Laser scanning three-dimensional imaging method, device, medium and equipment

Abstract

The present application belongs to the field of three-dimensional imaging and measurement technology, and specifically discloses a laser scanning three-dimensional imaging method, device, medium and equipment, the method comprising: obtaining laser scanning data of a target object, the laser scanning data including, for example, depth data including reflection intensity, scanning angle, timestamp and flight time; preprocessing the laser scanning data, and fusing time series and scanning path information to obtain multi-channel input features adapted to a neural network; constructing a three-dimensional geometric reconstruction model, and training the model; inputting the multi-channel input features into the trained three-dimensional geometric reconstruction model to obtain a dual-channel output tensor comprising a dense depth map of the target object and a corresponding uncertainty map; converting the dual-channel output tensor comprising the dense depth map of the target object and the corresponding uncertainty map into a three-dimensional point cloud in world coordinates; and post-processing the three-dimensional point cloud to obtain a three-dimensional imaging of the target object.

Description

Technical Field

[0001] The present application belongs to the field of three-dimensional imaging and measurement technology, and specifically relates to a laser scanning three-dimensional imaging method, device, medium and equipment. Background Art

[0002] The current mainstream laser 3D imaging methods mainly include time-of-flight (ToF), structured light projection, and phase-shift interferometry. Among them, the time-of-flight method obtains depth information by measuring the round-trip time of the laser, and has the ability to image at long distances, but its depth resolution is limited and it is sensitive to changes in surface reflectivity. The structured light projection method uses a coded grating in conjunction with a camera to achieve high-precision reconstruction, but the system is complex and has poor resistance to environmental interference, making it difficult to apply to dynamic or complex background scenes. The phase-shift interferometry method has submillimeter accuracy and is suitable for microscopic measurements, but the system is sensitive to vibration and noise and has a low scanning speed.

[0003] In general, existing three-dimensional imaging technologies often improve spatial resolution at the expense of imaging speed or system robustness. Especially when faced with dynamic scenes, complex structures or non-coplanar targets, traditional single-path scanning methods find it difficult to balance measurement accuracy, system stability and acquisition efficiency.

[0004] Therefore, there is an urgent need for a new laser three-dimensional imaging method that integrates multi-path information, has high resolution and high robustness, to meet the application needs of multiple scenarios such as precision manufacturing, medical imaging and dynamic perception. Summary of the Invention

[0005] In response to the deficiencies in the prior art, the purpose of this application is to provide a laser scanning three-dimensional imaging method, device, medium and equipment, which can improve the accuracy, robustness and application stability of laser scanning three-dimensional imaging.

[0006] To achieve the above objectives, this application provides the following technical solutions:

[0007] A laser scanning three-dimensional imaging method, the method comprising: acquiring laser scanning data of a target object, the laser scanning data including, for example, depth data including reflection intensity, scanning angle, timestamp, and flight time; preprocessing the laser scanning data and fusing time series with scanning path information to obtain multi-channel input features adapted to a neural network; constructing a three-dimensional geometric reconstruction model and training the model; inputting the multi-channel input features into the trained three-dimensional geometric reconstruction model to obtain a dual-channel output tensor comprising a dense depth map of the target object and a corresponding uncertainty map; converting the dual-channel output tensor comprising the dense depth map of the target object and the corresponding uncertainty map into a three-dimensional point cloud in world coordinates; and post-processing the three-dimensional point cloud to obtain a three-dimensional imaging of the target object.

[0008] Optionally, the laser scanning data is preprocessed and the time series and scanning path information are fused to obtain multi-channel input features adapted to the neural network, including: point cloud filtering of the laser scanning data and projecting the filtered point cloud into a sparse depth map; extracting the laser echo intensity in the sparse depth map and generating a reflection intensity map; normalizing the timestamp and encoding it into a time feature channel to obtain a time feature map; encoding the spatial path position of each point to obtain a scanning path position map; and fusing the sparse depth map, reflection intensity map, time feature map and scanning path position map.

[0009] Optionally, the three-dimensional geometric reconstruction model includes: a multi-scale convolution feature extraction module, a voxel adaptive completion module, a dual-domain spatial attention fusion module, an edge-guided decoding module, and an uncertainty estimation and dual output head, wherein the multi-scale convolution feature extraction module is used to extract spatial semantic features with different receptive field ranges from the multi-channel input of the laser scanning data; the voxel adaptive completion module is used to receive spatial semantic features and dynamically adjust the convolution kernel parameters based on the voxel point density and local geometric prior to perform refined geometric completion on sparse or structurally broken areas; the dual-domain spatial attention fusion module is used to apply attention weights for joint modeling of spatial and frequency domains to the completed voxel features, and introduce a structural guidance mechanism to enhance the model's modeling ability for jump edges, high-frequency shapes, and long-distance dependencies; the edge-guided decoding module is used to fuse the edge explicit response map and semantic features during the feature sampling and decoding process, guiding the model to focus on structural contours and surface mutation areas; the uncertainty estimation and dual output head are used to simultaneously generate a target depth map and an uncertainty map and form a structured reconstruction result tensor.

[0010] Optionally, the multi-scale convolution feature extraction module includes: a dynamically deformable convolution path, a void frequency conversion path, a frequency domain activation path and a channel attention fusion mechanism arranged in parallel.

[0011] Optionally, the voxel adaptive completion module includes: a point density perception submodule, a local structure prior encoding submodule, an adaptive voxel convolution submodule and a multi-scale context fusion submodule.

[0012] Optionally, the dual-domain spatial attention fusion module includes: a spatial attention path, a frequency domain attention path and a structural prior path.

[0013] Optionally, the three-dimensional geometric reconstruction model is trained by the following steps: obtaining laser scanning raw data of multiple target objects with real three-dimensional shape labels, constructing a data set containing multi-channel input and geometric truth value pairs, and dividing the data set into a training set and a validation set; initializing training parameters, training the model with the training set until the training reaches a maximum number of iterations; inputting the validation set into the trained three-dimensional geometric reconstruction model, outputting the corresponding dense depth map and uncertainty map, and comparing them with the real three-dimensional labels, and evaluating the model based on performance indicators including the root mean square error of the dense depth map, the Chamfer distance of the three-dimensional point cloud reconstruction, and the accuracy of the point cloud reconstruction; when each performance indicator meets the set threshold at the same time, the model verification is passed; otherwise, the training parameters are adjusted until the model verification is passed.

[0014] The present application also provides a laser scanning three-dimensional imaging device, which includes: an acquisition module for acquiring laser scanning data of a target object, wherein the laser scanning data includes, for example, depth data including reflection intensity, scanning angle, timestamp and flight time; a preprocessing module for preprocessing the laser scanning data and fusing time series with scanning path information to obtain multi-channel input features adapted to a neural network; a model construction and training module for constructing a three-dimensional geometric reconstruction model and training the model; an output module for inputting the multi-channel input features into the trained three-dimensional geometric reconstruction model to obtain a dense depth map of the target object in the image space; a conversion module for converting the dense depth map of the target object into a three-dimensional point cloud in world coordinates; and a post-processing module for post-processing the three-dimensional point cloud to obtain a three-dimensional imaging of the target object.

[0015] The present application also provides a storage medium comprising instructions, which, when executed on a computer, enables the computer to execute the method as described in any of the preceding items.

[0016] The present application also provides an electronic device, characterized in that the electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] This application significantly improves the accuracy and robustness of three-dimensional imaging of target objects by combining the multi-channel feature input of laser scanning data and the innovative design of the three-dimensional geometric reconstruction model. By pre-processing the laser scanning data and fusing the time series and scanning path information, a multi-channel input feature adapted to the neural network is constructed, which effectively enhances the model's ability to understand complex scenes. The trained three-dimensional geometric reconstruction model can generate high-quality dense depth maps and uncertainty maps, which are further converted into three-dimensional point clouds in world coordinates, ensuring high accuracy and consistency of the reconstruction results. Finally, through post-processing of the three-dimensional point cloud, the imaging effect is further improved, ensuring the system's accurate reconstruction and imaging application of target objects in dynamic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a laser scanning three-dimensional imaging method provided by an embodiment of the present application;

[0020] Figure 2 is a structural schematic diagram of a three-dimensional geometric reconstruction model provided by another embodiment of the present application;

[0021] Figure 3 is a three-dimensional imaging schematic diagram of a handle assembly provided by another embodiment of the present application;

[0022] Figure 4 is a schematic diagram of a three-dimensional imaging of a handle assembly generated based on an existing method;

[0023] Figure 5 is a schematic structural diagram of a laser scanning three-dimensional imaging device provided by another embodiment of the present application; DETAILED DESCRIPTION

[0024] Specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Although specific embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to enable a more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0025] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present application, but the description is based on the general principles of the specification and is not intended to limit the scope of the present application. The scope of protection of this application shall be as defined by the attached claims.

[0026] To facilitate understanding of the embodiments of the present application, further explanation will be given below using specific embodiments as examples in conjunction with the accompanying drawings, and the various drawings do not constitute a limitation on the embodiments of the present application.

[0027] Figure 1 This is a flow chart of a laser scanning three-dimensional imaging method provided by an embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0028] S100: Acquire laser scanning data of a target object, wherein the laser scanning data includes, for example, depth data including reflection intensity, scanning angle, timestamp, and flight time;

[0029] S200: Preprocessing the laser scanning data and fusing time series and scanning path information to obtain multi-channel input features;

[0030] S300: Build a 3D geometric reconstruction model and train the model;

[0031] S400: Inputting the multi-channel input features into the trained three-dimensional geometric reconstruction model to obtain a dense depth map of the target object in the image space;

[0032] S500: Converting the dense depth map of the target object into a three-dimensional point cloud in world coordinates;

[0033] S600: Post-processing the three-dimensional point cloud to obtain a three-dimensional image of the target object.

[0034] In another exemplary embodiment, in step S200, preprocessing the laser scanning data and fusing the time series and scanning path information to obtain multi-channel input features adapted to the neural network includes the following steps:

[0035] S201: performing point cloud filtering processing on the laser scanning data, and projecting the laser scanning data after the point cloud filtering processing into a sparse depth map;

[0036] In this step, the original laser scanning data obtained often contains noise points and outliers caused by environmental interference, system errors or reflection anomalies. In order to improve the accuracy and robustness of subsequent modeling, the point cloud needs to be filtered to remove the interference information in the point cloud. For example, the statistical outlier removal method (SOR) can be used to calculate the mean distance between each point in the point cloud and other points in its neighborhood, and remove points with excessive mean deviations based on a set threshold. Furthermore, in order to reduce the density of the point cloud and unify the scale, the voxel grid sampling method can be used to divide the space into voxels of a fixed size, retaining a representative point in each voxel.

[0037] Through the above filtering processing, isolated points, redundant points and systematic error points in the point cloud can be effectively filtered out, thereby obtaining a point cloud with uniform spatial distribution and clear structure.

[0038] Furthermore, after point cloud filtering, the unstructured point cloud needs to be spatially mapped to adapt the laser scanning data to the model's tensor input format. For example, the laser scanning data can be projected onto a two-dimensional image plane in scanning order based on the lidar's calibration parameters (including laser beam angle, scanning resolution, and installation posture). This generates a sparse depth map, where each pixel location stores the depth value of its corresponding point. The resolution of the sparse depth map can be set to commonly used resolutions such as 64×1024 or 128×2048, depending on the performance of the scanning device.

[0039] S202: extracting laser echo intensity from the sparse depth map and generating a reflection intensity map;

[0040] In this step, the laser scanner not only records the flight time when detecting the target, but also captures the reflected light energy, that is, the reflectance / intensity (the echo intensity is closely related to the target surface material, color, and roughness). In order to enhance the model's ability to recognize the texture and material of the target object, this step synchronously maps the reflection intensity value of each point and its depth value to the two-dimensional image plane or voxel grid, corresponding one-to-one with the depth information, thereby constructing a "depth + reflection" dual-channel input feature. Specifically, in the sparse depth map, each pixel point can be attached with a reflection intensity value to form a second channel. This feature channel helps the model distinguish between targets of different materials (such as metal and fabric), thereby helping to improve the accuracy of edge detection and shape recovery.

[0041] S203: normalize the timestamp and encode it into a time feature channel to obtain a time feature map;

[0042] This step takes into account the well-defined temporal order of laser scanning, particularly when using rotary or linear motion scanning devices. Each laser point is assigned a unique timestamp. To enable the model to perceive dynamic objects, scene order, or scanning delays, this step normalizes the raw sample timestamps to a uniform mapping to the interval [0, 1], thereby forming a standardized temporal feature. This time encoding can be added as a separate channel to the reflection depth map, helping the model learn the temporal order and scanning history of the points.

[0043] S204: Encoding the spatial path position of each point to obtain a scanning path position map;

[0044] In this step, in a laser scanning system, different laser beams, scan lines, or paths have a fixed spatial emission order and angular structure. For example, in a 32-line or 64-line lidar, each laser line has a fixed elevation angle; in a MEMS or rotating mirror system, the scanning path has a fixed pattern. To preserve this spatial structure prior, this step encodes the beam number, line number, scanning angle, or emission direction of each point as "scanning path position information" into each point. This encoded information can form a two-dimensional or three-dimensional position channel for use with a depth map. For example, the beam number can be mapped to the vertical coordinate of the image and the scanning order can be used as the channel input; or the direction vector can be projected into a tensor to provide geometric constraints. The model can use these prior relationships to improve its understanding of spatial context and its ability to model structural associations between points.

[0045] S205: Fusing the sparse depth map, the reflection intensity map, the time feature map, and the scanning path position map.

[0046] In this step, the multiple channels obtained from the aforementioned processing (i.e., the sparse depth map, reflection intensity map, temporal feature map, and scan path position map) are fused and concatenated along the channel dimension to form a unified tensor representation. Furthermore, to ensure data scale consistency and model training stability, all channels are normalized (e.g., Z-score normalization or min-max normalization), and sparse regions are filled (e.g., by filling empty values ​​with zeros or interpolation).

[0047] The multi-channel input features output after normalization processing in this step have multiple information features such as structure, temporal and spatial semantics, which can provide a complete and semantically rich input data foundation for subsequent three-dimensional reconstruction.

[0048] In another exemplary embodiment, in step S300, Figure 2As shown, the three-dimensional geometric reconstruction model includes: a multi-scale convolutional feature extraction module, a voxel adaptive completion module, a dual-domain spatial attention fusion module, an edge-guided decoding module and an uncertainty estimation and dual output head, wherein the multi-scale convolutional feature extraction module is used to extract semantic features with different receptive fields and frequency responses from multi-channel input features; the voxel adaptive completion module is used to dynamically adjust the convolution method according to the voxel point density and local geometric prior to complete the structure of sparse and broken areas; the dual-domain spatial attention fusion module is used to guide the model to pay attention to edges, high-frequency areas and long-distance structural dependencies by introducing structural priors; the edge-guided decoding module is used to guide the decoding process through the edge response map to enhance the model's reconstruction ability on boundaries, curvature mutations and structural details; the uncertainty estimation and dual output head are used to generate a target depth map and estimate the uncertainty score of each predicted point.

[0049] Specifically, the multi-scale convolution feature extraction module includes a dynamic deformable convolution path, a void frequency conversion path and a frequency domain activation path arranged in parallel, wherein the dynamic deformable convolution path adopts a deformable convolution, which introduces a learnable spatial offset so that the sampling position of each convolution kernel can be adaptively adjusted according to the input features, breaking through the limitations of traditional convolution fixed sampling. As far as the present application is concerned, the dynamic deformable convolution path can dynamically adjust the sampling position and receptive field range of the convolution kernel according to the point density distribution, texture complexity and geometric mutation characteristics of each area in the laser scanning point cloud, thereby realizing adaptive modeling of the spatial structure. In addition, the path guides the convolution operation to focus on the key geometric change direction at the edge, surface mutation and sparse area by predicting the learnable spatial offset, which can not only enhance the model's perception of high-frequency structures and irregular shapes, but also improve the restoration accuracy and completion effect in high-risk areas such as structural fractures and blurred contours, so that the features extracted by the model have stronger geometric expression and contextual adaptability.

[0050] The dilated frequency conversion approach combines one-dimensional dilated convolution with wavelet transforms (Haar or Daubechies) to jointly model point cloud features in both spatial and frequency dimensions. Specifically, this approach first uses a set of shared-parameter one-dimensional dilated convolutions to perform preliminary multi-scale extraction of multi-channel input features. The dilated convolution dilates the original features by inserting gaps, expanding the receptive field without increasing the number of parameters. This allows the model to simultaneously capture local details and mid- to long-range contextual structure, thereby enhancing its ability to capture structure at edges and sparse regions. Furthermore, a fast wavelet transform is introduced to perform multi-resolution frequency analysis on the dilated convolution results. The wavelet transform constructs low-frequency (approximation) and high-frequency (detail) components, performing a hierarchical decomposition of the feature map. This effectively reveals structural variations at different scales. The high-frequency component is particularly sensitive to extracting structures such as object outlines and sharp corners, while the low-frequency component maintains overall contour continuity. This approach, with its dual capabilities of spatial scale enhancement and local frequency sensitivity, helps enhance the model's response accuracy and modeling stability in areas such as edges and surface breaks.

[0051] The frequency-domain activation pathway performs a fast Fourier transform (FFT) on multi-channel input features within a local window, mapping them from the spatial domain to the frequency domain. This allows for the extraction of global structural features from a frequency perspective. Specifically, this pathway first performs an FFT within a local sliding window of the feature tensor to obtain the corresponding spectral tensor. A designed complex-modulated convolution kernel is then used to weight high-response frequency regions in the frequency domain. This complex-domain convolution allows the model to explicitly distinguish and weight high-frequency structures (such as edges and texture changes) from low-frequency background (such as flat surfaces or background regions), enhancing its response to regions of geometric transitions. Subsequently, this pathway performs an inverse Fourier transform (IFFT) to reproject the processed spectra into the spatial domain, generating a set of frequency-enhanced feature maps. This process preserves the original structural information while providing the model with consistent geometric awareness across both the frequency and spatial domains, thereby improving feature representation and edge detail restoration.

[0052] The multi-scale convolutional feature extraction module also includes a channel attention fusion mechanism, which first performs global average pooling on the output of each path to obtain the channel vector of each path, and then performs nonlinear modeling on each channel vector through a shared two-layer multi-layer perceptron (MLP) network to obtain the weight vector of each path; finally, based on the weight vector, the feature map of each path is weighted channel by channel and summed and fused to obtain an enhanced multi-scale semantic feature map.

[0053] In summary, the multi-scale convolutional feature extraction module is divided into three sub-paths: a dynamic deformable path, a void frequency conversion path, and a frequency domain activation path. These sub-paths focus on spatial adaptive modeling, multi-scale geometric structure analysis, and frequency-sensitive feature capture in point clouds, respectively. This allows for comprehensive perception of multi-level, multi-scale, and multi-domain structures in sparse laser data. The dynamic deformable path improves the model's adaptability to structural mutation areas, the void frequency conversion path enhances the multi-scale expression of high-frequency details such as edges and corners, and the frequency domain activation path compensates for the blind spots in traditional spatial modeling's recognition of high-frequency changes and global periodic structures. Furthermore, the outputs of the three paths are dynamically weighted fused under the guidance of global context perception via a channel attention fusion mechanism. This not only suppresses redundant feature interference, but also enhances the response strength of key channels, ultimately forming a set of deep semantic features that combine spatial position sensitivity, frequency response capability, and structural expression, helping to improve the modeling accuracy and robustness of this application in complex structural areas and sparse data scenarios.

[0054] The voxel adaptive completion module includes a point density perception submodule, a local structure prior encoding submodule, an adaptive voxel convolution submodule and a multi-scale context fusion submodule. Among them, the point density perception submodule is used to count the density of laser points in each voxel and normalize it into a density map. This submodule includes a point counter and a density normalization unit. The point counter maps the input three-dimensional laser point cloud to a discrete voxel grid, counts the points falling into each voxel grid to evaluate its spatial density. Exemplarily, the point counter first converts the coordinates of each point into the corresponding voxel index according to the set voxel size, and accumulates the number of points of each voxel in a three-dimensional tensor that is initially zero, thereby constructing a point distribution map in the entire space. The density normalization unit is used to linearly normalize the point count distribution map (for example, using a Min-Max approach) and output a normalized density tensor M, which reflects the density of the laser point distribution within each voxel. Low density values ​​indicate missing areas, requiring a larger complement range, while high density values ​​indicate regions with clear structure, where a small receptive field can be used to avoid blurring. Ultimately, the density tensor M serves as a conditional input for the subsequent parameter control process of the adaptive voxel convolution, dynamically adjusting the convolution kernel size, dilation rate, and attention mechanism weights to achieve differentiated perception of different structural regions and structure-sensitive modeling.

[0055] The local structure prior encoding submodule extracts voxel-level geometric structural features from the input 3D laser point cloud, including structural descriptors such as curvature and normal vectors. This information guides subsequent convolution operations to specifically complete complex structures such as edges, abrupt changes, and fractures. This submodule first constructs a fixed-size local neighborhood window (e.g., a 3×3×3 voxel grid) centered on each voxel and extracts the coordinate distribution of all valid points within the window. If the original point coordinates are available, this submodule uses principal component analysis (PCA) to calculate the covariance of the neighborhood point set, extracting the eigenvector of the first principal direction, curvature estimates, and normal vectors as metrics for regional structural stability and boundary changes. If the input is a dense voxel tensor, the local gradient response can also be estimated based on the voxel intensity values ​​using a 3D Sobel operator, voxel gradient filtering, or a 3D Laplace operator. The extracted structural descriptors are mapped to fixed-length vectors and channel-encoded using 1×1×1 convolutions, ultimately generating a structural feature map that corresponds one-to-one to the original voxel locations. This structural feature map can characterize the geometric complexity of local voxels, such as surface flatness, edge jumps, normal discontinuities, etc., and serves as an important basis for dynamically adjusting the convolution kernel parameters and sampling direction in subsequent adaptive voxel convolution, thereby achieving differentiated processing and high-fidelity completion of spatial heterogeneous structures.

[0056] The adaptive voxel convolution submodule dynamically adjusts the sampling method and parameters of its convolution kernel based on the point cloud density and local structural complexity of each voxel, enabling on-demand feature completion and geometric reconstruction. This unit first concatenates the density map output by the point density perception submodule with the structural feature map output by the local structure prior submodule to form conditional input features. Using a lightweight multi-layer perceptron (MLP) or conditional convolution generator, this conditional information is mapped into a set of dynamic convolution kernel parameters or sampling offsets for a specific voxel location. The convolution kernel then operates on the local neighborhood centered at that voxel in the voxel tensor, aggregating and updating local 3D features. In low-density areas, the convolution kernel enlarges the receptive field and expands the sampling area. At edge structures, the convolution focuses on the direction of geometric changes to enhance boundary detail recovery. This unit is essentially an extension of the conditional convolution mechanism in 3D voxel space. It is structure-aware and position-adaptive, effectively improving the completion accuracy and boundary coherence of incomplete structures in sparse point clouds.

[0057] The multi-scale context fusion submodule integrates contextual information from different receptive fields within the voxel-completed feature space to enhance the synergy between local detail modeling and global structural consistency. This submodule utilizes a parallel multi-branch 3D dilated convolution architecture. For example, it comprises four branches, each performing a dilated convolution operation on the same voxel feature map using a different dilation rate (e.g., 1, 2, 3, or 5). This architecture forms feature receptive fields at multiple scales (branch 1 has a dilation rate of 1, resulting in a small receptive field suitable for detail edges and low-curvature regions; branch 2 has a dilation rate of 2, resulting in a medium receptive field capable of capturing local geometric changes; branch 3 has a dilation rate of 3, resulting in a large receptive field capable of identifying medium- and long-range dependencies; and branch 4 has a dilation rate of 4, resulting in a cross-region receptive field capable of combating occlusions and structural discontinuities). This allows the capture of multi-level spatial information, from microscopic geometric details to macroscopic structural trends, without increasing the number of parameters. Furthermore, this submodule introduces a residual connection mechanism that directly skips input features to the final fused output, forming an input-to-fusion architecture that helps prevent feature degradation caused by multi-scale convolution. Subsequently, the feature maps output by each branch are spliced ​​through the channel dimension to form a multi-scale joint feature representation before fusion, which is sent to the fusion layer with an attention mechanism (such as SE-block or channel weighting module). The fusion layer calculates the importance weight of each channel by modeling the response speed of different channels in the current task, and adaptively weights the multi-scale features accordingly, and finally outputs a voxel feature map that is optimized both spatially and semantically.

[0058] In summary, the above fusion strategy significantly improves the model's performance when handling non-uniform point cloud structures, complex surfaces, and occluded areas, effectively overcoming the shortcomings of single-scale convolution in its ability to perceive diverse geometric structures. In particular, in areas with drastic structural changes, such as building boundaries, human joints, and pipe joints, this submodule significantly improves the geometric accuracy, structural coherence, and edge restoration of the reconstruction results.

[0059] The dual-domain spatial attention fusion module comprises a spatial attention pathway, a frequency-domain attention pathway, and a structural prior pathway. The spatial attention pathway extracts position-sensitive salient regions from input features and guides the model's attention to key geometric regions in the laser scan data by combining local fine-grained modeling with global non-local dependency modeling. This pathway consists of two parallel sub-pathways: a local directional attention pathway, which employs multi-scale separable convolutions (e.g., 1×3, 3×1, and 3×3) to construct a directionally selective receptive field, enhancing the model's response to local structural abrupt changes such as edges and corners. A global non-local attention pathway constructs a lightweight non-local attention mechanism (e.g., a simplified Transformer Block) that leverages a query-key-value architecture to capture correlations between distant voxels and enhance contextual coherence between occluded and sparse regions. The outputs of the two sub-pathways are concatenated channel by channel and fed into a fusion layer, where they undergo sigmoid activation to generate a spatial attention map, which serves as the basis for subsequent feature weighting. This path effectively combines local detail sensitivity with global structural consistency, which can improve the modeling ability of key shape areas in sparse laser point clouds. It is particularly suitable for the identification and recovery of broken edges, surface changes and occluded areas.

[0060] The proposed frequency-domain attention pathway addresses the limited responsiveness of traditional spatial attention mechanisms to fine-grained structures (such as edge transitions and high-frequency noise) by incorporating the frequency distribution of laser scan features. This pathway first performs a Fast Fourier Transform (FFT) on a local sliding window of the feature tensor, mapping local voxel features to a frequency-domain representation. The transformed spectral tensor is decomposed into high-frequency and low-frequency components. The high-frequency components represent regions with edges, cracks, or surface abrupt changes in space, while the low-frequency components represent overall structural contours or background regions. Subsequently, two channels are designed: a high-frequency enhancement branch that emphasizes significant edge responses through frequency masking weighting; and a low-frequency modulation branch that suppresses background noise and enhances stability through smoothing filtering. The two branches are then fused and processed through an inverse IFFT or spectral attention mapping to generate an attention map consistent with the input space. This attention map guides the model's improved responsiveness to regions of structural variation in a frequency-sensitive manner. This submodule shifts the attention mechanism's reliance on spatial correlation and incorporates a frequency-domain sparsity prior, making it particularly suitable for handling the overlapping of local discontinuities and measurement noise in sparse point clouds.

[0061] The structure prior pathway incorporates geometric information explicitly extracted from the input laser point cloud (such as curvature, normal vectors, and edge strength) into the attention computation process, enhancing attention's understanding and guidance of spatial geometric structure. This pathway, in pre-processing modules (e.g., the local structure encoder), estimates the principal direction, normal distribution, and curvature within the local neighborhood of each voxel, forming a structure description tensor G. In the structure prior pathway, this structure tensor is first reduced in dimension by a 1×1 convolution. An activation function (e.g., ReLU+Sigmoid) is then used to generate a structure guidance mask, ranging from 0 to 1, representing the confidence level at each location as geometrically salient structure. This mask is then element-wise multiplied with the preliminary attention map generated by the spatial or frequency-domain attention pathway to explicitly enhance transition edges and suppress flat regions. The introduction of the structure prior pathway imbues the model with greater geometric selectivity in selecting regions of attention, significantly improving the accuracy of edge recovery and the consistency of depth completion.

[0062] The edge-guided decoding module is used during the decoding phase of 3D geometric reconstruction to focus on recovering edge blur, structural fractures, and depth jump distortion in laser point clouds caused by sparse sampling, occlusion, or reflection loss. This module utilizes a parallel setup consisting of a main decoding path, an auxiliary edge-guided path, and a feature fusion mechanism. The main path utilizes a top-down decoder structure consisting of multiple levels of 3D deconvolution or upsampling layers (e.g., 3 to 5 levels) to gradually restore low-resolution deep semantic feature maps to a high-resolution space. Shallow features from the encoder can be fused at each level of recovery to preserve spatial positioning information.

[0063] The auxiliary edge guidance pathway specifically models and guides edge regions, independently extracting and enhancing their geometric features. Specifically, the auxiliary edge guidance pathway first applies a 3D Sobel filter or spatial gradient operator to the feature tensor from the encoder to extract an explicit edge response map. This map captures geometric variations such as depth jumps, areas of high curvature, or surface abrupt changes. This edge response map is then enhanced through a series of lightweight convolutional layers and activation functions (such as ReLU) to extract more discernible local boundary semantic information. The extracted edge features further guide the main pathway in two ways: first, as an attention mask, they are directly multiplied with the main features to enhance edge responses; second, as channel-by-channel fusion features, they are concatenated after each decoding stage and fed into the next decoding stage, forming an edge-preserving information flow. At the final output stage of the decoding process, this module also uses the edge response map to perform fine-grained edge restoration on the predicted depth map or point cloud, further correcting blurred edges and connecting broken surfaces to ensure that the model retains important geometric jump information when reconstructing the 3D structure. Overall, this module can improve the performance of laser point cloud reconstruction in terms of edge clarity, structural continuity and detail fidelity through explicit edge modeling and dual-path fusion strategy.

[0064] The uncertainty estimation and dual output head is designed to enhance the 3D geometric reconstruction system's ability to model the reliability of prediction results. This allows the model to estimate the reliability of each predicted point while generating a depth map or point cloud, thereby enabling adaptive optimization and fault-tolerant control during post-processing or decision-making. The uncertainty estimation and dual output head employs a dual-branch output structure, consisting of a main output branch (depth prediction head) and an auxiliary output branch (uncertainty estimation head). Both branches share the encoded results of the backbone feature extraction network and incorporate targeted design features within their respective branches. Specifically, the main output branch is responsible for regressing or classifying the depth map (or 3D point position) of the target object. This branch comprises a series of consecutive 3D convolutional or deconvolutional layers (e.g., four 3D convolutional layers), each followed by BatchNormalization and ReLU activation functions. After layer-by-layer feature restoration and spatial upsampling, this branch ultimately outputs a dense depth map with dimensions of H×W×1 (H and W represent the height and width of the predicted image, respectively). In addition, the parallel uncertainty estimation branch uses a lightweight 3D convolutional network with a similar structure. This branch receives the same encoded feature input as the main output branch. After shallow convolution, activation, and normalization, it outputs an uncertainty map, also H×W×1. The value at each position in the map represents the variance, confidence, or expected error of the predicted value at that position, which is used to measure the reliability of the reconstruction result at that point. Depending on the modeling approach, this branch can adopt the Maximum Likelihood Estimation (MLE) framework, treating each predicted point as a sample from a Gaussian distribution and outputting the mean and variance of the depth value.

[0065] Finally, the uncertainty estimation and dual output head jointly outputs the dense depth map and uncertainty map as a dual-channel result, forming a structured reconstruction result tensor (i.e., the two are concatenated in the channel dimension to form an H×W×2 dual-channel prediction tensor). This joint output not only provides a 3D reconstruction value (such as depth or coordinate) for each spatial location, but also provides a confidence measure for that value, enabling the model to make smarter and safer decisions based on the confidence level of the prediction.

[0066] In another exemplary embodiment, in step S300, the 3D geometric reconstruction model is trained by the following steps:

[0067] S301: Acquire laser scanning raw data of multiple target objects with real three-dimensional shape labels, construct a dataset containing multi-channel input and geometric truth value pairs, and divide the dataset into a training set and a validation set (e.g., divided in a 6:4 ratio).

[0068] S302: Initialize training parameters, for example, set the maximum number of iterations to 100, the learning rate to 0.001, and the dropout ratio to 0.3, and train the model using the training set until the training reaches the maximum number of iterations;

[0069] S303: Input the validation set into the trained 3D geometric reconstruction model, output the corresponding dense depth map and uncertainty map, and compare them with the real 3D labels. Evaluate the model based on performance indicators including the root mean square error (RMSE) of the dense depth map, the Chamfer distance of the 3D point cloud reconstruction, and the accuracy of the point cloud reconstruction. When all performance indicators meet the set thresholds at the same time (for example, when RMSE ≤ 0.08m, Chamfer distance ≤ 0.02m, Accuracy ≥ 90%), the model is verified. Otherwise, adjust the training parameters (for example, adjust the maximum number of iterations to 200) until the model is verified.

[0070] In another exemplary embodiment, in step S500, converting the dual-channel output tensor including the dense depth map of the target object and the corresponding uncertainty map into a three-dimensional point cloud in world coordinates includes the following steps:

[0071] S501: constructing an image coordinate network;

[0072] In this step, it is necessary to construct the two-dimensional coordinates of each pixel in the depth image on the image plane, that is, for each depth value Specify its corresponding image space position ( represents the horizontal coordinate of the pixel in the image, Represents the vertical coordinate of the pixel in the image), these image coordinates will be combined with the depth value in the subsequent steps to infer its three-dimensional spatial position in the camera coordinate system.

[0073] Specifically, if the image resolution is H×W, then construct the pixel coordinate set:

[0074]

[0075] Finally, the image coordinate grid is formed , this coordinate grid corresponds to the position of each pixel in the depth image.

[0076] S502: combining the image coordinates of each pixel with its corresponding depth value, and converting them into points in the three-dimensional camera coordinate system through the camera's intrinsic parameter matrix;

[0077] In this step, let the camera intrinsic parameter matrix for:

[0078]

[0079] For each pixel and depth value , its camera coordinates are:

[0080]

[0081] in, Indicates the horizontal focal length, Indicates the vertical focal length, Indicates that in the image coordinate system, the optical axis is x The direction is projected onto the pixel position of the image, Indicates that in the image coordinate system, the optical axis is y The direction is projected onto the pixel position of the image, Indicates the camera coordinate system x coordinate, Indicates the camera coordinate system y coordinate, Indicates the camera coordinate system z coordinate.

[0082] S503: screening three-dimensional points;

[0083] In this step, the uncertainty map is given. , set the threshold T, then the screening condition for valid pixels is:

[0084]

[0085] This application only performs back-projection operations on pixels that meet this condition to form a trusted point set.

[0086] S504: Mapping the filtered 3D points from the camera coordinate system to world coordinates.

[0087] In this step, the point in the camera coordinate system The 3D position is based on the camera as the reference system, but the actual 3D reconstruction needs to be unified to the world coordinate system. Therefore, the camera extrinsic matrix is ​​used to perform coordinate transformation to map the points from the camera coordinate system to the world coordinate system.

[0088] Assume the camera extrinsic transformation matrix is , represents the rigid transformation from the world coordinate system to the camera coordinate system, then its inverse transformation Convert the camera coordinate system to world coordinates.

[0089] Homogeneous form:

[0090]

[0091] in, Represents the point in the world coordinate system x coordinate; Represents the point in the world coordinate system y coordinate; Represents the point in the world coordinate system z Coordinates; 1 represents a homogeneous coordinate constant term; Represents a 4×4 matrix, which is represented by the rotation matrix and translation vectors constitute:

[0092]

[0093] Finally, all points that meet the screening conditions are converted into three-dimensional coordinate points in the world coordinate system. ,These points together constitute structured or unstructured three-dimensional point cloud data.

[0094] In another exemplary embodiment, in step S600, post-processing the three-dimensional point cloud to obtain a three-dimensional image of the target object includes the following steps:

[0095] S601: performing weighted filtering on the three-dimensional point cloud according to the uncertainty map of the prediction result;

[0096] In this step, the reconstructed 3D point cloud is adaptively filtered using the pixel-level uncertainty map predicted by the model as a weighted basis. Specifically, the uncertainty values ​​in the image space are first mapped to the position of each point in the 3D point cloud, forming a point-level confidence distribution. Subsequently, the corresponding filtering weights are set based on the uncertainty value of each point. In areas with low confidence (i.e., points with high uncertainty), the receptive field of the filter kernel is expanded and the denoising strength is enhanced to suppress reconstruction errors or artifacts. In areas with high confidence, the filtering operation is appropriately weakened to preserve the original structural details and edge geometry. The filtering method can use Gaussian filtering or mean filtering based on the joint weighting of Euclidean distance and confidence to ensure a balance between global consistency and local accuracy.

[0097] Specifically, for point , its neighborhood mean filtering is expressed as:

[0098] ,

[0099]

[0100] in, Represents the point after weighted average new location; Indicates a point and point The weight between Represents point The set of related neighborhood points; and Represent points and point uncertainty value; represents the uncertainty threshold; Indicates a point and point The square of the Euclidean distance between them; Represents the influence degree between control points based on the distance Gaussian function; Represents a Gaussian attenuation function based on uncertainty, which is used to adjust the impact of points with large uncertainty on the results.

[0101] S602: In the filtered point cloud, calculate the covariance matrix of the neighborhood of each point to extract local geometric structure information;

[0102] In this step, in order to further explore the local geometric structure features in the point cloud, a fixed-scale neighborhood point set is constructed for each point, and the covariance matrix of the neighborhood point set is calculated. The specific calculation method is: for each point , determine its nearest neighbor in three-dimensional space , then perform mean centering on these neighborhood points and calculate the covariance matrix of their spatial distribution:

[0103]

[0104] in, represents the covariance matrix of point i; Represents the central coordinates of all points in the neighborhood of point i; Indicates the The offset vector of each neighborhood point relative to the neighborhood center; Represents the transpose of the offset vector.

[0105] S603: Using the spatial, intensity, temporal and uncertainty characteristics of the points, density adaptive clustering is performed in the voxel space to achieve structurally consistent semantic segmentation;

[0106] In this step, after completing the local geometric structure extraction, this step uses the fused multi-dimensional point cloud features to perform density-adaptive spatial clustering in the voxel grid, constructing a feature vector for each point to form a joint feature:

[0107]

[0108] in, Indicates a point The three-dimensional space coordinates of Indicates a point Laser reflection intensity; Indicates a point timestamp; Indicates a point uncertainty score.

[0109] S604: Apply multi-scale normal interpolation to cluster boundary areas to enhance the edge sharpness and geometric continuity of the point cloud;

[0110] In this step, based on the clustering results, multi-scale normal interpolation and geometric structure sharpening operations are performed on the edge areas. First, the boundary area of ​​each cluster block is identified and its normal information is extracted. Then, the interpolation position and direction of the boundary points are estimated based on multiple scales (such as local 3×3×3 voxel window and global feature response), using a bidirectional interpolation strategy:

[0111]

[0112] in, Indicates the Neighborhood point-to-point The weighting coefficient of Indicates the The neighboring point in s The estimated position in the iteration; Indicates a point The number of neighbors; Represents the normalization coefficient with its own weight 1 added.

[0113] S605: Based on the shape prior structure reprojection, the preliminary reconstructed point cloud is aligned to the 3D shape template of the known object category through non-rigid registration, achieving structured reconstruction and semantic integrity optimization of sparse data.

[0114] In this step, to further improve the consistency of the reconstruction results at the geometric and semantic levels, this application uses shape templates for non-rigid registration, aligning the processed point cloud with a predefined 3D model. First, a shape prior matching the current target category (such as a CAD model or statistical model) is selected and an initial correspondence between the point cloud and the template is established. Then, based on the Euclidean distance between points, the local normal direction, and the similarity of the structural boundaries, a non-rigid registration optimization objective is constructed:

[0115]

[0116] in, Represents the first point in the template point cloud points; represents the deformation field; represents the rotation matrix; represents the translation vector; Express The non-rigid deformation vector of represents the regularization coefficient, which controls the penalty weight of non-rigid deformation; Represents the energy of non-rigid deformation.

[0117] Below, the present application takes the handle assembly of the power switch cabinet as an example to exemplify the technical solution described in the present application.

[0118] First, the handle assembly is scanned within a ±30° range from the front. The laser scanning data of the handle assembly obtained (one frame of data is selected) is shown in Table 1:

[0119] Table 1

[0120]

[0121] After processing based on the method described in this application, the final three-dimensional point cloud information can be obtained, as shown in Table 2:

[0122] Table 2

[0123]

[0124] Based on the above data, we can obtain Figure 3 In addition, the present application also provides a three-dimensional image of the handle assembly generated based on the existing method (such as Figure 4 By comparison Figure 3 and Figure 4 It can be clearly observed that the three-dimensional imaging generated by the existing method shows the characteristics of dense point cloud but blurred edges. Although the overall structural outline is discernible, there are slight point cloud breaks or deformations at the bends and corners of the handle, and there is a certain amount of noise in the background point cloud, which affects the clarity of the target outline. The three-dimensional imaging generated by this method shows significantly improved geometric accuracy and structural integrity: the handle, base and contact contours with the door panel of the handle are highly continuous, with clear boundaries and no cracks, and the three-dimensional shape is complete and full of details. In addition, in the method described in this application, the background noise has been basically eliminated, and the reconstruction result focuses on the target structure itself, showing higher density consistency and semantic integrity, reflecting the effective enhancement of the model in key steps such as uncertainty-aware filtering, edge-guided decoding and shape prior registration. Overall, the method described in this application is significantly superior to existing methods in terms of the accuracy of three-dimensional structure restoration, edge preservation ability and target separation effect.

[0125] In another exemplary embodiment, the present application also provides a laser scanning three-dimensional imaging device, such as Figure 5As shown, the device includes: an acquisition module 100, which is used to acquire laser scanning data of the target object, and the laser scanning data includes, for example, depth data including reflection intensity, scanning angle, timestamp and flight time; a preprocessing module 200, which is used to preprocess the laser scanning data and fuse the time series and scanning path information to obtain multi-channel input features adapted to the neural network; a model construction and training module 300, which is used to construct a three-dimensional geometric reconstruction model and train the model; an output module 400, which is used to input the multi-channel input features into the trained three-dimensional geometric reconstruction model to obtain a dense depth map of the target object in the image space; a conversion module 500, which is used to convert the dense depth map of the target object into a three-dimensional point cloud in world coordinates; and a post-processing module 600, which is used to post-process the three-dimensional point cloud to obtain a three-dimensional imaging of the target object.

[0126] In another exemplary embodiment, the present application further provides a storage medium comprising instructions, which, when executed on a computer, enables the computer to execute any of the methods described above.

[0127] In another exemplary embodiment, the present application also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described in any of the preceding items when executing the program.

[0128] The above description is merely an optional example of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A laser scanning three-dimensional imaging method, characterized in that: The method comprises: Acquiring laser scanning data of the target object, the laser scanning data including depth data including reflection intensity, scanning angle, timestamp, and flight time; The laser scanning data is preprocessed and the time series and scanning path information are fused to obtain multi-channel input features adapted to the neural network, including: Filter the laser scanning data into a point cloud and project the filtered point cloud into a sparse depth map; Extracting laser echo intensity from the sparse depth map and generating a reflection intensity map; Normalize the timestamps and encode them into temporal feature channels to obtain temporal feature maps; Encode the spatial path position of each point to obtain a scanning path position map; Fuse the sparse depth map, reflection intensity map, time feature map and scanning path position map; Build a 3D geometric reconstruction model and train the model; The three-dimensional geometric reconstruction model includes: Multi-scale convolutional feature extraction module, voxel adaptive completion module, dual-domain spatial attention fusion module, edge-guided decoding module and uncertainty estimation and dual output head, among which, The multi-scale convolution feature extraction module is used to extract spatial semantic features with different receptive field ranges from the multi-channel input of laser scanning data; The voxel adaptive completion module is used to receive spatial semantic features and dynamically adjust convolution kernel parameters based on voxel point density and local geometric priors to perform refined geometric completion on sparse or structurally broken areas; The dual-domain spatial attention fusion module is used to apply attention weights for joint modeling in the spatial and frequency domains to the completed voxel features, and introduces a structure-guided mechanism to enhance the model's ability to model transition edges, high-frequency shapes, and long-range dependencies. The edge-guided decoding module is used to fuse the edge explicit response map and semantic features during the feature sampling and decoding process, guiding the model to focus on the structural contours and surface mutation areas; The uncertainty estimation and dual output heads are used to simultaneously generate a target depth map and an uncertainty map and form a structured reconstruction result tensor; Inputting the multi-channel input features into a trained 3D geometric reconstruction model to obtain a dual-channel output tensor including a dense depth map of the target object and a corresponding uncertainty map; Converting the dual-channel output tensor containing the dense depth map of the target object and the corresponding uncertainty map into a three-dimensional point cloud in world coordinates; The three-dimensional point cloud is post-processed to obtain a three-dimensional image of the target object.

2. The method according to claim 1, characterized in that The multi-scale convolution feature extraction module includes: The dynamically deformable convolution path, the void frequency conversion path, the frequency domain activation path and the channel attention fusion mechanism are set in parallel.

3. The method according to claim 1, characterized in that The voxel adaptive completion module includes: Point density perception submodule, local structure prior encoding submodule, adaptive voxel convolution submodule and multi-scale context fusion submodule.

4. The method according to claim 1, wherein The dual-domain spatial attention fusion module includes: Spatial attention pathway, frequency domain attention pathway and structural prior pathway.

5. The method according to claim 1, wherein The three-dimensional geometric reconstruction model is trained by the following steps: Obtain multiple laser scanning raw data of target objects with real 3D shape labels, construct a dataset containing multi-channel input and geometric truth value pairs, and divide the dataset into training set and validation set; Initialize the training parameters and train the model using the training set until the training reaches the maximum number of iterations; The validation set is input into the trained 3D geometric reconstruction model, and the corresponding dense depth map and uncertainty map are output. They are compared with the real 3D labels, and the model is evaluated based on performance indicators including the root mean square error of the dense depth map, the Chamfer distance of the 3D point cloud reconstruction, and the accuracy of the point cloud reconstruction. When all performance indicators meet the set thresholds at the same time, the model verification passes; otherwise, the training parameters are adjusted until the model verification passes.

6. A laser scanning three-dimensional imaging device, characterized in that: The device comprises: An acquisition module is used to acquire laser scanning data of a target object, wherein the laser scanning data includes depth data including reflection intensity, scanning angle, timestamp, and flight time; The preprocessing module is used to preprocess the laser scanning data and fuse the time series and scanning path information to obtain multi-channel input features suitable for the neural network, including: Filter the laser scanning data into a point cloud and project the filtered point cloud into a sparse depth map; Extracting laser echo intensity from the sparse depth map and generating a reflection intensity map; Normalize the timestamps and encode them into temporal feature channels to obtain temporal feature maps; Encode the spatial path position of each point to obtain a scanning path position map; Fuse the sparse depth map, reflection intensity map, time feature map and scanning path position map; Model building and training module, used to build a 3D geometric reconstruction model and train the model; The three-dimensional geometric reconstruction model includes: Multi-scale convolutional feature extraction module, voxel adaptive completion module, dual-domain spatial attention fusion module, edge-guided decoding module and uncertainty estimation and dual output head, among which, The multi-scale convolution feature extraction module is used to extract spatial semantic features with different receptive field ranges from the multi-channel input of laser scanning data; The voxel adaptive completion module is used to receive spatial semantic features and dynamically adjust convolution kernel parameters based on voxel point density and local geometric priors to perform refined geometric completion on sparse or structurally broken areas; The dual-domain spatial attention fusion module is used to apply attention weights for joint modeling in the spatial and frequency domains to the completed voxel features, and introduces a structure-guided mechanism to enhance the model's ability to model transition edges, high-frequency shapes, and long-range dependencies. The edge-guided decoding module is used to fuse the edge explicit response map and semantic features during the feature sampling and decoding process, guiding the model to focus on the structural contours and surface mutation areas; The uncertainty estimation and dual output heads are used to simultaneously generate a target depth map and an uncertainty map and form a structured reconstruction result tensor; An output module, configured to input the multi-channel input features into a trained three-dimensional geometric reconstruction model to obtain a dense depth map of the target object in the image space; A conversion module, configured to convert the dense depth map of the target object into a three-dimensional point cloud in world coordinates; The post-processing module is used to perform post-processing on the three-dimensional point cloud to obtain a three-dimensional image of the target object.

7. A storage medium, characterized in that: The method comprises instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 5.

8. An electronic device, characterized in that: The electronic device comprises: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images