Multi-source data fusion CAD drawing method and system based on deep learning
By using deep learning technology to fuse multi-source data, the problems of spatiotemporal alignment, feature extraction and vectorization in multi-source data fusion were solved, realizing high-precision CAD drawing, improving drawing efficiency and accuracy, and generating navigation maps that meet engineering specifications.
Patent Information
- Application Number
- CN202510971773.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-15
- Publication Date
- 2025-10-31
- Estimated Expiration
- 2045-07-15
AI Technical Summary
Existing multi-source data fusion methods suffer from low spatiotemporal alignment accuracy, poor noise filtering effect in point cloud data, lack of effective mechanisms for cross-modal feature fusion, and lack of engineering constraint optimization in the vectorization process, resulting in low efficiency and insufficient accuracy in CAD drawing.
Deep learning technology is used for multi-source data fusion. Spatiotemporal benchmark alignment is achieved through the ICP algorithm and Zhang Zhengyou calibration method. PointNet++ and MaskRCNN networks are used to extract features. Cross-modal feature fusion is performed by combining graph convolutional networks. Vector maps are generated by DBSCAN and B-spline fitting. Finally, engineering constraint optimization is performed by Hausdorff distance and Dijkstra algorithm.
It improves the automation and accuracy of CAD drafting, reduces labor costs, and generates efficient and reliable navigation maps that meet engineering specifications.
Smart Images

Figure CN120876645A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer-aided design technology, and more specifically, to a deep learning-based multi-source data fusion CAD drawing method and system. Background Technology
[0002] With the rapid development of intelligent buildings and automatic navigation technologies, the demand for precise mapping based on computer-aided design is increasing. Traditional CAD mapping mainly relies on manual surveying and 2D paper processing, which suffers from inefficiency and insufficient accuracy when dealing with large and complex scenarios. While existing multi-source data fusion technologies can integrate data from different sensors, they still face technical bottlenecks in spatiotemporal alignment, feature extraction, and vectorization. Laser point cloud data possesses high-precision 3D information, but is prone to noise interference in complex reflection environments; video data contains rich semantic information but lacks accurate spatial positioning; CAD base maps provide structured topological information, but are difficult to effectively integrate with real-time sensing data.
[0003] Existing multi-source data fusion methods mainly suffer from the following problems: low spatiotemporal alignment accuracy, with time synchronization errors and spatial registration errors between different sensors affecting the quality of data fusion; poor noise filtering effect of point cloud data in complex environments such as metal reflective surfaces, affecting the accuracy of subsequent feature extraction; lack of effective attention mechanisms in cross-modal feature fusion, failing to fully utilize the complementarity of different modal data; and lack of engineering constraint optimization in the vectorization process, resulting in vector maps that are difficult to meet the accuracy and specification requirements of practical applications.
[0004] The pressing technical challenges are how to achieve high-precision spatiotemporal alignment of multi-source data, how to effectively extract and fuse cross-modal features in complex environments, how to generate high-quality vector maps that meet engineering specifications, and how to build intelligent mapping systems suitable for practical applications. Summary of the Invention
[0005] This invention provides a CAD drawing method and system based on deep learning for multi-source data fusion, which solves the technical problems of low accuracy of multi-source data fusion, poor feature extraction effect, and insufficient vectorization quality in related technologies.
[0006] This invention provides a deep learning-based multi-source data fusion CAD drawing method, comprising the following steps:
[0007] Multi-source spatiotemporal reference alignment is performed by acquiring architectural CAD base maps, laser point cloud sequences, and high-definition video streams. Point cloud stitching is performed using the ICP algorithm, and the spatial correspondence of video point clouds is established using the Zhang Zhengyou calibration method to generate a spatiotemporal alignment dataset.
[0008] Reflectivity adaptive point cloud enhancement involves classifying the aligned point cloud by reflectivity material, employing a reflectivity correction method based on Lambert's law, and using adaptive radius filtering for dynamic noise filtering to generate enhanced point clouds.
[0009] Cross-modal feature fusion is achieved by using PointNet++ network to extract point cloud geometric features, MaskRCNN network to extract video semantic features, graph convolutional network to process CAD topological features, and multi-head attention mechanism to realize cross-modal feature fusion.
[0010] Progressive vector generation is achieved by extracting lane centerlines based on fusion features, connecting breakpoints using the DBSCAN clustering algorithm, generating lane vectors through B-spline fitting, vectorizing parking space boundaries using principal component analysis, and identifying parking space numbers using a CRNN network.
[0011] Engineering constraint optimization is performed, geometric accuracy is verified by Hausdorff distance, lane connectivity is verified by Dijkstra's algorithm, and constraint optimization is performed using the L-BFGS solver; the optimized map is output.
[0012] In a preferred embodiment, in the multi-source spatiotemporal reference alignment step, the point cloud timing calibration addresses the ±100ms clock drift problem of the lidar by correcting it through GPS timing deviation and device clock drift; the convergence threshold of the ICP algorithm is set to 0.01m; the camera intrinsic and extrinsic parameter matrices obtained by Zhang Zhengyou calibration method are used to establish the mapping relationship between the point cloud and the image plane.
[0013] In a preferred embodiment, the multi-source spatiotemporal reference alignment step further includes:
[0014] The method employs IMU data fusion and GPS-RTK high-precision positioning, where the IMU inertial measurement data includes data from a three-axis accelerometer and gyroscope, with a sampling frequency of 100Hz; the GPS-RTK high-precision positioning data achieves centimeter-level accuracy.
[0015] Multi-sensor data is fused using an extended Kalman filter, and the state vector includes position, velocity, attitude, and sensor bias.
[0016] In a preferred embodiment, in the reflection adaptive point cloud enhancement step, a lightweight MobileNetv3 neural network is used to perform pixel-level material classification on each frame of the image. The network adopts depthwise separable convolution and SE attention mechanism, and the input image size is 224×224×3. The material classification results are mapped to the point cloud according to the projection relationship, and each point is assigned a material label, including concrete, metal and paint.
[0017] In a preferred embodiment, in the reflection adaptive point cloud enhancement step, the dynamic noise filtering adopts an adaptive radius filtering method to calculate the K nearest neighbor density of each point, with K set to 20; the dynamic threshold is set to the mean point cloud density minus twice the standard deviation of the point cloud density; when the density of a point is less than the threshold, it is identified as a noise point and deleted; and histogram equalization is performed on the video frame to improve the visibility of details in dark areas.
[0018] In a preferred embodiment, in the cross-modal feature fusion step, a PointNet++ network is used to extract point cloud geometric features, which includes four sampling abstraction layers with a sampling point sequence of 4096→1024→256→64; a MaskRCNN network is used to extract video semantic features, based on ResNet50 as the backbone network; and a three-layer graph convolutional network is used to process CAD topological features, outputting topological features with a dimension of 128.
[0019] In a preferred embodiment, the cross-modal feature fusion step employs a multimodal Transformer architecture, including three independent modality encoders and a fusion decoder; each modality is processed by an independent Transformer encoder, and a multi-head attention mechanism enhances expressive power by computing multiple attention heads in parallel; positional encoding is added to point cloud and image features, and the fusion decoder learns the dependencies between modalities through self-attention and cross-attention mechanisms.
[0020] In a preferred embodiment, in the progressive vector generation step, the lane centerline is extracted by selecting a set of points with a confidence level greater than 0.7 and labeled as lane center from the fusion features; the DBSCAN clustering algorithm is used to connect breakpoints, with parameters set to a neighborhood radius of 1.5m and a minimum sample size of 5; a smooth curve is generated by fitting a 3rd-order B-spline to each cluster; the parking space boundary is vectorized by sliding a 2.5m×5m window on the fusion features to detect rectangular candidate regions, and verifying that the parking space size meets the constraints of 2.3m≤width≤2.7m and 4.8m≤length≤5.5m.
[0021] In a preferred embodiment, the engineering constraint optimization step involves verifying geometric accuracy using Hausdorff distance, with an accuracy criterion of no more than 15cm; verifying lane connectivity using Dijkstra's shortest path algorithm, checking whether all lanes form a connected navigation network; verifying facility accessibility, ensuring that the shortest path distance from each elevator or staircase to a lane is no more than 20m; and initiating iterative optimization when the accuracy is not met, with a maximum of 50 iterations.
[0022] In a preferred embodiment, a deep learning-based multi-source data fusion CAD drawing system is used to perform the steps of the aforementioned deep learning-based multi-source data fusion CAD drawing method, including:
[0023] The data acquisition module is used to acquire architectural CAD base maps, laser point cloud sequences, and high-definition video streams. The laser point cloud is acquired by Velodyne HDL64E LiDAR with a sampling frequency of 10Hz, an effective range of 100m, and an accuracy of ±2cm. The high-definition video stream is acquired simultaneously by four 4K resolution cameras at a frame rate of 30fps.
[0024] The spatiotemporal alignment module is used to stitch point clouds together using the ICP algorithm and to establish the spatial correspondence of video point clouds using the Zhang Zhengyou calibration method.
[0025] The point cloud enhancement module is used to classify the reflectivity material of the aligned point cloud, adopts a reflectivity correction method based on Lambert's law, and performs dynamic noise filtering through adaptive radius filtering;
[0026] The feature fusion module is used to extract point cloud geometric features, video semantic features, and CAD topological features using PointNet++, MaskRCNN, and graph convolutional networks, respectively, and to achieve cross-modal feature fusion through a multi-head attention mechanism.
[0027] The vector generation module is used to extract lane centerlines based on fusion features, connect breakpoints using the DBSCAN clustering algorithm, generate lane vectors through B-spline fitting, and vectorize parking space boundaries using principal component analysis.
[0028] The constraint optimization module is used to verify geometric accuracy using Hausdorff distance and lane connectivity using Dijkstra's algorithm, ensuring compliance with specifications for parking space size, lane width, and traffic path.
[0029] The beneficial effects of this invention are as follows: it realizes the intelligent fusion of multi-source heterogeneous data through deep learning technology, improves the automation and accuracy of CAD drawing, reduces labor costs and drawing time, and provides an efficient and reliable technical solution for the generation of large parking lot navigation maps.
[0030] The accuracy and robustness of multi-source data fusion are improved through precise spatiotemporal alignment algorithms and deep learning feature extraction; the feature extraction effect in complex environments is improved by adopting adaptive point cloud enhancement technology and cross-modal attention mechanism; the geometric accuracy and topological consistency of vector maps are improved through progressive vector generation and engineering constraint optimization; and the systematic interface design and standardized output format enhance the scalability and engineering applicability of the system. Attached Figure Description
[0031] Figure 1 This is a flowchart of a deep learning-based multi-source data fusion CAD drawing method according to the present invention. Detailed Implementation
[0032] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.
[0033] At least one embodiment of the present invention discloses a multi-source data fusion CAD drawing method based on deep learning, such as... Figure 1 As shown, it includes the following steps:
[0034] Step 1: Multi-source spatiotemporal reference alignment. Obtain architectural CAD base map, laser point cloud sequence and high-definition video stream. Perform point cloud stitching through ICP algorithm. Use Zhang Zhengyou calibration method to establish the spatial correspondence of video point clouds and generate spatiotemporal alignment dataset.
[0035] Architectural CAD base drawing G CAD DWG format file, containing architectural floor plans, structural column grid, and equipment pipeline layers;
[0036] Laser point cloud sequence P t Data was collected using a Velodyne HDL64E lidar with a sampling frequency of 10Hz, an effective range of 100m, and an accuracy of ±2cm.
[0037] High-definition video stream F k Four 4K resolution cameras simultaneously captured data at a frame rate of 30fps.
[0038] Data processing methods:
[0039] CAD data preprocessing involves converting DWG files to a standardized world coordinate system, extracting key layers, and filtering text annotation layers.
[0040] Point cloud timing calibration addresses the ±100ms clock drift issue with LiDAR by establishing a hardware timestamp mapping:
[0041] t corrected =t raw +Δt GPS -Δt drift
[0042] Where t corrected For the corrected timestamp, traw For the original hardware timestamp, Δt GPS For GPS timing deviation, Δt drift This represents the device clock drift.
[0043] The ICP algorithm (Iterative Closest Point) is used for point cloud stitching. Its objective function is:
[0044]
[0045] pose t Let p be the pose of frame t, R be a 3×3 rotation matrix, t be a 3×1 translation vector, and p be the position of the target. t,i Let NN(p) be the 3D coordinates of the i-th point in the current t-th frame. t-1,i This indicates searching for p in the point cloud of the previous frame. t-1,i The nearest neighbor of the point. The convergence threshold is set to 0.01m.
[0046] The video point cloud spatial correspondence is used to obtain the camera intrinsic parameter matrix K and extrinsic parameter matrix [R|t] through the Zhang Zhengyou calibration method. The Zhang Zhengyou calibration method is a camera calibration method based on a checkerboard pattern, which establishes the mapping relationship between the world coordinate system and the image coordinate system by observing the calibration board from multiple perspectives.
[0047] The transformation formula for projecting a point cloud onto an image plane is:
[0048] [u,v,1] T =K·[R cam |t cam [x,y,z,1] T
[0049] Where (u,v) are the image pixel coordinates, (x,y,z) are the 3D world coordinates of the point cloud, K is the 3×3 camera intrinsic parameter matrix, and R cam For a 3×3 camera rotation matrix, t cam Let T be the 3×1 camera translation vector, and let T denote the transpose operation.
[0050] In one embodiment of the present invention, this step may also employ IMU data fusion and GPS-RTK high-precision positioning methods, as detailed below:
[0051] Enhanced input data sources:
[0052] IMU inertial measurement data I t It includes data from a triaxial accelerometer and gyroscope, with a sampling frequency of 100Hz;
[0053] GPS-RTK high-precision positioning data G t The accuracy reaches the centimeter level;
[0054] IMU pre-integration processing:
[0055] IMU pre-integration is used to process high-frequency motion data; IMU pre-integration is a method that integrates IMU measurements between two keyframes to avoid redundant integration calculations.
[0056] Let there be two times t. i and t j The pre-integral quantities between them are:
[0057]
[0058] Where t i and t j These represent the start and end times of the IMU pre-integration, respectively; Δt is the IMU sampling time interval; k is the time index; and ΔR... ik From time t i to t k The cumulative rotation matrix between t is used to transform t k Acceleration measurements at time t converted to t i In the coordinate system at time t, ΔR ij Let Δv be the rotational change from time i to time j. ik From time t i to t k The cumulative change in velocity between them, Δv ij Let Δp be the change in velocity. ij Let ω be the change in position, Exp(·) be the exponential mapping function, and ω be the position change. k Let a be the measured angular velocity at time k. k Let Δt be the linear acceleration measurement at time k. k Let k be the time interval at time k.
[0059] GPS-RTK enhanced positioning establishes an RTK correction model; RTK is a carrier phase differential technique that eliminates most system errors by differentiating carrier phase observations between the base station and the rover.
[0060] RTK positioning solution formula:
[0061]
[0062] in The values are the double-difference carrier phase observations of satellites p and q from base station 1 and rover station 2, where λ is the carrier wavelength. For the corresponding double-difference pseudo-range, For double-difference integer ambiguity, To observe noise.
[0063] Multi-sensor fusion employs an extended Kalman filter (EKF) to fuse data from multiple sensors.
[0064] The state vector includes position, velocity, attitude, and sensor bias:
[0065] x=[p,v,q,b a ,b g ] T
[0066] Where p is a 3×1 position vector, v is a 3×1 velocity vector, q is a 4×1 quaternion attitude vector, and b a For a 3×1 acceleration bias, b g The bias is set for a 3×1 gyroscope;
[0067] The prediction step updates the state estimate based on the motion model, while the observation update step integrates sensor data from GPS, IMU and other sensors to correct the state estimate, achieving centimeter-level positioning accuracy.
[0068] Prediction equation:
[0069] x k|k-1 =f(x) k-1 ,u k-1 ,w k-1 )
[0070] Update equation:
[0071] x k|k =x k|k-1 +K k (z k -h(x k|k-1 ))
[0072] Where x k|k-1 The state prediction value at time k is based on the information at time k-1, that is, the current state vector estimated based on the IMU pre-integration result and the state at the previous time, which includes the predicted values of the position, velocity, attitude, acceleration bias and gyroscope bias of the lidar carrier.
[0073] x k-1 The optimal state estimate at time k-1 is the optimal pose estimate obtained by fusing observation data from multiple sensors such as GPS-RTK and IMU.
[0074] x k|k The optimal state estimate at time k is the final pose estimate obtained by fusing the current GPS-RTK observations.
[0075] u k-1 The control input vector at time k-1 specifically refers to the IMU measurement value in this system, including triaxial accelerometer data a. k-1 =[ax ,a y ,a z ] T and three-axis gyroscope data ω k-1 =[ω x ,ω y ,ω z ] T Sampling frequency 100Hz;
[0076] w k-1 The process noise vector at time k-1 represents uncertainties such as IMU sensor noise, system modeling error, and environmental interference. It is usually modeled as zero-mean Gaussian white noise.
[0077] f(·) is a nonlinear state transition function, constructed based on the IMU kinematic model and pre-integration theory, which describes the motion law of the carrier in three-dimensional space and includes a set of nonlinear differential equations for position integration, velocity integration and attitude update.
[0078] h(·) is a nonlinear observation function that maps the position, velocity, and attitude parameters in the state space to the GPS-RTK observation space, taking into account the lever effect and installation error between the GPS antenna and the IMU.
[0079] h(x k|k-1 ) represents the observation prediction value based on the predicted state, which is the GPS position prediction value obtained after transforming the state prediction value through the observation function h(·), and is used to calculate the observation residual;
[0080] z k This is the observation vector at time k, mainly containing GPS-RTK high-precision positioning data, in the format of three-dimensional position coordinates [x...]. GPS ,y GPS ,z GPS ] T Positioning accuracy reaches the centimeter level, with an update frequency of 10Hz;
[0081] K k The Kalman gain matrix at time k is calculated by minimizing the covariance of the posterior state estimate. It is used to balance the state prediction and GPS observation values. Its magnitude reflects the degree of confidence in the prediction model and the observation data. The gain is larger when the GPS signal quality is good and smaller when the GPS signal is blocked.
[0082] Output: Generate a spatiotemporally aligned dataset Includes registration transformation matrix Time synchronization mapping table and spatially calibrated point cloud P aligned =T init ·P t The format is an N×6 array (x,y,z,I,r,g);
[0083] When using enhancement methods, high-precision spatiotemporally aligned datasets can be obtained. Registration accuracy has been improved to ±3cm.
[0084] Step 2, adaptive point cloud enhancement of reflection: The aligned point cloud is classified by reflectivity material, and a reflectivity correction method based on Lambert's law is adopted. Dynamic noise filtering is performed through adaptive radius filtering to generate enhanced point cloud.
[0085] Input data source: Alignment point cloud P output from step 1 aligned and synchronized video frames F k .
[0086] Data processing methods:
[0087] For reflectivity-based material classification, addressing the issue of abnormal laser return intensity caused by the reflectivity of metal parking spaces, a video-based material recognition method is employed. A lightweight MobileNetv3 neural network is used to perform pixel-level material classification for each frame of the image. The network employs depthwise separable convolution and SE attention mechanisms. The input image size is 224×224×3, and the output is the material classification result.
[0088] Based on the projection relationship, the material classification results are mapped to the point cloud, and a material label m is assigned to each point. i ∈{concrete,metal,paint}, where m i This represents the material category of the i-th point cloud.
[0089] Reflectivity correction based on Lambert's law of reflection:
[0090]
[0091] in The corrected reflection intensity at point i. The original reflection intensity, For material m i The corresponding compensation coefficients (concrete 0.9, metal 0.4, paint 0.7), θ i The incident angle between the laser and the surface normal vector is denoted as .
[0092] Dynamic noise filtering employs an adaptive radius filtering method, first calculating the K-nearest neighbor density for each point:
[0093]
[0094] Where ρ i Let r be the density of the i-th point, K be the number of nearest neighbors (let's say 20), and r be the density of the i-th point. K,i Let be the distance from the i-th point to the k-th nearest neighbor.
[0095] Calculate the global density statistic and set a dynamic threshold ρ. thresh =μ ρ -2σ ρ , where μ ρ σ is the mean density of the point cloud. ρ Let ρ be the standard deviation of point cloud density; when ρ i <ρ thresh When the i-th point is identified as a noise point, it is deleted.
[0096] Illumination homogenization involves histogram equalization of video frames to redistribute image grayscale values, thereby improving the visibility of details in dark areas and enhancing the effectiveness of subsequent video feature extraction.
[0097] Output: Enhanced point cloud P enhanced It includes the corrected reflection intensity value, material label vector, and filter quality identifier.
[0098] Step 3: Cross-modal feature fusion. PointNet++ network is used to extract point cloud geometric features, MaskRCNN network is used to extract video semantic features, graph convolutional network is used to process CAD topological features, and multi-head attention mechanism is used to achieve cross-modal feature fusion.
[0099] Input data source: Enhanced point cloud P from step 2 enhanced Corresponding video frame F k and CAD topology map G CAD ;
[0100] Data processing methods:
[0101] Branch feature extraction:
[0102] The point cloud geometry branch uses the PointNet++ network for point cloud feature extraction. The network processes point cloud data through hierarchical feature learning and contains 4 sampling abstraction layers with a sampling point sequence of 4096→1024→256→64.
[0103] Each layer uses the farthest point sampling to select representative points, the BallQuery method to construct local neighborhoods, and the PointNet module to extract features within the local regions;
[0104] Output geometric feature Φ g ∈R 64×512 , where 64 is the final number of sampling points and 512 is the feature dimension.
[0105] The video semantic branch employs a pre-trained Mask R-CNN network for semantic segmentation. The network uses ResNet50 as its backbone, generates candidate boxes through a region candidate network, performs feature alignment using the RoIAlign layer, completes object detection with a classification and regression head, and implements instance segmentation using a masking head. The output is a semantic feature map Φ. v ∈R H / 16×W / 16×256 Where H and W are the height and width of the input image, and 256 is the number of feature channels; the output is the detection bounding box B. det ;
[0106] The CAD topology branch employs a 3-layer graph convolutional network to process the CAD topology structure. It represents the connections between CAD elements using an adjacency matrix, and the graph convolutional layers learn the spatial propagation patterns of node features. The network outputs topology features Φ. t ∈R M×128 , where M is the number of nodes in the CAD drawing and 128 is the topological feature dimension;
[0107] Spatial correspondence is established by creating a point cloud-to-image projection matrix and a CAD-to-point cloud registration matrix, and constructing a spatial correspondence matrix C∈R between the three modalities. N′×H×V , where N′ is the number of points in the enhanced point cloud, and H and V are the height and width of the video frame, respectively;
[0108] Attention weight calculation: Cross-modal fusion weights are calculated using a multi-head attention mechanism.
[0109]
[0110] Where α i,j,k Let φ be the attention weights for the i-th point in the point cloud, the j-th pixel in the video, and the k-th node in the CAD. g,i Let φ be the geometric feature of the i-th point in the point cloud. v,j Let φ be the semantic feature of the j-th pixel in the video. t,k Let f be the topological feature of the k-th node in the CAD system. att It is a 3-layer MLP attention network with hidden layer dimensions of [896,512,1].
[0111] Optionally, this step can also employ a multimodal Transformer attention mechanism, as detailed below:
[0112] The multimodal Transformer architecture is designed with a dedicated multimodal Transformer network, which includes three modal encoders and one fusion decoder.
[0113] Modal encoders, each mode employing an independent Transformer encoder; for point cloud modes, the point cloud features are serialized into... in Let N be the feature vector of the i-th point, and N′ be the number of points in the augmented point cloud.
[0114] The self-attention mechanism calculates attention weights through a query, key, and value matrix:
[0115]
[0116] Where Q = X p W Q For querying the matrix, K = X p W K Let V be the bond matrix, and V = X p W V W is a value matrix. Q W K W V Let d be the learnable first, second, and third weight matrices. k is the dimension of the key vector.
[0117] Multi-head attention enhances expressive power by computing multiple attention heads in parallel:
[0118] Output projection matrix;
[0119] Cross-modal attention: Design a cross-modal attention mechanism to handle interactions between different modalities, and calculate cross-modal features of three pairs of modalities: point cloud-video, point cloud-CAD, and video-CAD;
[0120] The fusion decoder adopts a Transformer decoder structure to fuse multimodal features and learns the dependencies between modalities through self-attention and cross-attention mechanisms;
[0121] Position encoding adds positional codes to point cloud and image features; for point clouds, 3D positional encoding is used.
[0122]
[0123] Where (x, y, z) are the 3D coordinates of the point cloud, i is the feature dimension index, and d model This represents the feature dimension of the model.
[0124] Output: Fusion feature tensor Φ fuse ∈R N′×896 It includes spatial index and confidence score c. i ∈[0,1] and semantic label L i ∈{parking space boundary, lane center, obstacle, facility};
[0125] When using the Transformer method, enhanced fusion features can be obtained. The ability to express features has been improved.
[0126] Step 4: Progressive vector generation. Based on the fusion feature extraction, the lane centerline is extracted, the DBSCAN clustering algorithm is used to connect the breakpoints, the lane vector is generated by B-spline fitting, the parking space boundary is vectorized using principal component analysis, and the parking space number is identified using a CRNN network.
[0127] Input data source: Fusion feature Φ from step 3 fuse and the original video frame F k ;
[0128] Data processing methods:
[0129] Lane centerline extraction, from Φ fuse Extracting the lane center response P center ={p i |L i = Lane center, c i >0.7}.
[0130] The DBSCAN clustering algorithm is used to connect breakpoints. DBSCAN is a density-based clustering algorithm whose core idea is to group sufficiently dense points into one class.
[0131] The parameters are set to neighborhood radius ∈ = 1.5m and minimum number of samples min_samples = 5; the algorithm groups points with density connection into the same cluster, effectively connecting lane center line segments that are blocked or interfered with by noise;
[0132] For each cluster, perform B-spline fitting to generate a smooth curve:
[0133]
[0134] Where C(t) is the coordinate of the curve point at parameter t, and N i,p (t) is a p-th degree B-spline basis function, P i Let p = 3 be the coordinates of the i-th control point, p = 3 be the order of the B-spline, and n be the number of control points.
[0135] Parking space boundary vectorization: in Φ fuse A 2.5m × 5m window is slid upwards to detect rectangular candidate regions. Principal Component Analysis (PCA) is used to fit the principal direction of the parking space boundaries; PCA finds the main direction of data variation through eigenvalue decomposition.
[0136]
[0137] Where C is the covariance matrix, X is the decentralized data matrix, and n is the number of control points.
[0138] Geometric constraint verification: The parking space dimensions must satisfy 2.3m≤w≤2.7m and 4.8m≤l≤5.5m;
[0139] Parking space number recognition: A CRNN network is used for text recognition; CRNN combines the advantages of CNN and RNN.
[0140] The CNN part extracts image feature sequences;
[0141] In the RNN part, LSTM learns sequence dependencies;
[0142] In the CTC section, attention should be paid to temporal classification and sequence labeling issues should be addressed.
[0143] Facility location is determined by identifying elevator / staircase signs based on template matching, and the Euclidean distance to the nearest driveway is calculated.
[0144] Optionally, this step can also employ a graph neural network adaptive vectorization method, as detailed below:
[0145] Graph structure construction transforms the vectorization problem into a graph structure learning problem;
[0146] Construct a graph containing nodes and edges:
[0147]
[0148] The set of nodes Representing potential parking spaces or lane key points, each node represents a potential parking space or lane key point edge set: edge set Connect nodes whose distance is less than a threshold τ, where τ is the distance threshold.
[0149] Graph Neural Network Vectorization: A dedicated GNN architecture is designed for vectorization; node features are initialized using an MLP network.
[0150]
[0151] in Let i be the initial feature of node i. For the corresponding fusion features, MLP stands for Multilayer Perceptron.
[0152] Interaction information between computing nodes during the message passing phase:
[0153]
[0154] in The message passed from node i to node j in layer l. and The features of nodes i and j at layer l are respectively, e ij EdgeMLP represents an edge-multilayer perceptron, which contains edge features (including geometric attributes such as distance and angle).
[0155] During the node update phase, aggregate neighbor messages to update node characteristics:
[0156]
[0157] in Let represent the feature vector of node i in the (l+1)th layer, and let represent the node representation after one message passing and feature update. Let represent the feature vector of node i in the l-th layer, which serves as the input feature vector for the current node. NodeMLP represents a node multilayer perceptron, and AGG is the aggregation function (summation, mean, or maximum). Let i be the set of neighbors of node i.
[0158] Geometric constraint integration: Integrating geometric constraints into the GNN, including angle constraints (parking space boundaries are approximately perpendicular, angle difference |θ) ij -90°|<15°), distance constraint (center distance between adjacent parking spaces 2.3m≤d) center ≤2.8m) and connectivity constraints (lane formation connectivity graph structure).
[0159] The constraint loss function is a weighted sum of all constraint terms:
[0160]
[0161] Where λ1, λ2, and λ3 are the first, second, and third loss weight coefficients, respectively. These are the loss terms for angle, distance, and connectivity constraints, respectively.
[0162] Adaptive post-processing: based on the detection confidence level c i and local density ρ i Adaptive adjustment of post-processing parameters:
[0163] α i =σ(w1c i +w2ρ i +b)
[0164] Where α i For adaptive weights, σ is the sigmoid activation function, w1 and w2 are the first and second adaptive weight parameters, respectively, and b is the bias term;
[0165] Output: A collection of vector elements It includes lane network (road identifier, geometry, width, direction of travel), parking space set (identifier, geometry, number, center point, orientation), facility location (type, location coordinates, access road) and metadata (total number of parking spaces, coverage area, generation timestamp), stored in structured JSON format;
[0166] When using graph neural network methods, an adaptive set of vector elements can be obtained. It has higher geometric consistency and topological correctness.
[0167] Step 5: Engineering constraint optimization. Verify geometric accuracy using Hausdorff distance, verify lane connectivity using Dijkstra algorithm, perform constraint optimization using L-BFGS solver, and output the optimized map.
[0168] Input data source: Vector elements from step 4 And engineering specifications and constraints (Code for Design of Automobile Garages);
[0169] Data processing methods:
[0170] Accuracy verification involves projecting the vectorized parking spaces back into the point cloud space and calculating the Hausdorff distance to verify geometric accuracy.
[0171]
[0172] Where d H Let P be the Hausdorff distance. enhanced To enhance point clouds, P proj For vector projection point clouds, d(p,q) is the Euclidean distance between points p and q, and the accuracy criterion is d. H ≤15cm, sup is the supremum operator, representing the maximum value in the set, and inf is the infmum operator, representing the minimum value in the set. Represents the projection point set P from point p. proj The minimum distance to all points, i.e. the shortest distance from point p to the set of projected points;
[0173] The calculation of the Hausdorff distance involves two directions: the first term The second term measures the maximum nearest distance from the augmented point cloud to the projected point set. The maximum nearest distance from the projected point set to the augmented point cloud is measured, and the maximum value of the two is taken as the final Hausdorff distance.
[0174] Verify lane connectivity using Dijkstra's shortest path algorithm, checking if all lanes form a connected navigation network. Verify facility accessibility, ensuring that the shortest path distance from each elevator / staircase to a lane is ≤20m.
[0175] Connectivity verification, specifically, involves constructing an undirected graph of all lane nodes:
[0176] G road =(V road E road )
[0177] Where V road E is the set of lane intersections. road Given a set of lane segments connecting adjacent intersections; the algorithm starts from any node and calculates the shortest path to all other nodes. If there are unreachable nodes, it indicates that the lane network is not connected and the broken connections need to be corrected.
[0178] Facility accessibility verification, for each elevator and stair facility point f i Calculate its Euclidean distance to the nearest lane. like The facility is then deemed inaccessible, requiring adjustments to the lane layout or relocation of the facility.
[0179] The verification process includes: identifying the location coordinates of all facilities; calculating the distance from each facility to all nodes in the lane network; selecting the minimum distance as the accessibility index; and checking whether the constraint condition of ≤20m is met.
[0180] Constrained optimization is performed, and iterative optimization is initiated when the accuracy is insufficient. The objective function is established as follows:
[0181]
[0182] in For vector adjustment, Let Frobenius norm represent the overall magnitude of the adjustment, and λ be the penalty weight, ∈ i Let be the precision error of the i-th vector element.
[0183] The constraints include: parking space dimensions conforming to specifications, lane width ≥ 2.5m, and unobstructed passageways. The solution is obtained using the L-BFGS finite-memory quasi-Newton method, with a maximum of 50 iterations.
[0184] Optionally, this step can also employ a multi-objective intelligent constraint optimization method, as detailed below:
[0185] Multi-objective optimization: Transforming constrained optimization into a multi-objective optimization problem.
[0186]
[0187] in The geometric accuracy objective function is used to minimize the geometric deviation between the vectorized result and the original point cloud data. The calculation formula is as follows:
[0188]
[0189] in Let be the Hausdorff distance between the i-th vector element and the corresponding point cloud region. P represents the i-th vector element.i This represents the i-th corresponding point cloud region; Let be the topology consistency objective function, used to ensure the connectivity of the lane network and the logical correctness of the parking space layout. The calculation formula is as follows:
[0190]
[0191] Where N disconnect N represents the number of connections to be disconnected. overlap The number of overlapping conflicts is represented by α, and β is the first and second topological consistency weighting coefficients, respectively. The objective function for engineering specification compliance is used to evaluate whether map elements meet the requirements of the "Code for Design of Automobile Garages". The calculation formula is as follows:
[0192]
[0193] Where ∈ j Let be the normalization deviation of the j-th element, ∈ std To standardize the allowable standard deviation threshold.
[0194] Genetic Algorithm Solution: The NSGA-II non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem. The algorithm maintains population diversity through non-dominated sorting and crowding distance selection strategies, and uses simulated binary crossover and polynomial mutation operations to generate new individuals;
[0195] The fitness evaluation function is a weighted combination of all objectives:
[0196]
[0197] w3, w4, and w5 are the third, fourth, and fifth weighting coefficients, respectively. For geometric precision functions, For topological consistency functions, To standardize the compliance function;
[0198] The geometric accuracy function measures the degree of geometric error matching between the vectorized result and the original point cloud data. This function calculates the Hausdorff distance between all vector elements and the corresponding point cloud regions, compares the accumulated error with the theoretical maximum error, and derives a standardized accuracy assessment value.
[0199] Specifically, the function first calculates each vector element. Its corresponding point cloud region P i Hausdorff distance between Then, sum all the errors, divide by the product of the total number of vector elements and the maximum permissible error threshold for normalization, and subtract this ratio from 1 to obtain the accuracy function value.
[0200] The topology consistency function is used to evaluate whether the topological relationships between elements in a parking map are correct, including the connectivity of the lane network and the spatial logical relationships between elements. The function compares the number of disconnections and overlapping conflicts in the graph structure with the total number of topological relationships to obtain an evaluation value for topology consistency.
[0201] Specifically, the function calculates the number N of disconnected lanes in the lane network that cannot reach each other. disconnect The number N of elements that overlap and conflict with the existing space overlap Add the two together and divide by the total number of topological relationships N. total Finally, subtract the ratio from 1 to obtain the topology consistency function value;
[0202] The code compliance function is used to measure whether each element in the parking lot map meets the standard requirements of the "Code for Design of Automobile Garages". This function checks whether the design parameters of each map element meet the code requirements, calculates the total deviation of all elements that exceed the allowable deviation range, and compares it with the theoretical maximum deviation to obtain the code compliance evaluation value.
[0203] Specifically, the function calculates the normalization deviation ∈ for each map element j. j When the deviation exceeds the standard deviation threshold allowed by the specification ∈ std When the deviation exceeds the standard deviation threshold, the excess is added to the total deviation, then divided by the product of the total number of map elements and the standard deviation threshold for normalization, and finally the standard compliance function value is obtained by subtracting the ratio from 1.
[0204] Output: Optimized map M final It includes all vector elements that meet the accuracy requirements, quality reports (geometric accuracy, topological integrity, number of optimization iterations), and change logs (element ID, original geometry, optimized geometry, reason for change);
[0205] When using a multi-objective optimization method, a multi-objective optimized map M can be obtained. optimized It achieves the best balance between accuracy, topology, and normalization.
[0206] In one embodiment of the present invention, an application example of multi-source data fusion navigation map generation for an underground parking lot is provided;
[0207] Application scenario: Intelligent navigation system construction project in the underground parking lot of a shopping mall;
[0208] The parking lot has three underground levels, a total of 2,580 parking spaces, and a total building area of 18,450 square meters.
[0209] The space layout consists of three levels: B1 (980 parking spaces), B2 (850 parking spaces), and B3 (750 parking spaces), including 12 elevator lobbies and 8 stairwells.
[0210] Environmental characteristics: Lighting conditions: LED strip lighting, brightness variation range 50-800 lux, with alternating light and dark areas; Material distribution: 70% epoxy flooring (grayscale reflectivity 0.3), 25% metal anti-slip strips (reflectivity 0.85), 5% colored marking paint (reflectivity 0.6); Spatial complexity: dense column grid (6m×8m), turning radius 2.5-4.2m, floor height 3.6m;
[0211] Data acquisition and processing;
[0212] Input data:
[0213] Architectural CAD base map: 2018 architectural construction DWG file, containing 45 layers;
[0214] Laser point cloud: 240 million point cloud data were collected using Velodyne HDL64E LiDAR;
[0215] Video data: 180 minutes of high-definition video collected simultaneously by 4 4K cameras;
[0216] Auxiliary data: RTK-GPS positioning data (entrance area) and IMU inertial measurement data;
[0217] Detailed processing procedure:
[0218] Step 1: Alignment of multi-source spatiotemporal references;
[0219] CAD data preprocessing: converting DWG to a standardized world coordinate system, and extracting building outlines, column grids, and equipment layers;
[0220] Point cloud timing calibration solves the ±100ms clock drift of LiDAR and uses GPS-RTK for timing correction.
[0221] Point cloud spatial registration is performed using the ICP iterative nearest point algorithm, increasing the number of sampling points to 5000 points / frame in complex regions;
[0222] Multi-sensor fusion was achieved by using an extended Kalman filter to fuse IMU and point cloud data, ultimately controlling the registration error to within ±3.2cm.
[0223] Step 2, adaptive point cloud enhancement based on reflection;
[0224] For material classification and recognition, MobileNetv3 is used for pixel-level material classification to address the issue of metal parking space surfaces (with a reflectivity of up to 85%).
[0225] Reflectivity correction involves setting compensation coefficients for different materials (concrete, metal, paint) to correct the reflection intensity.
[0226] Dynamic noise filtering uses an adaptive radius filtering method, setting K=20 neighborhood points, and filters out approximately 120 million noise points;
[0227] Illumination equalization: Histogram equalization processing is performed on the captured video, increasing the brightness of low-light areas by approximately 42%.
[0228] Step 3, cross-modal feature fusion;
[0229] Point cloud geometric features were extracted using the PointNet++ network, with sampling levels of 4096→1024→256→64 points.
[0230] Video semantic features are extracted using a Mask R-CNN network (ResNet50 backbone network) to identify parking lines and signs;
[0231] CAD topology features are processed using a 3-layer graph convolutional network with a node feature dimension of 128.
[0232] Multimodal Transformer fusion is implemented, and a 6-head multimodal attention mechanism is designed to fuse features from three modalities and generate a 1024-dimensional feature vector.
[0233] Step 4, progressive vector generation;
[0234] Lane centerlines were extracted, and the DBSCAN clustering algorithm (parameter ε = 1.5m) was used to connect breakpoints. A smooth curve was fitted using 3rd-order B-splines.
[0235] Parking space boundaries are vectorized, a sliding 2.5m×5m window is used to detect candidate areas, PCA analysis is used to determine orientation, and 768 candidate parking spaces are identified.
[0236] Parking space number recognition uses the CRNN+CTC recognition algorithm to process parking space numbers in video frames, achieving an accuracy rate of 96%.
[0237] Graph neural network optimization involves constructing a connected graph structure, applying a 3-layer GNN architecture for vectorization, and integrating angle and distance constraints.
[0238] Step 5, Engineering Constraint Optimization;
[0239] Accuracy verification: The geometric accuracy was verified by calculating the Hausdorff distance, and 42 points (>15cm) were found to be substandard.
[0240] The constraints are set as follows: parking space dimensions must meet the following requirements: 2.3m ≤ width ≤ 2.7m, 4.8m ≤ length ≤ 5.5m; lane width ≥ 2.5m.
[0241] Multi-objective optimization was performed using the NSGA-II genetic algorithm with a population size of 200, resolving 32 constraint conflicts in 28 iterations.
[0242] Topology verification: Dijkstra's algorithm verifies lane connectivity to ensure all parking spaces are reachable and corrects 10 topology errors.
[0243] Navigation engine interface output;
[0244] Format conversion: internal JSON is converted to GeoJSON format, and a WGS84 coordinate reference system is added;
[0245] API encapsulation, establishing RESTful interfaces to support parking space query, route planning, and status push;
[0246] Performance optimization: Redis caches frequently accessed data, and the interface response time is controlled within 80ms;
[0247] Deploy and publish the final navigation map (23MB), which is deployed on a cloud server and supports 2000 concurrent connections.
[0248] The actual performance parameters of each processing module are shown in Table 1:
[0249] Table 1: Actual performance parameters of each processing module
[0250]
[0251] The actual application indicators are shown in Table 2:
[0252] Table 2: Practical Application Indicators of Underground Parking Lot Navigation Map Generation Method
[0253] Evaluation indicators Traditional manual surveying Single sensor automation This method Total processing time 14 person-days (112 hours) 2.5 person-days (20 hours) 4.3 hours Total number of parking spaces / Number of recognitions 762 / 762 762 / 594 762 / 748 Spatial registration error ±5cm ±30cm ±3.2cm Parking space boundary positioning accuracy ±3cm ±25cm ±4.8cm Accurate number of parking space numbers 762 528 732 generated map file size 52MB 18MB 23MB Cost of a single data acquisition device ¥45,000 ¥28,000 ¥83,000 Manual verification time 40 hours 12 hours 2.5 hours Data update cycle 90 days 30 days 7 days
[0254] The recognition results are shown in Table 3:
[0255] Table 3: Specific recognition performance data under different scenario conditions
[0256]
[0257] Application effect:
[0258] The underground parking lot navigation map has been successfully deployed as a real-time navigation service; the specific effects are as follows:
[0259] The navigation app has 2,800 daily active users, a single positioning accuracy of ±3.2cm, and an average navigation path calculation time of 280ms.
[0260] The average vehicle location time was reduced from 8.4 minutes to 5.3 minutes after deployment;
[0261] Parking space utilization increased from 72% to 85%;
[0262] The system supports updating map data every 7 days, which can dynamically reflect changes in parking space status such as temporary occupancy and maintenance;
[0263] The server has an average load of 23%, a peak load of 58%, and has been running stably for 94 days.
[0264] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.
Claims
1. A deep learning-based multi-source data fusion CAD drawing method, characterized in that, Includes the following steps: Multi-source spatiotemporal reference alignment is performed by acquiring architectural CAD base maps, laser point cloud sequences, and high-definition video streams. Point cloud stitching is performed using the ICP algorithm, and the spatial correspondence of video point clouds is established using the Zhang Zhengyou calibration method to generate a spatiotemporal alignment dataset. Reflectivity adaptive point cloud enhancement involves classifying the aligned point cloud by reflectivity material, employing a reflectivity correction method based on Lambert's law, and using adaptive radius filtering for dynamic noise filtering to generate enhanced point clouds. Cross-modal feature fusion is achieved by using PointNet++ network to extract point cloud geometric features, MaskRCNN network to extract video semantic features, graph convolutional network to process CAD topological features, and multi-head attention mechanism to realize cross-modal feature fusion. Progressive vector generation is achieved by extracting lane centerlines based on fusion features, connecting breakpoints using the DBSCAN clustering algorithm, generating lane vectors through B-spline fitting, vectorizing parking space boundaries using principal component analysis, and identifying parking space numbers using a CRNN network. Engineering constraint optimization is performed, geometric accuracy is verified by Hausdorff distance, lane connectivity is verified by Dijkstra's algorithm, and constraint optimization is performed using the L-BFGS solver; the optimized map is output.
2. The CAD drawing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In the multi-source spatiotemporal reference alignment step, the point cloud timing calibration addresses the ±100ms clock drift issue of lidar by correcting it using GPS timing deviation and device clock drift; the convergence threshold of the ICP algorithm is set to 0.01m. The camera intrinsic and extrinsic parameter matrices obtained by Zhang Zhengyou's calibration method are used to establish the mapping relationship between point clouds and image planes.
3. The deep learning-based multi-source data fusion CAD drawing method according to claim 2, characterized in that, The multi-source spatiotemporal reference alignment steps also include: The method employs IMU data fusion and GPS-RTK high-precision positioning, where the IMU inertial measurement data includes data from a three-axis accelerometer and gyroscope, with a sampling frequency of 100Hz; the GPS-RTK high-precision positioning data achieves centimeter-level accuracy. Multi-sensor data is fused using an extended Kalman filter, and the state vector includes position, velocity, attitude, and sensor bias.
4. The CAD drawing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In the reflection adaptive point cloud enhancement step, a lightweight MobileNetv3 neural network is used to perform pixel-level material classification for each frame of the image. The network adopts depthwise separable convolution and SE attention mechanism, and the input image size is 224×224×3. The material classification results are mapped to the point cloud according to the projection relationship, and each point is assigned a material label, including concrete, metal and paint.
5. The deep learning-based multi-source data fusion CAD drawing method according to claim 4, characterized in that, In the reflection adaptive point cloud enhancement step, the dynamic noise filtering adopts the adaptive radius filtering method, calculates the K nearest neighbor density of each point, and sets the K value to 20; sets the dynamic threshold to the mean point cloud density minus twice the standard deviation of the point cloud density; when the density of a point is less than the threshold, it is identified as a noise point and deleted; and performs histogram equalization on the video frame to improve the visibility of details in dark areas.
6. The CAD drawing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In the cross-modal feature fusion step, PointNet++ network is used to extract point cloud geometric features, which includes 4 sampling abstraction layers with a sampling point sequence of 4096→1024→256→64; MaskRCNN network is used to extract video semantic features, based on ResNet50 as the backbone network; CAD topological features are processed through a 3-layer graph convolutional network to output topological features with a dimension of 128.
7. The CAD drawing method based on deep learning and multi-source data fusion according to claim 6, characterized in that, The cross-modal feature fusion step adopts a multimodal Transformer architecture, including three independent modal encoders and one fusion decoder. Each modality is processed by an independent Transformer encoder, and the multi-head attention mechanism improves expressive power by computing multiple attention heads in parallel. Positional encoding is added to point cloud and image features, and the fusion decoder learns the dependencies between modalities through self-attention and cross-attention mechanisms.
8. The CAD drawing method based on deep learning and multi-source data fusion according to claim 1, characterized in that, In the progressive vector generation step, the lane centerline is extracted by selecting a set of points with a confidence level greater than 0.7 and labeled as lane center from the fusion features; the DBSCAN clustering algorithm is used to connect breakpoints with parameters set to a neighborhood radius of 1.5m and a minimum sample size of 5; a smooth curve is generated by fitting a 3rd-order B-spline to each cluster; the parking space boundary is vectorized by sliding a 2.5m×5m window on the fusion features to detect rectangular candidate regions and verifying that the parking space size meets the constraints of 2.3m≤width≤2.7m and 4.8m≤length≤5.5m.
9. A deep learning-based multi-source data fusion CAD drawing method according to claim 1, characterized in that, In the engineering constraint optimization step, Hausdorff distance is used to verify geometric accuracy, with an accuracy criterion of no more than 15cm; Dijkstra's shortest path algorithm is used to verify lane connectivity and check whether all lanes form a connected navigation network. Verify facility accessibility to ensure that the shortest path distance from each elevator or staircase to the driveway is no greater than 20m. When the accuracy is not up to standard, iterative optimization is initiated, with a maximum of 50 iterations.
10. A deep learning-based multi-source data fusion CAD drawing system, used to execute the steps of the deep learning-based multi-source data fusion CAD drawing method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire architectural CAD base maps, laser point cloud sequences, and high-definition video streams. The laser point cloud is acquired by Velodyne HDL64E LiDAR with a sampling frequency of 10Hz, an effective range of 100m, and an accuracy of ±2cm. The high-definition video stream is acquired simultaneously by four 4K resolution cameras at a frame rate of 30fps. The spatiotemporal alignment module is used to stitch point clouds together using the ICP algorithm and to establish the spatial correspondence of video point clouds using the Zhang Zhengyou calibration method. The point cloud enhancement module is used to classify the reflectivity material of the aligned point cloud, adopts a reflectivity correction method based on Lambert's law, and performs dynamic noise filtering through adaptive radius filtering; The feature fusion module is used to extract point cloud geometric features, video semantic features, and CAD topological features using PointNet++, MaskRCNN, and graph convolutional networks, respectively, and to achieve cross-modal feature fusion through a multi-head attention mechanism. The vector generation module is used to extract lane centerlines based on fusion features, connect breakpoints using the DBSCAN clustering algorithm, generate lane vectors through B-spline fitting, and vectorize parking space boundaries using principal component analysis. The constraint optimization module is used to verify geometric accuracy using Hausdorff distance and lane connectivity using Dijkstra's algorithm, ensuring compliance with specifications for parking space size, lane width, and traffic path.
Citation Information
Patent Citations
Structural semantic map construction method for underground weak-light weak-texture environment
CN116912440A
Pumped storage power station construction anomaly detection method and system based on unmanned aerial vehicle image analysis
CN119888507A
A system for automatically generating realistic elevation images of three-dimensional shaped structures using 3D point clouds and automatic overlay matching with CAD-based drawings, and method thereof
KR102590975B1
Computer-aided design sequence and knowledge inference from product images
WO2025097118A1
Cited By
Laser point cloud and video fusion method and device based on 3D deploy and control ball
CN121639482A
