Method for realizing multi-attitude simulation based on AI behavior model
By extracting behavioral features through a pre-trained AI behavior model, constructing a dynamic behavior map, and generating various simulated postures, the problem of low efficiency and poor quality in posture simulation in existing technologies is solved, achieving efficient, diverse, and logically sound posture simulation effects.
Patent Information
- Application Number
- CN202510997734.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-19
- Publication Date
- 2025-11-11
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies struggle to efficiently extract key behavioral features when simulating various poses based on AI behavior models. This results in insufficient accuracy of the generated behavioral feature vectors, discrepancies between the simulated poses and the actual behavioral logic of the target object, and a lack of effective logic optimization and diversity enhancement mechanisms, leading to low quality pose sets.
Behavioral features are extracted by pre-trained AI behavior models to construct dynamic behavior maps. Various simulated postures are generated using posture simulation algorithms. Through logical matching degree verification and smoothness optimization, a set of simulated postures that conform to preset behavioral logic is output.
It significantly improves the efficiency and quality of attitude simulation, and the generated simulated attitudes are logical, have good coherence and rationality, can meet diverse attitude simulation needs, and realize the intuitive presentation of attitude simulation results.
Smart Images
Figure CN120930467A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for simulating various postures based on AI behavior models. Background Technology
[0002] In the field of simulating various postures based on AI behavior models, existing technologies often struggle to efficiently extract key behavioral features when processing basic behavioral data. This results in insufficient accuracy of the generated behavioral feature vectors, which in turn affects the effectiveness of subsequent posture simulations. Furthermore, the description of dynamic behavioral logic relationships is not precise enough, causing discrepancies between the simulated postures and the actual behavioral logic of the target object, thus failing to meet the diverse and logically consistent posture simulation requirements.
[0003] Existing technologies lack effective logic optimization and diversity enhancement mechanisms in the pose generation process. The generated simulated poses are not only limited in number but also lack coherence and reasonableness. Furthermore, the verification methods for simulated poses are inadequate, making it difficult to accurately select poses that conform to preset behavioral logic. This results in a low-quality final pose set, which cannot efficiently support the high-quality pose simulation requirements of relevant application scenarios. Summary of the Invention
[0004] This invention provides a method for simulating multiple postures based on AI behavior models, the main purpose of which is to solve the problem of low time efficiency in simulating multiple postures based on AI behavior models.
[0005] To achieve the above objectives, the present invention provides a method for simulating various postures based on an AI behavior model, comprising: S1. Obtain basic behavioral data of the target object, wherein the basic behavioral data includes action sequences and corresponding environmental parameters; S2. Input the basic behavioral data into the pre-trained AI behavioral model, extract behavioral features through the AI behavioral model, and generate a behavioral feature vector; S3. Construct a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object; S4. Based on the dynamic behavior map, generate a variety of simulated postures through a posture simulation algorithm, wherein the posture simulation algorithm includes a behavior logic optimization module and a posture generation module; S5. Verify the generated simulated postures and output a set of simulated postures that conform to the preset behavioral logic.
[0006] In a preferred embodiment, the acquisition of basic behavioral data of the target object, the basic behavioral data including action sequences and corresponding environmental parameters, includes: The raw motion data of the target object is collected through sensors; The original motion data is denoised and normalized to generate a standardized motion sequence; Environmental parameters corresponding to the standardized action sequence are extracted to form the basic behavioral data.
[0007] In a preferred embodiment, the step of inputting the basic behavioral data into a pre-trained AI behavioral model, extracting behavioral features through the AI behavioral model, and generating a behavioral feature vector includes: Construct a training set containing multiple sets of behavioral data samples; Design a behavior feature extraction network, which includes a temporal convolutional layer and an attention mechanism layer; The behavior feature extraction network is trained using the training set to generate the pre-trained AI behavior model.
[0008] In a preferred embodiment, the step of constructing a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object, includes: Map the behavioral feature vectors to graph nodes; Calculate the connection weights between nodes based on behavioral logic relationships; The dynamic behavior graph is generated by adjusting the connection weights using a graph optimization algorithm.
[0009] In a preferred embodiment, the formula for the map optimization algorithm is: In the formula, W ij S represents the connection weights between nodes. ij Let represent the behavioral similarity between node i and node j, and k be the adjustment coefficient.
[0010] In a preferred embodiment, based on the dynamic behavior map, multiple simulated postures are generated through a posture simulation algorithm. The posture simulation algorithm includes a behavior logic optimization module and a posture generation module, comprising: The dynamic behavior graph is logically corrected through the behavior logic optimization module. Based on the corrected map, the attitude generation module is used to generate an initial simulated attitude. The initial simulated pose is subjected to diversity enhancement processing to generate the various simulated poses. In a preferred embodiment, the formula for the diversity enhancement process is: P d =P0 + αΔP; In the formula, P dThe enhanced attitude parameters are defined as follows: P0 is the initial attitude parameter, α is the disturbance intensity coefficient, and ΔP is the preset attitude disturbance vector.
[0011] In a preferred embodiment, the step of verifying the generated simulated postures and outputting a set of simulated postures that conform to preset behavioral logic includes: Calculate the logical matching degree between the simulated posture and the dynamic behavior map; Filter simulated poses with a matching degree higher than a threshold; The smoothness of the selected simulated postures is optimized to generate the set of simulated postures.
[0012] In a preferred embodiment, the formula for calculating the logical matching degree is: In the formula, M represents the logical matching degree, and F... i G is the feature vector that simulates the attitude. i is the feature vector of the corresponding node in the dynamic behavior graph, where n is the feature dimension.
[0013] In a preferred embodiment, after verifying the generated simulated poses and outputting a set of simulated poses that conform to preset behavioral logic, the process includes: The simulated posture set is input into the visualization engine to generate behavioral simulation animation.
[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The method for simulating multiple poses based on AI behavior models provided by this invention efficiently extracts behavioral features through a pre-trained AI behavior model. Combined with the synergistic effect of dynamic behavior graph construction and pose simulation algorithms, it can quickly generate multiple simulated poses that conform to behavioral logic, significantly improving the efficiency of pose simulation. Simultaneously, by leveraging the temporal convolutional layers and attention mechanism layers in the behavior feature extraction network, behavioral features can be accurately captured, providing a high-quality foundation for subsequent pose generation and further ensuring the efficiency of the simulation process.
[0015] 2. This method enriches the diversity of poses through diversity enhancement processing when generating simulated poses. Furthermore, it verifies logical matching and optimizes smoothness to ensure that the output set of simulated poses not only conforms to the preset behavioral logic but also possesses good coherence and rationality. In addition, converting the simulated pose set into behavioral simulation animations provides an intuitive presentation of the pose simulation results, enhancing the method's practicality and application value, and providing efficient and reliable technical support for pose simulation needs in related fields. Attached Figure Description
[0016] Figure 1This is a flowchart illustrating a method for simulating multiple postures based on an AI behavior model, as provided in an embodiment of the present invention. Figure 2 A schematic diagram of the data-driven posture simulation full-chain process of a method for realizing multiple posture simulation based on an AI behavior model provided in an embodiment of the present invention; Figure 3 The following is a logical progression flowchart of a method for simulating multiple postures based on an AI behavior model, provided as an embodiment of the present invention. Detailed Implementation
[0017] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0018] This application provides a method for simulating multiple postures based on an AI behavior model. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method for simulating multiple postures based on an AI behavior model can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.
[0019] Reference Figure 1 The diagram shown is a flowchart illustrating a method for simulating multiple postures based on an AI behavior model, according to an embodiment of the present invention. In this embodiment, the method for simulating multiple postures based on an AI behavior model includes: S1. Obtain basic behavioral data of the target object, wherein the basic behavioral data includes action sequences and corresponding environmental parameters; S2. Input the basic behavioral data into the pre-trained AI behavioral model, extract behavioral features through the AI behavioral model, and generate a behavioral feature vector; S3. Construct a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object; S4. Based on the dynamic behavior map, generate a variety of simulated postures through a posture simulation algorithm, wherein the posture simulation algorithm includes a behavior logic optimization module and a posture generation module; S5. Verify the generated simulated postures and output a set of simulated postures that conform to the preset behavioral logic.
[0020] S1. Obtain basic behavioral data of the target object, wherein the basic behavioral data includes action sequences and corresponding environmental parameters; In this embodiment of the invention, the acquisition of basic behavioral data of the target object, the basic behavioral data including action sequences and corresponding environmental parameters, includes: The raw motion data of the target object is collected through sensors; The original motion data is denoised and normalized to generate a standardized motion sequence; Environmental parameters corresponding to the standardized action sequence are extracted to form the basic behavioral data.
[0021] Specifically, the target object was determined to be the human body, and appropriate sensors were selected to collect raw motion data. Seventeen inertial measurement unit (IMU) sensors were selected and fixed to the head, neck, chest, waist, left and right shoulders, left and right upper arms, left and right forearms, left and right hands, left and right thighs, left and right calves, and left and right feet of the human body. Acceleration, angular velocity, and magnetic field data of each part of the human body were collected in real time at a sampling frequency of 100Hz. These data together constitute the raw motion data.
[0022] Furthermore, the raw motion data is denoised. A sliding window midpoint filtering method is used, with a window size of 5. That is, for each data point, two data points before and after it are taken, for a total of 5 data points. These 5 data points are sorted by size, and the median value is taken as the denoised value of the data point. This eliminates the impulse noise caused by slight vibrations or electronic interference during sensor acquisition.
[0023] Furthermore, the denoised motion data is normalized. For acceleration data, the value range is from -16g to 16g (g is the acceleration due to gravity). Subtracting -16g from each acceleration data value and then dividing by (16g - (-16g)) yields standardized acceleration data between 0 and 1. For angular velocity data, the value range is from -2000 degrees / second to 2000 degrees / second. Subtracting -2000 degrees / second from each angular velocity data value and then dividing by (2000 degrees / second - (-2000 degrees / second)) yields standardized angular velocity data between 0 and 1. Finally, a standardized motion sequence containing timestamps and standardized acceleration and angular velocities for each part is formed.
[0024] Furthermore, environmental parameters corresponding to the standardized action sequence are extracted. While collecting human motion data, light sensors, temperature sensors, and humidity sensors are deployed. The light sensor collects ambient light intensity, the temperature sensor collects ambient temperature, and the humidity sensor collects ambient humidity. A timestamp corresponding to each environmental parameter is recorded. Environmental parameters with the same timestamp as those in the standardized action sequence are extracted and combined with the standardized action sequence to form basic behavioral data.
[0025] In summary, by collecting raw motion data from multiple sensors, the movement state of various parts of the human body can be comprehensively captured, providing rich raw information for subsequent posture simulation. Noise reduction effectively eliminates interference data, ensuring the stability of motion data. Normalization unifies sensor data of different magnitudes to the same range, facilitating unified processing and analysis of data by subsequent models. Extracting corresponding environmental parameters ensures that the basic behavioral data not only includes motion information but also incorporates environmental influencing factors, laying the foundation for posture simulation that is more closely aligned with real-world scenarios.
[0026] S2. Input the basic behavioral data into the pre-trained AI behavioral model, extract behavioral features through the AI behavioral model, and generate a behavioral feature vector; In this embodiment of the invention, the step of inputting the basic behavioral data into a pre-trained AI behavioral model, extracting behavioral features through the AI behavioral model, and generating a behavioral feature vector includes: Construct a training set containing multiple sets of behavioral data samples; Design a behavior feature extraction network, which includes a temporal convolutional layer and an attention mechanism layer; The behavior feature extraction network is trained using the training set to generate the pre-trained AI behavior model.
[0027] Specifically, a training set was constructed. Basic behavioral data from 1000 volunteers of different ages, genders, heights, and weights were collected in various scenarios, including indoors, outdoors, stairs, and flat ground. The data covered 20 common behaviors, such as walking, running, climbing stairs, sitting, squatting, and jumping. Each behavior contained 100 sets of valid data samples. Each set of samples contained a 30-second standardized action sequence and corresponding environmental parameters. After numbering and labeling these samples with behavior categories, they were divided into a training set and a validation set in an 8:2 ratio.
[0028] Furthermore, a behavior feature extraction network is designed. The temporal convolutional layer contains three convolutional blocks. Each convolutional block consists of a one-dimensional convolutional layer with a kernel size of 3 and an output channel number of 64, a batch normalization layer, and a ReLU activation function, used to extract local temporal features of the action sequence. The attention mechanism layer adopts a multi-head self-attention mechanism with eight attention heads. After converting the feature map output by the temporal convolutional layer into a query, key, and value matrix, it calculates the association weights between the features of each time step and the features of other time steps, highlighting important action features. The input of the network is a feature matrix of the standardized action sequence in the basic behavior data (dimension 3000×51, where 3000 represents a time step of 30 seconds × 100Hz and 51 represents data from 17 sensors × 3 axes), and the output is a 256-dimensional feature vector.
[0029] Furthermore, a behavioral feature extraction network is trained to generate a pre-trained AI behavioral model. Using the training set data, the Adam optimizer is employed with an initial learning rate of 0.001. The learning rate is halved when the validation set loss does not decrease for five consecutive epochs. The cross-entropy loss function is used, with behavioral category labels as the supervision signal. Training is performed on an NVIDIA Tesla V100 graphics card for 50 epochs. After each epoch, the model performance is evaluated using the validation set, and the model with the highest validation set accuracy is saved as the pre-trained AI behavioral model. Basic behavioral data is input into this model, and the model's temporal convolutional layers and attention mechanism layers process the data sequentially, ultimately outputting a 256-dimensional behavioral feature vector, where each vector corresponds to a set of core features from the basic behavioral data.
[0030] In summary, constructing a large-scale and diverse training set enables the model to learn the common and unique features of different scenarios and behaviors, thereby improving the model's generalization ability. The combined design of temporal convolutional layers and attention mechanism layers captures both the local temporal features of actions and focuses on the features of key action moments, improving the accuracy of feature extraction. The pre-trained model generated through sufficient training can efficiently extract the core features of basic behavioral data and transform them into feature vectors, providing a structured feature representation for subsequent construction of behavioral maps.
[0031] S3. Construct a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object; In this embodiment of the invention, the step of constructing a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object, includes: Map the behavioral feature vectors to graph nodes; Calculate the connection weights between nodes based on behavioral logic relationships; The dynamic behavior graph is generated by adjusting the connection weights using a graph optimization algorithm.
[0032] The formula for the graph optimization algorithm is: In the formula, W ij S represents the connection weights between nodes. ij Let represent the behavioral similarity between node i and node j, and k be the adjustment coefficient.
[0033] Specifically, behavioral feature vectors are mapped to graph nodes. The behavioral feature vectors generated in step S2 are clustered using the K-means clustering algorithm. Twenty cluster centers are pre-defined based on the behavioral category, with each cluster center representing a typical behavioral state. Each behavioral feature vector is assigned to the cluster containing its nearest cluster center. Each cluster corresponds to a graph node, meaning each node represents a type of behavioral state with similar characteristics, such as the "walking on flat ground" node or the "climbing stairs" node. The node's attributes include the average value of all behavioral feature vectors within that cluster and the behavioral category label.
[0034] Furthermore, the behavioral similarity between nodes is calculated. For any two nodes i and j, the average behavioral feature vectors of node i and node j are taken, and the cosine similarity between the two vectors is calculated. That is, the dot product of the two vectors is divided by the product of the magnitudes of the two vectors. This gives the behavioral similarity between node i and node j. The value ranges from -1 to 1. The closer the value is to 1, the more similar the behavioral states represented by the two nodes are. For example, the behavioral similarity between the "walking on flat ground" node and the "jogging" node is 0.7, while the behavioral similarity between the "walking on flat ground" node and the "sitting" node is 0.2.
[0035] Furthermore, the initial connection weights are calculated based on behavioral similarity. With an adjustment coefficient of 2, the behavioral similarity between node i and node j is substituted into the graph optimization algorithm formula to calculate the initial connection weights between nodes. For example, when the behavioral similarity is 0.7, the initial connection weight is 1 divided by (1 plus the natural constant raised to the power of negative 2), approximately equal to 0.817; when the behavioral similarity is 0.2, the initial connection weight is approximately equal to 0.598.
[0036] Furthermore, connection weights are adjusted using a graph optimization algorithm. Combining prior knowledge of behavioral logic—such as the fact that "sitting" behavior typically transitions to "standing" rather than directly to "running"—connections that are illogical, such as the connection weight between "sitting" and "running" nodes being too high, are manually adjusted to 0.1. For logically strong connections, such as the connection weight between "standing" and "walking on flat ground" nodes, if the calculated value is 0.7, it remains unchanged. This process ultimately generates a dynamic behavioral graph where the connection weights between nodes accurately reflect the logical transitions between behavioral states.
[0037] Furthermore, the entire process of data-driven attitude simulation, such as... Figure 2 As shown.
[0038] In summary, this step transforms behavioral features from vector form to graph nodes, concretizing abstract feature vectors into basic units that can participate in graph construction. This lays a structured foundation for subsequently describing the behavioral logic relationships of target objects, enabling behavioral features to participate in the construction of behavioral logic relationships in the form of nodes. In summary, by calculating the connection weights between nodes, the strength of the association between different behavioral features can be quantitatively described, clearly reflecting the tightness of the target object's behavioral logic. This allows the dynamic behavioral graph to not only reflect the association between behaviors but also the strength of the association, thus enhancing the accuracy of the graph's description of behavioral logical relationships.
[0039] In summary, by adjusting connection weights using graph optimization algorithms, the relationships between nodes can be optimized, unreasonable weight allocations can be eliminated, and the dynamic behavior graph can more accurately reflect the real behavioral logic of the target object, thereby improving the reliability and effectiveness of the graph and providing a better logical basis for subsequent graph-based attitude simulation.
[0040] In summary, mapping behavioral feature vectors to graph nodes enables a structured representation of behavioral features, transforming abstract feature vectors into intuitively understandable behavioral state nodes. Calculating connection weights based on behavioral similarity and optimization algorithms quantifies the degree of correlation between behavioral states. Adjusting weights using prior knowledge ensures that the dynamic behavioral graph accurately describes the behavioral logic relationships of the target object, providing a clear framework for subsequent posture simulation.
[0041] S4. Based on the dynamic behavior map, generate a variety of simulated postures through a posture simulation algorithm, wherein the posture simulation algorithm includes a behavior logic optimization module and a posture generation module; In this embodiment of the invention, based on the dynamic behavior map, multiple simulated postures are generated through a posture simulation algorithm. The posture simulation algorithm includes a behavior logic optimization module and a posture generation module, comprising: The dynamic behavior graph is logically corrected through the behavior logic optimization module. Based on the corrected map, the attitude generation module is used to generate an initial simulated attitude. The initial simulated posture is subjected to diversity enhancement processing to generate the various simulated postures.
[0042] The formula for the diversity enhancement process is: P d =P0 + αΔP; In the formula, P d The enhanced attitude parameters are defined as follows: P0 is the initial attitude parameter, α is the disturbance intensity coefficient, and ΔP is the preset attitude disturbance vector.
[0043] Specifically, the dynamic behavior graph is logically corrected through a behavior logic optimization module. This module has a built-in library of common behavior conversion rules. For example, the "lie down" behavior can only be converted from "sit" or "stand" behavior, and the conversion probability is low; the "go up or down stairs" behavior requires an environment parameter with stairs present. The module iterates through all connections in the dynamic behavior graph. When a connection is detected between the "lie down" node and the "run" node with a weight of 0.3, the weight of this connection is corrected to 0 according to the rule library. When a "go up stairs" node is detected with a weight of 0.5 in a connection without stairs, its weight is corrected to 0.2, ensuring that the behavior logic of the graph conforms to the actual situation.
[0044] Furthermore, based on the corrected graph, an initial simulated posture is generated using a posture generation module. The "walking on flat ground" node is selected as the starting node. The posture generation module employs a generative adversarial network (GAN). The generator receives the feature vector and random noise vector of this node and outputs the angle parameters of 17 joints of the human body (each joint angle ranges from -180 degrees to 180 degrees). The discriminator judges whether the generated joint angle parameters conform to the posture characteristics of "walking on flat ground." Through adversarial training between the generator and the discriminator, the generator can output a natural initial simulated posture, such as a walking posture with the left leg stepping forward, the right leg supporting the body, and the arms swinging naturally. The specific joint angle parameters are: hip flexion 30 degrees, knee flexion 45 degrees (left leg), hip extension 10 degrees, and knee extension 5 degrees (right leg), etc.
[0045] Furthermore, a preset posture perturbation vector is set. The vector contains angular perturbation values for 17 joints. The perturbation value for each joint is randomly generated within the range of -5 degrees to 5 degrees. For example, the perturbation value for the left hip joint is 2 degrees, and the perturbation value for the right knee joint is -3 degrees, etc., to ensure that the perturbation values are within the range of human physiological activity.
[0046] Furthermore, the perturbation intensity coefficient is set to 0.8. According to the diversity enhancement processing formula, each joint angle in the initial posture parameters is multiplied by 0.8 by the perturbation value of the corresponding joint to obtain the enhanced posture parameters. For example, the initial angle of the left hip joint is 30 degrees, and after enhancement, it is 30 degrees plus 0.8 multiplied by 2 degrees, which equals 31.6 degrees; the initial angle of the right knee joint is 5 degrees, and after enhancement, it is 5 degrees plus 0.8 multiplied by (-3 degrees), which equals 2.6 degrees.
[0047] Furthermore, the process of generating initial simulated postures and enhancing diversity is repeated. Different random noise vectors are generated each time to obtain different initial postures, resulting in a total of 50 sets of initial simulated postures. Each set of initial postures is enhanced three times with different preset posture perturbation vectors and perturbation intensity coefficients (0.6, 0.8, and 1.0), ultimately generating 150 diverse simulated postures, covering different walking posture variations in stride length, arm swing amplitude, and body forward lean angle.
[0048] Furthermore, the logically progressive process is as follows: Figure 3 As shown.
[0049] In summary, this step can correct logical contradictions, redundant associations, or inaccurate behavioral relationships that may exist in the dynamic behavior graph, ensuring that the behavioral logic described by the graph is more rigorous and reasonable, providing a reliable logical basis for subsequent attitude generation, and reducing invalid simulated attitude generation caused by graph logic errors.
[0050] In summary, relying on the corrected and accurate dynamic behavior map, the posture generation module can generate the initial simulated posture by following the correct behavioral logic, ensuring the rationality and accuracy of the initial posture in terms of behavioral logic, avoiding the generation of postures that do not conform to actual behavioral rules based on incorrect logic, and improving the basic quality of posture simulation.
[0051] In summary, by enhancing diversity, more diverse simulated poses can be generated based on the initial simulated poses, enriching the diversity of poses, meeting the needs for simulating multiple poses of the target object, avoiding the problem of uniform simulated poses, and improving the method's adaptability to different pose scenarios.
[0052] In summary, the behavior logic optimization module corrects the dynamic behavior graph, ensuring the logical rationality of subsequent pose generation and avoiding unrealistic pose transitions; the pose generation module uses a GAN network to generate initial simulated poses, producing natural and behaviorally consistent basic poses; and the diversity enhancement processing, through reasonable perturbation settings, increases the number and diversity of simulated poses while ensuring pose rationality, meeting the needs of various pose simulations.
[0053] S5. Verify the generated simulated postures and output a set of simulated postures that conform to the preset behavioral logic.
[0054] In this embodiment of the invention, the step of verifying the generated simulated postures and outputting a set of simulated postures that conform to preset behavioral logic includes: Calculate the logical matching degree between the simulated posture and the dynamic behavior map; Filter simulated poses with a matching degree higher than a threshold; The smoothness of the selected simulated postures is optimized to generate the set of simulated postures.
[0055] The formula for calculating the logical matching degree is: In the formula, M represents the logical matching degree, and F... i G is the feature vector that simulates the attitude. i is the feature vector of the corresponding node in the dynamic behavior graph, where n is the feature dimension.
[0056] After verifying the generated simulated poses and outputting a set of simulated poses that conform to the preset behavioral logic, the process includes: The simulated posture set is input into the visualization engine to generate behavioral simulation animation.
[0057] Specifically, feature vectors of simulated postures are extracted. The joint angle parameters of each simulated posture are converted into feature matrices in the same format as the basic behavior data (assuming a 100×51 matrix, corresponding to 1 second of action data), input into the pre-trained AI behavior model, and output a 256-dimensional simulated posture feature vector.
[0058] Furthermore, the corresponding nodes in the dynamic behavior graph are determined. For simulated postures related to "walking on flat ground", the corresponding node is the "walking on flat ground" node, and the 256-dimensional average feature vector of this node is obtained as the feature vector of the corresponding node in the dynamic behavior graph.
[0059] Further, the matching value for each feature dimension is calculated. For the first feature dimension, the product of the simulated pose feature vector value in that dimension and the corresponding node feature vector value in that dimension is calculated, and then divided by the product of the magnitudes of the two vectors in that dimension to obtain the matching value for the first dimension; the matching values for the first to the 256th feature dimensions are calculated in this way.
[0060] Further, the logical matching degree is calculated. The matching values of the 256 feature dimensions are added together and then divided by 256 to obtain the logical matching degree of the simulated pose.
[0061] For example, if the sum of 256 matching values is 205, then the logical matching degree is 205 divided by 256, which is approximately 0.801.
[0062] Furthermore, the logical matching threshold is set to 0.7. 150 simulated postures are traversed, and simulated postures with a logical matching degree greater than or equal to 0.7 are selected. It is assumed that 120 postures meet the condition and these postures have a strong logical correlation with the "walking on flat ground" node in terms of behavioral characteristics.
[0063] Furthermore, the smoothness of the selected simulated postures is optimized. Using cubic spline interpolation, for continuous simulated posture sequences (such as walking postures arranged chronologically), a smooth curve is constructed between the joint angle parameters of adjacent postures, ensuring that the rate of change of joint angles does not exceed 10 degrees per frame (assuming an animation frame rate of 30 frames per second). For example, if the left knee joint angle is 45 degrees in the previous frame and 55 degrees in the next frame, the optimized angles in the intermediate frames are 48 degrees, 51 degrees, and 54 degrees respectively, avoiding abrupt angle changes and ensuring a natural posture transition.
[0064] Furthermore, the 120 optimized simulated postures are clustered according to the similarity of behavioral features, and postures with high repetition are removed. Finally, 80 representative postures are retained to form a set of simulated postures that conform to the preset behavioral logic.
[0065] In summary, by calculating the logical matching degree and filtering, the output simulated postures are ensured to conform to the behavioral logic described by the dynamic behavior graph, and unreasonable postures are eliminated. Smoothness optimization makes the transition of the posture sequence more natural and smooth, improving the realism of the simulated postures. The final set of simulated postures not only meets the behavioral logic requirements, but also has diversity and representativeness, providing high-quality posture data for subsequent applications.
[0066] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.
[0067] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0068] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.
[0069] Finally, 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.
Claims
1. A method for simulating multiple postures based on an AI behavior model, characterized in that, The method includes: S1. Obtain basic behavioral data of the target object, wherein the basic behavioral data includes action sequences and corresponding environmental parameters; S2. Input the basic behavioral data into the pre-trained AI behavioral model, extract behavioral features through the AI behavioral model, and generate a behavioral feature vector; S3. Construct a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object; S4. Based on the dynamic behavior map, generate a variety of simulated postures through a posture simulation algorithm, wherein the posture simulation algorithm includes a behavior logic optimization module and a posture generation module; S5. Verify the generated simulated postures and output a set of simulated postures that conform to the preset behavioral logic.
2. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, The acquisition of basic behavioral data of the target object includes action sequences and corresponding environmental parameters, including: The raw motion data of the target object is collected through sensors; The original motion data is denoised and normalized to generate a standardized motion sequence; Environmental parameters corresponding to the standardized action sequence are extracted to form the basic behavioral data.
3. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, The step of inputting the basic behavioral data into a pre-trained AI behavioral model, extracting behavioral features through the AI behavioral model, and generating a behavioral feature vector includes: Construct a training set containing multiple sets of behavioral data samples; Design a behavior feature extraction network, which includes a temporal convolutional layer and an attention mechanism layer; The behavior feature extraction network is trained using the training set to generate the pre-trained AI behavior model.
4. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, The step of constructing a dynamic behavior graph based on the behavior feature vector, wherein the dynamic behavior graph is used to describe the behavioral logical relationships of the target object, includes: Map the behavioral feature vectors to graph nodes; Calculate the connection weights between nodes based on behavioral logic relationships; The dynamic behavior graph is generated by adjusting the connection weights using a graph optimization algorithm.
5. The method for simulating multiple postures based on an AI behavior model as described in claim 4, characterized in that, The formula for the graph optimization algorithm is: In the formula, W ij S represents the connection weights between nodes. ij Let represent the behavioral similarity between node i and node j, and k be the adjustment coefficient.
6. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, Based on the dynamic behavior map, a variety of simulated postures are generated through a posture simulation algorithm. The posture simulation algorithm includes a behavior logic optimization module and a posture generation module, comprising: The dynamic behavior graph is logically corrected through the behavior logic optimization module. Based on the corrected map, the attitude generation module is used to generate an initial simulated attitude. The initial simulated posture is subjected to diversity enhancement processing to generate the various simulated postures.
7. The method for simulating multiple postures based on an AI behavior model as described in claim 6, characterized in that, The formula for the diversity enhancement process is: P d =P0+αΔP; In the formula, P d The enhanced attitude parameters are defined as follows: P0 is the initial attitude parameter, α is the disturbance intensity coefficient, and ΔP is the preset attitude disturbance vector.
8. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, The step of verifying the generated simulated postures and outputting a set of simulated postures that conform to preset behavioral logic includes: Calculate the logical matching degree between the simulated posture and the dynamic behavior map; Filter simulated poses with a matching degree higher than a threshold; The smoothness of the selected simulated postures is optimized to generate the set of simulated postures.
9. The method for simulating multiple postures based on an AI behavior model as described in claim 8, characterized in that, The formula for calculating the logical matching degree is: In the formula, M represents the logical matching degree, and F... i G is the feature vector that simulates the attitude. i is the feature vector of the corresponding node in the dynamic behavior graph, where n is the feature dimension.
10. The method for simulating multiple postures based on an AI behavior model as described in claim 1, characterized in that, After verifying the generated simulated poses and outputting a set of simulated poses that conform to the preset behavioral logic, the process includes: The simulated posture set is input into the visualization engine to generate behavioral simulation animation.