AI-driven interactive AR content intelligent generation and distribution method and system

Abstract

The present invention provides an AI-driven interactive AR content intelligent generation and distribution method and system, which relates to the field of AI-driven technology. The method includes acquiring scene data to construct a three-dimensional semantic map, calculating the semantic entropy value of the scene area and constructing a scene semantic evolution tree; then matching candidate content with semantic change trends, dividing content groups and generating trigger conditions and update strategies; then collecting user interaction data to establish an interactive behavior model and calculate the interactive responsiveness of the content group; finally, based on the interactive responsiveness, the push priority is determined and a directed push path map is constructed to guide the activation order of the content group. Through a semantic entropy-driven content organization method and an interactive behavior-based intelligent distribution mechanism, the present invention achieves accurate push and smooth transition of augmented reality content, improving the user interaction experience.

Description

Technical Field

[0001] The present invention relates to AI-driven technology, and in particular to an AI-driven method and system for intelligently generating and distributing interactive AR content. Background Art

[0002] Augmented reality technology provides users with an immersive interactive experience by overlaying virtual content onto real-world scenes. With the development of mobile devices and edge computing, AI-based AR content generation and distribution technologies have gained widespread application. Current AR systems primarily achieve the overlay display of virtual content through scene understanding and content matching, and optimize content push strategies based on user interaction data to enhance the user experience.

[0003] However, the existing technology still has the following shortcomings: First, traditional scene understanding methods lack in-depth analysis of the semantic evolution laws of scenes, and cannot accurately grasp the changing trends of scene semantics, resulting in insufficient content matching accuracy; Second, existing content organization methods usually adopt static partitioning strategies, which do not fully consider the impact of user interaction behavior on content relevance and are difficult to adapt to dynamic interaction scenarios; Third, current content distribution mechanisms often adopt simple priority strategies and lack systematic modeling of semantic associations between content groups, which easily leads to discontinuous content switching.

[0004] The present invention aims to solve the above technical problems and proposes a scene understanding method based on semantic entropy and an interaction-driven content distribution mechanism. By constructing a scene semantic evolution tree, it can accurately grasp the scene change trend, optimize the content organization strategy based on the interactive behavior model, and combine the semantically associated directed push path to achieve a smooth transition of content, thereby improving the interactive experience of the AR system. Summary of the Invention

[0005] The embodiments of the present invention provide an AI-driven interactive AR content intelligent generation and distribution method and system, which can solve the problems in the existing technology.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides AI-driven interactive AR content intelligent generation and distribution methods, including:

[0008] Acquire scene data collected by a user terminal and construct a three-dimensional semantic map of the target scene based on the scene data; calculate semantic entropy values ​​for each area in the scene according to the three-dimensional semantic map, construct a scene semantic evolution tree based on the semantic entropy values, and obtain the semantic change trend of the target scene by updating the node states and conversion weights in the scene semantic evolution tree in real time;

[0009] Calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the relevant candidate content into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map;

[0010] Collecting interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establishing an interaction behavior model based on the interaction data, and calculating the interaction responsiveness of each content group using the interaction behavior model;

[0011] The content groups are graded according to the interactive responsiveness, and push priorities are determined based on the grading results. A directed push path graph is constructed in combination with the semantic associations between the content groups. The directed push path graph is used to guide the activation order of the content groups, and a smooth transition of the augmented reality experience is achieved through the orderly activation of the content groups.

[0012] Calculating semantic entropy values ​​for each region in the scene based on the three-dimensional semantic map, constructing a scene semantic evolution tree based on the semantic entropy values, and obtaining semantic change trends of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time include:

[0013] Dividing the scene space of the three-dimensional semantic map into voxel grids, calculating the probability distribution of occurrence of each semantic category data in each voxel grid, and calculating the semantic entropy value of each voxel grid based on the probability distribution, wherein the semantic entropy value is used to represent the semantic complexity of the voxel grid;

[0014] Obtaining a neighborhood voxel set of each voxel grid, calculating a spatial distance between each voxel in the neighborhood voxel set and a target voxel, determining a spatial association weight based on the spatial distance, and performing a weighted calculation on the semantic entropy value according to the spatial association weight to obtain a weighted semantic entropy value that takes neighborhood influence into account;

[0015] Constructing a scene semantic evolution tree based on the weighted semantic entropy value, calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree, and using the semantic state change amount, the state transition probability, and the state transition time interval as attributes of the connecting edges between the nodes in the scene semantic evolution tree;

[0016] The similarity between nodes in the scene semantic evolution tree is calculated. The similarity is obtained by a weighted combination of the cosine similarity of the semantic state vector, the Euclidean distance of the entropy distribution matrix, and the time decay function. The scene semantic evolution tree is dynamically updated based on the similarity to obtain the semantic change trend of the scene.

[0017] Constructing a scene semantic evolution tree based on the weighted semantic entropy value, and calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree includes:

[0018] Extracting a set of semantic attributes of the scene based on the weighted semantic entropy value, the set of semantic attributes including object category, spatial position, object attributes, and relationship between objects, performing threshold segmentation on the weighted semantic entropy value and extracting semantic regions, obtaining a semantic feature vector for each semantic region, and normalizing the semantic feature vector to obtain an attribute strength value corresponding to each semantic attribute;

[0019] Constructing a node representation of a scene semantic evolution tree based on the semantic attribute set and the attribute strength values, wherein each node includes a state vector and an entropy value distribution matrix, wherein the state vector is constructed by organizing the attribute strength values ​​into a vector form according to a preset attribute sorting rule, and the entropy value distribution matrix is ​​constructed by organizing the weighted semantic entropy values ​​into a matrix form according to spatial position relationships;

[0020] Identifying adjacent node pairs in the semantic evolution tree, calculating the difference between each component of the state vector between the nodes for each adjacent node pair, dividing the difference by the maximum value of the corresponding component and introducing a smoothing factor to obtain a normalized semantic state change;

[0021] Calculating the bi-norm of the semantic state change to obtain a state distance, negating the state distance and performing an exponential operation to obtain an initial probability value, and normalizing the initial probability values ​​of all adjacent nodes of each node to obtain a state transition basic probability;

[0022] The cosine similarity of the corresponding semantic attributes in the state vector is calculated to obtain a similarity value of each attribute, and the similarity value is weighted and summed with a preset importance weight to obtain a state transition time interval.

[0023] Calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the candidate content with relevance into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map include:

[0024] Collecting feature information of candidate content in an augmented reality content library, and converting the feature information into a feature vector, wherein the feature vector includes a visual feature component, a semantic feature component, and an interactive feature component;

[0025] Performing a matching calculation on the feature vector of the candidate content and the semantic change trend, and calculating the degree of consistency between the change direction of the feature vector of the candidate content in the spatial dimension, attribute dimension, and interaction dimension and the semantic change trend, to obtain a matching score for each candidate content;

[0026] Using the vector inner product of the matching scores as the similarity value between the candidate contents to construct a similarity matrix; calculating the mean and standard deviation of the similarity matrix, setting a dynamic clustering threshold based on the mean and the standard deviation, and grouping candidate contents with similarities higher than the dynamic clustering threshold into the same content group;

[0027] extracting scene features from the three-dimensional semantic map, the scene features including spatial topology, semantic attribute distribution, and interaction hotspot distribution, and generating corresponding trigger conditions for each content group based on the scene features;

[0028] Monitoring scene changes in the three-dimensional semantic map, and when the scene changes meet the triggering conditions of the content group, calculating an update index for the corresponding content group, the update index combining the access frequency, timeliness, and scene relevance of the content group;

[0029] An update priority of the content group is determined according to the update index. When the update priority exceeds a preset priority threshold, the content with the highest matching degree with the current semantic change trend is selected from the candidate content library as the update strategy.

[0030] Establishing an interactive behavior model based on the interactive data, and calculating the interactive responsiveness of each content group using the interactive behavior model includes:

[0031] Extracting interaction time series features, interaction space features, and interaction modal features from the interaction data, and fusing the interaction time series features, the interaction space features, and the interaction modal features to construct a behavior feature vector;

[0032] Establishing an interactive behavior model based on the behavior feature vector, wherein the interactive behavior model adopts a hybrid structure of a recursive neural network and an attention mechanism, capturing the temporal dependency of the interactive behavior model through the recursive neural network, and highlighting the importance of the interactive features using the attention mechanism to generate a dynamic representation of the interactive behavior model;

[0033] The interactive responsiveness of the content group is calculated based on the interactive behavior model, the feature vector of the content group is similarly matched with the dynamic representation, and the result of the similarity matching is dynamically weighted by a temporal smoothing factor to obtain a responsiveness score representing the degree of matching between the content group and the interactive behavior.

[0034] Constructing a directed push path graph based on the semantic association between the content groups, wherein the directed push path graph is used to guide the activation sequence of the content groups. Achieving a smooth transition of the augmented reality experience through the orderly activation of the content groups includes:

[0035] Normalizing the concept weights, entity vectors, and context features in the content group to obtain a standardized feature vector, and constructing a semantic representation of the content group based on the standardized feature vector;

[0036] Calculating the centrality, density, and betweenness centrality of the content groups, and obtaining node importance based on a weighted combination of the centrality, density, and betweenness centrality; constructing a directed push path graph based on the node importance, wherein the directed push path graph is used to represent activation dependency relationships between content groups;

[0037] In the directed push path graph, an activation path score is calculated using a conditional probability function that takes into account historical states, wherein the conditional probability function takes the semantic association strength, edge weight, and resource consumption of adjacent content groups as input parameters, and selects the path with the highest score as the optimal activation path;

[0038] Preloading the content groups to be activated into corresponding cache levels according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representations of the preloaded content groups, and calculating the context relevance in combination with real-time scene information;

[0039] The content difference and the context relevance are input into an adaptive state smoothing function to dynamically generate transition time between adjacent content groups. The activation process of the content group is controlled according to the transition time to achieve smooth state transition.

[0040] Preloading the content groups to be activated into the corresponding cache level according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representation of the preloaded content groups, and calculating the context relevance in combination with the real-time scene information includes:

[0041] Constructing a time series feature vector for the content group, the time series feature vector including an access frequency feature, a time interval feature, and a duration feature, and generating a dynamic access pattern for the content group based on the time series feature vector;

[0042] Calculating a preloading priority score according to the dynamic access pattern, the preloading priority score being a weighted combination of an access frequency score, a time urgency, and a resource requirement, and determining a loading order of the content groups based on the preloading priority score;

[0043] Allocating cache tiers for the content groups whose loading order is determined, determining an optimal cache tier by calculating the difference between the utility value and storage cost of the content groups at each cache tier, loading the content groups into the corresponding cache tiers, and generating a cache allocation plan;

[0044] Acquiring multimodal features of the content group that has completed cache allocation, performing weighted fusion on the multimodal features to obtain a fusion feature matrix, and generating a semantic embedding vector based on the fusion feature matrix through a nonlinear mapping function;

[0045] The scene state information, user behavior information and environmental information corresponding to the semantic embedding vector are collected, the scene state information, the user behavior information and the environmental information are input into a context encoding function to generate a context feature representation, and the context relevance between content groups is calculated based on the context feature representation.

[0046] According to a second aspect of the embodiments of the present invention,

[0047] Provides an AI-driven interactive AR content intelligent generation and distribution system, including:

[0048] The first unit is configured to obtain scene data collected by a user terminal and construct a three-dimensional semantic map of a target scene based on the scene data; calculate semantic entropy values ​​of each region in the scene according to the three-dimensional semantic map, construct a scene semantic evolution tree based on the semantic entropy values, and obtain a semantic change trend of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time;

[0049] A second unit is configured to calculate a matching degree between candidate content in the augmented reality content library and the semantic change trend, divide the candidate content with relevance into content groups according to the matching degree, and generate a corresponding trigger condition and update strategy for each content group based on the three-dimensional semantic map;

[0050] a third unit configured to collect interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establish an interaction behavior model based on the interaction data, and calculate the interaction responsiveness of each content group using the interaction behavior model;

[0051] The fourth unit is used to classify the content groups according to the interactive responsiveness, determine the push priority based on the classification results, and construct a directed push path graph based on the semantic association between the content groups. The directed push path graph is used to guide the activation order of the content groups, and achieve a smooth transition of the augmented reality experience through the orderly activation of the content groups.

[0052] According to a third aspect of the embodiments of the present invention,

[0053] An electronic device is provided, comprising:

[0054] processor;

[0055] a memory for storing processor-executable instructions;

[0056] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0057] According to a fourth aspect of the embodiments of the present invention,

[0058] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0059] The beneficial effects of this application are as follows:

[0060] 1. By constructing a three-dimensional semantic map and semantic evolution tree, it can reflect the semantic change trend of the target scene in real time, thereby improving the relevance and adaptability of augmented reality content.

[0061] 2. The interactive behavior model established through user interaction data can accurately evaluate the interactive responsiveness of content groups, thereby optimizing content push strategies and improving user experience and satisfaction.

[0062] 3. Building a directed push path map based on the semantic relevance of content groups can achieve orderly activation of content, ensure a smooth transition of the augmented reality experience, and enhance user immersion and participation. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of the process of an AI-driven interactive AR content intelligent generation and distribution method according to an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram comparing the accuracy of semantic entropy calculation in different scenarios according to an embodiment of the present invention;

[0065] Figure 3 This is a complete flow chart of the processing mechanism based on semantic attributes and state transitions according to an embodiment of the present invention;

[0066] Figure 4 This is a flow chart of the interactive behavior model and interactive responsiveness calculation according to an embodiment of the present invention;

[0067] Figure 5 The figure is a schematic diagram comparing preload cache hit rates in different application scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0069] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG is a flow chart of an AI-driven interactive AR content intelligent generation and distribution method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] Acquire scene data collected by a user terminal and construct a three-dimensional semantic map of the target scene based on the scene data; calculate semantic entropy values ​​for each area in the scene according to the three-dimensional semantic map, construct a scene semantic evolution tree based on the semantic entropy values, and obtain the semantic change trend of the target scene by updating the node states and conversion weights in the scene semantic evolution tree in real time;

[0072] Calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the relevant candidate content into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map;

[0073] Collecting interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establishing an interaction behavior model based on the interaction data, and calculating the interaction responsiveness of each content group using the interaction behavior model;

[0074] The content groups are graded according to the interactive responsiveness, and push priorities are determined based on the grading results. A directed push path graph is constructed in combination with the semantic associations between the content groups. The directed push path graph is used to guide the activation order of the content groups, and a smooth transition of the augmented reality experience is achieved through the orderly activation of the content groups.

[0075] In an optional embodiment, the semantic entropy value of each area in the scene is calculated according to the three-dimensional semantic map, a scene semantic evolution tree is constructed based on the semantic entropy value, and the semantic change trend of the target scene is obtained by updating the node status and conversion weight in the scene semantic evolution tree in real time, including:

[0076] Dividing the scene space of the three-dimensional semantic map into voxel grids, calculating the probability distribution of occurrence of each semantic category data in each voxel grid, and calculating the semantic entropy value of each voxel grid based on the probability distribution, wherein the semantic entropy value is used to represent the semantic complexity of the voxel grid;

[0077] Obtaining a neighborhood voxel set of each voxel grid, calculating a spatial distance between each voxel in the neighborhood voxel set and a target voxel, determining a spatial association weight based on the spatial distance, and performing a weighted calculation on the semantic entropy value according to the spatial association weight to obtain a weighted semantic entropy value that takes neighborhood influence into account;

[0078] Constructing a scene semantic evolution tree based on the weighted semantic entropy value, calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree, and using the semantic state change amount, the state transition probability, and the state transition time interval as attributes of the connecting edges between the nodes in the scene semantic evolution tree;

[0079] The similarity between nodes in the scene semantic evolution tree is calculated. The similarity is obtained by a weighted combination of the cosine similarity of the semantic state vector, the Euclidean distance of the entropy distribution matrix, and the time decay function. The scene semantic evolution tree is dynamically updated based on the similarity to obtain the semantic change trend of the scene.

[0080] An adaptive voxel partitioning strategy is used to divide the scene space of the 3D semantic map into a voxel grid. The initial voxel size is set to 1m × 1m × 1m, and the voxel size is dynamically adjusted based on scene complexity. For areas with rich details, the voxel size can be refined to 0.25m × 0.25m × 0.25m; for open, uniform areas, the voxel size can be increased to 2m × 2m × 2m.

[0081] For example, in an indoor scene, the furniture-dense area uses a voxel size of 0.5m × 0.5m × 0.5m, while the corridor and lobby areas use a voxel size of 1m × 1m × 1m. After the scene is divided, a total of 5,267 voxel units are generated.

[0082] To calculate the probability distribution of each semantic category within each voxel grid, we count the number of point clouds or volume percentages with different semantic labels within each voxel. Preset semantic categories include 20 common indoor object categories, such as walls, floors, ceilings, doors, windows, tables and chairs, and electrical appliances. For voxel V1, the ground point cloud accounts for 45%, the table point cloud accounts for 30%, the chair point cloud accounts for 15%, and other object point clouds account for 10%, forming a probability distribution vector [0, 0.45, 0, 0, 0, 0.3, 0.15, ..., 0.1].

[0083] The semantic entropy of each voxel grid was calculated based on the probability distribution. Using the Shannon entropy formula, we sum the product of the class probabilities and their logarithms and then take the negative of the sum. For voxel V1, the calculated semantic entropy value was 1.58, indicating that the voxel had medium semantic complexity.

[0084] To obtain the neighborhood voxel set for each voxel grid, a 26-neighborhood model is used. This considers all voxels within a 3×3×3 cubic region surrounding the target voxel (excluding the target voxel itself). For voxel V1, its neighborhood voxel set consists of 26 voxels, labeled V2, V3, ..., V27. The spatial distance between each voxel in the neighborhood voxel set and the target voxel is calculated using the Euclidean distance formula based on the coordinates of the voxel center. For example, if the center coordinates of voxel V1 are (10.5, 8.2, 3.0) and the center coordinates of voxel V2 are (11.5, 8.2, 3.0), the spatial distance between them is 1.0.

[0085] When determining spatial association weights based on spatial distance, a Gaussian decay function was used with a standard deviation parameter of 1.2, assigning higher weights to closer distances. For voxels V1 and V2, the spatial distance was 1.0, resulting in a calculated spatial association weight of 0.66. Semantic entropy values ​​were weighted according to the spatial association weights, combining the raw entropy of the target voxel with the weighted entropy of its neighboring voxels. For voxel V1, the raw entropy was 1.58, and the weighted semantic entropy after accounting for the influence of the neighborhood was 1.72, indicating that neighboring voxels increase the semantic complexity of the target region.

[0086] When constructing a scene semantic evolution tree based on weighted semantic entropy, scene states are treated as nodes and state transitions as edges. Each node represents the semantic state of the scene at a specific moment and contains two pieces of information: a semantic state vector and an entropy distribution matrix. The semantic state vector has a dimension of 128 and contains the semantic category distribution and spatial structural characteristics of each voxel in the scene. The entropy distribution matrix has the same size as the voxel grid used to divide the scene, with each element corresponding to the weighted semantic entropy value of a voxel. For example, node N1 represents the scene state at time t1, with its semantic state vector being [0.32, 0.45, 0.28, ..., 0.51], and the value of voxel V1 in the entropy distribution matrix is ​​1.72.

[0087] The Euclidean distance between state vectors is used to calculate the semantic state change between adjacent nodes in the scene semantic evolution tree. For adjacent nodes N1 and N2, the semantic state vectors are [0.32, 0.45, 0.28, ..., 0.51] and [0.35, 0.42, 0.30, ..., 0.48], respectively. The calculated semantic state change is 0.26, indicating a moderate difference between the two scene states.

[0088] The state transition probability is calculated based on historical state transition data and current state characteristics. For the transition from node N1 to N2, 15 similar transitions were observed in historical data, resulting in a transition probability of 0.65 under current conditions. The state transition interval is calculated based on state variance and historical transition duration data. The average historical transition duration from node N1 to N2 is 2.3 seconds, and the predicted transition time, taking into account the current state variance, is 2.1 seconds.

[0089] The semantic state change, state transition probability, and state transition interval are used as attributes of the edges connecting nodes in the scene semantic evolution tree. For example, edge E12 connects nodes N1 and N2, and its attribute values ​​are (0.26, 0.65, 2.1), representing the semantic state change, state transition probability, and state transition interval, respectively. These edge attributes are used to guide scene change prediction and resource allocation optimization.

[0090] When calculating the similarity between nodes in the scene semantic evolution tree, the cosine similarity of the semantic state vectors, the Euclidean distance of the entropy distribution matrix, and the time decay function are comprehensively considered. The cosine similarity of the semantic state vectors calculates the similarity of the semantic content between nodes; the Euclidean distance of the entropy distribution matrix measures the difference in entropy distribution between nodes; and the time decay function takes into account the influence of time, reducing the contribution of similarity as the time interval increases. For nodes N1 and N3, the cosine similarity of the semantic state vectors is 0.78, the Euclidean distance of the entropy distribution matrix is ​​0.45, the time interval is 5 seconds, and the time decay coefficient is 0.85, resulting in a combined similarity of 0.62.

[0091] The weighted combination used in the similarity calculation uses three weight coefficients: a semantic state vector weight of 0.5, an entropy distribution matrix weight of 0.3, and a time decay weight of 0.2. The time decay function uses an exponential decay model with a decay rate of 0.05, which reduces the contribution of state similarity from 10 seconds ago by approximately 40% and from 20 seconds ago by approximately 65%.

[0092] When dynamically updating the scene semantic evolution tree based on similarity, node merging, edge weight adjustment, and path optimization are performed. Node merging rules: Nodes with a similarity exceeding 0.85 are merged into a single node, and the attributes of the merged node are a weighted average of the original nodes. Edge weight adjustment rules: The state transition probabilities and time intervals of edges are updated based on newly added observations. Path optimization rules: Edges with a transition probability greater than 0.2 are retained, and low-probability edges are deleted to simplify the tree structure. For example, if the similarity between nodes N4 and N5 is 0.92, exceeding the threshold, they are merged into a new node N45, and the weighted average of the state vector and entropy distribution matrix is ​​taken.

[0093] The dynamically updated scene semantic evolution tree reveals the semantic change trends of the scene. These trends include: major state transition paths, high-frequency transition patterns, and key state nodes. A major state transition path is a sequence of paths connected by edges with high transition probabilities, such as N1→N2→N45→N7, with transition probabilities of 0.65, 0.72, and 0.58, respectively. A high-frequency transition pattern is a state cycle that occurs repeatedly within a short period of time, such as the N3 to N6 cycle, with transition probabilities of 0.53 and 0.61. A key state node is a node connected to multiple high-probability transition edges, such as N2, which is connected to three edges with a transition probability greater than 0.6.

[0094] Based on the predicted state transition path N1→N2→N45→N7, the visual resources corresponding to the nodes on the path are preloaded. For high-frequency transition patterns N3 to N6, resources for both states are cached simultaneously. For the critical state node N2, rendering resources are prioritized and its preloading priority is increased. Experimental results show that compared with traditional methods, this resource preloading strategy based on semantic change trends reduces resource loading wait time by 42.7%, improves scene switching smoothness by 38.5%, and improves user experience scores by 25.3%.

[0095] Figure 2 This is a schematic diagram comparing the accuracy of semantic entropy calculation in different scenarios according to an embodiment of the present invention:

[0096] This figure compares the roaming experience accuracy of three different technical solutions in different scenario types. The figure is presented as a line graph, with the horizontal axis representing five scenario types of varying complexity (simple indoor, complex indoor, open outdoor, mixed urban, and dynamic multi-person scenarios), and the vertical axis showing the semantic entropy calculation accuracy (%). Our technical solution (marked with triangles) exhibits the best performance across all scenarios, ranging from 93% in simple indoor scenarios to 78% in dynamic multi-person scenarios. Although performance decreases with increasing scene complexity, the decline is minimal. The cluster analysis method (marked with squares) performs second best, with accuracy dropping from 85% to 60%. Traditional statistical methods (marked with circles) perform the worst, with a sharp drop from 80% to 55%. The performance differences among the three solutions become more pronounced as scene complexity increases. In particular, in mixed urban and dynamic multi-person scenarios, our technical solution demonstrates a clear advantage over the other two methods, maintaining high accuracy of 83% and 78%, respectively, while the other methods experience significant declines. This demonstrates the improved adaptability and stability of our technical solution in complex and dynamic scenarios.

[0097] The existing technology mainly uses simple state diagrams or Markov models in the analysis of scene semantic changes, which lacks quantitative analysis of the scene semantic complexity and consideration of spatial correlation. Traditional methods usually only focus on the overall state transition of the scene, ignoring the differences in semantic complexity and semantic change laws of different regions within the scene. The starting point of the improvement of this application is to introduce semantic entropy theory to quantify scene complexity, combine spatial correlation analysis to build a more accurate scene semantic representation, and capture the dynamic change laws of scene semantics through the evolutionary tree model. Compared with the existing technology, this application has improved the scene change prediction accuracy by 37.2%, reduced the state transition time prediction error by 43.5%, greatly improved the accuracy and timeliness of resource preloading, and provided key technical support for immersive interactive experience.

[0098] In an optional embodiment, constructing a scene semantic evolution tree based on the weighted semantic entropy value, and calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree includes:

[0099] Extracting a set of semantic attributes of the scene based on the weighted semantic entropy value, the set of semantic attributes including object category, spatial position, object attributes, and relationship between objects, performing threshold segmentation on the weighted semantic entropy value and extracting semantic regions, obtaining a semantic feature vector for each semantic region, and normalizing the semantic feature vector to obtain an attribute strength value corresponding to each semantic attribute;

[0100] Constructing a node representation of a scene semantic evolution tree based on the semantic attribute set and the attribute strength values, wherein each node includes a state vector and an entropy value distribution matrix, wherein the state vector is constructed by organizing the attribute strength values ​​into a vector form according to a preset attribute sorting rule, and the entropy value distribution matrix is ​​constructed by organizing the weighted semantic entropy values ​​into a matrix form according to spatial position relationships;

[0101] Identifying adjacent node pairs in the semantic evolution tree, calculating the difference between each component of the state vector between the nodes for each adjacent node pair, dividing the difference by the maximum value of the corresponding component and introducing a smoothing factor to obtain a normalized semantic state change;

[0102] Calculating the bi-norm of the semantic state change to obtain a state distance, negating the state distance and performing an exponential operation to obtain an initial probability value, and normalizing the initial probability values ​​of all adjacent nodes of each node to obtain a state transition basic probability;

[0103] The cosine similarity of the corresponding semantic attributes in the state vector is calculated to obtain a similarity value of each attribute, and the similarity value is weighted and summed with a preset importance weight to obtain a state transition time interval.

[0104] When extracting the semantic attribute set of a scene based on weighted semantic entropy, we perform multi-level threshold segmentation on the weighted semantic entropy. We set an initial threshold of 0.35, a step size of 0.1, and three thresholds of 0.35, 0.45, and 0.55. We then segment the weighted semantic entropy into regions based on these three thresholds, yielding multiple semantic regions.

[0105] For a scene image, threshold segmentation is performed to obtain five semantic regions, labeled R1, R2, R3, R4, and R5. A feature extractor is applied to each semantic region to obtain a semantic feature vector. This feature extraction uses a deep convolutional neural network consisting of eight convolutional layers and three fully connected layers, outputting a feature vector with a dimension of 128.

[0106] For example, the original feature vector extracted for region R1 is [0.75, 0.23, 0.56, ..., 0.41]. The extracted semantic feature vector is normalized so that each component value falls between 0 and 1, obtaining the attribute strength value corresponding to each semantic attribute. This normalization uses the maximum-minimum normalization method. For region R1, the normalized attribute strength value vector is [0.82, 0.15, 0.64, ..., 0.53].

[0107] The semantic attribute set includes four categories: object category, spatial location, object attributes, and inter-object relationships. The object category attribute includes 64 predefined common object categories, such as people, cars, and buildings. Each category is assigned an attribute strength value, indicating the confidence that the region contains an object of that category. In region R1, the attribute strength value for the "people" category is 0.92, and the attribute strength value for the "building" category is 0.18.

[0108] Spatial position attributes are expressed in relative coordinates, including center coordinates (x, y) and dimensions (width, height), normalized to the range 0–1. The spatial position attribute values ​​for region R1 are [0.35, 0.42, 0.28, 0.31], indicating that its center is located at (0.35, 0.42) of the image, its width is 0.28 of the image width, and its height is 0.31 of the image height.

[0109] Object attributes include characteristics such as color, texture, and shape, and are represented as multidimensional vectors. The color feature of region R1 is [0.75, 0.25, 0.45], representing the intensities of the red, green, and blue channels. Inter-object relationship attributes describe the spatial and semantic connections between regions, including relative positional relationships and interactions. The relationship attribute values ​​between regions R1 and R2 are [0.8, 0.1, 0.05, 0.05], representing the probability distribution of the four relationships: "contains," "adjacent," "overlap," and "separated."

[0110] When constructing a node representation of a scene semantic evolution tree based on a set of semantic attributes and attribute strength values, each node contains two pieces of information: a state vector and an entropy distribution matrix. When constructing the state vector, attribute strength values ​​are organized into a vector form according to a pre-set attribute sorting rule. This sorting rule is: object category first, then spatial position, then object attributes, and finally, relationships between objects.

[0111] For a specific scene node N1, the first 64 components of its state vector represent the object category strength value, such as [0.92 (person), 0.05 (car), 0.18 (building), ...]; the next 4 components represent the spatial position information [0.35, 0.42, 0.28, 0.31]; followed by object attribute information [0.75, 0.25, 0.45, ...]; and finally the relationship information between objects [0.8, 0.1, 0.05, 0.05, ...].

[0112] When constructing the entropy distribution matrix, the weighted semantic entropy values ​​are organized into a matrix based on their spatial positions. The matrix size matches the original scene image, and each element represents the semantic entropy value at the corresponding position. For node N1, the entropy distribution matrix is ​​16×16, with some values ​​being [[0.42, 0.38, 0.51, ...], [0.37, 0.43, 0.48, ...], ...].

[0113] To identify adjacent node pairs in the semantic evolution tree, a strategy combining temporal proximity and spatial similarity is employed. Temporal proximity refers to the proximity of two nodes in a temporal sequence, such as frame t and frame t+1. Spatial similarity is determined by calculating the cosine similarity of the node state vectors. The similarity threshold is set at 0.85, and node pairs exceeding the threshold are considered to have spatial similarity.

[0114] For each identified pair of adjacent nodes (N1, N2), calculate the difference between each component of the state vector. For example, if the state vector of N1 is [0.92, 0.05, 0.18, ...] and the state vector of N2 is [0.88, 0.08, 0.22, ...], the resulting difference vector is [0.04, -0.03, -0.04, ...]. Divide the difference by the maximum value of the corresponding component and introduce a smoothing factor to obtain the normalized semantic state change. The smoothing factor is set to 0.05 to avoid division by zero. For the first component of the difference vector, 0.04, the corresponding maximum value is 0.92, and the normalized change is 0.04 / (0.92 + 0.05) = 0.041.

[0115] To calculate the state distance using the square root of the sum of squares, we calculated the normalized semantic state change vector [0.041, -0.032, -0.042, ...] by taking the sum of squares of each component and taking the square root, we obtained a state distance of 0.217. The initial probability value was obtained by negating the state distance and performing an exponential operation.

[0116] For the state distance of 0.217, negate it to -0.217, and perform an exponential operation to obtain an initial probability value of 0.805. For each node, the initial probability values ​​of all its neighboring nodes are normalized to obtain the basic probability of state transition. For example, node N1 has three neighboring nodes, N2, N3, and N4, with corresponding initial probabilities of 0.805, 0.723, and 0.652, respectively. The normalized state transition probabilities are 0.37, 0.33, and 0.30, respectively.

[0117] When calculating the cosine similarity of the corresponding semantic attributes in the state vector to obtain the similarity value of each attribute, the cosine similarity of four semantic attributes is calculated separately: object category, spatial position, object attribute, and inter-object relationship. For the node pair (N1, N2), the object category similarity is 0.95, the spatial position similarity is 0.92, the object attribute similarity is 0.87, and the inter-object relationship similarity is 0.81.

[0118] The state transition interval is calculated by weighting the similarity values ​​with the preset importance weights. The importance weights for the four attribute categories are set to 0.4, 0.3, 0.2, and 0.1, respectively. The total similarity obtained by weighted summation is 0.95 × 0.4 + 0.92 × 0.3 + 0.87 × 0.2 + 0.81 × 0.1 = 0.911. The total similarity is converted to the state transition interval using an inverse proportional relationship. The base interval value is 1.0, and the actual interval is the base value divided by the total similarity, i.e., 1.0 / 0.911 = 1.098, indicating that the expected time to transition from state N1 to state N2 is 1.098 time units.

[0119] Figure 3 The following is a complete flow chart of the processing mechanism based on semantic attributes and state transitions according to an embodiment of the present invention:

[0120] This flowchart describes a processing mechanism based on semantic attributes and state transitions. First, the scene's semantic attribute set is extracted using weighted semantic entropy, including factors such as object category, spatial location, object attributes, and inter-object relationships. The weighted semantic entropy is then segmented using a threshold to obtain semantic regions. Semantic feature vectors are extracted for each semantic region, and attribute strength values ​​corresponding to each semantic attribute are normalized. Second, a node representation of the scene's semantic evolution tree is constructed based on the semantic attribute set and attribute strength values. Each node contains a state vector and an entropy distribution matrix. The attribute strength values ​​are organized into vector form according to preset rules to construct the state vector, and the entropy distribution matrix is ​​organized into a matrix form based on spatial location relationships. Next, adjacent pairs of nodes in the semantic evolution tree are identified, and the difference between each component of the state vector between the nodes is calculated. This difference is divided by the maximum value of the corresponding component and a smoothing factor is introduced to obtain the normalized semantic state change. Next, the bi-norm of the semantic state change is calculated to obtain the state distance. This state distance is then subjected to exponential operation to obtain the initial probability value. For each node, the initial probability values ​​of all its neighboring nodes are normalized to obtain the base probability of state transition. Finally, the cosine similarity of the corresponding semantic attributes in the state vector is calculated to obtain the similarity value of each attribute. The similarity value is then weighted and summed with the preset importance weight to obtain the state transition interval. This entire process achieves precise expression and smooth transition of semantic states through rigorous mathematical processing.

[0121] Existing technologies mainly use simple state transition diagrams or Markov chain models in scene semantic evolution analysis. These methods usually only consider the overall state of the scene and ignore the fine-grained changes and complex associations of semantic elements within the scene. Traditional methods such as rule-based scene understanding systems or statistical-based state transition models are difficult to accurately capture the dynamic evolution laws of object categories, spatial positions, object attributes, and relationships between objects in the scene. The starting point of the improvement of this application is to construct a multi-level, fine-grained scene semantic evolution tree, integrating semantic entropy theory and attribute strength modeling to achieve accurate quantification of the scene semantic state and its changes.

[0122] The main improvements of this application compared to the prior art include: introducing weighted semantic entropy as a measure of the amount of information in the scene content to achieve adaptive recognition of important areas of the scene; using a multi-dimensional semantic attribute set to represent the scene content, including object categories, spatial positions, object attributes, and relationships between objects, providing a more comprehensive scene description; designing a dual node representation of state vectors and entropy distribution matrices to simultaneously capture the structured semantic information and unstructured entropy distribution characteristics of the scene; proposing a state transition measurement method based on normalized semantic state change to accurately quantify the degree of difference between scene states; and establishing a state transition time model based on multi-attribute similarity weighting to achieve accurate prediction of transition timing characteristics.

[0123] In an optional embodiment, calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the relevant candidate content into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map includes:

[0124] Collecting feature information of candidate content in an augmented reality content library, and converting the feature information into a feature vector, wherein the feature vector includes a visual feature component, a semantic feature component, and an interactive feature component;

[0125] Performing a matching calculation on the feature vector of the candidate content and the semantic change trend, and calculating the degree of consistency between the change direction of the feature vector of the candidate content in the spatial dimension, attribute dimension, and interaction dimension and the semantic change trend, to obtain a matching score for each candidate content;

[0126] Using the vector inner product of the matching scores as the similarity value between the candidate contents to construct a similarity matrix; calculating the mean and standard deviation of the similarity matrix, setting a dynamic clustering threshold based on the mean and the standard deviation, and grouping candidate contents with similarities higher than the dynamic clustering threshold into the same content group;

[0127] extracting scene features from the three-dimensional semantic map, the scene features including spatial topology, semantic attribute distribution, and interaction hotspot distribution, and generating corresponding trigger conditions for each content group based on the scene features;

[0128] Monitoring scene changes in the three-dimensional semantic map, and when the scene changes meet the triggering conditions of the content group, calculating an update index for the corresponding content group, the update index combining the access frequency, timeliness, and scene relevance of the content group;

[0129] An update priority of the content group is determined according to the update index. When the update priority exceeds a preset priority threshold, the content with the highest matching degree with the current semantic change trend is selected from the candidate content library as the update strategy.

[0130] Extract feature information for candidate content from the AR content library. This feature information includes visual, semantic, and interactive features. Visual features can be obtained through image processing techniques, such as using convolutional neural networks (CNNs) to extract image color histograms and texture features. Semantic features are obtained by analyzing the content description text using natural language processing techniques to extract keywords and their semantic relationships. Interactive features are obtained by analyzing user interaction data, such as click-through rate and dwell time.

[0131] Convert this feature information into a feature vector. The feature vector should contain three components: visual feature component, semantic feature component, and interaction feature component. The value of each component can be normalized so that it can be compared on the same scale.

[0132] Calculate the degree of match between the candidate content's feature vector and its semantic change trend. By analyzing the direction of change of the candidate content's feature vector in the spatial, attribute, and interaction dimensions, assess its consistency with the semantic change trend. Specifically, the degree of match can be determined by comparing the rate of change and trend direction of the feature vector.

[0133] If the visual feature vector of a candidate content matches the current popular trend in terms of color, the matching score of the content will be relatively high. This score can be obtained by comprehensively considering the changes in various dimensions.

[0134] The matching scores of all candidate content are organized into a vector, and the similarity between these scores is calculated. Similarity can be derived by taking the inner product of the score vectors. After constructing the similarity matrix, the mean and standard deviation of the matrix are calculated. Based on the mean and standard deviation, a dynamic clustering threshold is set, and candidate content with similarity above this threshold is grouped together.

[0135] If the mean matching score of a group of candidate content is 0.8 and the standard deviation is 0.1, the threshold can be set to 0.9 to group the content with a score higher than 0.9.

[0136] Scene features are extracted from the 3D semantic map, primarily including spatial topology, semantic attribute distribution, and interaction hotspot distribution. Spatial topology can be obtained by analyzing the relative positions and connectivity of objects in the scene. Semantic attribute distribution is obtained through statistical analysis of the semantic labels of objects in the scene. Interaction hotspot distribution analyzes user interaction behavior data within the scene to identify key areas of user attention.

[0137] Based on the extracted scene features, corresponding trigger conditions are generated for each content group. For example, if the scene features of a content group show that users interact more frequently in a specific area, changes in that area can be set as a trigger condition.

[0138] Monitor scene changes in the 3D semantic map. When a scene change meets the trigger conditions for a content group, calculate the update index for that content group. This update index should be comprehensively evaluated based on the content group's access frequency, timeliness, and scene relevance. For example, if a content group's access frequency has increased significantly over the past week and is highly correlated with the current scene change, its update priority will be increased accordingly.

[0139] Based on the calculated update metrics, the update priority of each content group is determined. When the update priority of a content group exceeds the preset priority threshold, the content with the highest degree of match to the current semantic change trend is selected from the candidate content library as the update strategy. This content should be selected to ensure that it effectively meets user needs and improves the user experience.

[0140] For example, if the update priority of a content group is 95 and the score of the candidate content with the highest matching degree is 0.9, the candidate content can be pushed as the update strategy.

[0141] In an optional implementation, establishing an interactive behavior model based on the interactive data, and calculating the interactive responsiveness of each content group using the interactive behavior model includes:

[0142] Extracting interaction time series features, interaction space features, and interaction modal features from the interaction data, and fusing the interaction time series features, the interaction space features, and the interaction modal features to construct a behavior feature vector;

[0143] Establishing an interactive behavior model based on the behavior feature vector, wherein the interactive behavior model adopts a hybrid structure of a recursive neural network and an attention mechanism, capturing the temporal dependency of the interactive behavior model through the recursive neural network, and highlighting the importance of the interactive features using the attention mechanism to generate a dynamic representation of the interactive behavior model;

[0144] The interactive responsiveness of the content group is calculated based on the interactive behavior model, the feature vector of the content group is similarly matched with the dynamic representation, and the result of the similarity matching is dynamically weighted by a temporal smoothing factor to obtain a responsiveness score representing the degree of matching between the content group and the interactive behavior.

[0145] To extract interaction temporal features, spatial features, and modal features from interaction data, the user's interaction logs with content groups are parsed and feature extracted. Interaction temporal features include temporal information such as the time of interaction, duration, and interval. For example, a user clicks on content group A at 10:23:45 on March 15, 2025, with a dwell time of 45 seconds, and then clicks on content group B at 10:25:12 with a dwell time of 30 seconds. This yields a timestamp sequence, a duration sequence, and an interaction interval sequence. Spatial features include spatial information such as the physical location of the interaction, device posture, and touch point coordinates. For example, a user clicks on content group C at coordinates (320, 450), the device is in landscape mode, and the touch point pressure is 0.7. Modal features include interaction types, such as clicks, swipes, voice commands, eye tracking, and other features of different interaction modes. For example, a user long-presses content group D with a pressure of 0.85 and a duration of 2 seconds.

[0146] The extracted temporal, spatial, and modal interaction features are fused to construct a behavioral feature vector. This feature fusion employs a multi-layer perceptron architecture, normalizing different feature types before mapping them to a common feature space via a fully connected layer. Specifically, temporal features are mapped to 128-dimensional vectors, spatial features to 64-dimensional vectors, and modal features to 32-dimensional vectors. These features are then concatenated to form a 224-dimensional joint representation. Finally, a fully connected layer with a dimension of 256 is used to generate the final behavioral feature vector. For a single user interaction with content group E, the extracted temporal feature values ​​might be [0.25, 0.78, 0.45, ...], spatial feature values ​​[0.62, 0.31, 0.58, ...], and modal feature values ​​[0.9, 0.2, 0.4, ...]. The fused behavioral feature vector is [0.53, 0.71, 0.38, 0.42, ..., 0.65].

[0147] When building an interactive behavior model based on behavioral feature vectors, a hybrid architecture combining a recurrent neural network and an attention mechanism is employed. The recurrent neural network utilizes a long short-term memory (LSTM) unit, consisting of an input layer, two hidden layers, and an output layer. The input layer receives a sequence of behavioral feature vectors. Each LSTM unit contains 64 neurons, which are used to capture the temporal dependencies of interactive behaviors. For example, a user may quickly browse content group F and then return to carefully read content group G. This behavioral sequence pattern can be learned by the LSTM unit. The attention mechanism uses a self-attention structure to calculate correlation scores between behavioral feature vectors, assigning higher weights to important features. For example, in an article reading scenario, the dwell time feature might have a weight of 0.6, while in an image browsing scenario, the click location feature might have a weight of 0.7. The attention mechanism can adaptively adjust the importance of each feature in different scenarios.

[0148] During the training process of the interactive behavior model, a training set is constructed using historical interaction data, with a sequence of behavioral feature vectors as input and the user's subsequent interactive behavior as the prediction target. Training utilizes the backpropagation algorithm, a cross-entropy loss function, and the Adam optimizer. The initial learning rate is set to 0.001, and a learning rate decay strategy is implemented, decreasing the learning rate by 10% every 10 training rounds. After 500 training rounds, the prediction accuracy on the validation set reached 85.6%. After training, the model is able to generate a dynamic representation vector that represents the user's current interaction intent based on the sequence of historical interactive behavior feature vectors.

[0149] When calculating the interactive responsiveness of a content group using the interactive behavior model, the feature vector of the content group to be evaluated is matched against the dynamic representation generated by the model for similarity. The content group feature vector contains attributes such as content subject, format type, and complexity, and is represented using the same dimensionality as the behavioral feature vector. Similarity matching uses the cosine similarity calculation method, with the cosine angle between the content group feature vector and the dynamic representation vector as the raw match score. For example, if the feature vector of content group H is [0.65, 0.28, 0.73, ...] and the current dynamic representation vector is [0.70, 0.31, 0.68, ...], the calculated cosine similarity is 0.91, indicating that content group H is highly compatible with the current interactive behavior.

[0150] To mitigate the impact of short-term interaction fluctuations, a temporal smoothing factor is introduced to dynamically weight the similarity matching results. Based on an exponential moving average algorithm, the temporal smoothing factor assigns decreasing weights to the matching scores of the most recent N interactions, resulting in a smoothed responsiveness score. In the calculation of the smoothing factor, the decay rate α is set to 0.8, and the weights of the most recent five interactions are 0.8^0 = 1, 0.8^1 = 0.8, 0.8^2 = 0.64, 0.8^3 = 0.512, and 0.8^4 = 0.4096, respectively. For content group I, the raw matching scores of the most recent five interactions were 0.85, 0.78, 0.82, 0.79, and 0.73, respectively. After weighted averaging, the resulting responsiveness score is 0.81.

[0151] In the responsiveness score application strategy, the high responsiveness threshold is set at 0.75 and the medium responsiveness threshold is set at 0.5. Content groups with a responsiveness score above 0.75 are marked as "high responsiveness" and given the highest priority; content groups with scores between 0.5 and 0.75 are marked as "medium responsiveness"; and content groups with scores below 0.5 are marked as "low responsiveness." Content group I, with a responsiveness score of 0.81, is marked as "high responsiveness" and receives priority in resource allocation and content presentation.

[0152] To validate the effectiveness of the interactive responsiveness calculation, a test was conducted on 100 users, each of whom interacted with the system for 30 minutes, generating a total of 28,634 interaction records. The analysis results showed that users tended to engage more with the "highly responsive" content group. The click-through rate for this group was 2.7 times higher than that for the "lowly responsive" group, and the dwell time was 3.2 times longer. This demonstrates that the interactive responsiveness calculation method can effectively predict user interaction interest.

[0153] Figure 4 This is a flow chart of the interactive behavior model and interactive responsiveness calculation according to an embodiment of the present invention:

[0154] This flowchart illustrates the complete process of feature extraction and responsiveness calculation based on interaction data. First, three types of features are extracted from the interaction data: aggregated interaction features, interaction space features, and interaction scene features. These feature relationships are then structured into behavioral feature warnings, which are then processed in two branches: The interdependencies and associated behavioral patterns between neural network interactions are derived; The importance of interaction features is then highlighted through focus analysis. The processing results of these two branches, focusing on key interactions, are combined to generate a dynamic representation of the interaction behavior model, which is then used to calculate interaction responsiveness. This multi-dimensional feature extraction combined with deep learning ensures accurate modeling of the interactivity of interaction behaviors and precisely captures key interaction behavior features, enabling more accurate evaluation of user-content interaction effects. This complete closed-loop system, from raw interaction data to final responsiveness calculation, forms a complete system.

[0155] The existing technology mainly relies on static feature matching or simple collaborative filtering methods when calculating content responsiveness, which cannot adapt to complex and changing interactive scenarios. Traditional methods such as rule-based scoring systems or statistical-based recommendation algorithms often ignore the temporal and multimodal characteristics of interactive behaviors, resulting in inaccurate responsiveness calculations. The starting point of the improvement of this application is to construct a behavioral model that can capture the dynamic characteristics of interactive behaviors, and integrate temporal features, spatial features, and modal features for comprehensive analysis. By introducing recursive neural networks to capture long-term dependencies and combining attention mechanisms to highlight key features, accurate modeling of complex interactive behaviors is achieved. Compared with the existing technology, this application has improved the accuracy of interactive behavior prediction by 23.5%, and the adaptability to different interactive modes by 18.7%, significantly improving the accuracy and stability of content responsiveness evaluation, providing a reliable basis for accurate preloading and real-time rendering of content groups, and optimizing the user interaction experience.

[0156] In an optional embodiment, a directed push path graph is constructed based on the semantic association between the content groups. The directed push path graph is used to guide the activation sequence of the content groups. Achieving a smooth transition of the augmented reality experience through the orderly activation of the content groups includes:

[0157] Normalizing the concept weights, entity vectors, and context features in the content group to obtain a standardized feature vector, and constructing a semantic representation of the content group based on the standardized feature vector;

[0158] Calculating the centrality, density, and betweenness centrality of the content groups, and obtaining node importance based on a weighted combination of the centrality, density, and betweenness centrality; constructing a directed push path graph based on the node importance, wherein the directed push path graph is used to represent activation dependency relationships between content groups;

[0159] In the directed push path graph, an activation path score is calculated using a conditional probability function that takes into account historical states, wherein the conditional probability function takes the semantic association strength, edge weight, and resource consumption of adjacent content groups as input parameters, and selects the path with the highest score as the optimal activation path;

[0160] Preloading the content groups to be activated into corresponding cache levels according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representations of the preloaded content groups, and calculating the context relevance in combination with real-time scene information;

[0161] The content difference and the context relevance are input into an adaptive state smoothing function to dynamically generate transition time between adjacent content groups. The activation process of the content group is controlled according to the transition time to achieve smooth state transition.

[0162] Normalize the concept weights, entity vectors, and contextual features within the content group. This step ensures that different features are compared on the same scale. Specifically, concept weights are normalized by subtracting the minimum weight from each concept and dividing the result by the difference between the maximum and minimum values. This method is also applied to entity vectors and contextual features. The resulting normalized feature vectors are used for subsequent semantic representation.

[0163] After obtaining the standardized feature vectors, a semantic representation of the content group needs to be constructed based on these vectors. A weighted average approach can be used to weight different feature vectors according to their importance to generate a comprehensive semantic representation. This representation will reflect the overall characteristics of the content group, facilitating subsequent analysis and processing.

[0164] The centrality, density, and betweenness centrality of a content group are calculated to assess the importance of each node. Centrality is determined by analyzing the number of connections a node has in the network, density indicates the closeness of connections between nodes, and betweenness centrality reflects the key role a node plays in information flow. A weighted combination of these three metrics yields a node importance score, which in turn provides a basis for constructing a directed push path map.

[0165] Based on node importance, a directed push path graph is constructed. This graph represents the activation dependencies between content groups. Each node represents a content group, and the direction of the edges indicates the activation order. This graph ensures that high-importance nodes are activated first to ensure effective content push.

[0166] In a directed push path graph, a conditional probability function that takes into account historical state is used to calculate the score of the activation path. Inputs to this function include the semantic association strength of adjacent content groups, edge weights, and resource consumption. By analyzing these factors, the path with the highest score is selected as the optimal activation path. This process ensures the rationality and efficiency of the activation sequence.

[0167] Preload the content groups to be activated into the corresponding cache tiers based on the access order of the optimal activation path. The preloaded content groups should be calculated based on their semantic representations to determine the degree of content differentiation between adjacent content groups. This differentiation should take into account the content's theme, style, and other relevant features to ensure coherence between adjacent content groups.

[0168] Contextual relevance is calculated based on real-time scene information. This reflects the impact of the current environment on content group activation and can help adjust activation strategies. By analyzing user behavior patterns and current environmental characteristics, content push strategies can be dynamically adjusted.

[0169] Content diversity and contextual relevance are input into an adaptive state smoothing function to dynamically generate transition times between adjacent content groups. This function adjusts the activation process of content groups based on content diversity and environmental changes to achieve smooth state transitions. The transition time should be set based on user experience to avoid abrupt content transitions.

[0170] Control the activation of content groups based on dynamically generated transition times. By precisely adjusting the activation timing of each content group, users can ensure a smooth and consistent augmented reality experience. This process requires real-time monitoring of user feedback to adjust activation strategies and improve user satisfaction.

[0171] Suppose there are three content groups, A, B, and C, each representing a different augmented reality scenario. Content group A has a weight of 0.8, B 0.6, and C 0.4. After normalization, the normalized feature vectors of A, B, and C are [0.8, 0.7, 0.6], [0.6, 0.5, 0.4], and [0.4, 0.3, 0.2], respectively. Based on the calculated centrality, density, and betweenness centrality, assume the node importance is 0.9 for A, 0.7 for B, and 0.5 for C. The resulting directed push path graph prioritizes activation of content group A, followed by B and C.

[0172] In an optional embodiment, preloading the content groups to be activated into corresponding cache levels according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representations of the preloaded content groups, and calculating the context relevance in combination with real-time scene information includes:

[0173] Constructing a time series feature vector for the content group, the time series feature vector including an access frequency feature, a time interval feature, and a duration feature, and generating a dynamic access pattern for the content group based on the time series feature vector;

[0174] Calculating a preloading priority score according to the dynamic access pattern, the preloading priority score being a weighted combination of an access frequency score, a time urgency, and a resource requirement, and determining a loading order of the content groups based on the preloading priority score;

[0175] Allocating cache tiers for the content groups whose loading order is determined, determining an optimal cache tier by calculating the difference between the utility value and storage cost of the content groups at each cache tier, loading the content groups into the corresponding cache tiers, and generating a cache allocation plan;

[0176] Acquiring multimodal features of the content group that has completed cache allocation, performing weighted fusion on the multimodal features to obtain a fusion feature matrix, and generating a semantic embedding vector based on the fusion feature matrix through a nonlinear mapping function;

[0177] The scene state information, user behavior information and environmental information corresponding to the semantic embedding vector are collected, the scene state information, the user behavior information and the environmental information are input into a context encoding function to generate a context feature representation, and the context relevance between content groups is calculated based on the context feature representation.

[0178] To construct the temporal feature vector for a content group, we collect access records for each content group, including a list of access timestamps and access durations. For content group C_i, we count the number of accesses within the past T time window to obtain the access frequency feature f_i; we calculate the time difference sequence between two consecutive accesses to obtain the time interval feature i_i; and we record the duration sequence of each access to obtain the duration feature d_i. These three features are combined to form the temporal feature vector [f_i, i_i, d_i].

[0179] When generating dynamic access patterns for content groups based on time series feature vectors, a sliding time window approach is used to segment the continuous time series into multiple overlapping time windows. Pattern features are extracted from access behavior within each time window, such as periodic changes in access frequency, peak access times, and typical access sequences. By comparing access patterns across different time windows, stable patterns and changing trends are identified, forming a description of the dynamic access patterns for the content groups. For example, content group C_1 is accessed significantly more frequently between 9:00 AM and 11:00 AM on weekdays than during other times and is typically accessed before access to C_2.

[0180] When calculating the preloading priority score based on dynamic access patterns, we first calculate the access frequency score S_f, which reflects the likelihood of a content group being accessed. Next, we calculate the time urgency S_t, which indicates how urgent the next access is expected to be. Finally, we calculate the resource requirement S_r, which indicates the amount of resources required to load the content group. The preloading priority score P is calculated through a weighted combination: P = w_f × S_f + w_t × S_t + w_r × S_r, where w_f, w_t, and w_r are the weight coefficients of the three factors, and w_f + w_t + w_r = 1.

[0181] For example, if content group C_1 has an access frequency score of S_f = 0.85, a time urgency score of S_t = 0.92, a resource requirement score of S_r = 0.4, and weight coefficients w_f = 0.4, w_t = 0.4, and w_r = 0.2, then the preloading priority score P = 0.4 × 0.85 + 0.4 × 0.92 + 0.2 × 0.4 = 0.78. Based on the calculated preloading priority scores, all content groups are sorted from highest to lowest to determine the loading order.

[0182] When allocating cache levels for content groups that determine the loading order, assume that the system has n cache levels {L_1, L_2, ..., L_n}, each with different access speed and storage capacity.

[0183] For each content group C_i and each cache level L_j, we calculate the utility value U_ij, which represents the performance benefit of storing C_i in L_j. We also calculate the storage cost C_ij, which represents the cost of occupying space in L_j. We determine the optimal cache level for each content group by calculating the difference between the utility value and the storage cost: D_ij = U_ij - C_ij.

[0184] The utility values ​​of content group C_2 at the three cache levels are U_21 = 0.95, U_22 = 0.75, and U_23 = 0.45, respectively. The storage costs are C_21 = 0.8, C_22 = 0.5, and C_23 = 0.2, respectively. The differences D_21 = 0.15, D_22 = 0.25, and D_23 = 0.25, respectively. Therefore, L_2 or L_3 is selected as the cache level for C_2. The cache level allocation results for all content groups are comprehensively considered to generate an overall cache allocation plan.

[0185] When acquiring multimodal features for a content group that has completed cache allocation, text, image, and audio features are extracted. For text features, a text encoder is used to extract semantic representations; for image features, a visual feature extractor is used to obtain visual representations; and for audio features, an audio feature extractor is used to obtain acoustic representations. A weighted fusion is performed on the extracted multimodal features to generate the fused feature matrix M.

[0186] For content group C_3, its text feature vector is [0.5, 0.8, 0.3], its image feature vector is [0.7, 0.2, 0.9], and its audio feature vector is [0.4, 0.6, 0.5]. We use weight coefficients [0.5, 0.3, 0.2] for weighted fusion, resulting in a fused feature vector of [0.54, 0.56, 0.52]. Based on the fused feature matrix M, a nonlinear mapping function is used to generate a semantic embedding vector S. This nonlinear mapping function uses a multi-layer perceptron structure to map high-dimensional features into a low-dimensional semantic space, capturing the essential semantic information of the content group.

[0187] When collecting scene state information, user behavior information, and environmental information corresponding to the semantic embedding vector, scene state information includes the current application state and active page type; user behavior information includes user operation sequence, dwell time, and click behavior; and environmental information includes network status, device performance, and battery level. This information is input into the context encoding function to generate a context feature representation C. The context encoding function uses an attention mechanism to assign different weights to different types of context information, highlighting key factors related to the current scene.

[0188] Based on the contextual feature representation C, the contextual relevance R between content groups is calculated. For a content group pair (C_i, C_j), the contextual relevance R_ij is obtained by calculating the conditional similarity of its semantic embedding vectors S_i and S_j under the contextual feature representation C. For example, in the scenario where a user browses an electronic product page, the contextual relevance between the product description content group and the product review content group is 0.86, while the contextual relevance between the product description content group and the payment process content group is only 0.35.

[0189] Based on the calculated content diversity and contextual relevance, the system can predict the sequence of content groups that users are most likely to access and optimize preloading strategies accordingly, improving resource utilization efficiency and user experience. For example, if a user is browsing the details page of product A and stays there for more than 30 seconds, the system predicts that the user is likely to view product reviews next. Therefore, it preloads the review content group into the L_1 cache in advance, allowing the user to see the review content immediately after clicking the "View Reviews" button, without having to wait for the review to load.

[0190] Figure 5 The following is a schematic diagram comparing preload cache hit rates in different application scenarios according to an embodiment of the present invention:

[0191] This figure compares the cache hit rates of four different technical solutions across seven application scenarios. The horizontal axis represents different application scenarios, including VR exhibition halls, large-scale open-world games, interactive educational applications, smart home control, multiplayer collaborative environments, high-precision medical visualization, and autonomous driving simulation; the vertical axis represents cache hit rate (%). Our technical solution (triangle markers) performs best in all scenarios, with a hit rate consistently between 87% and 96%, reaching a peak of 96% in the smart home control scenario. Semantic-driven caching (plus markers) performs second best, with a hit rate fluctuating between 75% and 88%. Preset path preloading (square markers) performs in the middle, with a hit rate between 68% and 83%. The LRU cache algorithm (circle markers) performs worst, with a hit rate dropping to around 59%. All four solutions exhibit similar fluctuations across different application scenarios, but our technical solution exhibits the smallest fluctuation, demonstrating its superior stability and adaptability. In particular, in computationally intensive scenarios such as high-precision medical visualization and autonomous driving simulation, our technical solution maintains a high hit rate, demonstrating a significant performance advantage.

[0192] Existing technologies usually adopt simple strategies based on LRU (least recently used) or LFU (least frequently used) for content preloading and cache management, which only considers access frequency and time factors, and ignores content semantics and contextual information. The starting point of the improvement of this application is to integrate the temporal features, semantic features and contextual information of the content group to achieve more accurate preloading decisions. Specific improvements include: constructing a multi-dimensional temporal feature vector containing access frequency, time interval and duration to capture richer access patterns; introducing multimodal feature fusion and semantic representation to understand the essential semantics of the content group; combining real-time scene status, user behavior and environmental information to calculate context relevance and implement a context-aware preloading strategy. Through these improvements, this application significantly improves the precision and recall rate of content preloading. Experimental results show that compared with traditional methods, the preloading hit rate is improved by 35%, resource utilization efficiency is improved by 28%, and user waiting time is reduced by 43%, providing users with a smoother interactive experience.

[0193] According to a second aspect of the embodiments of the present invention,

[0194] Provides an AI-driven interactive AR content intelligent generation and distribution system, including:

[0195] The first unit is configured to obtain scene data collected by a user terminal and construct a three-dimensional semantic map of a target scene based on the scene data; calculate semantic entropy values ​​of each region in the scene according to the three-dimensional semantic map, construct a scene semantic evolution tree based on the semantic entropy values, and obtain a semantic change trend of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time;

[0196] A second unit is configured to calculate a matching degree between candidate content in the augmented reality content library and the semantic change trend, divide the candidate content with relevance into content groups according to the matching degree, and generate a corresponding trigger condition and update strategy for each content group based on the three-dimensional semantic map;

[0197] a third unit configured to collect interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establish an interaction behavior model based on the interaction data, and calculate the interaction responsiveness of each content group using the interaction behavior model;

[0198] The fourth unit is used to classify the content groups according to the interactive responsiveness, determine the push priority based on the classification results, and construct a directed push path graph based on the semantic association between the content groups. The directed push path graph is used to guide the activation order of the content groups, and achieve a smooth transition of the augmented reality experience through the orderly activation of the content groups.

[0199] According to a third aspect of the embodiments of the present invention,

[0200] An electronic device is provided, comprising:

[0201] processor;

[0202] a memory for storing processor-executable instructions;

[0203] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0204] According to a fourth aspect of the embodiments of the present invention,

[0205] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0206] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0207] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. AI-driven interactive AR content intelligent generation and distribution method, characterized by: include: Acquire scene data collected by the user terminal, and construct a three-dimensional semantic map of the target scene based on the scene data; Calculating semantic entropy values ​​for each region in the scene based on the three-dimensional semantic map, constructing a scene semantic evolution tree based on the semantic entropy values, and obtaining a semantic change trend of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time; Calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the relevant candidate content into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map; Collecting interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establishing an interaction behavior model based on the interaction data, and calculating the interaction responsiveness of each content group using the interaction behavior model; The content groups are graded according to the interactive responsiveness, and push priorities are determined based on the grading results. A directed push path graph is constructed in combination with the semantic associations between the content groups. The directed push path graph is used to guide the activation order of the content groups, and a smooth transition of the augmented reality experience is achieved through the orderly activation of the content groups.

2. The method according to claim 1, characterized in that Calculating semantic entropy values ​​for each region in the scene based on the three-dimensional semantic map, constructing a scene semantic evolution tree based on the semantic entropy values, and obtaining semantic change trends of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time include: Dividing the scene space of the three-dimensional semantic map into voxel grids, calculating the probability distribution of occurrence of each semantic category data in each voxel grid, and calculating the semantic entropy value of each voxel grid based on the probability distribution, wherein the semantic entropy value is used to represent the semantic complexity of the voxel grid; Obtaining a neighborhood voxel set of each voxel grid, calculating a spatial distance between each voxel in the neighborhood voxel set and a target voxel, determining a spatial association weight based on the spatial distance, and performing a weighted calculation on the semantic entropy value according to the spatial association weight to obtain a weighted semantic entropy value that takes neighborhood influence into account; Constructing a scene semantic evolution tree based on the weighted semantic entropy value, calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree, and using the semantic state change amount, the state transition probability, and the state transition time interval as attributes of the connecting edges between the nodes in the scene semantic evolution tree; The similarity between nodes in the scene semantic evolution tree is calculated. The similarity is obtained by a weighted combination of the cosine similarity of the semantic state vector, the Euclidean distance of the entropy distribution matrix, and the time decay function. The scene semantic evolution tree is dynamically updated based on the similarity to obtain the semantic change trend of the scene.

3. The method according to claim 2, characterized in that Constructing a scene semantic evolution tree based on the weighted semantic entropy value, and calculating the semantic state change amount, state transition probability, and state transition time interval between adjacent nodes in the scene semantic evolution tree includes: Extracting a set of semantic attributes of the scene based on the weighted semantic entropy value, the set of semantic attributes including object category, spatial position, object attributes, and relationship between objects, performing threshold segmentation on the weighted semantic entropy value and extracting semantic regions, obtaining a semantic feature vector for each semantic region, and normalizing the semantic feature vector to obtain an attribute strength value corresponding to each semantic attribute; Constructing a node representation of a scene semantic evolution tree based on the semantic attribute set and the attribute strength values, wherein each node includes a state vector and an entropy value distribution matrix, wherein the state vector is constructed by organizing the attribute strength values ​​into a vector form according to a preset attribute sorting rule, and the entropy value distribution matrix is ​​constructed by organizing the weighted semantic entropy values ​​into a matrix form according to spatial position relationships; Identifying adjacent node pairs in the semantic evolution tree, calculating the difference between each component of the state vector between the nodes for each adjacent node pair, dividing the difference by the maximum value of the corresponding component and introducing a smoothing factor to obtain a normalized semantic state change; Calculating the bi-norm of the semantic state change to obtain a state distance, negating the state distance and performing an exponential operation to obtain an initial probability value, and normalizing the initial probability values ​​of all adjacent nodes of each node to obtain a state transition basic probability; The cosine similarity of the corresponding semantic attributes in the state vector is calculated to obtain a similarity value of each attribute, and the similarity value is weighted and summed with a preset importance weight to obtain a state transition time interval.

4. The method according to claim 1, wherein Calculating the matching degree between candidate content in the augmented reality content library and the semantic change trend, dividing the candidate content with relevance into content groups according to the matching degree, and generating corresponding trigger conditions and update strategies for each content group based on the three-dimensional semantic map include: Collecting feature information of candidate content in an augmented reality content library, and converting the feature information into a feature vector, wherein the feature vector includes a visual feature component, a semantic feature component, and an interactive feature component; Performing a matching calculation on the feature vector of the candidate content and the semantic change trend, and calculating the degree of consistency between the change direction of the feature vector of the candidate content in the spatial dimension, attribute dimension, and interaction dimension and the semantic change trend, to obtain a matching score for each candidate content; Using the vector inner product of the matching scores as the similarity value between the candidate contents to construct a similarity matrix; calculating the mean and standard deviation of the similarity matrix, setting a dynamic clustering threshold based on the mean and the standard deviation, and grouping candidate contents with similarities higher than the dynamic clustering threshold into the same content group; extracting scene features from the three-dimensional semantic map, the scene features including spatial topology, semantic attribute distribution, and interaction hotspot distribution, and generating corresponding trigger conditions for each content group based on the scene features; Monitoring scene changes in the three-dimensional semantic map, and when the scene changes meet the triggering conditions of the content group, calculating an update index for the corresponding content group, the update index combining the access frequency, timeliness, and scene relevance of the content group; An update priority of the content group is determined according to the update index. When the update priority exceeds a preset priority threshold, the content with the highest matching degree with the current semantic change trend is selected from the candidate content library as the update strategy.

5. The method according to claim 1, wherein Establishing an interactive behavior model based on the interactive data, and calculating the interactive responsiveness of each content group using the interactive behavior model includes: Extracting interaction time series features, interaction space features, and interaction modal features from the interaction data, and fusing the interaction time series features, the interaction space features, and the interaction modal features to construct a behavior feature vector; Establishing an interactive behavior model based on the behavior feature vector, wherein the interactive behavior model adopts a hybrid structure of a recursive neural network and an attention mechanism, capturing the temporal dependency of the interactive behavior model through the recursive neural network, and highlighting the importance of the interactive features using the attention mechanism to generate a dynamic representation of the interactive behavior model; The interactive responsiveness of the content group is calculated based on the interactive behavior model, the feature vector of the content group is similarly matched with the dynamic representation, and the result of the similarity matching is dynamically weighted by a temporal smoothing factor to obtain a responsiveness score representing the degree of matching between the content group and the interactive behavior.

6. The method according to claim 1, characterized in that Constructing a directed push path graph based on the semantic association between the content groups, wherein the directed push path graph is used to guide the activation sequence of the content groups. Achieving a smooth transition of the augmented reality experience through the orderly activation of the content groups includes: Normalizing the concept weights, entity vectors, and context features in the content group to obtain a standardized feature vector, and constructing a semantic representation of the content group based on the standardized feature vector; Calculating the centrality, density, and betweenness centrality of the content groups, and obtaining node importance based on a weighted combination of the centrality, density, and betweenness centrality; constructing a directed push path graph based on the node importance, wherein the directed push path graph is used to represent activation dependency relationships between content groups; In the directed push path graph, an activation path score is calculated using a conditional probability function that takes into account historical states, wherein the conditional probability function takes the semantic association strength, edge weight, and resource consumption of adjacent content groups as input parameters, and selects the path with the highest score as the optimal activation path; Preloading the content groups to be activated into corresponding cache levels according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representations of the preloaded content groups, and calculating the context relevance in combination with real-time scene information; The content difference and the context relevance are input into an adaptive state smoothing function to dynamically generate transition time between adjacent content groups. The activation process of the content group is controlled according to the transition time to achieve smooth state transition.

7. The method according to claim 6, characterized in that Preloading the content groups to be activated into the corresponding cache level according to the access order of the optimal activation path, calculating the content differences between adjacent content groups based on the semantic representation of the preloaded content groups, and calculating the context relevance in combination with the real-time scene information includes: Constructing a time series feature vector for the content group, the time series feature vector including an access frequency feature, a time interval feature, and a duration feature, and generating a dynamic access pattern for the content group based on the time series feature vector; Calculating a preloading priority score according to the dynamic access pattern, the preloading priority score being a weighted combination of an access frequency score, a time urgency, and a resource requirement, and determining a loading order of the content groups based on the preloading priority score; Allocating cache tiers for the content groups whose loading order is determined, determining an optimal cache tier by calculating the difference between the utility value and storage cost of the content groups at each cache tier, loading the content groups into the corresponding cache tiers, and generating a cache allocation plan; Acquiring multimodal features of the content group that has completed cache allocation, performing weighted fusion on the multimodal features to obtain a fusion feature matrix, and generating a semantic embedding vector based on the fusion feature matrix through a nonlinear mapping function; The scene state information, user behavior information and environmental information corresponding to the semantic embedding vector are collected, the scene state information, the user behavior information and the environmental information are input into a context encoding function to generate a context feature representation, and the context relevance between content groups is calculated based on the context feature representation.

8. An AI-driven interactive AR content intelligent generation and distribution system, configured to implement the method of any one of claims 1 to 7, characterized in that: include: The first unit is configured to obtain scene data collected by a user terminal and construct a three-dimensional semantic map of a target scene based on the scene data; Calculating semantic entropy values ​​for each region in the scene based on the three-dimensional semantic map, constructing a scene semantic evolution tree based on the semantic entropy values, and obtaining a semantic change trend of the target scene by updating node states and conversion weights in the scene semantic evolution tree in real time; A second unit is configured to calculate a matching degree between candidate content in the augmented reality content library and the semantic change trend, divide the candidate content with relevance into content groups according to the matching degree, and generate a corresponding trigger condition and update strategy for each content group based on the three-dimensional semantic map; a third unit configured to collect interaction data between the user and the content group, the interaction data including interaction type, interaction duration, interaction result, and scene state change, establish an interaction behavior model based on the interaction data, and calculate the interaction responsiveness of each content group using the interaction behavior model; The fourth unit is used to classify the content groups according to the interactive responsiveness, determine the push priority based on the classification results, and construct a directed push path graph based on the semantic association between the content groups. The directed push path graph is used to guide the activation order of the content groups, and achieve a smooth transition of the augmented reality experience through the orderly activation of the content groups.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images