individual and the corresponding action path are extracted. By calculating the selection probability of each individual on multiple paths, the execution effect and priority of different paths are evaluated. When calculating the probability of path selection, factors such as the execution time, resource consumption, and success rate of the path need to be considered, and the priority of the path is determined by comprehensive weight. For the path selection of each individual, the individual with excellent path performance is determined by accumulating the path score. The excellent path has high adaptability and stability. Then, according to the path score, the path with the best priority is selected to enter the candidate path pool for path combination. Through the combination selection strategy, multiple high-priority paths are integrated to generate the overall optimal path. In this process, the compatibility and continuity between paths are considered, and a seamless action execution path is formed by adjusting the gene transmission between individuals. Finally, the optimal action strategy file is generated, which contains the optimized motion steps, path priority and resource allocation scheme, providing a complete and optimized solution for the execution and scheduling of digital human actions. Through the optimization of the path selection submodule, the accuracy and efficiency of the system in executing complex tasks are ensured, providing reliable technical support for the generation of digital human actions in a multi-task environment.
[0086] In this embodiment, the action scheduling and execution module includes:
[0087] The plan generation submodule plans the action execution sequence and resource allocation based on the optimal action strategy, builds a plan by judging the time requirements and resource constraints of the action strategy, and then generates a preliminary plan by analyzing the execution time window and resource availability;
[0088] The resource allocation submodule allocates resources based on the preliminary plan. It allocates resources by judging the resource requirements of each action and combining the resource usage priority of the execution queue to generate a resource scheduling plan.
[0089] The plan adjustment submodule makes plan adjustments based on the resource scheduling plan. It adjusts the action execution time by analyzing the dependencies and priorities in the action execution plan, optimizes resource allocation, corrects timing constraints, and obtains the action execution plan.
[0090] In the plan generation submodule, the action execution sequence and resource allocation are planned through the optimal action strategy. The input data format is an action sequence file generated by the optimal action strategy, which contains information such as action steps, execution time, and required resources. First, the time requirements and resource constraints of each action in the action strategy are analyzed to identify the critical path and bottleneck resources. The time requirement analysis includes calculating the execution time window of each action, that is, the feasible start time and end time of the action in the sequence; the resource constraint analysis evaluates the physical and computing resources (such as CPU, memory, bandwidth, etc.) required for the execution of each action. Then, the scheduling algorithm is used to refine the action strategy and build a plan. By permuting and combining the time window and resource constraints of the action, multiple feasible execution plans are calculated, and the optimal plan is selected. Finally, the execution time window and resource availability are analyzed to generate a preliminary plan. The plan file contains the specific execution time of each action, the required resources, and the dependencies between the actions. The preliminary plan provides a framework for subsequent resource allocation to ensure the operability and efficiency of the system during execution.
[0091] In the resource allocation submodule, resources are allocated through a preliminary plan. The input data format is a preliminary plan file, which lists the execution time and resource requirements of all actions. First, the resource usage priority is analyzed according to the resource requirements of each action to identify resource conflicts and competition. By establishing a resource requirement matrix, the resource occupancy of each action is quantified and arranged. Then, the resource scheduling algorithm is used to allocate available resources in combination with the resource usage priority in the execution queue. The resource scheduling algorithm optimizes the resource allocation scheme through methods such as linear programming or dynamic programming to maximize resource utilization and reduce resource waste and waiting time. For actions with higher priority, the required resources are allocated in advance to ensure that they are executed as planned. For actions with resource conflicts, their execution time or resource requirements are adjusted to avoid deadlock or resource preemption problems. Finally, a resource scheduling plan file is generated, which lists the resource allocation and scheduling order of each action in detail, providing a basis for subsequent plan adjustments and ensuring the effective utilization and reasonable allocation of system resources during execution.
[0092] In the plan adjustment submodule, the plan is adjusted through the resource scheduling plan. The input data format is the resource scheduling plan file, which contains the resource allocation and execution order of the actions. First, by analyzing the dependencies in the action execution plan, the order constraints between the actions are identified. Dependency analysis includes checking which actions must be executed after other actions are completed, thereby building a dependency graph for action execution. Next, the priority in the execution plan is analyzed, and the execution order of the actions is optimized through the task scheduling algorithm. The task scheduling algorithm uses priority queues and dynamic adjustment strategies to schedule high-priority tasks in advance to ensure that key actions are completed in time. For low-priority actions, their execution will be delayed when resources are limited to ensure the rationality and efficiency of the overall plan. On this basis, the execution time of the action is adjusted, the reallocation of resources is optimized, and by correcting the timing constraints, all actions are ensured to proceed as planned, eliminating potential conflicts and delays. Finally, the action execution plan file is obtained, which contains the optimized execution schedule, resource scheduling details, and action priority information, providing a precise execution guide for the system at runtime to ensure that all actions proceed smoothly in a complex environment.
[0093] A method for generating digital human motions, which is performed based on the digital human motion generation system, comprises the following steps:
[0094] S1, based on the input action data, uses edge detection algorithm to extract the motion trajectory features of the action, including the key positions of the joints. At the same time, the change rate and trend characteristics of the joint angle, velocity and acceleration are identified through morphological transformation analysis. The extracted features are combined into feature vectors, input into the feature encoder, and the feature vector representation is generated through the feature encoding algorithm;
[0095] S2, based on the feature vector representation, maps the input feature vector to the parameters of multiple candidate architectures through the hypernetwork mapping algorithm, and uses the dynamic adjustment algorithm to generate the main model structure parameters, including the number of layers, number of neurons, and activation function type of the main model, and then generates dynamic structure parameters, and finally generates network parameter configuration through parameter sharing and adjustment technology;
[0096] S3, based on the network parameter configuration, applies the network parameter configuration in the main model to generate an initial action sequence, generates the initial action sequence through a linear combination algorithm and a nonlinear activation function, then compares the difference between the generated action sequence and the input feature, calculates the adjustment value and adjusts the generated action sequence through a difference minimization algorithm, and finally generates the action sequence path by comparing the action sequence paths between multiple modules and selecting the path with the smallest difference;
[0097] S4, based on the action sequence path, encodes the action strategy into a gene sequence, uses a genetic algorithm to randomly initialize and generate a population, calculates the fitness value of each individual, generates adaptive individuals through crossover and mutation operations, and selects individuals that enter the next generation through fitness value sorting to generate the optimal individual, and finally initializes through pheromone, applies the ant colony algorithm to calculate the action path selection probability, selects the optimal path, and generates the optimal action strategy;
[0098] S5, based on the optimal action strategy, plans the timing and resource allocation of action execution, generates an execution plan through the time requirement judgment algorithm and resource consumption, integrates the dependencies and priorities between action strategies, uses the execution queue construction technology to sort the action execution order, determines the action execution time and resource allocation, and generates an action execution plan by calculating the dependencies and timing constraints of action execution;
[0099] S6, executes actions based on the action execution plan, collects execution data and uses feedback control algorithms to correct deviations in action execution, and continuously adjusts execution strategies through optimized feedback loops to generate optimized execution results.
[0100] First, by introducing edge detection algorithms and morphological transformation analysis, this method can accurately extract the motion trajectory characteristics of the action and the key positions of the joints. Combined with the change rate and trend characteristics of the joint angle, velocity and acceleration, the generated feature vector representation is more comprehensive and detailed. This multi-dimensional feature extraction and encoding helps to improve the accuracy and precision of action recognition, making the subsequent action generation more in line with actual needs.
[0101] Secondly, the hypernetwork mapping algorithm and dynamic adjustment algorithm are used to optimize the main model structure parameters. The number of layers, number of neurons and activation function type of the model can be adaptively adjusted according to the input features to generate dynamic structure parameters. Compared with the traditional fixed structure model, this method has stronger flexibility and adaptability and can handle more complex action generation tasks. At the same time, the application of parameter sharing and adjustment technology further improves the efficiency and accuracy of the model and reduces the computational cost.
[0102] In addition, this method generates the initial action sequence through a linear combination algorithm and a nonlinear activation function, and uses a difference minimization algorithm to adjust the action sequence, which can effectively reduce the difference between the generated action and the input features, and improve the accuracy and coherence of action generation. By comparing the action sequence paths between multiple modules and selecting the path with the smallest difference, it helps to improve the overall coordination and fluency of action generation.
[0103] In the optimization of action strategies, by encoding action strategies into gene sequences and applying genetic algorithms and ant colony algorithms, the optimal path can be selected from multiple candidate paths. This combined optimization method not only improves the effectiveness and efficiency of action strategies, but also ensures the balance between diversity and adaptability of the generated action strategies through pheromone initialization and calculation of path selection probabilities.
[0104] Finally, in the action execution phase, this method generates an execution plan through time requirement judgment algorithm and resource consumption, integrating the dependencies and priorities between action strategies. By using the execution queue construction technology, the action execution order can be effectively sorted, time and resource allocation can be optimized, and the overall efficiency and effect of action execution can be improved. By collecting execution data and optimizing feedback loops, deviations in action execution can be corrected in a timely manner to ensure the accuracy and reliability of execution results.
[0105] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.
Claims
1. A digital human motion generation system, characterized in that: The system comprises: The feature extraction and encoding module extracts task-related features from the input action data, including the motion trajectory, joint angle, velocity and acceleration of the action. It calculates the key positions of the joints by performing edge detection on the motion trajectory, identifies and calculates the rate of change and trend characteristics by performing morphological transformation analysis on the angle and velocity, combines the features into feature vectors, inputs them into the feature encoder, and encodes the feature vectors to generate feature vector representations. The hypernetwork optimization module uses the feature vector representation as input, generates the main model structure parameters by mapping the input feature vector with the parameters of multiple candidate architectures in the hypernetwork, dynamically adjusts the main model structure, calculates the number of main model layers and neurons, compares the performance of the architecture, selects the optimal network architecture, generates dynamic structure parameters, adapts the activation function between the input features and the network layer, selects the activation function, generates the activation parameters, adjusts the network parameters according to the characteristics of the task and the complexity of the data, generates the structure parameters and the network architecture, and generates the network parameter configuration by sharing and adjusting the parameters generated by the hypernetwork; The action generation and adjustment module applies the network parameter configuration in the main model to generate an action sequence. It generates an initial action sequence by calculating the linear combination of the input features and the network parameters and applying a nonlinear activation function. It calculates the adjustment value by comparing the difference between the generated action sequence and the input features and adjusts the generated action sequence. It compares the action sequence paths, selects the path with the minimum difference, and combines the action strategies to obtain the action sequence path. The bio-inspired optimization module applies the action sequence path, encodes the action strategy into a gene sequence, generates a population with random initialization, calculates the fitness value of each individual, generates adaptive individuals by performing crossover and mutation operations on the gene sequence, sorts the fitness values, selects adaptive individuals to enter the next generation, generates the optimal individual, calculates the action path selection probability through pheromone initialization, selects the optimal path, and generates the optimal action strategy; The action scheduling and execution module uses the optimal action strategy to plan the timing and resource allocation of action execution. It generates an execution plan by judging the time requirements and resource consumption of the action strategy, integrates the dependencies and priorities between action strategies, builds an action execution queue, determines the action execution time and resource allocation by sorting the action execution order, and adjusts the action execution plan by calculating the dependencies and timing constraints of the action execution to obtain the action execution plan.
2. A digital human motion generation system according to claim 1, characterized in that: The feature vector representation includes trajectory information, joint angle data, velocity change rate and acceleration characteristics, the network parameter configuration includes neuron connection weights, hierarchy configuration and activation function type, the action sequence path includes joint motion trajectory, posture adjustment sequence and continuous action chain, the optimal action strategy includes optimized motion steps, path priority and resource allocation plan, and the action execution plan includes timing scheduling, resource allocation and execution priority.
3. A digital human motion generation system according to claim 1, characterized in that: The feature extraction and encoding module includes: The trajectory processing submodule analyzes the motion trajectory data based on the input action data, obtains the three-dimensional coordinate information from the sensor data, performs trajectory analysis, processes the trajectory in segments, calculates the key positions of the joints through edge detection, identifies the key turning points in the trajectory change, extracts the spatial coordinate information, and generates the trajectory position features; The angle analysis submodule analyzes the joint angle and speed based on the trajectory position characteristics. It calculates the rotation angle of the joint from the position change, calculates the speed change rate based on the time difference, identifies the angle change trend through morphological transformation processing, and combines the angle change and speed change rate to form the angular velocity feature data describing the joint movement. The feature encoding submodule performs feature vector combination processing based on the angular velocity feature data. It vectorizes the angular velocity, position and acceleration features, integrates the features into feature vectors, maintains data consistency through vector normalization, optimizes the data structure by rearranging feature components, encodes them into feature vector matrices, and generates feature vector representations.
4. A digital human motion generation system according to claim 1, characterized in that: The super network optimization module includes: The structural parameter generation submodule performs parameter mapping between the input feature vector and the hypernetwork architecture based on the feature vector representation. By judging the response effects of various architectures to the feature vector, the number of main model layers and neuron configuration are generated, and parameter combination tests are performed. By aggregating the mapping results, the configuration with the best response effect is extracted to generate an architecture parameter mapping table. The dynamic architecture selection submodule performs performance analysis of candidate network architectures based on the architecture parameter mapping table. By calculating the efficiency and accuracy scores of multiple architectures under the target task, the architecture with the best score is selected as the main model infrastructure. The optimal network layer configuration is determined by analyzing the inter-layer connection relationship in the architecture, and the network layer structure parameters are generated. The parameter configuration submodule performs parameter optimization based on the network layer structure parameters. It determines the activation function by analyzing the neuron activation mode of each layer, selects the weight setting by comparing the performance of the weight configuration, applies the parameter sharing strategy to improve the computing efficiency, obtains the optimal combination through optimization and adjustment, and generates the network parameter configuration.
5. The digital human motion generation system according to claim 1, characterized in that: The action generation and adjustment module includes: The sequence generation submodule generates action sequences by applying parameters in the main model based on the network parameter configuration, generates neural activation values by calculating the linear combination of input features and network parameters, obtains the initial action sequence by applying activation operations, and generates initial sequence difference data by comparing the initial action sequence with the target action features; The sequence adjustment submodule adjusts the action sequence based on the initial sequence difference data. It adjusts the elements of the initial action sequence by calculating the deviation value of each action element in the sequence, and applies the adjustment strategy to correct the deviation in the sequence to obtain the corrected action sequence and generate the adjusted action sequence. The path optimization submodule performs sequence path optimization based on the adjusted action sequence. It judges the path smoothness and fluency by comparing the action sequence paths generated by the module, selects the path with the lowest difference as the final output, and selects the optimal path combination through the path integration strategy to obtain the action sequence path.
6. A digital human motion generation system according to claim 1, characterized in that: The bio-inspired optimization module includes: The gene encoding submodule encodes the action strategy based on the action sequence path, decomposes the action strategy into gene sequence fragments, initializes the gene library for population generation, and forms a diversified initial population by combining gene fragments to generate the encoding gene library; The fitness value sorting submodule calculates the fitness value based on the coding gene library, calculates the matching value by judging the performance of each individual's gene sequence in the task, generates an individual fitness value list, sorts it by fitness value, and selects the individual with the best matching degree for gene transfer to generate a sequence of adapted individuals; The path selection submodule selects the optimal path based on the adaptation individual sequence. It calculates the probability of individual action path selection, judges the performance of multiple paths, selects the path with the best priority, and generates the optimal path and the optimal action strategy by combining the selection strategies.
7. A digital human motion generation system according to claim 1, characterized in that: The action scheduling and execution module includes: The plan generation submodule plans the action execution sequence and resource allocation based on the optimal action strategy, builds a plan by judging the time requirements and resource constraints of the action strategy, and then generates a preliminary plan by analyzing the execution time window and resource availability; The resource allocation submodule allocates resources based on the preliminary plan. It allocates resources by judging the resource requirements of each action and combining the resource usage priority of the execution queue to generate a resource scheduling plan. The plan adjustment submodule makes plan adjustments based on the resource scheduling plan. It adjusts the action execution time by analyzing the dependencies and priorities in the action execution plan, optimizes resource allocation, corrects timing constraints, and obtains the action execution plan.
8. A method for generating digital human motions, characterized in that: According to any one of claims 1 to 7, the digital human motion generation system is implemented, comprising the following steps: Based on the input action data, the edge detection algorithm is used to extract the motion trajectory features of the action, including the key positions of the joints. At the same time, the change rate and trend characteristics of the joint angle, velocity and acceleration are identified through morphological transformation analysis. The extracted features are combined into feature vectors, input into the feature encoder, and the feature vector representation is generated through the feature encoding algorithm; Based on the feature vector representation, the input feature vector is mapped to the parameters of multiple candidate architectures through the hypernetwork mapping algorithm, and the main model structure parameters are generated by the dynamic adjustment algorithm, including the number of layers, number of neurons, and activation function type of the main model, and then the dynamic structure parameters are generated. Finally, the network parameter configuration is generated through parameter sharing and adjustment technology; Based on the network parameter configuration, the network parameter configuration is applied in the main model to generate an initial action sequence, the initial action sequence is generated by a linear combination algorithm and a nonlinear activation function, and then the difference between the generated action sequence and the input feature is compared, the adjustment value is calculated and the generated action sequence is adjusted by a difference minimization algorithm, and finally the action sequence path is generated by comparing the action sequence paths and selecting the path with the smallest difference; Based on the action sequence path, the action strategy is encoded into a gene sequence, and the genetic algorithm is used to randomly initialize the population. The fitness value of each individual is calculated, and the adaptive individuals are generated through crossover and mutation operations. The individuals entering the next generation are selected by fitness value sorting to generate the optimal individuals. Finally, the pheromone is initialized, and the ant colony algorithm is used to calculate the action path selection probability, select the optimal path, and generate the optimal action strategy. Based on the optimal action strategy, plan the timing and resource allocation of action execution, generate an execution plan through time requirement judgment algorithm and resource consumption, integrate the dependencies and priorities between action strategies, use execution queue construction technology to sort the action execution order, determine the action execution time and resource allocation, and generate an action execution plan by calculating the dependencies and timing constraints of action execution; Execute actions based on the action execution plan, collect execution data and use feedback control algorithms to correct deviations in action execution, and continuously adjust execution strategies through optimized feedback loops to generate optimized execution results.
