Intelligent indoor design method based on point cloud model

Abstract

The application discloses an intelligent indoor design method based on a point cloud model, which comprises three serial modules. Module A detects a defect area based on a triple judgment mechanism of point density anomaly, reflection intensity anomaly and image semantic auxiliary by fusing multi-modal data of laser point cloud, images and depth maps, generates point cloud repair by using a cross-modal cross attention network, and corrects mirror area error by using a neural radiation field depth priori. Module B simultaneously performs segmentation and parameterized reconstruction on building structures and non-structural elements such as furniture by using a hybrid transformer three-dimensional instance segmentation algorithm, and constructs a full-element building information model base. Module C extracts a real space constraint rasterization coding into a three-channel conditional image, generates a candidate layout scheme by using a diffusion model, and outputs a feasible indoor design scheme by double closed-loop verification of collision detection and compliance knowledge graph.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and building information technology, specifically involving 3D point cloud data processing technology, indoor scene semantic understanding and 3D instance segmentation technology, automatic construction technology of parametric building information models, and automatic generation and multi-dimensional compliance verification technology of indoor design schemes based on multi-agent collaboration framework and generative diffusion model. Background Technology

[0002] 3D point cloud is a massive collection of 3D coordinate points obtained by using a LiDAR scanner to collect data on an interior space. It is an important data foundation for the digitization of interior spaces and the construction of building information models. The technical process of converting point cloud data obtained from on-site scanning into parametric building information models is the core workflow in the field of digital twin architecture. However, this technical process currently faces three fundamental technical bottlenecks in practical engineering applications, which seriously restrict the realization of intelligent interior design throughout the entire process.

[0003] The first bottleneck is the low quality of point cloud data under complex indoor materials and occlusion environments. Transparent or mirrored materials such as glass curtain walls, mirrors, and high-gloss floor tiles, which are widely present in indoor scenes, can cause two typical defects in laser point clouds: transmission perforation defects, where the laser penetrates the glass and hits objects behind it, resulting in large-area holes in the glass itself; and scattering noise defects, where the laser is reflected multiple times on the mirror, generating a large number of false points with severely distorted 3D coordinates, causing serious contamination of the geometric information in that area. In addition, the shadow areas formed by indoor furniture can also cause severe local point cloud loss. Existing point cloud repair methods rely only on geometric smoothing algorithms (such as Poisson reconstruction algorithm and moving least squares method), lacking an understanding of the physical causes of the missing areas and unable to introduce semantic prior information to distinguish between different types of defects such as transmission perforation and occlusion loss. This results in large-area geometric distortion in the repaired model, with deviations in the size of door and window openings reaching tens of centimeters. All subsequent design work based on this point cloud is built on a distorted spatial base, producing systemic errors that are difficult to eliminate.

[0004] The second bottleneck is the lack of fully automated modeling of all elements from point cloud to building information model. The existing scanning to building information model technology can only automatically extract the main building structure such as walls, floors, and ceilings. However, non-structural elements such as furniture, lighting, and electrical switches, which occupy most of the usable value of the interior space, cannot be automatically identified, semantically classified, and parametrically reconstructed. They still rely heavily on manual operation. Specifically, operators must manually draw the model of the main building components based on point cloud slices, and then manually select matching parametric furniture models from the family library one by one for placement and adjustment. The manual modeling cycle for a single ordinary residential unit usually takes three to seven working days, which seriously restricts delivery efficiency and causes the intelligent end-to-end of the digital operation of interior design to break down at the non-structural element modeling stage.

[0005] The third bottleneck is the lack of hard constraints and compliance verification of real physical space in generative interior design schemes. Existing generative interior design algorithms (including text-driven algorithms based on large language models and generative algorithms based on diffusion models) all operate in an idealized, simple rectangular room description, simplifying the room to three numerical parameters: length, width, and height. They completely discard key physical constraints contained in real scanned point clouds, such as the geometry of irregular walls, reserved areas for pipeline routing, and the positions of load-bearing walls and columns. This results in structural conflicts between the generated furniture layout schemes and the real space, such as furniture being placed in the positions occupied by load-bearing columns, or the layout completely ignoring the limitations of load-bearing walls and being unable to be implemented. At the same time, existing algorithms lack the ability to automatically verify building regulations such as the clear width of fire evacuation routes, ergonomic operating space, and barrier-free access. The generated schemes cannot be directly converted into executable construction blueprints, resulting in extremely low value for project implementation. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent interior design method based on point cloud models to solve three core technical problems in the existing technology: insufficient accuracy of point cloud data repair in transparent and mirrored material environments, inability to automatically parametrically model non-structural elements, and lack of real spatial constraints and multi-dimensional compliance verification in generative interior design schemes.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] An intelligent interior design method based on point cloud models includes a multimodal point cloud repair step, a full-element segmentation and reconstruction step, and a multi-agent design generation and verification step, all executed sequentially.

[0009] In the multimodal point cloud restoration step, three-dimensional laser point clouds of the indoor scene, RGB panoramic images calibrated with the laser scanner's extrinsic parameters, and depth maps from an RGB-D camera are simultaneously acquired. Using pre-calibrated extrinsic and camera intrinsic parameter matrices, the three-dimensional points are projected onto the RGB image plane to obtain color features, forming a rich attribute point cloud containing three-dimensional coordinates, laser reflection intensity, and RGB color information. A triple-judgment mechanism is employed, consisting of local point density anomaly analysis, laser reflection intensity anomaly analysis, and RGB image semantic-assisted classification. The three types of detection results are then combined and a three-dimensional morphological dilation operation is performed to determine which aspects require restoration. To address the defective region set, a multimodal perceptual point cloud repair network was designed. Using 3D geometric features as the query matrix and 2D semantic features extracted from RGB images as the key and value matrices, cross-modal cross-attention fusion was performed. The fused features were input into a point cloud generation decoder based on a Transformer decoder, and the output was a repaired point cloud of the missing region. This was then merged with the original complete region point cloud to obtain the repaired complete point cloud. For mirror regions with high reflectivity, a local neural radiation field network was additionally trained to reconstruct the depth prior of the region using multi-view RGB images, correcting the depth error caused by multiple laser reflections.

[0010] In the full-element segmentation and reconstruction step, a 3D point cloud instance segmentation algorithm based on a hybrid Transformer architecture is adopted. After the repaired complete point cloud is voxelized, point-level features are extracted through a multi-scale sparse convolutional backbone network. Through the iterative interaction of learnable instance query vectors and multi-layer Transformer decoders, the binary mask probability of each instance is output by the mask prediction head, and the semantic category probability is output by the category prediction head. After 3D non-maximum suppression post-processing, the 3D instance segmentation results of non-structural elements such as building structural elements and interior furniture are obtained. For the point cloud of building structural element instances, a random sampling consensus algorithm is used for plane fitting, boundary contours are extracted, and door and window openings are detected to generate industrial basic parametric building information model objects. For the point cloud of non-structural element instances, principal component analysis is used to calculate the oriented bounding box. The model is matched in the parametric CAD model database using semantic category labels and bounding box size as search keys. The spatial topological relationship between objects is modeled through a graph attention network. An interior space topology map carrying geometric attributes is constructed with rooms as nodes and door opening connections as edges, which serves as the input for rigid spatial constraints generated in subsequent design.

[0011] In the multi-agent design generation and verification step, the master control large language model agent receives user requirements and schedules the spatial parsing agent to extract the floor outline polygon, prohibited placement areas, and door and window positions of the target room from the indoor space topology map. The three types of spatial information are encoded into binary masks and stitched together to form a three-channel spatial constraint image. The layout generation agent uses the spatial constraint image and user requirement embedding as dual conditions to generate multiple candidate furniture layout schemes through a denoising diffusion probability model. The compliance verification agent performs axis-aligned bounding box collision detection and compliance knowledge graph rule verification on each candidate scheme in sequence, and calculates a comprehensive score covering compliance, space utilization, and aesthetic quality. For schemes whose scores do not reach the passing threshold, the violation information is encoded into a conditional penalty vector and attached to the spatial constraint image to drive the diffusion model to resample, forming a dual closed-loop iterative optimization mechanism of design-verification-feedback-regeneration, and outputting a building information model file and a compliance verification report.

[0012] The beneficial effects of this invention are reflected in the following three aspects:

[0013] 1. This invention overcomes the serious distortion problem of pure geometric restoration methods in complex environments with transparent and mirror materials by deeply fusing the two-dimensional semantic prior of RGB images with the geometric features of three-dimensional point clouds through a cross-modal cross-attention mechanism, and introducing the depth prior of neural radiation field as a depth correction auxiliary means for mirror areas. This enables the geometric accuracy of the restored indoor space model to reach the millimeter level, providing an accurate and reliable spatial digital base for subsequent steps in the entire process.

[0014] 2. This invention is the first to incorporate non-structural elements such as furniture, lighting fixtures, and switches into a unified technical framework for end-to-end instance segmentation and parametric reconstruction, realizing a fully automated conversion from the original scanned point cloud to a parametric building information model that includes all elements of the building structure and interior. This significantly reduces the project modeling cycle from several days to within a few hours, greatly improving the efficiency of digital interior design operations and filling the automation gap in the non-structural element stage of the traditional scanning to building information model process.

[0015] 3. This invention uses the real irregular point cloud spatial topology as rigid geometric constraints to drive generative design, and constructs a compliance knowledge graph covering fire safety regulations, ergonomic standards, and accessibility design specifications. It introduces a dual closed-loop iterative optimization mechanism to ensure that the generated scheme meets multi-dimensional regulatory requirements, and realizes end-to-end intelligent processing from on-site scanning to a 3D interior design scheme that can be directly implemented in engineering. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the three core serial modules of the present invention;

[0017] Figure 2 This is a flowchart of the triple detection and repair process for point cloud defects in this invention;

[0018] Figure 3 This is a schematic diagram illustrating the dual-loop verification of the diffusion model layout of this invention. Detailed Implementation

[0019] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0020] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0021] I. System Overall Architecture

[0022] The intelligent interior design method based on point cloud model proposed in this invention is divided into three serial technology modules: module A is a multimodal point cloud adaptive repair and completion module, module B is a full-element instance segmentation and parametric reconstruction module, and module C is a dual closed-loop module for multi-agent design generation and compliance verification.

[0023] The data flow of the three modules is as follows: the original scanned point cloud is repaired by module A to obtain a high-quality complete point cloud; module B performs full-element instance segmentation on the complete point cloud, performs parametric reconstruction of building structural elements and non-structural elements respectively, and constructs an interior space topology map with geometric attributes; module C uses the interior space topology map and the full-element parametric building information model base as hard constraints, drives the diffusion model to generate and iteratively verify the interior design scheme through a multi-agent collaboration framework, and finally outputs an interior design building information model file and a compliance verification report that can be used for engineering implementation. The three modules form an end-to-end intelligent operation link from the original point cloud scanning to the implementable design scheme.

[0024] II. Module A: Adaptive Repair and Completion of Multimodal Point Clouds

[0025] (I) Multimodal data acquisition and preprocessing

[0026] During the data acquisition phase, this invention requires the simultaneous acquisition of the following three types of sensor data: the first type is three-dimensional laser point cloud. ,in , , For three-dimensional coordinates in space, The intensity of laser reflection. The first category represents the total number of point clouds, while the second category comprises RGB panoramic images calibrated with laser scanners using extrinsic parameters. ,in The total number of viewpoints is represented by the first category, and the third category consists of depth maps captured by RGB-D cameras. Each frame corresponds to an RGB image.

[0027] By completing the sensor calibration process in advance, the rigid body transformation matrix (extrinsic parameter matrix) from the camera coordinate system to the laser coordinate system is obtained. and camera intrinsic parameter matrix Using the above calibration parameters, each three-dimensional point By projecting the pinhole camera model onto the corresponding RGB image plane, its color features are obtained. This results in a rich attribute point cloud containing three-dimensional coordinates, laser reflection intensity, and RGB colors. This provides a multimodal data foundation for the subsequent detection and repair of transparent and mirror-like areas.

[0028] For multi-frame point clouds obtained from multi-station scanning, a point cloud registration method based on the iterative nearest point algorithm is used for precise alignment. The point clouds obtained from each station are unified to the global coordinate system and stitched together to form a complete scene point cloud covering the entire indoor scene. Then, a rich attribute point cloud is used. The form is used as input for subsequent processing.

[0029] (ii) Automatic detection of transparent and mirrored areas

[0030] This invention designs a triple judgment mechanism to automatically identify defect areas from three complementary dimensions: point density, laser reflection intensity, and image semantics. Finally, it fuses the three types of signals to determine the target area for repair.

[0031] Local point density anomaly analysis: analysis of attribute-rich point clouds Construct a KD-Tree spatial index for each point. Calculate the local point density in the spherical neighborhood Statistical global average point density and global density standard deviation Local density below the threshold The region is marked as a candidate void region. This criterion identifies low-density areas that deviate from the normal density distribution by more than two standard deviations in a statistical sense as potential defect areas, which can effectively cover large-area voids caused by laser penetration of glass and local missing areas caused by furniture obstruction.

[0032] Anomaly analysis of reflection intensity: laser reflection intensity at all points in a rich attribute point cloud Perform global statistical analysis to identify areas with reflectance exceeding the saturation threshold. Points marked with the 95th percentile of the reflection intensity distribution are designated as the high-reflectivity candidate point set. The mirror material causes multiple mirror reflections of the laser, resulting in a significantly higher received echo intensity than normal diffuse reflection points. Furthermore, the calculated three-dimensional coordinates suffer from severe positional deviations due to the superposition of multiple reflection paths. Such false points can be effectively identified by using a reflection intensity threshold.

[0033] RGB semantic-assisted classification: classifying RGB panoramic images Input a pre-trained 2D semantic segmentation network with SegFormer as the backbone, output pixel-wise semantic labels, and extract semantic regions belonging to the glass and mirror classes from them, denoted as 2D semantic masks. By using the inverse projection transformation of the camera's intrinsic and extrinsic parameter matrices, the two-dimensional semantic mask is... Back-projecting onto the 3D point cloud space yields 3D semantic candidate regions. Image semantic segmentation networks have a strong ability to recognize glass and mirror materials, providing semantic-level supplementary verification for pure geometric statistical judgments and effectively reducing the false negative rate.

[0034] Multimodal fusion decision-making: for candidate void regions High reflectivity candidate point set and three-dimensional semantic candidate regions Take the union of the sets and apply a radius-wise operation to the merged result. A three-dimensional morphological dilation operation with parameters is performed to ultimately determine the set of defect regions that need to be repaired. Morphological dilation is used to appropriately expand the boundary range of the defect area, ensuring that the repair operation can fully cover the transition zone between the defect area and the surrounding normal point cloud, avoiding geometric jumps in the repair results at the boundary. The fusion decision mechanism of the three-judgment fully utilizes the different sensitivities of each judgment dimension to different types of defects, and enhances the robustness and completeness of the overall detection through complementarity.

[0035] (III) Multimodal Perception Point Cloud Repair Network

[0036] This invention designs a multimodal perceptual point cloud repair network. Its core architecture consists of three sub-networks: a two-dimensional semantic feature extractor, a three-dimensional geometric feature extractor, and a cross-modal attention fusion decoder. It is specifically designed for high-precision semantic-guided repair of point cloud defects caused by transparent and specular materials.

[0037] For the set of defect regions Each missing sub-region Using its geometric center as the origin, extract the surrounding radius. Using the complete neighborhood point cloud within the range as context input, the neighborhood point cloud is downsampled to a fixed number using the farthest point sampling algorithm. (Value is 1024) points, each containing six dimensions of information including three-dimensional coordinates and RGB color, to ensure the consistency of network input dimensions.

[0038] The 2D semantic feature extractor crops the corresponding local regions of the RGB image from several viewpoints closest to the defect sub-region and inputs them into a pre-trained 2D semantic segmentation network with SegFormer as the backbone (pre-trained on a large-scale image segmentation dataset and fine-tuned on an indoor scene dataset). It extracts a multi-scale semantic feature map with a spatial resolution of one-quarter of the original image. Through the back projection operation of the camera intrinsic and extrinsic parameter matrices, the feature vector of each pixel position in the feature map is mapped to three-dimensional space. Then, trilinear interpolation is used to assign a corresponding semantic feature vector to each point in the sampling point set, forming an enhanced point set carrying semantic context information.

[0039] The 3D geometric feature extractor takes the enhanced point set as input to a sparse convolution-based U-Net backbone network, which contains four downsampling layers and corresponding skip connections, and extracts the 3D geometric feature vector for each point. Then, the geometric features of all points are globally aggregated through max pooling to obtain a global geometric feature vector that represents the overall shape context of the entire local neighborhood. It is used to provide macroscopic shape guidance information during the decoding stage.

[0040] The cross-modal cross-attention fusion module obtains the query matrix by linearly projecting the 3D geometric features. The two-dimensional semantic features are linearly projected to obtain the key matrix. Sum matrix Perform scaling dot product attention calculation:

[0041] ;

[0042] in, The query matrix is ​​obtained by linear projection of three-dimensional geometric features, with dimension . , The key matrix is ​​obtained by linear projection of two-dimensional semantic features, with dimension . , The value matrix is ​​obtained by linear projection of two-dimensional semantic features, with dimension . , is the dimension of the key vector, used for scaling to prevent the dot product value from becoming too large. This is the transpose of the key matrix. For normalized exponential functions, As a fused multimodal feature matrix, this attention mechanism enables each 3D point to adaptively retrieve the most relevant semantic information from the 2D semantic features while extracting the geometric context. This allows the point cloud restoration network to be guided by both geometric consistency constraints and semantic priors when generating point clouds of missing regions, which has significant advantages over pure geometric restoration methods.

[0043] For mirror regions with high reflectivity, this invention additionally trains a local neural radiation field network (using an instantaneous neural graph primitive method implemented with hash coding technology to improve training speed), and reconstructs the continuous radiation field of the region using multi-view RGB images covering the region to obtain the depth prior of the region. Depth Prior After being converted to 3D point coordinates, these coordinates are used as additional geometric priors and merged with the original neighborhood point cloud. They are then input into the 3D geometric feature extractor to effectively correct the false depth values ​​caused by multiple laser reflections, making the point cloud restoration results of the mirror area closer to the real physical geometry.

[0044] Fusion features With global geometric features After concatenation, the input is a point cloud generation decoder based on a Transformer decoder structure (containing several decoder layers, each containing self-attention, cross-attention, and feedforward network sub-layers). The decoder outputs a set of newly generated point coordinates within the repaired area. This will generate a point cloud. Merging the point cloud with the original complete region yields a fully restored point cloud. This provides an accurate three-dimensional digital foundation for subsequent full-element instance segmentation and parametric reconstruction.

[0045] The network training uses a weighted combination of Chamfer distance loss and Earth movement distance loss as the total training loss, with the weight coefficients of the two losses being... and By controlling the contribution ratio of the corresponding losses, the network is trained under supervision on a large-scale indoor point cloud dataset to enable it to complete point clouds with high accuracy in real and complex indoor scenes.

[0046] Module B: Full-Feature Instance Segmentation and Parametric Reconstruction

[0047] (a) Hybrid Transformer 3D Instance Segmentation

[0048] This invention proposes a 3D point cloud instance segmentation algorithm based on a hybrid Transformer architecture. It introduces a multi-scale local-global attention mechanism designed specifically for indoor scenes on the basis of the Transformer mask instance segmentation framework, which can simultaneously complete instance segmentation and semantic classification of architectural structural elements and non-structural elements in a single forward inference.

[0049] The repaired complete point cloud Uniform voxelization with a side length of 2 cm is performed to obtain a sparse voxel representation. For each non-empty voxel, the mean coordinates and mean color of all points within the voxel are used as initial voxel features to form an initial voxel feature matrix. The initial voxel features are input into a U-Net backbone network based on sparse convolution. This backbone network contains four downsampling stages (each with a stride of 2, resulting in a final voxel resolution of 1 / 16 of the original). Four multi-scale feature maps with different resolutions are extracted, with feature dimensions of 64, 128, 256, and 512, respectively. The upsampling branch fuses the features of the corresponding downsampling branch through skip connections. The final output is a point-level feature with the same resolution as the original voxel and 256 feature channels. This feature integrates local geometric details and global semantic context information.

[0050] Initialize 200 learnable instance query vectors, each with a dimension of 256. Each query vector represents the model's search anchor point for a potential indoor object instance. Perform 6 layers of Transformer decoding iterations. In each decoder layer, perform three sub-operations sequentially: instance query self-attention computation, enabling different instance queries to perceive each other and effectively preventing multiple query vectors from converging to the same object instance; cross-scale cross-attention computation; and query vectors using point cloud features. Cross-attention is performed on key-value pairs to extract the most relevant local geometric information and global semantic context from point cloud features. A feedforward network transformation is performed on the updated query vector, and a two-layer fully connected transformation is added with residual connections to enhance feature representation. The iterative refinement process of the 6-layer Transformer decoding enables each query vector to gradually focus on a specific object instance in the point cloud, ultimately forming a precise receptive field for each instance.

[0051] Based on the query vector output by the final decoder layer, mask prediction and class prediction are performed by two independent prediction heads. The mask prediction head combines each query vector with global point features. A pointwise inner product operation is performed, and after Sigmoid activation, the binary mask probability of each instance query for all voxel points is obtained, which represents the confidence that the voxel point belongs to the current instance. The category prediction head performs a fully connected classification layer on each query vector and outputs the semantic category probability distribution of the indoor objects corresponding to the query. The semantic category system covers 31 categories of indoor objects (including common building components and home furnishings such as walls, floors, ceilings, doors, windows, beds, sofas, tables, chairs, lamps, and switches).

[0052] All predicted instances are sorted in descending order of confidence score. Three-dimensional non-maximum suppression post-processing is performed. For any two candidate instances, if the three-dimensional intersection-union ratio of their point cloud masks exceeds the threshold of 0.5, the instance with higher confidence is retained and the other instance is removed. The final output instance contains a point cloud mask, semantic label and confidence score, which can be used for subsequent parameterized reconstruction steps for different categories.

[0053] (ii) Parametric reconstruction of building structure

[0054] For instance point clouds identified as building structure categories such as walls, floors, ceilings, doors, and windows, the following parametric reconstruction process is performed:

[0055] Taking the wall instance as an example, the point cloud is fitted with the optimal plane equation using the random sampling consensus algorithm. The number of iterations is dynamically calculated to ensure a reliable set of interior points is obtained under the set success probability (valued at 0.99). The points in the wall plane are projected onto the local two-dimensional coordinate system of the fitted plane, and the convex hull profile or alpha shape profile of the point set is extracted to completely describe the two-dimensional boundary shape of the wall in the plane. The spatial position, orientation, height and thickness of the wall are determined by combining the plane equation parameters, and an IfcWall parametric object conforming to the industrial basic class standard is generated. The floor and ceiling instances are also reconstructed using the plane fitting method, and corresponding IfcSlab parametric objects are generated respectively.

[0056] Within the wall plane, by detecting locally connected regions where the point cloud density is significantly lower than that of the wall background, the positions of door and window openings are automatically identified, and the position coordinates, width, and height parameters of each opening relative to the bottom baseline of the wall are extracted. If the opening is a door, an IfcDoor parameterized object is generated; if the opening is a window, an IfcWindow parameterized object is generated. The corresponding opening object is then associated with the wall to ensure semantic consistency within the building information model.

[0057] Construct an interior space topology map using the identified room areas as nodes and the doorway connections between adjacent rooms as edges. ,in This is a set of room nodes, where each node carries the sequence of vertices of the polygon representing the floor outline of the room and its spatially enclosing geometry. The set of edges connecting doorways carries the spatial coordinates, net width, and opening direction of the corresponding doorway. The interior space topology map describes the geometric topological constraints of the real interior space in a structured form. It will serve as the rigid spatial constraint input for the multi-agent design of module C, ensuring the feasibility of the generated scheme in the real physical space.

[0058] (III) Parametric Reconstruction of Unstructured Elements

[0059] For instance point clouds identified as unstructured elements such as furniture, lighting fixtures, and switches, the following parametric reconstruction process is performed:

[0060] First, a statistical outlier filtering algorithm is used to remove noise points for each unstructured instance point cloud. For each point, the mean of its 20 nearest neighbor distances is calculated. Calculate the global mean of the mean for all points. and global standard deviation , will satisfy Points that are deemed outliers are identified and deleted, while the set of interior points with a reasonable geometric distribution is retained.

[0061] The covariance matrix of the preprocessed point cloud computing The covariance matrix is ​​then decomposed into eigenvalues ​​to obtain three principal direction unit vectors (arranged in descending order of eigenvalues), corresponding to the longest axis, second longest axis, and thinnest axis of the object, respectively. The point cloud is then transformed into a local coordinate system with the three principal directions as coordinate axes, and the extreme values ​​of the coordinates in the three principal directions are calculated to obtain the center coordinates of the oriented bounding box. Orientation Matrix And the half-length vector, its three components These represent the length, width, and height of the object, respectively.

[0062] Using semantic category labels and orientation bounding box size parameters Using the retrieval key, the system searches the parametric CAD model database (indexed by category and size) for the best matching parametric CAD model. The retrieval distance is defined as the Euclidean distance between the size parameters of the candidate model and the size parameters of the current instance. The system returns several candidate models with the smallest distance. The final matching model is selected by the post-processing module or the user. This completes the parametric modeling of non-structural elements and places the matching parametric CAD object in the corresponding position and orientation of the oriented bounding box.

[0063] Using the center of the oriented bounding box of all unstructured instances as nodes, and spatial proximity (distance between nodes is less than a preset threshold) as the basis for selection. Using () as the edges, construct a spatial relationship diagram of indoor objects. The system uses the category embedding vector and spatial coordinates of each node as initial node features, and the relative position vector and distance between two nodes as edge features. The input graph attention network performs multiple rounds of message passing, and the output is updated node features that incorporate the semantic context of the neighborhood. These features are used to verify the physical rationality of the relationship between adjacent objects. When an abnormal spatial relationship that violates physical laws is detected (such as the chair's enclosure being detected as floating above the table instead of being placed on the ground), the system triggers an abnormal alarm and provides feedback to the user for manual confirmation, thus ensuring the physical credibility of the object placement relationship in the building information model.

[0064] IV. Module C: Dual Closed Loop of Multi-Agent Design Generation and Compliance Verification

[0065] (I) Multi-agent system architecture

[0066] Module C outputs the indoor space topology map from Module B. Using a fully parameterized building information model as input, an interior design scheme that meets spatial constraints and multi-dimensional regulatory requirements is automatically generated through a multi-agent collaborative framework.

[0067] The system comprises four types of intelligent agents with clearly defined roles and collaborative operation: the master control big language model agent is responsible for receiving users' natural language descriptions of their needs, extracting structured information such as functional area requirements, style preferences, and budget constraints, decomposing the complete design task, and distributing it to other intelligent agents; the spatial analysis agent is responsible for analyzing the building information model base, extracting structured spatial information such as the layable space polygons, door and window coordinates, and streamline skeleton diagrams of the target room, and quantifying and encoding it into conditional representations that can be directly consumed by the subsequent diffusion model; the layout generation agent is responsible for calling the diffusion model and generating a sequence of candidate interior furniture layout schemes based on both structured spatial constraints and user requirement embedding; and the compliance verification agent is responsible for calling the collision detection engine and compliance knowledge graph to perform multi-dimensional scoring and violation identification on the candidate schemes, and feeding back the violation details to the layout generation agent to drive a new round of sampling optimization. The four types of intelligent agents communicate through standardized message interfaces, forming a closed collaborative loop.

[0068] (ii) Structured Representation of Spatial Constraints

[0069] Spatial analysis agents analyze indoor spatial topology maps Extract the floor outline polygon of the target room. The system normalizes the furniture to a unit bounding box with the global coordinate origin as a reference, and records the original scaling factor for subsequent inverse normalization to ensure that the actual size and position information of the furniture remain consistent throughout the process.

[0070] Within the floor outline, identify and mark three types of prohibited areas where furniture is not allowed: fire evacuation areas, i.e., corridors within 0.5 meters on either side of the center line of the doorway, to ensure unobstructed evacuation routes and comply with fire safety regulations; hard constraint areas, i.e., the locations of mechanical and electrical pipe shafts and load-bearing columns read from the structural layer of the building information model, where furniture occupying this area will prevent construction; and door opening areas, i.e., the fan-shaped areas calculated based on the net width of each doorway and the opening direction, to ensure that the doors can be fully opened without obstruction by furniture. The union of these prohibited areas is denoted as... .

[0071] Layoutable polygons Prohibited areas Together with the sequence of door and window positions, they are encoded into a structured spatially constrained image: the two-dimensional space is discretized into... The grid is used to generate binary masks for the layable regions. Forbidden area binary mask Binary mask for door and window positions The three masks are stitched together along the channel dimension to form a three-channel spatially constrained image. The three-channel image fully encodes the geometric constraint information of the real point cloud space into the conditional input of the diffusion model, so that the generated layout scheme naturally meets the physical feasibility constraints of the real space.

[0072] (III) Generation of 3D Layout of Diffusion Model

[0073] This invention represents an interior layout scheme as an ordered sequence of furniture. ,in For the first The semantic category encoding of each piece of furniture, The center coordinates of the furniture orientation box. The orientation angle of the furniture in the horizontal plane (in radians). The length, width, and height dimensions of the furniture. This represents the total number of furniture pieces in the layout plan.

[0074] Continuous attributes of each piece of furniture Normalized to the [-1, 1] interval, categorical attributes Mapped to learnable embedding tables The 3D vectors are concatenated to obtain the complete feature vector of each piece of furniture. ,in The complete layout sequence is represented as a matrix. , as the modeling object of the diffusion model.

[0075] During the training phase, the actual layout sequence is used. Gaussian noise is gradually added using a cosine-based scheduling strategy. The total number of diffusion steps is [number missing]. The value is set to 1000, and a time step is used to train a program. Spatial constrained images Embedded with user needs Conditional Denoising Transformer Network The network employs a classifier-free guidance mechanism, randomly discarding conditional inputs with a probability of 0.1 (replacing them with empty conditions), enabling the model to simultaneously possess both unconditional and conditional generation capabilities. The training objective is to minimize the following mean squared error loss:

[0076] ;

[0077] in, For the training loss of the diffusion model, For the actual layout sequence, In the first The noisy layout sequence after adding noise step by step. Standard Gaussian noise, For parameters A denoising Transformer network whose output is a prediction of the added noise. It is the square of the L2 norm (mean square error). For time steps Real layout sequence and noise The joint expectations.

[0078] During the inference phase, from pure Gaussian noise Starting with the DDIM sampler, the number of inference steps is reduced to 50 to improve inference efficiency at each time step. A classifier-free guidance strategy is applied, with a guidance weight coefficient of 7.5, to control the influence of spatial constraints and user requirements on the generated layout. After inference, 10 candidate layout schemes are generated. The compliance verification agent will then conduct a multi-dimensional evaluation.

[0079] (iv) Compliance verification dual closed-loop mechanism

[0080] The compliance verification agent performs two levels of evaluation for each candidate layout scheme: collision detection and specification verification.

[0081] Collision detection: For each piece of furniture in the candidate layout, its orientation bounding box is approximated as an axis-aligned bounding box. Geometric collision interference between all pieces of furniture and between furniture and the building structure (walls, load-bearing columns) is detected using two axis-aligned bounding boxes. and The condition for a collision is that the projection intervals along all three coordinate axes overlap.

[0082] ;

[0083] in, This is a collision detection function; an output value of 1 indicates a collision, and an output value of 0 indicates no collision. and respectively enclosed box and The center point is Coordinates along the axis, and respectively enclosed box and exist The length of half a side along the axial direction, This is an indicator function (1 if the condition is true, 0 otherwise). This means that the logical product of the three coordinate axes is taken. A collision is determined only if all three directions meet the overlap condition. This collision detection algorithm has high computational efficiency and can quickly filter out all infeasible layouts with geometric interference.

[0084] At the standard verification level: Interior design standards such as fire protection standards (GB 50016), accessibility design standards (GB 50763), and ergonomics standards (GB / T 14775) are structured and stored as a compliance knowledge graph. Each standard is represented in the form of rule triplets (subject, relation, object or value). For example, regulations are stored as triplets such as (evacuation corridor, minimum clear width not less than 0.9 meters) and (bed, not directly facing the entrance door). This allows for the verification of candidate layouts. Query each applicable rule and assess the violation, generating a violation type vector. ,in The total number of rules is represented by the vector, where the components with a value of 1 correspond to the specific rule entries that were violated.

[0085] Overall score calculation: for candidate layouts Calculate the overall score:

[0086] ;

[0087] in, For comprehensive scoring, the value ranges from [0,1], with higher values ​​indicating better solution quality. , , These are the weighting coefficients for compliance score, space utilization score, and aesthetic score, respectively. The default values ​​are 0.5, 0.3, and 0.2, which can be adjusted by the user according to actual needs. (Compliance score...) ,in The number of rule violations determines the space utilization score. The ratio of furniture-covered area to total available area reflects the adequacy of space utilization and is used for aesthetic scoring. The aesthetic quality score output by the pre-trained aesthetic scoring model (a reward model fine-tuned based on human preference data) reflects the degree of visual harmony of the layout.

[0088] Dual-loop iterative optimization: When the overall score of all candidate solutions is lower than the passing threshold of 0.85, the compliance verification agent will report the violation. Together with the natural language description text obtained by the large language model from the violation encoding, it is returned to the layout generation agent. The layout generation agent encodes the violation information into a conditional penalty vector. Attach it to the spatially constrained image Then, the DDIM sampling process is re-executed to generate a new batch of candidate solutions, which then enter the next round of compliance verification. This forms a double closed-loop iterative optimization mechanism of generation-verification-feedback-regeneration, which is executed cyclically until a candidate solution with a comprehensive score reaches the passing threshold or the maximum number of iterations (5 times) is reached. The solution with the highest comprehensive score among all historical candidate solutions is selected as the final design solution.

[0089] The parameters (category, center position, orientation angle, size) of each piece of furniture in the final design scheme are converted into the corresponding industrial basic class parametric building information model object (IfcFurnishingElement), which is then merged with the building structure building information model generated by module B to output a complete interior design building information model file (IFC format). At the same time, a compliance verification report is generated, which details all testing rules and their pass or violation status for designers and relevant parties to review and verify the project.

[0090] V. Examples

[0091] Taking a two-bedroom apartment with a floor area of ​​approximately 90 square meters as an example, the complete execution process of this invention will be explained.

[0092] During the data acquisition phase, one or two scanning stations were set up in the living room, master bedroom, secondary bedroom, kitchen, and bathroom using a terrestrial 3D laser scanner. Multimodal data was acquired simultaneously using an RGB panoramic camera and an RGB-D depth camera. All sensors underwent joint calibration before acquisition. After scanning, iterative nearest-point registration was performed on the point clouds obtained from each station, and the points were stitched together to form a rich attribute point cloud in a unified coordinate system for the entire house, with approximately 300 million points.

[0093] In the point cloud restoration stage, the algorithm detected candidate void regions and high-reflectivity candidate point sets at the locations of the living room floor-to-ceiling glass door, bathroom mirror, and entryway mirror. Combining the glass-like and mirror-like pixel regions identified by the semantic segmentation network, three defect regions were determined. Neighborhood point cloud context was extracted for each of the three defect regions, and a multimodal perception point cloud restoration network was constructed and inferred. For the bathroom mirror region, a local neural radiation field network was additionally trained using multi-view RGB images to generate a depth prior, correcting the error of multiple reflections of the mirror. After restoration, the geometric accuracy of the three defect regions reached the millimeter level.

[0094] In the full-element segmentation stage, hybrid Transformer 3D instance segmentation was performed on the repaired complete point cloud of the whole house, identifying a total of 26 wall instances, 5 floor instances, 5 ceiling instances, 8 doorway instances, 6 windowway instances, 1 sofa instance, 1 coffee table instance, 2 bed frame instances, 2 wardrobe instances, and several lighting instances. Parametric reconstruction was performed on the building structure instances to generate a complete set of industrial basic parametric objects. An interior space topology map containing 5 nodes (living room, master bedroom, secondary bedroom, kitchen, and bathroom) and corresponding doorway edges was constructed. Oriented bounding box estimation and CAD database retrieval were performed on non-structural element instances such as sofa, coffee table, bed frame, and wardrobe to complete parametric matching modeling. The entire process required no manual intervention and the modeling time was approximately 3 hours.

[0095] In the interior design generation phase, the living room is the target room. The user inputs a living room in a modern Chinese style, with the design requirement of meeting four guests. The spatial analysis agent extracts the polygonal outline of the living room floor (hexagonal, including one protruding area), marks the positions of the front door, balcony door, and two windows, and sets three types of prohibited placement areas, which are encoded as a three-channel spatial constraint image. The layout generation agent uses the spatial constraint image and the user's requirements as conditions to generate 10 candidate layout schemes by a diffusion model. The compliance verification agent verifies through collision detection and knowledge graph. In the first round, 7 schemes violated the rule that the sofa would block the evacuation corridor. The violation feedback was encoded into a conditional penalty vector to drive the second round of resampling. In the second round, 3 out of the 10 candidate schemes passed all verification rules. The scheme with the highest comprehensive score scored 0.91. This scheme was selected to export a complete IFC format building information model file and a compliance verification report.

[0096] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent interior design method based on point cloud models, characterized in that, The following steps are included and executed sequentially: Multimodal point cloud restoration steps: Simultaneously acquire 3D laser point clouds of the indoor scene, RGB panoramic images calibrated with the laser scanner's extrinsic parameters, and RGB-D depth maps. Project 3D points onto the RGB image plane using the extrinsic parameter matrix and camera intrinsic parameter matrix to obtain the color features of each point, forming a rich attribute point cloud. Based on a triple judgment mechanism of local point density anomaly analysis, laser reflection intensity anomaly analysis, and RGB image semantic auxiliary classification, take the union of the three types of detection results and perform a 3D morphological dilation operation to determine the defect region set. Use a multimodal perception point cloud restoration network to restore the defect region. The multimodal perception point cloud restoration network extracts neighborhood point clouds and corresponding viewpoint RGB image features for each missing sub-region. Use the multimodal perception point cloud restoration network to fuse the 3D geometric features of the neighborhood point cloud with the 2D semantic features of the corresponding viewpoint RGB image, guide the point cloud generation decoder to output the restored missing region point cloud, and merge it with the original complete region point cloud to obtain the restored complete point cloud. Full-element segmentation and reconstruction steps: Voxelization and multi-scale sparse convolution feature extraction are performed on the repaired complete point cloud. Through the iterative interaction between the learnable instance query vector and the multi-layer Transformer decoder, the binary mask probability of each instance is output by the mask prediction head, and the semantic category probability is output by the category prediction head. After three-dimensional non-maximum suppression post-processing, the three-dimensional instance segmentation results of building structural elements and non-structural elements are obtained simultaneously. The point cloud of building structural element instances is fitted with a random sampling consensus algorithm to extract the boundary contour and detect door and window openings, generating industrial basic class parametric building information model objects. The point cloud of non-structural element instances is calculated by principal component analysis. The matching model is retrieved in the parametric CAD model database with semantic category and oriented bounding box size as the retrieval key. The spatial topological relationship between objects is modeled by graph attention network. The indoor space topology map carrying geometric attributes is constructed with rooms as nodes and door opening connections as edges. Multi-agent design generation and verification steps: The master control large language model agent receives user requirements and schedules the spatial parsing agent to extract the floor outline polygon, prohibited placement areas, and door and window positions of the target room from the indoor space topology map. These three are encoded into binary masks and then stitched together to form a three-channel spatial constraint image. The layout generation agent uses the spatial constraint image and user requirement embedding as dual conditions to generate multiple candidate furniture layout schemes through a denoising diffusion probability model. The compliance verification agent performs axis-aligned bounding box collision detection and compliance knowledge graph rule verification on each candidate scheme in sequence, calculates the comprehensive score, and encodes the violation information into a conditional penalty vector for schemes that do not reach the passing threshold and attaches it to the spatial constraint image to drive the diffusion model to resample. The process is repeated until the comprehensive score reaches the passing threshold or the upper limit of the number of iterations is reached. The scheme with the highest comprehensive score is selected, and the interior design building information model file and compliance verification report in industrial basic class format are exported.

2. The method according to claim 1, characterized in that, In the aforementioned triple-judgment mechanism: local point density anomaly analysis constructs a KD-Tree spatial index for attribute-rich point clouds, calculates the local point density of each point within a spherical neighborhood, and identifies points with local densities below a threshold. The region is marked as a candidate cavity region, where The global average point density, For the global density standard deviation, the laser reflection intensity anomaly analysis marks points with reflection intensity higher than the 95th percentile of the reflection intensity distribution as high reflection candidate points. The RGB image semantic auxiliary classification backprojects the glass-like and mirror-like pixel regions identified by the RGB image through the two-dimensional semantic segmentation network to the three-dimensional space through inverse projection transformation to obtain the three-dimensional semantic candidate regions.

3. The method according to claim 1, characterized in that, The multimodal sensing point cloud restoration network also includes a neural radiation field depth prior auxiliary step for the mirror region: a local neural radiation field network is trained on the mirror region with high reflectivity, and the continuous radiation field of the region is reconstructed using multi-view RGB images covering the region to obtain the depth prior. After the depth prior is converted into three-dimensional point coordinates, it is used as an additional geometric prior and merged with the neighborhood point cloud into the three-dimensional geometric feature extractor to correct the depth error caused by multiple laser reflections.

4. The method according to claim 1, characterized in that, In the 3D instance segmentation, the voxel side length of the voxel is 2 cm. The multi-scale feature extraction network includes four downsampling stages with feature dimensions of 64, 128, 256, and 512 respectively. The final output is a point-level feature with 256 feature channels. There are 200 instance query vectors. The Transformer decoder has 6 layers. Each layer performs instance query self-attention, cross-scale cross-attention, and feedforward network transformation in sequence. The 3D nonmaximum suppression uses a mask with a 3D intersection-union threshold of 0.

5. The semantic category system covers 31 categories of indoor objects, including walls, floors, ceilings, doors, windows, and various household items.

5. The method according to claim 1, characterized in that, In the parametric reconstruction of the building structural elements, the door and window opening detection step detects locally connected regions in the wall fitting plane where the point cloud density is significantly lower than that of the wall background, and extracts the position coordinates, width, and height parameters of the opening: if the opening is a door, an IfcDoor parametric object is generated; if the opening is a window, an IfcWindow parametric object is generated, and a topological association is established with the wall to which it belongs. In the indoor space topology map, each room node carries the floor outline polygon vertex sequence and spatial enclosing geometric information of the room, and each door opening connecting edge carries the spatial position coordinates, net width dimension, and opening direction attribute of the corresponding door opening.

6. The method according to claim 1, characterized in that, In the graph attention network modeling step, an indoor object spatial relationship graph is constructed using the center of the oriented bounding box of all unstructured instances as nodes and the proximity relationship between nodes with a distance below a preset threshold as edges. The category embedding vector and spatial coordinates of each node are used as node features, and the relative position vector and distance between two nodes are used as edge features. These are input into the graph attention network for message passing. The updated node features are used to verify the physical rationality of the spatial relationship between adjacent objects. Alarms are triggered for detected abnormal spatial relationships and feedback is sent to the user for confirmation.

7. The method according to claim 1, characterized in that, The prohibited placement areas include: The fire evacuation area is the corridor zone within 0.5 meters on each side of the center line of the doorway; The load-bearing structural area refers to the area where the electromechanical pipe shafts and load-bearing columns are located, which are read from the structural layer of the building information model; The door opening area is a fan-shaped area calculated based on the net width of the doorway and the opening direction.

8. The method according to claim 1, characterized in that, The training process of the denoising diffusion probability model adopts a classifier-free guidance mechanism, randomly discarding conditional inputs with a probability of 0.1, so that the model has both unconditional generation and conditional generation capabilities. The training objective is the expected value of the mean square error between predicted noise and real noise. The total number of diffusion steps is 1000, and the noise intensity increases according to the cosine scheduling strategy. The inference process uses the DDIM sampler to reduce the number of inference steps to 50, and the guidance weight coefficient is 7.5, generating 10 candidate layout schemes.

9. The method according to claim 1, characterized in that, The compliance knowledge graph stores fire safety regulations, accessibility design regulations, and interior design regulations from ergonomic standards in the form of rule triples. The formula for calculating the comprehensive score is as follows: ,in For comprehensive scoring, the value range is [0,1]. , , These are the weighting coefficients for compliance score, space utilization score, and aesthetic score, with default values ​​of 0.5, 0.3, and 0.2 respectively. ,in For violation type vectors, The number of rule violations. The total number of rules is 0.85, the overall score threshold is 0.85, and the maximum number of iterations is 5.

10. The method according to claim 1, characterized in that, The collision detection criterion for the axis-aligned bounding boxes is that the projection intervals of the two axis-aligned bounding boxes overlap in all three coordinate axis directions, that is, for any coordinate axis... All meet ,in , respectively enclosed box , The center point is Coordinates along the axis, , respectively enclosed box , exist The half-side length along the axis, and the industrial base class parameterized object type corresponding to each piece of furniture in the final output building information model file is IfcFurnishingElement.

Images