An indoor space design method, system, device and medium based on line-of-sight attention mechanism

Abstract

The application relates to an indoor space design method, system, device and medium based on a line-of-sight attention mechanism. The method and scheme generate a multi-modal dynamic line sequence through multi-agent congestion feedback iteration by constructing an indoor space information model and dividing a space topological connected graph; after rendering a panorama along the dynamic line and encoding static and dynamic features, the space convolution double-state interruption recovery memory network is input, local freezing update is performed on a task-oriented state graph, decay update is performed on an interruption response state graph, and weighted fusion is performed, and a visual saliency prediction graph is output; then, the task-oriented weight is reset to zero to separate an interference saliency graph, spatial weighted integration and task key semantic region saliency integration are respectively performed, and a quantitative evaluation value is comprehensively obtained; when the evaluation value does not reach the standard, a mixed mechanism is used to iteratively optimize the model until the standard is reached. Therefore, attention interruption and recovery are explicitly modeled as two independent regulation states, and quantitative evaluation and automatic optimization of space orientation and visual experience under the disturbance of a human flow are realized.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and artificial intelligence, and in particular relates to an indoor space design method, system, device and medium based on gaze attention mechanism. Background Technology

[0002] In the field of architectural interior space design, using digital means to pre-evaluate the directional and aesthetic experience of spatial sequences has become an important way to improve design quality. A typical technical approach is to simulate the walking path of a virtual person in an interior space, use attention mechanism models to predict the visual salience distribution of the virtual person at different observation points, and then quantitatively evaluate key design indicators such as the visibility of directional signs and the overlap of landscape views, thereby achieving a pre-assessment of the spatial experience at the design stage.

[0003] The core of this technical approach lies in accurately modeling the continuous visual attention states formed as a virtual human moves along a path. Existing methods typically use convolutional neural networks to extract visual features from a single frame of a scene, and then use recurrent neural networks to integrate the feature sequences along the time axis to generate a continuous attention map. This type of architecture is based on the assumption that human attention shifting is a smooth process driven by static visual features and task objectives, and can better simulate the pattern of an observer's gaze being attracted to high contrast, high color saturation, or specific semantic elements in a distraction-free scene.

[0004] However, real indoor spaces are not static; the presence and movement of other pedestrians significantly disrupt the observer's attention. The impact of dynamic obstacles goes beyond simply introducing a new visual focus. A deeper cognitive effect lies in the fact that, as a sudden stimulus, it interrupts the observer's existing task-oriented visual exploration strategy, forcing a forced redirection of attention. Crucially, after this interruption, attention does not immediately return to its previous task trajectory but undergoes a brief process of relocation and refocusing. Existing models, failing to explicitly model this dynamic "attention interruption-recovery" process, generate a prediction residual after each dynamic event. This residual propagates along the time series and accumulates with subsequent errors, forming a continuously growing prediction bias that undermines the reliability of evaluations of directional sign visibility and landscape experience.

[0005] Therefore, there is currently no effective method in the technology to make stable and accurate digital predictions and evaluations of the orientation and aesthetic experience of indoor spaces under dynamic pedestrian flow disturbance conditions. Summary of the Invention

[0006] Therefore, it is necessary to provide an interior space design method, system, device, and medium based on the visual attention mechanism to address the aforementioned technical problems.

[0007] Firstly, this application provides an interior space design method based on a line-of-sight attention mechanism, including:

[0008] S1. Extract the geometric elements from the interior space design data of the target building and assign semantic labels to obtain the interior space information model;

[0009] S2. Divide the walkable area in the indoor spatial information model into a navigation grid to obtain a spatial topology connectivity graph;

[0010] S3. Based on the spatial topology connectivity graph and preset crowd composition parameters, generate a multimodal movement sequence through multi-agent crowd feedback iterative path search;

[0011] S4. Render the panoramic view from the main virtual human's viewpoint at each moment in the multimodal motion sequence, and determine the dynamic obstacles based on the positions of other virtual humans; perform visual feature encoding on the panoramic view and obstacle feature encoding on the dynamic obstacles to generate a panoramic environment state tensor sequence.

[0012] S5. Input the panoramic environment state tensor sequence into a spatially convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map. The forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input; it performs local region freezing and updating of the task-oriented state map based on the spatial interruption intensity, and performs decay updating of the interruption response state map based on the duration of no interruption; the two updated state maps are then weighted and fused to output a visual saliency prediction map.

[0013] S6. After resetting the weights corresponding to the task-oriented state map in the weighted fusion to zero, perform the forward calculation again to obtain the interference saliency map generated only by the interrupt response state map decoding; perform spatial weighted integration on the interference saliency map based on semantic labels to obtain the spatial anti-interference index, and perform saliency integration on the task key semantic region of the visual saliency prediction map based on semantic labels to obtain the saliency integral value. Combine the spatial anti-interference index and the saliency integral value to obtain the quantitative evaluation value.

[0014] S7. When the quantitative evaluation value does not meet the standard, the indoor space information model is modified by a hybrid mechanism that combines continuous parameter optimization and discrete structure selection. Then, S2 to S6 are executed again until the quantitative evaluation value meets the standard. The indoor space design scheme is generated based on the modified indoor space information model.

[0015] Secondly, this application also provides an interior space design system based on a gaze attention mechanism to implement the method described in the first aspect, the system comprising:

[0016] Semantic modeling unit is used to extract geometric elements from the target building's interior space design data and assign semantic labels to them to obtain an interior space information model;

[0017] The navigation topology building unit is used to divide the walkable area in the indoor spatial information model into a navigation mesh to obtain a spatial topology connectivity graph.

[0018] The multimodal path planning unit is used to generate a multimodal movement sequence by iteratively searching the path through multi-agent congestion feedback based on the spatial topology connectivity graph and preset crowd composition parameters.

[0019] The panoramic environment encoding unit is used to render a panoramic image from the viewpoint of the main virtual human at each moment in the multimodal motion sequence, determine dynamic obstacles based on the positions of other virtual humans, perform visual feature encoding on the panoramic image and obstacle feature encoding on the dynamic obstacles, and generate a panoramic environment state tensor sequence.

[0020] The dual-state saliency fusion unit is used to input the panoramic environment state tensor sequence into a spatially convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map. The forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input; it performs local region freezing and updating of the task-oriented state map according to the spatial interruption intensity, and performs decay updating of the interruption response state map according to the duration of no interruption; the two updated state maps are then weighted and fused to output a visual saliency prediction map.

[0021] The interference quantification evaluation unit is used to reset the weights corresponding to the task-oriented state map in the weighted fusion to zero and then perform forward calculation again to obtain an interference saliency map generated only by decoding the interruption response state map; the spatial anti-interference index is obtained by spatial weighted integration of the interference saliency map based on semantic labels, and the saliency integral value is obtained by saliency integration of the task key semantic region of the visual saliency prediction map based on semantic labels. The quantitative evaluation value is obtained by combining the spatial anti-interference index and the saliency integral value.

[0022] The closed-loop optimization generation unit is used to modify the indoor spatial information model by adopting a hybrid mechanism that combines continuous parameter optimization and discrete structure selection when the quantitative evaluation value fails to meet the standard. It re-activates the navigation topology construction unit, multimodal path planning unit, panoramic environment coding unit, dual-state saliency fusion unit, and interference quantification evaluation unit to perform corresponding operations until the quantitative evaluation value meets the standard. The indoor spatial design scheme is then generated based on the modified indoor spatial information model.

[0023] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an interior space design method based on a gaze attention mechanism as described in the first aspect.

[0024] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an interior space design method based on a gaze attention mechanism as described in the first aspect.

[0025] The aforementioned indoor space design method, system, device, and medium based on a gaze-attention mechanism constructs an indoor spatial information model containing semantic labels, divides walkable areas into navigation grids to generate a spatial topological connectivity graph, and generates a multimodal movement sequence carrying dynamic crowd features through multi-agent congestion feedback iterative path search. After rendering a panoramic image along the movement lines and encoding static visual features and dynamic obstacle features, the generated panoramic environment state tensor sequence is input into a spatial convolutional dual-state interruption recovery memory network. This network maintains a spatialized task-oriented state graph and an interruption response state graph in parallel. During forward computation, it performs local region freeze updates on the task-oriented state graph based on the detected spatial interruption intensity, and simultaneously updates the system based on the absence of... The interruption duration is used to perform a decay update on the interruption response state map, and then the two state maps are weighted and fused to output a visual saliency prediction map. Subsequently, by resetting the fusion weight corresponding to the task-oriented state map to zero and performing forward calculation again, the interference saliency map generated solely from the decoding of the interruption response state map is separated. Then, the spatial weighted integral of the interference saliency map is performed separately with semantic labels to obtain a spatial anti-interference index, and the saliency integral of the task-critical semantic region of the visual saliency prediction map is performed to obtain a saliency integral value. The combination of the two forms a quantitative evaluation value. When the quantitative evaluation value does not meet the standard, a hybrid mechanism combining continuous parameter optimization and discrete structure selection is used to iteratively modify the indoor spatial information model until the standard is met and a design scheme is output. Thus, this method explicitly models the attention interruption and recovery process caused by dynamic obstacles as two independently controllable states, and can quantitatively evaluate and automatically optimize the guidance clarity and visual experience quality of indoor space under conditions of human traffic disturbance during the scheme design stage. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart illustrating an interior space design method based on a gaze attention mechanism provided by this invention;

[0028] Figure 2 This is a flowchart illustrating the forward computation performed in step S5 of an optional embodiment of the present invention.

[0029] Figure 3 This is a structural schematic diagram of an interior space design system based on a gaze attention mechanism provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0031] This invention relates to an interior space design method based on a gaze-attention mechanism, applied to a computer-aided design system combining Building Information Modeling (BIM) and virtual reality simulation. A typical deployment environment for this system includes a high-performance graphics workstation or cloud server cluster for performing core calculations, and one or more terminal devices for presenting design results and receiving interactive input from designers. The workstation or server is equipped with a BIM parsing engine, a virtual human simulation engine, a panoramic rendering engine, a spatial convolutional dual-state interrupted recovery memory network calculation module, a quantitative evaluation module, and a design optimization module. The terminal devices connect to the workstation or server via a network, receiving target building interior space design data uploaded by designers. After processing the design data, the workstation or server returns the generated visual saliency prediction map, quantitative evaluation values, and optimized design scheme to the terminal devices for visualization. Designers adjust design parameters based on the presented results, and the terminal devices send adjustment instructions to the workstation or server, triggering a new round of simulation evaluation until the design scheme meets the standards.

[0032] The core of this invention lies in the explicit modeling of attention interruption and recovery processes in dynamic pedestrian flow scenarios by constructing and running a spatial convolutional dual-state interruption recovery memory network on the server side. This eliminates the cumulative prediction error caused by ignoring this cognitive dynamic in traditional models, thereby achieving accurate quantitative evaluation and automatic optimization of indoor space orientation and aesthetic experience during the scheme design stage.

[0033] Example 1:

[0034] refer to Figure 1 The document presents a flowchart illustrating an interior space design method based on visual attention mechanisms, as provided in this application. The method includes the following steps:

[0035] S1. Extract the geometric elements from the target building's interior space design data and assign semantic labels to obtain the interior space information model.

[0036] Specifically, the server-side Building Information Modeling (BIM) parsing engine reads the design data files of the target building's interior space in industrial basic formats or other proprietary formats, and parses out all the geometric elements that constitute the interior space. These geometric elements include, but are not limited to, the three-dimensional surface models of components such as walls, floors, ceilings, columns, door and window openings, fixed furniture, partitions, stairs, and ramps. For each type of geometric element, the parsing engine assigns a corresponding semantic label according to a preset semantic classification system. The categories in the semantic classification system include: assigning the ground surface constituting horizontal circulation space the label "walkable ground"; assigning the steps and staircases constituting vertical circulation the label "stairs"; assigning the inaccessible enclosure structure the label "wall"; assigning the light-permeable opening the label "window"; assigning the elements with specific aesthetic intentions placed by the designer, such as sculptures, water features, and art installations, the label "landscape node"; and assigning the key directional signs, safety exit signs, and functional area nameplates the label "guidance signs". The process of assigning semantic labels is achieved by matching the type name or attribute field of the components in the original design data. For components that cannot be automatically matched, designers can manually label them through the terminal. This results in an interior space information model containing accurate three-dimensional geometric information and complete semantic labels, which serves as the unified data foundation for all subsequent calculation steps.

[0037] It should be noted that the role of "semantic labels" is consistent throughout the entire technical solution in this step: in the path generation stage, walkable ground labels are used to extract passable areas, and guidance sign labels are used to evaluate the cost of obtaining path guidance information; in the visual feature encoding stage, semantic labels are encoded as key components of static visual feature vectors; in the quantitative evaluation stage, semantic labels are used to construct a spatial weight map and assign different evaluation weights to different types of areas.

[0038] S2. Divide the walkable area in the indoor spatial information model into a navigation grid to obtain a spatial topology connectivity graph.

[0039] Specifically, the server-side simulation engine first extracts all geometric surfaces marked as "walkable ground" from the indoor space information model and merges these surfaces into a continuous polygonal region. To eliminate the wall-hugging effect caused by the pedestrian's body radius and filter narrow gaps that are not wide enough for a single person to pass through, this polygonal region is offset inward by a distance corresponding to the average shoulder width of an adult. Then, the offset walkable region is segmented into convex polygons. In the basic implementation, the Delaunay triangulation algorithm can be used to triangulate the walkable region, and then adjacent and connected small triangles are merged into larger convex polygonal units. The centroid of each convex polygonal unit is regarded as a topological node. If two convex polygonal units share a sufficiently long boundary and this boundary is not blocked by impassable components such as "walls," then an undirected edge is established between the corresponding two nodes, with the edge weight set to the Euclidean distance between the centroids of the two nodes. Here, a sufficiently long boundary means that the boundary length is greater than a certain threshold to exclude tiny gaps caused by mesh segmentation. For multi-level spaces, cross-level connection edges also need to be added between the navigation mesh nodes corresponding to the starting and ending floors of stairs or elevators. The weight of the cross-level connecting edge is equal to the physical length of the transportation facility multiplied by a time cost coefficient. The time cost coefficient for stairs can be set to 1.5, and the time cost coefficient for elevators can be set to 1.0, to reflect the characteristic that going up and down stairs takes more time than walking horizontally. The resulting graph structure is the spatial topology connectivity graph of the indoor space, which explicitly encodes the accessibility and path distance information between any two points in the space, providing a data structure foundation for subsequent path search.

[0040] The set of nodes in a spatial topologically connected graph is denoted as ,in Indicates the first Topological nodes Let be the total number of nodes. Let the edge set be denoted as . Each edge corresponds to a weight. , representing the travel cost between the two nodes. The formula for calculating the weight of a cross-layer connection edge is:

[0041]

[0042] in, Represents a node With nodes The weights of cross-layer connection edges between them Indicates the physical length of the stairs or elevator; This is a time cost coefficient; for example, 1.5 for stairs and 1.0 for elevators.

[0043] S3. Based on the spatial topology connectivity graph and preset crowd composition parameters, generate a multimodal movement sequence through multi-agent crowd feedback iterative path search.

[0044] Specifically, the server first loads preset crowd composition parameters based on the building type, including the proportion of virtual humans, walking speed distribution, and height distribution for different task modes. Task modes include at least three types: directional pathfinding, roaming and appreciation, and purposeful movement. After configuring the virtual human agents, an iterative multi-agent movement planning mechanism is used to generate movement sequences. The data for each frame time step includes the three-dimensional spatial coordinates, body orientation, and head deflection angle of each virtual human. The head deflection angle is updated every 2 to 3 frames, with a random deflection angle superimposed on the body orientation during the update. The deflection angle ranges from 30 degrees to the left and right of the body orientation. The deviation angle variance is smaller for directional pathfinding tasks to simulate the characteristic of focusing more on the front during pathfinding; the deviation angle variance is larger for roaming tasks to simulate the characteristic of frequently scanning the surroundings during random movement.

[0045] In the first iteration, the cost of regional congestion estimation is initialized to zero, meaning that congestion factors are not considered in the first round of path search. For the orientation pathfinding task, the path search algorithm is run on the spatial topology connected graph, using the nearest reachable node from the virtual human's starting position as the source node and the decision area node in front of the target room entrance as the sink node. In the basic implementation, the A* algorithm is used for path search, and the cost function comprehensively considers the physical length of the path and the cost of obtaining guidance information.

[0046] The total cost function for path search is defined as:

[0047]

[0048] in, For the total cost, The cost is the physical length of the path, which is the sum of the weights of all edges on the path. The cost of acquiring guiding information; This is the distance cost weighting coefficient. The guidance information cost weighting coefficient controls the relative importance of distance and guidance factors in path selection. The formula for calculating the cost of obtaining guidance information is:

[0049]

[0050] in, This indicates the cost of obtaining guidance information. To guide the area of ​​the sign in the virtual human's field of vision, The distance from the virtual person to the marker. The distance is used as a reference. The larger the marker area and the closer the distance, the lower the cost of obtaining navigation information, and the more the path search algorithm tends to choose the path with clear navigation information.

[0051] For the roaming and appreciation task, an attraction-repulsion model is adopted, with landscape nodes acting as attractors and other virtual humans as repulsion sources. Paths are generated by solving Newton's equations of motion. The net force acting on a virtual human at any given moment is the vector sum of the attractive and repulsive forces. For the purposeful passage task, the shortest path search is directly employed. After all virtual humans generate their initial movement paths, the congestion statistics phase begins. The walkable area is divided into navigation grid units, and the number of times each grid unit is traversed by virtual humans in each time interval is counted along the time axis, forming a time-varying congestion field. In time period to Crowding is defined as:

[0052]

[0053] in, Represents navigation grid cells At any moment The congestion value, This indicates the number of virtual people who passed through this unit during that time period. Represents navigation grid cells The area of ​​all units. The crowding degree of all units constitutes a time-varying crowding field.

[0054] Next, the congestion field is fed back into the pathfinding algorithm: for the directed pathfinding task, the congestion cost of the path nodes at the corresponding time points is added to the original cost function, and the corrected total cost function is:

[0055]

[0056] in, For congestion cost weighting coefficient, The cost of estimating regional congestion is equal to the sum of the congestion levels of all nodes on the path at the corresponding time points. After rerunning the path search, a revised path is obtained, and the weighted Hausdorff distance between the revised path and the previous path is calculated. In the calculation of the weighted Hausdorff distance, key paths connecting main entrances / exits and core functional areas are assigned twice the weight of ordinary paths to reflect the importance of key experience paths. The weighted Hausdorff distance is compared with a preset convergence threshold, which is typically set to 1% to 3% of the average length of the initial path. An upper limit is set for the maximum number of iterations, for example, no more than 5 times; if this limit is exceeded, the current best result is taken as the output. If the rate of change is greater than the convergence threshold, the iteration returns to the congestion statistics step until convergence or the upper limit is reached, outputting the final multimodal path sequence.

[0057] S4. Render the panoramic view from the main virtual human's viewpoint at each moment in the multimodal motion sequence, and determine the dynamic obstacles based on the positions of other virtual humans; perform visual feature encoding on the panoramic view and obstacle feature encoding on the dynamic obstacles to generate a panoramic environment state tensor sequence.

[0058] Specifically, the main virtual human refers to the virtual human who carries out the task of evaluating the spatial experience. It is usually selected from the multimodal movement sequence to select the virtual human corresponding to the typical task mode. For example, the virtual human who performs the orientation and pathfinding task is selected as the main virtual human, or the designer specifies the virtual human corresponding to the key experience path on the terminal.

[0059] The server-side panoramic rendering engine renders the panoramic image at each time step, along the position and head orientation of the main virtual human's viewpoint, using an equidistant cylindrical projection format. During rendering, the effective vertical field of view is selected at certain angles above and below the main virtual human's horizon line to cover the main areas of human visual focus.

[0060] Dynamic obstacles are determined as follows: each virtual human agent in the scene other than the main virtual human is considered a potential dynamic obstacle, and its spatial occupation is represented by an axis-aligned bounding box. The size of the bounding box is determined by its height and shoulder width parameters. In a preferred embodiment, dynamic obstacles can be represented by oriented bounding boxes, whose orientation is consistent with the virtual human's body orientation, which can more accurately match the actual spatial occupation of the obstacle and reduce misjudgment of the projection area. The vertices of each bounding box are projected onto the current panoramic image to obtain the coverage area of ​​the obstacle in the image. When the projection areas of multiple obstacles overlap, the overlapping pixels are processed according to the "nearest priority" principle—only the features of the obstacle closest to the main virtual human are retained; if the distances are the same, the features of the obstacle with the largest projection area are retained.

[0061] For each pixel direction in the panoramic image, a static visual feature vector is constructed based on the surface of the spatial component first struck in that direction. Its components include: the normalized position of the struck 2D pixel coordinates in image space, the depth value of the struck surface in screen space, the one-hot encoded vector of the semantic tag to which the surface belongs, and the brightness and contrast components calculated from the material color and local illumination. For pixel directions covered by obstacle projection, a dynamic obstacle feature vector is superimposed on the static visual feature vector. Its components include: a flag indicating the presence of an obstacle in that direction, the normalized distance to the obstacle, the area ratio of the obstacle's bounding box in image space, the trigonometric function encoded values ​​of the obstacle's relative azimuth and elevation angles, and the normalized direction and velocity of the obstacle's motion vector. The obstacle motion vector is obtained by calculating the position difference between the virtual person in the current frame and the previous frame divided by the time step. In a preferred embodiment, a Kalman filter can be used to predict the obstacle's motion trend. The filter uses the trajectory of the previous three frames as observations to estimate its position and velocity state, and predicts the position one time step forward. The image plane offset between the predicted and current positions is encoded into the dynamic obstacle feature vector to simulate human ability to predict obstacle movement. The environmental state vector in each pixel direction is arranged and combined according to the spatial arrangement of the panoramic image to obtain the panoramic environmental state tensor at the current moment. The tensors at all moments along the movement path constitute the panoramic environmental state tensor sequence.

[0062] Panoramic Environment State Tensor The dimension is ,in This represents the discrete resolution of the panoramic image in the vertical direction. This represents the discrete resolution of the panoramic image in the horizontal direction. The total dimension of the environment state vector in each pixel direction is equal to the dimension of the static features. With dynamic feature dimensions The sum of The dimension of the static visual feature vector. Let be the dimension of the feature vectors of dynamic obstacles. Along the entire movement path, obtain the environmental state tensor sequence. ,in This represents the total number of time steps for the movement route.

[0063] S5. Input the panoramic environment state tensor sequence into a spatial convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map. The forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input, performs local region freezing update on the task-oriented state map according to the spatial interruption intensity, and performs decay update on the interruption response state map according to the duration of no interruption. The two updated state maps are then weighted and fused to output a visual saliency prediction map.

[0064] Specifically, the Spatial Convolutional Dual-State Disruption Recovery Memory Network is the core computational model of this invention used to simulate changes in human attention in dynamic scenes. This network is deployed and inferred on the server side.

[0065] The network first inputs the tensor of the panoramic environment state tensor sequence at each time step into the visual encoder, extracts visual features through several convolutional layers and downsampling layers, but does not perform global pooling, thus preserving the visual feature map of spatial structure. ,in This represents the height of the feature map after downsampling. This represents the width of the downsampled feature map. This represents the number of channels in the feature map. This visual feature map serves as the common input to two parallel state update paths.

[0066] The network maintains two spatialized hidden states: a task-oriented state graph. It is used to encode and transmit smooth attention strategies when users perform pathfinding or appreciation tasks under undisturbed conditions, where Number of channels in the task-oriented state diagram; interrupt response state diagram It is used to capture and process attention shifts caused by sudden dynamic obstacle events, in which This represents the number of channels in the interrupt response state graph. Both state graphs are multidimensional arrays that maintain spatial dimensions on their respective channels, with the vector at each spatial location independently encoding the attention history features of that local region.

[0067] Forward computation includes the following key operations within a time step.

[0068] Firstly, interruption detection. The network extracts dynamic obstacle-related feature channels from the input panoramic environment state tensor to construct a dynamic event feature map. Through an interrupt detector Pixel-by-pixel processing is performed to output a spatial interruption intensity map. In the basic implementation, the interruption detector can consist of a single 3×3 convolutional layer and a sigmoid activation function. The input is a dynamic event feature map, and the output is a single-channel spatial interruption intensity map. The detection process can be summarized as a mapping function:

[0069]

[0070] time Each spatial location in the spatial interruption intensity map numerical value This indicates the degree to which the line of sight is threatened by dynamic obstacles. The interruption detector's internal calculation of threat intensity comprehensively considers multiple components of the dynamic obstacle feature vector along the pixel direction corresponding to that spatial location, specifically including the normalized value of the obstacle distance. Area ratio of obstacles in image space and the speed of obstacle movement The formula for calculating threat intensity can be:

[0071]

[0072] in, The sigmoid activation function maps the output to the interval between 0 and 1. This is the normalized distance from the obstacle to the main virtual character, with a value ranging from 0 to 1. The smaller the value, the closer the obstacle is. This represents the percentage of the projected area of ​​the obstacle bounding box in that pixel direction. The normalized rate of the obstacle's motion vector; This is the distance weighting coefficient. This is the area weighting coefficient. These are the velocity weight coefficients. All three weight coefficients are learnable parameters that are automatically adjusted through backpropagation during training. To prevent small constants from being divided by zero; This is the bias term. As can be seen from the formula, the closer the obstacle (...), the more biased the bias term. The smaller the size, the larger the projected area, and the higher the movement speed, the higher the threat intensity value. The larger the value, the more alert humans are to obstacles that are close-range, large-area, and approaching at high speed. If a single convolutional layer and a sigmoid activation function are used to implement the interruption detector, the aforementioned weight coefficients and bias terms are implicitly expressed by the kernel parameters and bias parameters of that convolutional layer.

[0073] Secondly, the task-oriented state graph undergoes partial freeze-up updates. The network generates a spatial suppression gate based on the spatial interruption intensity graph. In the basic implementation, the value of the spatial suppression gate at each location is determined by the complement of the interruption strength value at that location—the gate value is maximum in the uninterrupted region, indicating normal state updates; the gate value tends to be minimum in the strongly interrupted region, indicating that state updates are strongly suppressed. The formula for calculating the spatial suppression gate is:

[0074]

[0075] During the update process of the task-oriented state graph, the network first calculates the candidate update content at that moment. Update Gate and reset gate Both are generated by convolutional operations between the current visual feature map and the historical task-oriented state map. Specifically, the visual feature map... Historical mission-oriented state diagram The layers are concatenated along the channel dimension, then passed through a 3×3 convolutional layer and a sigmoid activation function to obtain the update gates. and reset gate The reset gate controls the influence of historical states on candidate update content. When the reset gate value is close to zero, historical states are ignored, and candidate content is generated solely based on the current visual feature map. The formula for calculating candidate update content is:

[0076]

[0077] in, This represents the convolution operation. This indicates element-wise multiplication. For input The convolution kernel weight matrix to the candidate state, For historical state The convolution kernel weight matrix to the candidate state, is the bias vector of the candidate state.

[0078] Then, the spatial suppression gate is broadcast to the same size as the candidate content, and the suppressed candidate features are obtained by multiplying them element-wise. Simultaneously, the historical task-oriented state graph is multiplied by the suppression completion coefficients to obtain the suppression completion term. The suppression completion coefficients are determined by the spatial suppression gate, and their calculation method is to take the complement of the value of the spatial suppression gate at each spatial location, i.e.:

[0079]

[0080] in, Indicates time The suppression completion coefficient in spatial location The value at the given location. As can be seen from this formula, in the uninterrupted region, the spatial inhibition threshold is approximately 1, the inhibition completion coefficient is approximately 0, and the historical state is hardly used for filling; in the strongly interrupted region, the spatial inhibition threshold tends to 0, the inhibition completion coefficient tends to 1, and the suppressed candidate update content is filled by the historical state at a ratio of approximately 1:1, thereby numerically achieving the "freezing" of task attention in this region.

[0081] The suppressed candidate features are added to the suppressed completion terms to form the total candidate update after spatially selective suppression. The complete update formula for the task-oriented state graph is:

[0082]

[0083] in, This is the update gate for the task-oriented state graph, where each element has a value between 0 and 1, controlling the fusion ratio between historical states and candidate content.

[0084] This mechanism simulates the local interruption of human attention: when a moving pedestrian suddenly appears in a certain part of the field of vision, only the established gaze plan for that area is temporarily frozen, while other areas can continue their original visual exploration.

[0085] Third, the decay update of the interrupt response state graph. The network maintains a continuous, uninterrupted time step counter. At each time step, check if the global maximum value of the spatial interruption intensity map is lower than a preset interruption threshold—if so, increment the counter; otherwise, reset it to zero. Based on the current value of the counter, dynamically calculate the attenuation factor. The attenuation factor decreases monotonically as the counter increases, and its calculation formula is as follows:

[0086]

[0087] in, Based on the attenuation rate, The attenuation coefficient is... This represents the current number of uninterrupted steps. The interruption response state map is updated by extracting feature sub-maps related to dynamic events from the visual feature map and obtaining candidate states through convolution. Then, the retention rate of the historical interruption response state diagram is modulated using an attenuation factor. The complete update formula for the interruption response state diagram is:

[0088]

[0089] This mechanism simulates the cognitive process of attention gradually recovering from distractions: after the obstacle disappears, its influence on attention does not instantly drop to zero, but gradually weakens over time, much like the gradual fading of "distraction aftereffects".

[0090] Fourth, weighted fusion and decoding output. The network concatenates the visual feature map, the updated task-oriented state map, and the updated interruption response state map along the channel dimension, and generates a spatial fusion gate through a convolutional layer. :

[0091]

[0092] in, This indicates concatenation along the channel dimension. The sigmoid activation function maps the output to the interval between 0 and 1. This represents a 1×1 convolution operation. The fusion gate is used to perform a pixel-by-pixel weighted summation of the two state maps to obtain the fused state map. :

[0093]

[0094] Finally, the decoder, consisting of several transposed convolutional layers and convolutional layers, upsamples the fused state map to the original panoramic image resolution and outputs a single-channel visual saliency prediction map. —The value of each pixel in the image represents the probability of the observer allocating attention to that direction.

[0095] S6. After resetting the weights corresponding to the task-oriented state map in the weighted fusion to zero, perform forward calculation again to obtain the interference saliency map generated only by the interrupt response state map decoding; perform spatial weighted integration on the interference saliency map based on semantic labels to obtain the spatial anti-interference index, and perform saliency integration on the task key semantic region of the visual saliency prediction map based on semantic labels to obtain the saliency integral value. Combine the spatial anti-interference index and the saliency integral value to obtain the quantitative evaluation value.

[0096] Specifically, the server-side quantization evaluation module utilizes the internal decomposition capability of the two-state network to separate the attention interference components purely caused by dynamic obstacles from the visual saliency prediction map. The specific operation involves performing ablation-based forward computation, i.e., keeping the input panoramic environment state tensor sequence unchanged, rerunning the network forward computation, but forcibly setting all values ​​of the spatial fusion gate to zero during this process—this means the fused state map will only be contributed by the interrupt response state map, i.e. The decoder outputs an interference saliency map after decoding the state diagram. This figure reveals the pure shift in attention caused by dynamic pedestrian flow. Similarly, a pure task attention map can be obtained by performing another ablation-based forward computation: keeping the input unchanged, forcing all values ​​of the spatial fusion gate to 1, so that the fusion state map is contributed only by the task-oriented state map, and the decoder output is the pure task attention map. ,in This indicates a clean task attention map in the pixel direction. The value at that location.

[0097] The quantitative evaluation consists of two parallel computational paths. The first path assesses the space's resilience: using the semantic label categories of each geometric element in the indoor spatial information model, a spatial weight value is assigned to each pixel orientation in the panoramic image. Based on the fundamental implementation method, different semantic labels correspond to different weights, with wayfinding signs and landscape nodes assigned higher weights, and ordinary areas assigned a baseline weight. This spatial weighting map is then used. The cumulative attentional shift is obtained by performing frame-by-frame spatial weighted integration of the interference saliency map along the entire motion path. :

[0098]

[0099] in, This represents the cumulative attention offset. Indicates the total number of time steps. Indicates the domain of the panoramic image. The spatial weight map represents the pixel orientation. The weight value at that point, Indicates time The saliency map of interference in the pixel direction The value at this location. The larger this value, the greater the degree to which human attention is continuously disturbed by crowds on the critical task path. Based on this cumulative attention shift, a spatial anti-interference index is further derived. The spatial robustness index can be defined as the ratio of the task-oriented attention integral value to the interference attention integral value within a critical task area. Specifically, it is calculated as follows: first, the integral of the clean task attention map in the critical task area is used as the task attention index; second, the integral of the interference saliency map in the same area is used as the interference attention index. The ratio of these two is the spatial robustness index for that area. The higher the ratio, the stronger the spatial robustness.

[0100] The second approach is used to evaluate the inherent directional and aesthetic effectiveness of the space: geometric elements labeled "wayfinding signs" and "landscape nodes" in the indoor spatial information model are projected onto a visual saliency prediction map using perspective projection, resulting in the projection areas for wayfinding signs and landscape nodes, collectively referred to as the task-critical semantic regions. The saliency values ​​of these regions on the visual saliency prediction map are then integrated to obtain the saliency integral value. This value reflects the design scheme's ability to effectively guide attention to key objectives; a higher value indicates stronger visual appeal of wayfinding signs and landscape nodes.

[0101] After completing the calculations for the two lines, the spatial anti-interference index and significance integral value are first normalized to the [0,1] interval, and then the comprehensive quantitative evaluation value is calculated according to the following formula:

[0102]

[0103] in, This represents the comprehensive quantitative evaluation value. This represents the normalized spatial anti-interference index. This represents the significance integral value after normalization. The anti-interference index weighting coefficient is... This is the weighting coefficient for the significance integral value, which can be adjusted according to project needs. The default value is 0.5. This value comprehensively describes the guidance and aesthetic experience quality of the interior space sequence from two dimensions: "static design guidance" and "dynamic pedestrian flow resistance".

[0104] Furthermore, an attention-guided continuity index for spatial sequences can be calculated. Specifically, using a sequence of clean task attention maps generated by decoding the task-guided state map separately, the centroid displacement vectors of adjacent frames of clean task attention maps are calculated, and then the average absolute angle of the displacement vectors over five consecutive frames is calculated. The average absolute angle is calculated by averaging the angles between the displacement vectors of every two adjacent frames over five consecutive frames. Centroid coordinates of time The calculation formula is:

[0105]

[0106]

[0107] in, In order to be in The clean task attention map at any given moment in the pixel direction The value at that location, and These are the horizontal and vertical coordinates of the pixel's orientation in the panoramic image, respectively. displacement vector at time t The smaller the average absolute angle, the more coherent the visual guidance in the space; less than 30° is considered excellent continuity, 30° to 60° is good, and greater than 60° indicates a significant break in the guidance. This indicator can pinpoint the weak points in the visual guidance sequence of the design scheme.

[0108] S7. When the quantitative evaluation value does not meet the standard, the indoor space information model is modified by a hybrid mechanism that combines continuous parameter optimization and discrete structure selection. Then, S2 to S6 are executed again until the quantitative evaluation value meets the standard. The indoor space design scheme is generated based on the modified indoor space information model.

[0109] Specifically, the server-side design optimization module determines whether the quantitative evaluation values ​​meet the various compliance conditions preset by the designer through the terminal. If the standards are not met, the optimization module first divides the design parameters associated with the non-compliant indicators into two categories: one category is continuous design parameters that do not change the spatial topological connectivity, such as the spatial coordinates, three-dimensional orientation, geometric dimensions, material color, and self-illumination intensity of directional signs; the other category is discrete design parameters that change the spatial topological connectivity, such as wall positions, corridor widths, and stair opening positions.

[0110] In the basic implementation, continuous parameters are optimized using an approximate gradient method. Small positive and negative perturbations are applied sequentially to the parameters to be optimized, with the perturbation step size typically set to 1% of the current value of the parameter. For each perturbation, the entire process from trajectory generation to quantitative evaluation is re-executed on the model, and the design target loss after each perturbation is calculated. The design target loss is defined as the negative value of the comprehensive quantitative evaluation.

[0111]

[0112] in, To design the target loss, This is the comprehensive quantitative evaluation value. The optimization objective is to maximize the comprehensive evaluation value by minimizing this loss. The ratio of the loss difference obtained from two perturbations to the perturbation amount is used as the approximate gradient of this parameter, and the parameter is updated along the gradient descent direction. The learning rate of gradient descent is set between 0.1 and 0.5, and an upper limit is set on the number of iterations for continuous parameter optimization in a single round, for example, no more than 20 times. Linear inequality constraints are applied during the optimization process to ensure that the guide signs are located above the walkable area and do not overlap with fixed components such as walls, ensuring the physical feasibility of design modifications.

[0113] For discrete parameters, an exhaustive search method is used to select the best option: a finite number of discrete modification alternatives that meet the requirements are generated according to the design specifications. For each alternative, the entire process from generating the movement line to quantitative evaluation is performed, and the option that makes the quantitative evaluation value optimal is selected as the modification direction.

[0114] After each round of parameter modifications, changes in spatial layout or signage location may alter pedestrian flow distribution and individual movement patterns. The optimization module re-executes movement pattern generation and crowding field calculations based on the modified indoor spatial information model to obtain evaluation indicators under the new movement pattern conditions, forming a nested iterative closed loop of "design modification - movement pattern regeneration - re-evaluation." When multiple conflicting evaluation indicators exist, the Pareto optimality principle from multi-objective optimization is used. Specifically, the NSGA-II algorithm can be employed, generating the next generation of design schemes through selection, crossover, and mutation operations. After several generations of evolution, the Pareto optimal frontier is obtained. Designers can select the most suitable design scheme from the Pareto frontier based on the specific needs of the project. The iterative process continues until all quantitative indicators meet the criteria, at which point the indoor spatial information model at this point is output as the final optimized scheme to the terminal.

[0115] The aforementioned indoor space design method based on a gaze-attention mechanism constructs an indoor spatial information model containing semantic labels, divides walkable areas into navigation grids to generate a spatial topological connectivity graph, and generates a multimodal movement sequence carrying dynamic features of the crowd through multi-agent congestion feedback iterative path search. After rendering a panoramic image along the movement path and encoding static visual features and dynamic obstacle features, the generated panoramic environment state tensor sequence is input into a spatial convolutional dual-state interruption recovery memory network. This network maintains a spatialized task-oriented state graph and an interruption response state graph in parallel. In the forward computation, the task-oriented state graph is updated by freezing local regions based on the detected spatial interruption intensity, and the update is also performed based on the duration of no interruption. The interruption response state map is updated using a decay process, and then the two state maps are weighted and fused to output a visual saliency prediction map. Subsequently, by resetting the fusion weights corresponding to the task-oriented state map to zero and performing forward computation again, the interference saliency map generated solely from the interruption response state map decoding is separated. Then, spatial weighted integration is performed on the interference saliency map using semantic labels to obtain a spatial anti-interference index, and saliency integration is performed on the task-critical semantic region of the visual saliency prediction map to obtain a saliency integral value. The combination of these two values ​​forms a quantitative evaluation value. When the quantitative evaluation value fails to meet the standard, a hybrid mechanism combining continuous parameter optimization and discrete structure optimization is used to iteratively modify the indoor spatial information model until the standard is met and a design scheme is output. Thus, this method explicitly models the attention interruption and recovery process caused by dynamic obstacles as two independently controllable states, allowing for quantitative evaluation and automatic optimization of the guidance clarity and visual experience quality of the indoor space under conditions of pedestrian disturbance during the scheme design stage.

[0116] Example 2:

[0117] refer to Figure 2 Step S5 includes the following steps:

[0118] S11. Input the panoramic environment state tensor at each time step in the panoramic environment state tensor sequence into the visual encoder. Through convolution and downsampling processing, obtain a visual feature map that preserves the spatial structure. Perform convolution operation on the visual feature map to generate the update gate and candidate update content of the task-oriented state map.

[0119] Specifically, the visual encoder consists of multiple stacked convolutional and max-pooling layers. In a preferred embodiment, the encoder includes three convolutional layers and two max-pooling layers, with convolutional kernel sizes of 5×5, 3×3, and 3×3, respectively, all with a stride of 1. The pooling kernel size is 2×2 with a stride of 2. No global pooling or fully connected layers are placed at the end of the encoder; therefore, the output visual feature map retains the spatial structure corresponding to the input panoramic image, with a spatial size that is a fraction of the input size. The number of output channels of the visual encoder can be selected from 64 to 128, and the number of channels in the task-oriented state map... The number of channels in the interrupt response status diagram can be set to 64. It can be set to 32. This visual feature map that preserves the spatial structure is the basis for all subsequent spatialized state operations of the entire network.

[0120] After obtaining the visual feature map, the network applies a 3×3 convolution operation to the feature map. The number of convolution kernels is set to the sum of the number of channels in the task-oriented state map and 2. The number of channels in the task-oriented state map is used to generate candidate update content, and the additional 2 channels are used to generate the update gate and the reset gate, respectively. Both the reset gate and the update gate use the sigmoid activation function, and their output value range is [0,1].

[0121] It should be noted that in Example 1, the network uses a shallow encoder and a large stride to achieve basic attention prediction with low computational cost. This example, by increasing the convolution depth and using fine convolutions with small strides, obtains richer visual feature representations, thereby improving the accuracy of subsequent state updates.

[0122] S12. Extract dynamic obstacle components from the panoramic environment state tensor to form a dynamic event feature map. Input the dynamic event feature map into the interruption detector and obtain the spatial interruption intensity map through convolution operation.

[0123] Specifically, the panoramic environment state tensor is composed of static visual feature components and dynamic obstacle feature components stitched together along the channel dimension. The network extracts only those channels containing dynamic obstacle feature components from the panoramic environment state tensor according to a preset channel index range, forming a dynamic event feature map. This feature map preserves the spatial layout of the panoramic image.

[0124] The interruption detector is a small fully convolutional network consisting of two 3×3 convolutional layers. The first convolutional layer maps the input channels to several intermediate feature channels, and the second convolutional layer compresses the intermediate features into a single-channel output. The output layer uses a sigmoid activation function to ensure that each pixel value falls within the [0,1] interval. This single-channel output is the spatial interruption intensity map, and the values ​​at each spatial location in the map collectively reflect the threat level of dynamic obstacles in that direction.

[0125] S13. Generate a spatial suppression gate based on the spatial interruption intensity map, use the spatial suppression gate to spatially modulate the candidate update content, and use the update gate to temporally fuse the historical task-oriented state map with the modulated candidate content to obtain the updated task-oriented state map.

[0126] Specifically, the network generates a spatial suppression gate based on the spatial interruption intensity map. Each element of the spatial suppression gate is determined by subtracting the interruption intensity value at the corresponding location from 1, and its calculation formula has been given in Example 1. Then, the spatial suppression gate is broadcast along the channel dimension to a size identical to the candidate update content, ensuring that locations with strong spatial threats correspond to lower suppression gate values ​​across all channels. The broadcast spatial suppression gate is then multiplied element-wise with the candidate update content to obtain suppressed candidate features. Simultaneously, suppression completion coefficients are calculated, which involves inverting each value in the spatial suppression gate—the completion coefficient for locations without interruption tends to 0, and the completion coefficient for locations with strong interruption tends to 1. The historical task-oriented state map is then multiplied element-wise with these suppression completion coefficients to obtain a suppression completion term. The purpose of this completion term is to fill the gaps in regions where candidate updates are strongly suppressed with historical states, simulating a "freeze" of task attention.

[0127] The suppressed candidate features are added element-wise to the suppression completion terms to form the total candidate update quantity after spatially selective suppression. Finally, an update gate is used to interpolate and fuse the historical task-oriented state map and the total candidate update quantity along the time dimension—at positions where the update gate value is close to 1, the new total candidate update quantity is favored; at positions where the update gate value is close to 0, the historical state is favored to be preserved—the update is performed independently at each spatial location to obtain the updated task-oriented state map. The complete formula for this update has been given in Example 1.

[0128] Through the combined operation of spatial modulation and temporal fusion described above, this step achieves local selective control of task-oriented attention: task state updates in areas with dynamic obstacles are suppressed, and the attention strategy in these areas remains stable; while the attention strategy in barrier-free areas can continue to evolve based on new visual information.

[0129] S14. Extract dynamic feature sub-maps from the visual feature map, perform convolution operations on the dynamic feature sub-maps, and generate candidate states for the interrupt response state map.

[0130] Specifically, the network selects several channels related to dynamic event perception from the visual feature map based on a pre-defined dynamic feature channel index, forming a dynamic feature sub-map. Since the channels of the visual encoder spontaneously differentiate response preferences to different visual attributes during training, channels related to dynamic features such as motion and depth changes are pre-identified and have fixed indices. A 3×3 convolution operation is applied to this dynamic feature sub-map to generate candidate states for the interruption response state map. This candidate state encodes information about newly appearing dynamic obstacles at the current moment.

[0131] S15. Based on the comparison between the global maximum value of the spatial interruption intensity map and the preset threshold, the continuous uninterrupted time step counter is incremented or cleared to obtain the current uninterrupted duration step count; the attenuation factor is calculated based on the current uninterrupted duration step count, and the historical interruption response state map and candidate state are weighted and updated using the attenuation factor to obtain the updated interruption response state map.

[0132] Specifically, the network maintains a continuous, uninterrupted time step counter, which is updated in each forward calculation. The update rule is as follows: calculate the maximum value of all spatial locations in the spatial interruption intensity map at the current moment; if the maximum value is less than a preset interruption judgment threshold, it indicates that there are no significant dynamic obstacles in the current field of view, and the counter is incremented by 1; if the maximum value is greater than or equal to the judgment threshold, it indicates that there are still dynamic obstacles, and the counter is reset to zero. The typical value of the interruption judgment threshold can be selected in the range of 0.1 to 0.3. The value of this counter is the current number of uninterrupted time steps. .

[0133] The attenuation factor is calculated based on the current number of uninterrupted steps, and its calculation formula has been given in Example 1. When the dynamic obstacle persists, the counter is 0, and the attenuation factor remains at a relatively high base attenuation level—at this time, the historical interruption response state map is sufficiently preserved, and new candidate state information is continuously incorporated. When the obstacle disappears, the counter begins to accumulate, and the attenuation factor gradually decreases—the contribution of the attenuated historical state obtained by multiplying the historical interruption response state map by the attenuation factor gradually weakens, while the contribution of the current candidate state correspondingly strengthens. After the uninterrupted state has lasted for a sufficiently long time, for example, when the number of uninterrupted steps reaches 20, the attenuation factor is clamped to a preset lower limit value, such as 0.5, to retain a certain amount of historical interference traces and avoid the complete loss of historical information.

[0134] In the update operation, the historical interruption response state map is first multiplied element-wise by the decay factor to obtain the decayed historical state; then, the decayed historical state and the candidate state are weighted and summed according to a preset ratio to obtain the updated interruption response state map. The complete formula for this update has been given in Example 1. This update mechanism based on a dynamic decay factor explicitly models the time dynamic process of human attention gradually recovering from an interruption event, eliminating the prediction error introduced by the instantaneous return of attention to the task trajectory in traditional models.

[0135] S16. After concatenating the visual feature map, the updated task-oriented state map, and the updated interruption response state map along the channel dimension, a spatial fusion gate is generated through convolution operation. The spatial fusion gate is used to perform pixel-by-pixel weighted fusion of the updated task-oriented state map and the updated interruption response state map to obtain a fused state map. The fused state map is input into the decoder and processed by transpose convolution and convolution to obtain a visual saliency prediction map.

[0136] Specifically, the visual feature map, the updated task-oriented state map, and the updated interruption response state map are concatenated into a composite feature map along the channel dimension. A 1×1 convolution operation is applied to this composite feature map, mapping the concatenated multi-channel information to a single-channel output. The output layer uses a sigmoid activation function, and this single-channel feature map serves as the spatial fusion gate. The formula for generating the spatial fusion gate has been given in Example 1. The spatial fusion gate is used to perform pixel-wise weighted fusion of the two state maps, and the formula for calculating the fused state map is: .

[0137] The decoder consists of alternating transposed convolutional layers and regular convolutional layers. The transposed convolutional layers are responsible for progressively increasing the spatial size of the feature map, while the regular convolutional layers smooth and refine the features during the size restoration process. By using a preset upsampling rate, the decoder progressively restores the spatial size of the fused state map to a resolution completely consistent with the original panoramic environment state tensor. The final output of the decoder is a single-channel image, which, after normalization, becomes the visual saliency prediction image for the current moment—the higher the value in the image, the more likely that direction is to become the observer's focus of attention at the current moment.

[0138] Example 3:

[0139] The candidate update content is spatially modulated using a spatial suppression gate, and the historical task-oriented state graph is temporally fused with the modulated candidate content using an update gate to obtain the updated task-oriented state graph. This process includes the following steps:

[0140] S21. Multiply the spatial suppression gate element by element with the candidate update content to obtain the suppressed candidate features.

[0141] Specifically, the spatial interruption intensity map obtained in step S12 is transformed into a spatial suppression gate as follows: the value of each position of the spatial suppression gate is determined by subtracting the spatial interruption intensity value at that position from 1. During training or inference, the spatial interruption intensity map is output by the interruption detector after convolving the dynamic event feature map, and its value range is [0,1]. Positions with values ​​close to 1 indicate the presence of a high-intensity dynamic obstacle threat, while those close to 0 indicate the opposite. The formula for calculating the spatial suppression gate is: .in, Indicates the spatial location of the spatial inhibition gate The value at that location, Indicates the spatial interruption intensity map at the same spatial location The threat intensity value at a given location. As can be seen from the formula, the stronger the threat in a region, the smaller its spatial inhibition threshold, and the stronger the subsequent inhibition of candidate update content.

[0142] Subsequently, the spatial suppression gate is replicated and broadcast along the channel dimension, making its channel number the same as the candidate update content. The broadcast spatial suppression gate and each channel of the candidate update content are then multiplied element-wise at their corresponding spatial locations. In regions of high threat intensity, the candidate update content is significantly reduced or even zeroed out, preventing new information from the current visual input from being written into the task-oriented state; in regions of low threat intensity, the candidate update content is almost completely preserved. The result is the suppressed candidate feature.

[0143] S22. Calculate the suppression completion coefficient based on the spatial suppression gate, and multiply the historical task-oriented state diagram with the suppression completion coefficient element by element to obtain the suppression completion term.

[0144] Specifically, the suppression completion coefficient is derived from the spatial suppression gate. Its calculation involves taking the complement of the spatial suppression gate value at each location, i.e., subtracting the spatial suppression gate value at that location from 1. Physically, the suppression completion coefficient determines the degree to which the historical task-oriented state fills the suppressed information in the corresponding region—the more candidate updates are suppressed at a location, the larger the suppression completion coefficient, requiring more historical states to fill the gaps and maintain the consistency of the state at that location. The suppression completion term is obtained by element-wise multiplying the historical task-oriented state graph with the suppression completion coefficient broadcast to it, matching its size.

[0145] S23. Add the suppressed candidate features to the suppression completion terms to obtain the total candidate update amount; use the update gate to interpolate and fuse the historical task-oriented state graph and the total candidate update amount to obtain the updated task-oriented state graph.

[0146] The formula for calculating the spatial suppression gate is as follows:

[0147]

[0148] in, express The spatial inhibition gate of time in spatial position The value at that location, express Spatial interruption intensity map at time of moment in spatial location Threat intensity value at the location.

[0149] Specifically, the suppressed candidate features and the suppression complements are added element-wise at their corresponding spatial locations. After addition, in areas severely threatened by dynamic obstacles, the total candidate update is mainly contributed by the suppression complements composed of historical states, while the influence of the current visual input is marginalized; in areas without threat, the total candidate update is mainly composed of the current candidate update, with historical states participating very little. This design numerically achieves local selective freezing of task-oriented attention.

[0150] Finally, the update gate acts as a fusion controller in the temporal dimension. The update gate is learned by the network through convolutional operations based on the current visual feature map and historical states; each element's value is between 0 and 1. The update gate is used to interpolate and fuse the historical task-oriented state map and the total candidate update values—positions with update gate values ​​close to 1 tend to adopt the total candidate update values, while positions with update gate values ​​close to 0 tend to retain historical states. Interpolation is performed independently at each spatial location, ultimately obtaining the updated task-oriented state map. The complete formula for the entire update process has been given in Example 1 and will not be repeated here.

[0151] Example 4:

[0152] The historical interrupt response state diagram and candidate states are weighted and updated using a decay factor to obtain the updated interrupt response state diagram, including the following steps:

[0153] S31. Multiply the historical interruption response state diagram element by element with the attenuation factor to obtain the attenuated historical state.

[0154] Specifically, the historical interruption response state map is multiplied element-wise by a decay factor. The decay factor is a scalar that changes dynamically over time, calculated by the network at each time step based on the number of uninterrupted steps. After being broadcast to the same spatial size and number of channels as the historical interruption response state map, it is applied element-wise to each component of the historical state. When the decay factor is large, more activation values ​​from the historical interruption response state map are retained, indicating that the network still maintains a strong memory of the most recent interruption event; when the decay factor is small, the historical interruption response state map is suppressed, and interference traces at corresponding locations fade more quickly.

[0155] S32. The attenuated historical state and the candidate state are weighted and summed to obtain the updated interrupt response state diagram.

[0156] The formula for calculating the attenuation factor is as follows:

[0157]

[0158] in, Indicates time The attenuation factor, Indicates the basic attenuation rate. Indicates the attenuation coefficient. This indicates the number of steps that can continue without interruption.

[0159] Specifically, the calculation of the decay factor is based on a continuous uninterrupted time step counter. Based on the current number of uninterrupted steps obtained from the counter, the decay factor is dynamically calculated using the following formula: .in, Indicates time The attenuation factor, Indicates the basic attenuation rate. Indicates the attenuation coefficient. This represents the current number of uninterrupted steps. As can be seen from the formula, when dynamic obstacles continuously appear, the counter... Keep it at 0, the exponent term is The attenuation factor is equal to the basic attenuation rate. When the obstacle disappears, the counter... Start by gradually accumulating, exponential term Gradual decay, leading to a decay factor The decay rate gradually decreases from the baseline. Parameters Control the sensitivity of the decay rate relative to the number of counting steps.

[0160] The complete update of the interrupt response state diagram is expressed by the following formula:

[0161]

[0162] in, This represents the updated interrupt response status diagram. This represents a historical interruption response state diagram. Indicates the candidate state. As the attenuation factor, This indicates element-wise multiplication. In the formula, the attenuation factor... It directly determines the fusion ratio of historical states and candidate states: when When the value is large, more historical interruption response state diagrams are retained; when When the attenuation factor is smaller, the contribution of the candidate state is more significant. By dynamically adjusting this attenuation factor, the network intrinsically simulates the time process of interference information gradually fading from the interrupted response state, achieving a gradual recovery of attention rather than an instantaneous jump.

[0163] Example 5:

[0164] The training process of a spatial convolutional dual-state interrupted recovery memory network includes the following steps:

[0165] S41. Obtain the visual environment sample sequence and real eye movement fixation data, perform saliency map transformation on the real eye movement fixation data, and generate a baseline saliency map sequence.

[0166] Specifically, training data was collected in a virtual reality environment. Participants wore eye-tracking devices and performed designated tasks in a virtual indoor space constructed according to a pre-designed scheme. The participants' head posture and the points of eye fixation in three-dimensional space were continuously recorded, forming a sequence of real eye-tracking fixation data. The fixation point at each time step was used as the center of a two-dimensional Gaussian kernel, with the kernel's diffusion range matching the physiological uncertainty of fixation. This Gaussian kernel was superimposed on a blank image of the same resolution as the panoramic image to obtain the baseline saliency map for that moment. The baseline saliency maps for all time steps were arranged chronologically to form a baseline saliency map sequence.

[0167] S42. Render and encode the visual environment sample sequence according to the method in S4 to generate a panoramic environment state sample tensor sequence.

[0168] Specifically, the visual environment sample sequences used for training employ the same rendering and encoding process as the inference phase. This involves rendering isometric cylindrical projection panoramic images along the movement path at each viewpoint, identifying dynamic obstacles based on the positions of other virtual humans, encoding obstacle features, and generating a panoramic environment state sample tensor sequence. This step ensures the consistency of the distribution between training and inference data. During training, to improve the model's generalization ability, data augmentation processing can be performed on the rendered and encoded panoramic images, including random rotation, brightness adjustment, and contrast adjustment, to expand the diversity of training samples.

[0169] S43. Input the tensor sequence of panoramic environment state samples into the spatial convolutional dual-state interruption recovery memory network for forward computation to obtain the corresponding prediction saliency map sequence, spatial interruption intensity map and prediction fusion gate.

[0170] Specifically, the forward computation process during the training phase is the same as that during the inference phase. The network processes the tensor sequence of panoramic environment state samples, outputs a sequence of predicted saliency maps, and simultaneously outputs the spatial interruption intensity map at each time step and the predicted fusion gate generated during the fusion phase. .

[0171] S44. Calculate the similarity loss between the predicted saliency map sequence and the benchmark saliency map sequence, and use the similarity loss as the basic loss term; generate the target fusion gate based on the spatial interruption intensity map, and calculate the deviation between the predicted fusion gate and the target fusion gate as the decoupling constraint loss term in the region where the interruption intensity exceeds the preset threshold.

[0172] The calculation formula for the target fusion gate is as follows:

[0173]

[0174] in, express The target of the moment merges the door in spatial position The value at that location, Indicates the impact coefficient of interruption intensity. express Spatial interruption intensity map at time of moment in spatial location Threat intensity value at the location.

[0175] Specifically, the composite loss function consists of two parts: a basic loss term and a decoupling constraint loss term. The basic loss term measures the difference between the network's predicted attention distribution and the actual behavior. It uses Kullback-Leibler divergence as a metric, calculating the distribution difference between the predicted saliency map sequence and the baseline saliency map sequence frame by frame, and averaging over all time steps. The formula for the basic loss term is:

[0176]

[0177] in, Indicates the basic loss term. Indicates the total number of time steps. Indicates in The baseline saliency map at time step in the pixel direction The value at that location, Indicates in Predictive saliency map at time step in pixel direction The value at that location, To prevent small constants from being divided by zero.

[0178] The design of the decoupling constraint loss term is the core innovation of the training method in this invention. Its goal is to force the network to reasonably separate the contributions of the task-oriented state and the interruption response state in spatiotemporal regions with dynamic obstacles, preventing the network from fitting all scenarios based solely on one state. First, a target fusion gate is generated based on the spatial interruption intensity map at each time step. The calculation formula for the target fusion gate is:

[0179]

[0180] in, express The target of the moment merges the door in spatial position The value at that location, This represents the influence coefficient of the interruption intensity on the fusion target. express Spatial interruption intensity map at time of moment in spatial location The threat intensity value at the location. As can be seen from the formula, in the uninterrupted region, the threat intensity value is 0, the target fusion threshold is approximately 1, and the desired task-oriented state is completely dominant; in the strongly interrupted region, the threat intensity value is close to 1, and the target fusion threshold is approximately... That is, when the interruption intensity is high, the target fusion gate is lowered, and the interruption response state is expected to contribute a larger proportion.

[0181] Subsequently, only pixel locations within the panoramic image space that have a spatial interruption intensity value greater than a preset significant interruption threshold are selected. A typical value for the significant interruption threshold can be set to 0.2. These locations form the spatiotemporal region. Internally, the deviation between the predicted fusion gate and the target fusion gate is calculated as a decoupling constraint loss term:

[0182]

[0183] in, This represents the decoupling constraint loss term. This represents the set of significantly interrupted spatiotemporal regions, that is, the spatiotemporal regions consisting of all pixel locations whose spatial interruption intensity value is greater than a preset threshold. This represents the total number of elements in the set. express Predicting the fusion gate in time in spatial location The value at that location, express The target fusion gate's value at the same spatial location at any given moment.

[0184] S45. The basic loss term and the decoupling constraint loss term are weighted and summed to obtain the composite loss function. The composite loss function is then used to optimize the parameters of the spatial convolutional dual-state interrupted recovery memory network.

[0185] Specifically, the basic loss term and the decoupling constraint loss term are weighted and summed according to a preset balance coefficient to obtain the final composite loss function:

[0186]

[0187] in, Represents the composite loss function. Indicates the basic loss term. This represents the decoupling constraint loss term. This is a balancing coefficient, ranging from 0.1 to 1.0, with a default value of 0.5, used to adjust the strength of the decoupling constraint. A gradient-based optimization algorithm is employed to iteratively optimize all learnable parameters of the network end-to-end on the training set until the loss converges.

[0188] Example 6:

[0189] The spatial robustness index is obtained by spatially weighting the integral of the interference saliency map based on semantic labels, including the following steps:

[0190] S51. Based on the semantic label category of each geometric element in the indoor spatial information model, assign spatial weight values ​​to each pixel direction of the panoramic image corresponding to the panoramic environment state tensor sequence, and construct a spatial weight map.

[0191] Specifically, the process of constructing the spatial weight map utilizes the semantic label information assigned to geometric elements in step S1. In each pixel direction of the panoramic image, the surface of the spatial component it first hits is traced back to determine the semantic label category of that component in the indoor spatial information model. According to a pre-defined weight mapping rule, different weight values ​​are assigned to different semantic label categories. The weights of the semantic labels can be divided into three levels: Level 1 guiding identifiers are assigned weights. First-level landscape nodes are assigned weights Secondary guide icons are assigned weights Secondary landscape nodes are assigned weights Ordinary walls, floors, etc. are assigned a baseline weight. High weights are three times the baseline weights, and medium weights are twice the baseline weights. Through this method, the entire panoramic image is transformed into a spatial weight map. The weight value at each pixel location reflects the importance of that region in the spatial design evaluation.

[0192] S52. Use the spatial weight map to perform frame-by-frame spatial weighted integration on the interference saliency map to obtain the cumulative attention offset, and calculate the spatial anti-interference index based on the cumulative attention offset.

[0193] The formula for calculating the cumulative attention offset is:

[0194]

[0195] in, This represents the cumulative attention offset. Indicates the total number of time steps. The panoramic image domain represents the panoramic image. The spatial weight map represents the pixel orientation. The weight value at that point, Indicates time The saliency map of interference in the pixel direction The value at that location.

[0196] Specifically, using the spatial weight map constructed in the previous sub-step, a spatially weighted integral is performed on the interference saliency map at each time step. Pixel directions assigned high weights will have their contribution amplified even if their interference saliency values ​​are not extremely high—because attention shifts occurring near directional signs or landscape nodes have a more severe impact on the spatial experience; while the interference contribution of ordinary areas assigned baseline weights is relatively weakened. The cumulative attention shift is obtained by summing the full-map weighted integrals from all time steps, and its calculation formula is: .in, This represents the cumulative attention offset. Indicates the total number of time steps. Indicates the domain of the panoramic image. The spatial weight map represents the pixel orientation. The weight value at that point, Indicates time The saliency map of interference in the pixel direction The value at that point. The integral in the domain. Perform on, product This reflects the spatial importance-weighted interference strength at all times. From 1 to The summation yields the total cumulative attention offset.

[0197] The spatial robustness index is calculated based on the cumulative attentional shift. The spatial robustness index can be determined by the ratio of the integral value of the clean task attention map to the integral value of the interfering saliency map within the critical task region.

[0198]

[0199] in, Indicates the spatial anti-interference index. This represents the set of pixels representing the critical task region, specifically the pixel orientations corresponding to the semantic labels assigned high and medium weights. For a moment The clean task attention map in the pixel direction The value at that location, For a moment The saliency map of interference in the pixel direction The value at that location, This is a small constant to prevent division by zero. The higher this ratio, the better the space can maintain a stable direction of attention under the disturbance of people's flow, that is, the stronger its anti-interference ability.

[0200] Significance integral value The calculation method is as follows:

[0201]

[0202] in, This represents the significance integral value. For a moment Visual saliency prediction map in pixel direction The value at the location. Compared with the uniform weight evaluation that does not distinguish semantic labels in Example 1, this example uses spatial weighting to make the evaluation results closer to the actual design requirements—the cost of damage to signage and landscape experience is reasonably amplified, and the detection sensitivity and practical reference value of design defects are significantly improved.

[0203] Based on Example 1, this embodiment provides a detailed explanation of the spatial weighting method in the quantitative evaluation process, introducing a spatial weighting graph based on semantic tags to make the evaluation indicators more sensitive to key task areas.

[0204] The aforementioned indoor space design method based on a gaze-attention mechanism constructs an indoor spatial information model containing semantic labels, divides walkable areas into navigation grids to generate a spatial topological connectivity graph, and generates a multimodal movement sequence carrying dynamic features of the crowd through multi-agent congestion feedback iterative path search. After rendering a panoramic image along the movement path and encoding static visual features and dynamic obstacle features, the generated panoramic environment state tensor sequence is input into a spatial convolutional dual-state interruption recovery memory network. This network maintains a spatialized task-oriented state graph and an interruption response state graph in parallel. In the forward computation, the task-oriented state graph is updated by freezing local regions based on the detected spatial interruption intensity, and the update is also performed based on the duration of no interruption. The interruption response state map is updated using a decay process, and then the two state maps are weighted and fused to output a visual saliency prediction map. Subsequently, by resetting the fusion weights corresponding to the task-oriented state map to zero and performing forward computation again, the interference saliency map generated solely from the interruption response state map decoding is separated. Then, spatial weighted integration is performed on the interference saliency map using semantic labels to obtain a spatial anti-interference index, and saliency integration is performed on the task-critical semantic region of the visual saliency prediction map to obtain a saliency integral value. The combination of these two values ​​forms a quantitative evaluation value. When the quantitative evaluation value fails to meet the standard, a hybrid mechanism combining continuous parameter optimization and discrete structure optimization is used to iteratively modify the indoor spatial information model until the standard is met and a design scheme is output. Thus, this method explicitly models the attention interruption and recovery process caused by dynamic obstacles as two independently controllable states, allowing for quantitative evaluation and automatic optimization of the guidance clarity and visual experience quality of the indoor space under conditions of pedestrian disturbance during the scheme design stage.

[0205] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0206] Based on the same inventive concept, this application also provides a system for implementing the aforementioned interior space design method based on the eye-attention mechanism. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the interior space design system based on the eye-attention mechanism provided below can be found in the limitations of the interior space design method based on the eye-attention mechanism described above, and will not be repeated here.

[0207] In one exemplary embodiment, such as Figure 3 As shown, an interior space design system 30 based on a gaze attention mechanism is provided to implement the methods in the above-described embodiments. The system includes:

[0208] Semantic modeling unit 31 is used to extract geometric elements from the target building's interior space design data and assign semantic labels to them, thereby obtaining an interior space information model.

[0209] The navigation topology construction unit 32 is used to divide the walkable area in the indoor spatial information model into a navigation grid to obtain a spatial topology connectivity graph.

[0210] The multimodal path planning unit 33 is used to generate a multimodal movement sequence by iterative path search through multi-agent congestion feedback based on the spatial topology connectivity graph and preset crowd composition parameters.

[0211] The panoramic environment encoding unit 34 is used to render a panoramic image from the viewpoint of the main virtual human at each moment in the multimodal motion sequence, determine dynamic obstacles based on the positions of other virtual humans, perform visual feature encoding on the panoramic image and obstacle feature encoding on the dynamic obstacles, and generate a panoramic environment state tensor sequence.

[0212] The dual-state saliency fusion unit 35 is used to input the panoramic environment state tensor sequence into a spatially convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map. The forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input, performs local region freezing update on the task-oriented state map according to the spatial interruption intensity, and performs decay update on the interruption response state map according to the duration of no interruption; the two updated state maps are then weighted and fused to output a visual saliency prediction map.

[0213] Interference quantification evaluation unit 36 ​​is used to reset the weights corresponding to the task-oriented state map in the weighted fusion to zero and then perform forward calculation again to obtain an interference saliency map generated only by decoding the interruption response state map; the spatial anti-interference index is obtained by spatial weighted integration of the interference saliency map based on semantic labels, and the saliency integral value is obtained by saliency integration of the task key semantic region of the visual saliency prediction map based on semantic labels. The quantitative evaluation value is obtained by combining the spatial anti-interference index and the saliency integral value.

[0214] The closed-loop optimization generation unit 37 is used to modify the indoor spatial information model by adopting a hybrid mechanism that combines continuous parameter optimization and discrete structure selection when the quantitative evaluation value fails to meet the standard. It re-activates the navigation topology construction unit 32, multimodal path planning unit 33, panoramic environment coding unit 34, dual-state saliency fusion unit 35, and interference quantification evaluation unit 36 ​​to perform corresponding operations until the quantitative evaluation value meets the standard. The indoor spatial design scheme is then generated based on the modified indoor spatial information model.

[0215] Embodiments of this application also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the aforementioned method embodiments.

[0216] Embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0217] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0218] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An interior space design method based on visual attention mechanisms, characterized in that, The method includes: S1. Extract the geometric elements from the interior space design data of the target building and assign semantic labels to obtain the interior space information model; S2. Divide the walkable area in the indoor space information model into a navigation grid to obtain a spatial topology connectivity graph; S3. Based on the spatial topology connectivity graph and preset crowd composition parameters, generate a multimodal movement sequence through multi-agent crowd feedback iterative path search; S4. Render a panoramic view along the main virtual human's viewpoint at each moment in the multimodal motion sequence, and determine dynamic obstacles based on the positions of other virtual humans; perform visual feature encoding on the panoramic view and obstacle feature encoding on the dynamic obstacles to generate a panoramic environment state tensor sequence. S5. Input the panoramic environment state tensor sequence into a spatial convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map; wherein, the forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input, performs local region freezing update on the task-oriented state map according to the spatial interruption intensity, and performs decay update on the interruption response state map according to the duration of no interruption; the two updated state maps are weighted and fused to output a visual saliency prediction map; S6. After resetting the weights corresponding to the task-oriented state map in the weighted fusion to zero, perform the forward calculation again to obtain an interference saliency map generated solely by decoding the interruption response state map; perform spatial weighted integration on the interference saliency map based on the semantic labels to obtain a spatial anti-interference index; perform saliency integration on the task-critical semantic region of the visual saliency prediction map based on the semantic labels to obtain a saliency integral value; and combine the spatial anti-interference index and the saliency integral value to obtain a quantitative evaluation value. S7. When the quantitative evaluation value fails to meet the standard, the indoor space information model is modified by a hybrid mechanism combining continuous parameter optimization and discrete structure selection. Then, S2 to S6 are executed again until the quantitative evaluation value meets the standard. An indoor space design scheme is generated based on the modified indoor space information model.

2. The method according to claim 1, characterized in that, S5 includes: S11. Input the panoramic environment state tensor of each time step in the panoramic environment state tensor sequence into the visual encoder, and obtain a visual feature map that preserves the spatial structure through convolution and downsampling processing; perform convolution operation on the visual feature map to generate the update gate and candidate update content of the task-oriented state map. S12. Extract dynamic obstacle components from the panoramic environment state tensor to form a dynamic event feature map, input the dynamic event feature map into the interruption detector, and obtain the spatial interruption intensity map through convolution operation. S13. Generate a spatial suppression gate based on the spatial interruption intensity map, use the spatial suppression gate to spatially modulate the candidate update content, and use the update gate to temporally fuse the historical task guidance state map with the modulated candidate content to obtain the updated task guidance state map. S14. Extract dynamic feature sub-maps from the visual feature map, perform convolution operation on the dynamic feature sub-maps, and generate candidate states for the interrupt response state map; S15. Based on the comparison result between the global maximum value of the spatial interruption intensity map and the preset threshold, the continuous uninterrupted time step counter is incremented or cleared to obtain the current uninterrupted duration step number; the attenuation factor is calculated based on the current uninterrupted duration step number, and the historical interruption response state map and the candidate state are weighted and updated using the attenuation factor to obtain the updated interruption response state map. S16. After concatenating the visual feature map, the updated task-oriented state map, and the updated interruption response state map along the channel dimension, a spatial fusion gate is generated through convolution operation; the spatial fusion gate is used to perform pixel-by-pixel weighted fusion of the updated task-oriented state map and the updated interruption response state map to obtain a fused state map; the fused state map is input into the decoder, and the visual saliency prediction map is obtained through transpose convolution and convolution processing.

3. The method according to claim 2, characterized in that, The step of spatially modulating the candidate update content using the spatial suppression gate and temporally fusing the historical task-oriented state graph with the modulated candidate content using the update gate to obtain the updated task-oriented state graph includes: S21. Multiply the spatial suppression gate element by element with the candidate update content to obtain the suppressed candidate features; S22. Calculate the suppression completion coefficient based on the spatial suppression gate, and multiply the historical task guidance state diagram with the suppression completion coefficient element by element to obtain the suppression completion term; S23. Add the suppressed candidate features to the suppression completion term to obtain the total candidate update amount; use the update gate to interpolate and fuse the historical task-oriented state graph and the total candidate update amount to obtain the updated task-oriented state graph; The calculation formula for the spatial suppression gate is as follows: in, express The spatial inhibition gate at time of the specified position in space The value at that location, express The spatial interruption intensity map at time of the specified location in space. Threat intensity value at the location.

4. The method according to claim 2, characterized in that, The step of using the attenuation factor to weight and update the historical interruption response state diagram and the candidate states to obtain the updated interruption response state diagram includes: S31. Multiply the historical interruption response state diagram element by element with the attenuation factor to obtain the attenuated historical state; S32. The attenuated historical state and the candidate state are weighted and summed to obtain the updated interrupt response state diagram. The formula for calculating the attenuation factor is as follows: in, Indicates time The attenuation factor, Indicates the basic attenuation rate. Indicates the attenuation coefficient. This indicates the current number of uninterrupted steps.

5. The method according to claim 2, characterized in that, The training process of the spatial convolutional dual-state interrupted recovery memory network includes: S41. Obtain a visual environment sample sequence and real eye movement fixation data, perform saliency map transformation on the real eye movement fixation data, and generate a baseline saliency map sequence; S42. Render and encode the visual environment sample sequence according to the method in S4 to generate a panoramic environment state sample tensor sequence. S43. Input the tensor sequence of the panoramic environment state sample into the spatial convolutional dual-state interruption recovery memory network for the forward calculation to obtain the corresponding prediction saliency map sequence, spatial interruption intensity map and prediction fusion gate; S44. Calculate the similarity loss between the predicted saliency map sequence and the benchmark saliency map sequence, and use the similarity loss as the basic loss term; generate a target fusion gate based on the spatial interruption intensity map, and calculate the deviation between the predicted fusion gate and the target fusion gate as the decoupling constraint loss term in the region where the interruption intensity exceeds a preset threshold. The calculation formula for the target fusion gate is as follows: in, express The target of the moment merges the door in spatial position The value at that location, Indicates the impact coefficient of interruption intensity. express The spatial interruption intensity map at time of the specified location in space. Threat intensity value at the location; S45. The basic loss term and the decoupling constraint loss term are weighted and summed to obtain a composite loss function. The composite loss function is then used to optimize the parameters of the spatial convolutional dual-state interruption recovery memory network.

6. The method according to claim 1, characterized in that, The spatial anti-interference index is obtained by performing a spatial weighted integral on the interference saliency map based on the semantic tags, including: S51. Based on the semantic label category of each geometric element in the indoor spatial information model, assign spatial weight values ​​to each pixel direction of the panoramic image corresponding to the panoramic environment state tensor sequence, and construct a spatial weight map. S52. Use the spatial weight map to perform frame-by-frame spatial weighted integration on the interference saliency map to obtain the cumulative attention offset, and calculate the spatial anti-interference index based on the cumulative attention offset; The formula for calculating the cumulative attention offset is: in, This represents the cumulative attention offset. Indicates the total number of time steps. This represents the panoramic image domain of the panoramic image. This indicates that the spatial weight map is in the pixel direction. The weight value at that point, Indicates time The interference saliency map in the pixel direction The value at that location.

7. An interior space design system based on a gaze-attention mechanism, used to implement the method of any one of claims 1 to 6, characterized in that, The system includes: Semantic modeling unit is used to extract geometric elements from the target building's interior space design data and assign semantic labels to them to obtain an interior space information model; The navigation topology construction unit is used to divide the walkable area in the indoor spatial information model into a navigation grid to obtain a spatial topology connectivity graph. A multimodal path planning unit is used to generate a multimodal movement sequence based on the spatial topology connectivity graph and preset crowd composition parameters through multi-agent congestion feedback iterative path search. A panoramic environment encoding unit is used to render a panoramic image along the viewpoint of the main virtual human at each moment in the multimodal motion sequence, determine dynamic obstacles based on the positions of other virtual humans, perform visual feature encoding on the panoramic image and obstacle feature encoding on the dynamic obstacles, and generate a panoramic environment state tensor sequence. A dual-state saliency fusion unit is used to input the panoramic environment state tensor sequence into a spatially convolutional dual-state interruption recovery memory network that maintains a spatialized task-oriented state map and an interruption response state map, perform forward computation, and output a visual saliency prediction map. The forward computation process is as follows: the network detects the spatial interruption intensity and the duration of no interruption based on the input; it performs local region freezing updates on the task-oriented state map according to the spatial interruption intensity; and it performs decay updates on the interruption response state map according to the duration of no interruption. The two updated state maps are then weighted and fused to output a visual saliency prediction map. The interference quantification evaluation unit is used to reset the weights corresponding to the task-oriented state map in the weighted fusion to zero and then perform the forward calculation again to obtain an interference saliency map generated solely by decoding the interruption response state map; to obtain a spatial anti-interference index by performing spatial weighted integration on the interference saliency map based on the semantic labels, and to obtain a saliency integral value by performing saliency integration on the task-critical semantic region of the visual saliency prediction map based on the semantic labels; and to obtain a quantitative evaluation value by combining the spatial anti-interference index and the saliency integral value. The closed-loop optimization generation unit is used to modify the indoor space information model by adopting a hybrid mechanism combining continuous parameter optimization and discrete structure selection when the quantitative evaluation value fails to meet the standard. The unit readjusts the navigation topology construction unit, the multimodal path planning unit, the panoramic environment coding unit, the dual-state saliency fusion unit, and the interference quantification evaluation unit to perform corresponding operations until the quantitative evaluation value meets the standard. The unit then generates an indoor space design scheme based on the modified indoor space information model.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images