Earthwork metering and acceptance method based on BIM and unmanned aerial vehicle point cloud

Abstract

The invention discloses an earthwork metering and acceptance method based on BIM and unmanned aerial vehicle point cloud, and relates to the technical field of engineering construction monitoring. The method comprises the following steps: creating a 4D BIM, planning the flight of an unmanned aerial vehicle to collect a high-resolution image, recognizing and removing a non-terrain target on a construction site by adopting a pre-training segmentation model, and recovering the terrain and reconstructing an interference-free 3D point cloud through an image restoration model. And the terrain point cloud and the BIM are fused to calculate the earthwork digging and filling volume, a metering result is obtained, and project acceptance is completed. According to the method, point cloud distortion and volume errors caused by interference of construction machinery, temporary facilities and the like are effectively eliminated, the earthwork volume calculation accuracy and stability are improved, the measurement efficiency is improved, high-frequency fine management is supported, and a reliable data basis is provided for engineering settlement and acceptance.

Description

Technical Field

[0001] This invention relates to the field of engineering construction monitoring technology, specifically to a method for earthwork measurement and acceptance based on BIM and UAV point cloud. Background Technology

[0002] With the continuous expansion of infrastructure construction, earthwork engineering accounts for a significant proportion of road, building, and water conservancy projects. Accurate measurement of earthwork volume is not only directly related to project cost control and construction schedule management, but also has a crucial impact on project settlement, quality acceptance, and risk management.

[0003] Existing methods for measuring earthwork volume mainly include manual measurement, cross-sectional calculation, and digital measurement methods based on measuring equipment such as total stations and GPS. These methods typically rely on manual on-site operations, resulting in low measurement efficiency. Furthermore, in complex terrain or large-scale construction scenarios, they are easily affected by human error and environmental factors, making it difficult to meet the needs for high-frequency and precise earthwork measurement.

[0004] With the development of UAV photogrammetry technology, some projects have begun to use UAVs to acquire construction site image data and reconstruct 3D models based on the image data for earthwork volume estimation. Chinese patent CN119648902A discloses an earthwork measurement method based on BIM and UAV photogrammetry. This method can convert a 3D model constructed based on site image data into a quantity calculation model and determine the earthwork volume for the current construction stage based on the difference between the quantity calculation model and the reference model. This type of technology reduces the cost of manual measurement and improves measurement efficiency to some extent. However, in actual construction, earthwork construction sites often contain a large number of non-terrain targets, such as construction machinery, transport vehicles, temporary facilities, and construction personnel. These non-terrain targets are inevitably acquired simultaneously during UAV image acquisition, introducing noise points or abnormal structures, leading to deviations between the reconstructed point cloud model and the actual terrain. When non-terrain targets are densely distributed or frequently change in the construction site, it can easily cause local point cloud distortion, terrain surface discontinuities, or volume anomalies, thus adversely affecting the calculation results of earthwork excavation and filling volumes, reducing the accuracy and stability of earthwork volume calculation. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of low accuracy in earthwork volume calculation caused by non-terrain targets mentioned in the background art, and to propose an earthwork engineering measurement and acceptance method based on BIM and UAV point cloud.

[0006] This invention provides a method for earthwork measurement and acceptance based on BIM and UAV point cloud, the method comprising: Create a 4D BIM model based on the design drawings, including terrain surfaces, cut and fill work areas, and progress nodes; Based on the excavation and filling work area, plan the flight path of the drone and collect high-resolution images of the construction site; A pre-trained segmentation and removal model is used to identify and segment non-terrain targets in high-resolution images to obtain a binary mask. The original image is then masked based on the binary mask to obtain a masked image. A pre-trained image inpainting model is used to inpaint the occluded image, resulting in an interference-free terrain image. Reconstructing the 3D point cloud of the construction site based on non-intrusive terrain imagery yields the current terrain point cloud. By integrating the existing terrain point cloud with BIM, the volume of earthwork excavation and filling is calculated to obtain the earthwork measurement results. Based on the earthwork measurement results and BIM design requirements, the project was accepted.

[0007] By adopting the above technical solutions, combining BIM and UAV point cloud technology, and using pre-trained segmentation models to identify and remove non-terrain targets at the construction site, and then restoring the real terrain through image restoration, point cloud distortion and volume errors caused by interference from construction machinery, temporary facilities, etc. can be effectively eliminated, achieving interference-free 3D point cloud reconstruction, improving the accuracy and stability of earthwork volume calculation, while also improving measurement efficiency, supporting high-frequency and refined management, and providing reliable data for project settlement and acceptance.

[0008] Optionally, the segmentation removal model is a segmentation model based on the U-Net architecture.

[0009] U-Net is a classic and efficient architecture in the field of semantic segmentation, capable of accurately segmenting non-terrain targets in images.

[0010] Optionally, the step of masking the original image using a binary mask to obtain a masked image includes: The binary mask is dilated by expanding the mask area outward by a preset number of pixels to obtain an expanded mask image. The original image is masked using the dilated mask image to obtain a masked image.

[0011] By employing the above technical solution and performing appropriate morphological dilation on the segmented mask, a wider-coverage buffer can be created. This effectively solves the pixel-level uncertainty problem that may exist in the segmentation model at the target boundary, ensuring that all pixels related to non-terrain targets are fully included in the occlusion range. This fundamentally avoids the introduction of interfering textures or structures due to incomplete occlusion during subsequent restoration, thus improving the purity and reliability of the restoration results.

[0012] Optionally, the step of masking the original image with the dilated mask image to obtain a masked image includes: Perform connected component analysis on the dilated mask map to obtain the minimum bounding rectangle of each connected mask region; Expand outwards from the midpoint of the smallest bounding rectangle to obtain the coordinates of the expanded rectangle; Based on the expanded rectangular coordinates, the corresponding image block is cropped from the original image as a sub-image; at the same time, the corresponding mask block is cropped from the same position in the dilated mask image as a sub-mask. The sub-image is masked using the sub-mask to obtain a masked image suitable for input to an image restoration model.

[0013] By adopting the above technical solution, cropping and repairing are performed on local images containing occluded areas, so that the image repair model focuses on the actual area that needs to be repaired, while retaining sufficient terrain context information and significantly reducing computational load and GPU memory consumption.

[0014] Optionally, the image inpainting model includes a structure extraction module, a coarse structure inpainting network, and an image completion network; wherein: The structure extraction module is used to perform grayscale processing and edge processing on the masked image to obtain a fragmented grayscale image and a fragmented edge image. The coarse structural repair network is used to coarsely repair the incomplete grayscale image and the incomplete edge image to obtain a grayscale prediction image and an edge prediction image; the input of the coarse structural repair network is composed of the incomplete grayscale image, the incomplete edge image and the binary mask spliced ​​along the channel; The image completion network is used to complete the occluded parts of the occluded image based on the grayscale prediction map and edge prediction map provided by the coarse structural repair network, so as to obtain an interference-free terrain image. The input of the image completion network is composed of the occluded image, the binary mask, the grayscale prediction map and the edge prediction map stitched together along the channel.

[0015] By employing the aforementioned technical solution, a two-stage restoration process guided by structural information was constructed. First, the underlying terrain structure information (grayscale and edges) of the occluded areas was explicitly extracted and restored, providing strong semantic priors and geometric constraints for the final color and texture restoration. This structure-first, texture-later restoration strategy ensures that the generated terrain image maintains a high degree of consistency with the surrounding environment in terms of geometric continuity, elevation trends, and terrain feature lines (such as ridges and gullies). This significantly reduces the risk of distortion in the three-dimensional geometry of the restoration results, thus providing a reliable guarantee for subsequent high-precision point cloud reconstruction.

[0016] Optionally, the coarse-grained repair network includes an encoder (Encoder1), a bottleneck layer (BR1), and two decoders (Decoder1 and Decoder2); wherein: The encoder Encoder1 includes four coding layers, which are used to extract and downsample the input feature map sequentially to obtain the encoded feature map of each layer; each coding layer includes two gated convolutional modules and a downsampling module connected in sequence. The bottleneck layer BR1 includes four cascaded receptive field enhancement modules, which receive the output of the last layer of encoder 1 as input, aggregate contextual information of different ranges, and output the first bottleneck feature map. The decoder Decoder1 comprises four decoding layers and one output convolutional layer, used to sequentially upsample the first bottleneck feature map and fuse the feature maps output by the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into a single-channel layer and obtains the grayscale prediction map I through the tanh activation function. gp Each decoding layer includes an upsampling module, a splicing module, and two gated convolutional modules. Decoder2 has the same structure as Decoder1. The final output convolutional layer compresses the high-dimensional feature map into a single-channel layer, and then uses a sigmoid activation function and binarization to obtain the edge prediction map I. ep .

[0017] By adopting the above technical solution, the dual-decoder parallel structure focuses on the generation of grayscale images and edge images respectively, realizing the separation and optimization of texture details and local structures, and avoiding the problem that a single output branch cannot take into account both types of information at the same time.

[0018] Optionally, the operation process of the gated convolution module includes: ; Where, x in It is the input of the gated convolution module; DConv 3×3 This indicates a depthwise convolutional layer with a kernel size of 3×3. 1×1 Represents pointwise convolution; sigmoid is a normalization function; g1 is the gate value; Conv 3×3 This indicates a convolutional layer with a kernel size of 3×3; x1 is the intermediate result generated during the operation; x2 is the output of the gated convolution module.

[0019] By adopting the above technical solution, a feature-adaptive gating mechanism is introduced into the gated convolution module, enabling the network to adaptively suppress invalid or masked features during feature extraction. At the same time, it enhances the effective feature response of real terrain regions, allowing the network to automatically weaken noisy features, incomplete structures, and abnormal responses without introducing additional explicit rules, thereby improving the robustness and stability of feature representation.

[0020] Optionally, the computation process of the receptive field enhancement module includes: ; Among them, y in y1 is the input to the receptive field enhancement module; AConv represents dilated convolution, the superscript 3×3 indicates the kernel size, and the subscripts r=1, r=2, r=4, r=8 indicate the dilation rate; concat represents channel concatenation; y1, y2, y3, y4, and y5 are intermediate results generated during the operation; sigmoid is a normalization function; g2 is the gating value; and y6 is the output of the receptive field enhancement module.

[0021] By employing the aforementioned technical solution, the receptive field enhancement module rapidly aggregates multi-scale contextual information, ranging from local details to distant regions, through parallel multi-path dilated convolutions with varying dilation rates, without downsampling or loss of spatial resolution. Combined with an adaptive gating fusion mechanism, this module intelligently fuses these features at different scales, enhancing the terrain structure semantics relevant to the current restoration location while weakening irrelevant information. This is crucial for understanding the reasonable morphology of large-area occluded regions within the overall terrain, providing the decoder with feature priors rich in global consistency.

[0022] Optionally, the image completion network includes an encoder (Encoder2), a bottleneck layer (BR2), and a decoder (Decoder3); wherein: The encoder Encoder2 includes four coding layers, which are used to extract and downsample the input feature map sequentially to obtain the encoded feature map of each layer; each coding layer includes a gated convolutional module, a Transformer module and a downsampling module connected in sequence. The bottleneck layer BR2 includes four cascaded Transformer modules, which receive the output of the last layer of the encoder Encoder2 as input, capture long-distance spatial dependencies, and output a second bottleneck feature map. The decoder Decoder3 includes four decoding layers and one output convolutional layer, which are used to sequentially upsample the second bottleneck feature map and fuse the feature maps output by the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into 3 channels and obtains the repaired image through the sigmoid function and inverse normalization. Each decoding layer includes an upsampling module, a stitching module, and two gated convolutional modules.

[0023] By employing the aforementioned technical solution, a Transformer module is embedded at key locations in the image completion network. Leveraging its powerful long-distance dependency modeling capability through its self-attention mechanism, the network can establish semantic relationships between pixels or regions across arbitrary distances within the image. This is particularly crucial for terrain restoration, ensuring that the restored textures and structures not only harmonize with neighboring areas but also maintain semantic continuity and naturalness with regions on the other side of the image that possess similar terrain features. This results in a highly consistent and seamlessly integrated restoration result, both visually and geometrically, significantly enhancing the realism and usability of the final terrain image.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. By removing non-terrain targets from the imagery and repairing the terrain, point cloud distortion is effectively eliminated, significantly improving the accuracy and stability of earthwork volume calculation.

[0025] 2. Utilizing drones for automated data collection and processing significantly improves operational efficiency and supports high-frequency, precise measurement of large-area or complex projects.

[0026] 3. By accurately comparing the existing point cloud with the BIM design model, quantitative and reliable data support is provided for project settlement and acceptance, thereby improving the scientific nature of management. Attached Figure Description

[0027] Figure 1 A flowchart illustrating a method for earthwork measurement and acceptance based on BIM and UAV point cloud, provided for an embodiment of the present invention; Figure 2 A network architecture diagram of a segmentation and removal model provided in an embodiment of the present invention; Figure 3 This is a network architecture diagram of an image restoration model provided in an embodiment of the present invention; Figure 4 This is an architecture diagram of a coarse-grained repair network provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a gated convolution module provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the receptive field enhancement module provided in an embodiment of the present invention; Figure 7This is an architecture diagram of an image completion network provided in an embodiment of the present invention. Detailed Implementation

[0028] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0029] This invention provides a method for earthwork measurement and acceptance based on BIM and UAV point cloud. See also... Figure 1 , Figure 1 A flowchart illustrating a method for earthwork measurement and acceptance based on BIM and UAV point clouds, provided as an embodiment of the present invention. The method includes the following steps: S1 creates a 4D BIM model based on design drawings, including terrain surfaces, cut and fill work areas, and progress nodes.

[0030] S2, based on the excavation and filling work area, plans the flight path of the drone and collects high-resolution images of the construction site.

[0031] S3 uses a pre-trained segmentation and removal model to identify and segment non-terrain targets in high-resolution images, obtaining a binary mask; S4. The original image is masked using a binary mask to obtain a masked image.

[0032] S5 uses a pre-trained image inpainting model to inpaint the occluded image, resulting in an interference-free terrain image.

[0033] S6 reconstructs the 3D point cloud of the construction site based on non-intrusive terrain images to obtain the current terrain point cloud.

[0034] S7 integrates the existing terrain point cloud with BIM, calculates the earthwork excavation and filling volume, and obtains the earthwork measurement results.

[0035] S8, combining earthwork measurement results with BIM design requirements, completed the project acceptance.

[0036] In one embodiment, see Figure 2 , Figure 2 This is a network architecture diagram of a segmentation and removal model provided in an embodiment of the present invention. This segmentation and removal model is a segmentation model based on the U-Net architecture. U-Net is a classic and efficient architecture in the field of semantic segmentation. Trained on a specialized dataset, it has the ability to identify multiple targets (such as dump trucks, excavators, backhoe loaders, wheel loaders, road rollers, bulldozers, etc.) in complex construction site environments, thereby accurately segmenting construction machinery of different sizes and shapes in the image and generating a binary mask where all non-terrain pixels are marked as 1 and background terrain pixels are marked as 0.

[0037] In one implementation, high-resolution images of the construction site, including targets such as excavators and dump trucks, are collected from the perspective of a drone. Pixel-level segmentation masks of various construction machinery, vehicles, and non-terrain targets in the images are then labeled to form a training dataset.

[0038] The loss function used during the training of this segmentation removal model is: ; Among them, L CE For cross-entropy loss, L Dice For Dice's loss, , It is a weighting coefficient that balances the two losses and can be adjusted flexibly.

[0039] In one embodiment, step S4, which involves masking the original image using a binary mask to obtain a masked image, includes: The binary mask is dilated by expanding the mask area outward by a preset number of pixels to obtain an dilated mask image. The original image is then covered using the dilated mask image to obtain a covered image.

[0040] The mask boundaries generated by the segmentation model may have slight deviations from the actual physical edges of the mechanical target at the pixel level. Directly using the original mask for coverage may result in a small number of pixels belonging to the mechanical edges (such as the edge of an excavator's bucket, parts in shadow, etc.) remaining uncovered. This embodiment ensures that all pixels possibly related to the target, including their blurred edges, are completely covered by extending the mask area outward with a conservative buffer (e.g., 30 pixels). This provides a clear repair area for the subsequent image inpainting model, avoiding the inpainting model misclassifying residual interference textures as part of the terrain due to incomplete coverage.

[0041] In one implementation, the original image is cropped based on the masked region to obtain a sub-image containing the masked region. The masked image of the sub-image is then input into an image inpainting model for inpainting. The specific cropping and masking process includes: Step 1: Perform connected component analysis on the dilation mask to obtain the minimum bounding rectangle of each connected mask region.

[0042] Step 2: Expand outwards from the midpoint of the smallest bounding rectangle to obtain the coordinates of the expanded rectangle.

[0043] Step 3: Based on the expanded rectangular coordinates, crop out the corresponding image block from the original image and denot it as a sub-image; at the same time, crop out the corresponding mask block from the same position in the dilation mask and denot it as a sub-mask.

[0044] Step 4: Use a sub-mask to mask the sub-image to obtain the masked image of the sub-image.

[0045] In one embodiment, see Figure 3 , Figure 3 This invention provides a network architecture diagram of an image restoration model. The image restoration model includes a structure extraction module, a coarse structure restoration network, and an image completion network; wherein: The structure extraction module is used for the masked image (I inc Grayscale and edge processing are performed to obtain a fragmented grayscale image (I). ginc ) and incomplete edge map (I einc ).

[0046] A coarse-grained repair network is used to coarsely repair incomplete grayscale images and incomplete edge images to obtain a grayscale prediction image (I). gp ) and edge prediction map (I ep The input to the coarse-grained repair network consists of a fragmented grayscale image, a fragmented edge image, and a mask concatenated along the channels.

[0047] An image completion network is used to complete the occluded parts of an occluded image based on the grayscale prediction map and edge prediction map provided by a coarse-grained inpainting network, resulting in an interference-free terrain image. The input to the image completion network consists of the occluded image, a binary mask, the grayscale prediction map, and the edge prediction map stitched together along the channels.

[0048] In one implementation, the structure extraction module first employs a weighted average method (I... g The input RGB image is converted to grayscale using the formula (=0.299R+0.587G+0.114B). Then, an edge detection algorithm (Canny edge detection algorithm) is applied to the grayscale image to extract the terrain boundary information, resulting in an edge map. Next, the pixel values ​​of corresponding masked regions in both the grayscale and edge maps are set to zero using a binary mask to eliminate interference from occluded areas on structural information. Finally, the processed grayscale image is normalized. Then input it into the coarse structural repair network for subsequent coarse repair of terrain structure.

[0049] In one embodiment, see Figure 4 , Figure 4 This is an architecture diagram of a coarse-grained repair network provided in an embodiment of the present invention. The coarse-grained repair network includes an encoder (Encoder1), a bottleneck layer (BR1), and two decoders (Decoder1 and Decoder2); wherein: The encoder Encoder1 consists of four coding layers, which sequentially extract and downsample the input feature map to obtain the encoded feature map for each layer. Each coding layer includes two gated convolutional modules and a downsampling module connected in sequence. The downsampling module uses max pooling with a stride of 2 for downsampling. For details, see [link to documentation]. Figure 5 , Figure 5 This is a schematic diagram of a gated convolution module provided in an embodiment of the present invention. The operation process of the gated convolution module includes: ; Where, x in It is the input of the gated convolution module; DConv 3×3 This indicates a depthwise convolutional layer with a kernel size of 3×3. 1×1 Represents pointwise convolution; sigmoid is a normalization function; g1 is the gate value; Conv 3×3 This indicates a convolutional layer with a kernel size of 3×3; x1 is an intermediate result produced during the operation; the operator ⊗ indicates element-wise multiplication (Hadamard product); x2 is the output of the gated convolution module.

[0050] The bottleneck layer BR1 consists of four cascaded receptive field enhancement modules. It receives the output of the last layer of Encoder1 as input, aggregates contextual information from different ranges, and outputs the first bottleneck feature map. For details, see [link to documentation]. Figure 6 , Figure 6 This is a schematic diagram of a receptive field enhancement module provided in an embodiment of the present invention. The operation process of the receptive field enhancement module includes: ; Among them, y in y1 is the input to the receptive field enhancement module; AConv represents dilated convolution, the superscript 3×3 indicates the kernel size, and the subscripts r=1, r=2, r=4, r=8 indicate the dilation rate; concat represents channel concatenation; y1, y2, y3, y4, and y5 are intermediate results generated during the operation; sigmoid is a normalization function; g2 is the gating value; and y6 is the output of the receptive field enhancement module.

[0051] Decoder1 consists of four decoding layers and one output convolutional layer. These layers sequentially upsample the first bottleneck feature map and fuse it with the feature maps output from the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into a single-channel layer and obtains the grayscale prediction image I through the tanh activation function. gp Each decoding layer includes an upsampling module, a concatenation module, and two gated convolutional modules. The upsampling module performs upsampling through deconvolution.

[0052] Decoder2 has the same structure as Decoder1. The final output convolutional layer compresses the high-dimensional feature map into a single-channel layer, and then uses a sigmoid activation function and binarization to obtain the edge prediction map I. ep .

[0053] This coarse-grained restoration network embodiment provides a network structure with shared coding features and dual-branch decoding, responsible for predicting grayscale and edge maps respectively. The combination of the encoder and receptive field enhancement module enables the understanding of terrain structure from a global context, while the parallel operation of the dual decoders achieves the separation, optimization, and collaborative generation of texture trends and local structures. This design allows the network to simultaneously and efficiently recover grayscale information reflecting macroscopic terrain undulations and edge information depicting subtle terrain changes, providing comprehensive and accurate structural guidance for the final image completion.

[0054] In one implementation, a clean, unobstructed terrain image of the construction site is used. A binary mask simulating the shape, size, and position of machinery is randomly generated by a program to mask and clear the grayscale and edge images of the image, simulating the incomplete grayscale and edge images encountered by the network during inference. The original, unmasked, clean grayscale and binary edge images are then used as the learning targets for the network.

[0055] The loss function used during the training of this coarse-grained repair network is: ; Among them, L gray The grayscale image loss consists of L1 loss and structural similarity loss (SSIM); L edge For edge graph loss, a weighted binary cross-entropy loss is used; L perc To perceive consistency loss, features are extracted using a pre-trained ImageNet classification network (such as VGG), and the L2 distance between the predicted complete grayscale image and the true complete grayscale image in the high-level feature space is calculated. , , These are weighting coefficients, which can be adjusted flexibly.

[0056] In one embodiment, see Figure 7 , Figure 7 This is an architecture diagram of an image completion network provided in an embodiment of the present invention. The image completion network includes an encoder (Encoder2), a bottleneck layer (BR2), and a decoder (Decoder3); wherein: The encoder Encoder2 consists of four coding layers, which are used to sequentially extract and downsample the input feature map to obtain the encoded feature map of each layer. Each coding layer includes a gated convolutional module, a Transformer module, and a downsampling module connected in sequence.

[0057] The bottleneck layer BR2 consists of four cascaded Transformer modules. It receives the output of the last layer of the encoder Encoder2 as input, captures long-range spatial dependencies, and outputs the second bottleneck feature map.

[0058] The decoder, Decoder3, consists of four decoding layers and one output convolutional layer. This output layer sequentially upsamples the second bottleneck feature map and fuses it with the feature maps output from the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into three channels and obtains the repaired image through a sigmoid function and inverse normalization. Each decoding layer includes an upsampling module, a stitching module, and two gated convolutional modules.

[0059] This image completion network embodiment achieves high-quality completion of terrain images by organically combining gated convolution with a Transformer structure. By introducing a Transformer module and utilizing its self-attention mechanism, the network can overcome the limitations of the local receptive field of traditional convolution, calculating the correlation weights between pixels or regions across the entire feature map. This is crucial for terrain restoration because it ensures that the restored textures (such as the continuity of land areas and the consistency of slope) not only coordinate with directly adjacent areas but also maintain harmony with semantically related distant areas at any distance in the image. This generates restoration results that are structurally continuous, naturally textured, and highly consistent with the surrounding terrain, providing high-quality image input for subsequent 3D point cloud reconstruction, thereby significantly improving the accuracy and reliability of earthwork measurement and engineering acceptance results.

[0060] In one implementation, the training set uses the same clean image set and binary mask as the structural coarse-grained restoration network. The trained structural coarse-grained restoration network generates corresponding grayscale prediction maps and edge prediction maps as input. The original clean images are used as the learning target.

[0061] The loss function used by this image completion network during training is: ; Among them, L rec To reconstruct the loss, L1 loss is used; L perc For perceived consistency loss; L adv To counteract the loss, a discriminator network is introduced to play a game against the generator (i.e., the image completion network); L struct It is the structural consistency loss, which consists of the L1 loss of the grayscale image and the binary cross-entropy loss of the edge image; , , , These are weighting coefficients, which can be adjusted flexibly.

[0062] In one embodiment, step S7 involves fusing the existing terrain point cloud with BIM, calculating the earthwork excavation and filling volume, and obtaining the earthwork measurement results, including: Step 1: Import the existing terrain point cloud into the BIM management platform and unify its coordinate system with the designed terrain surface in the created 4D BIM. By selecting feature points in the unchanged area surrounding the construction zone, perform iterative nearest-point registration to ensure that the spatial position of the existing point cloud is consistent with the design model.

[0063] Step 2: Based on the registered data, the earthwork volume is analyzed using the triangular mesh volume calculation method: an irregular triangular mesh (TIN) model is generated using the designed terrain surface and the existing point cloud respectively. The software automatically calculates the three-dimensional space volume enclosed between the two TIN models to obtain the actual excavation and filling volume.

[0064] In one embodiment, step S8, combining the earthwork measurement results with the BIM design requirements, completes the project acceptance by comparing the calculated actual excavation and filling volume with the theoretical engineering volume in the BIM design model. If the deviation is within the percentage (such as ±3%) agreed in the contract or specifications, it is considered qualified.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention are within the scope of the claims of the present invention.

Claims

1. A method for earthwork measurement and acceptance based on BIM and UAV point cloud, characterized in that, The method includes: Create a 4D BIM model based on the design drawings, including terrain surfaces, cut and fill work areas, and progress nodes; Based on the excavation and filling work area, plan the flight path of the drone and collect high-resolution images of the construction site; A pre-trained segmentation and removal model is used to identify and segment non-terrain targets in high-resolution images to obtain a binary mask. The original image is then masked based on the binary mask to obtain a masked image. A pre-trained image inpainting model is used to inpaint the occluded image, resulting in an interference-free terrain image. Reconstructing the 3D point cloud of the construction site based on non-intrusive terrain imagery yields the current terrain point cloud. By integrating the existing terrain point cloud with BIM, the volume of earthwork excavation and filling is calculated to obtain the earthwork measurement results. Based on the earthwork measurement results and BIM design requirements, the project was accepted.

2. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 1, characterized in that, The segmentation and removal model is a segmentation model based on the U-Net architecture.

3. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 1, characterized in that, The process of masking the original image using a binary mask to obtain a masked image includes: The binary mask is dilated by expanding the mask area outward by a preset number of pixels to obtain an expanded mask image. The original image is masked using the dilated mask image to obtain a masked image.

4. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 3, characterized in that, The process of masking the original image using the dilated mask image to obtain the masked image includes: Perform connected component analysis on the dilated mask map to obtain the minimum bounding rectangle of each connected mask region; Expand outwards from the midpoint of the smallest bounding rectangle to obtain the coordinates of the expanded rectangle; Based on the expanded rectangular coordinates, the corresponding image block is cropped from the original image as a sub-image; at the same time, the corresponding mask block is cropped from the same position in the dilated mask image as a sub-mask. The sub-image is masked using the sub-mask to obtain a masked image suitable for input to an image restoration model.

5. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 1, characterized in that, The image restoration model includes a structure extraction module, a coarse structure restoration network, and an image completion network; wherein: The structure extraction module is used to perform grayscale processing and edge processing on the masked image to obtain a fragmented grayscale image and a fragmented edge image. The coarse structural repair network is used to coarsely repair the incomplete grayscale image and the incomplete edge image to obtain a grayscale prediction image and an edge prediction image; the input of the coarse structural repair network is composed of the incomplete grayscale image, the incomplete edge image and the binary mask spliced ​​along the channel; The image completion network is used to complete the occluded parts of the occluded image based on the grayscale prediction map and edge prediction map provided by the coarse structural repair network, so as to obtain an interference-free terrain image. The input of the image completion network is composed of the occluded image, the binary mask, the grayscale prediction map and the edge prediction map stitched together along the channel.

6. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 5, characterized in that, The coarse-repair network comprises an encoder (Encoder1), a bottleneck layer (BR1), and two decoders (Decoder1 and Decoder2); wherein: The encoder Encoder1 includes four coding layers, which are used to extract and downsample the input feature map sequentially to obtain the encoded feature map of each layer; each coding layer includes two gated convolutional modules and a downsampling module connected in sequence. The bottleneck layer BR1 includes four cascaded receptive field enhancement modules, which receive the output of the last layer of encoder 1 as input, aggregate contextual information of different ranges, and output the first bottleneck feature map. The decoder Decoder1 comprises four decoding layers and one output convolutional layer, used to sequentially upsample the first bottleneck feature map and fuse the feature maps output by the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into a single-channel layer and obtains the grayscale prediction map I through the tanh activation function. gp Each decoding layer includes an upsampling module, a splicing module, and two gated convolutional modules. Decoder2 has the same structure as Decoder1. The final output convolutional layer compresses the high-dimensional feature map into a single-channel layer, and then uses a sigmoid activation function and binarization to obtain the edge prediction map I. ep .

7. The earthwork measurement and acceptance method based on BIM and UAV point cloud as described in claim 6, characterized in that, The operation process of the gated convolution module includes: ； Where, x in It is the input of the gated convolution module; DConv 3×3 This indicates a depthwise convolutional layer with a kernel size of 3×3. 1 ×1 Represents pointwise convolution; sigmoid is a normalization function; g1 is the gate value; Conv 3×3 This indicates a convolutional layer with a kernel size of 3×3; x1 is the intermediate result generated during the operation; x2 is the output of the gated convolution module.

8. A method for earthwork measurement and acceptance based on BIM and UAV point cloud as described in claim 6, characterized in that, The computation process of the receptive field enhancement module includes: ； Among them, y in y1 is the input to the receptive field enhancement module; AConv represents dilated convolution, the superscript 3×3 indicates the kernel size, and the subscripts r=1, r=2, r=4, r=8 indicate the dilation rate; concat represents channel concatenation; y1, y2, y3, y4, and y5 are intermediate results generated during the operation; sigmoid is a normalization function; g2 is the gating value; and y6 is the output of the receptive field enhancement module.

9. A method for earthwork measurement and acceptance based on BIM and UAV point cloud as described in claim 5, characterized in that, The image completion network includes an encoder (Encoder2), a bottleneck layer (BR2), and a decoder (Decoder3); wherein: The encoder Encoder2 includes four coding layers, which are used to extract and downsample the input feature map sequentially to obtain the encoded feature map of each layer; each coding layer includes a gated convolutional module, a Transformer module and a downsampling module connected in sequence. The bottleneck layer BR2 includes four cascaded Transformer modules, which receive the output of the last layer of the encoder Encoder2 as input, capture long-distance spatial dependencies, and output a second bottleneck feature map. The decoder Decoder3 includes four decoding layers and one output convolutional layer, which are used to sequentially upsample the second bottleneck feature map and fuse the feature maps output by the corresponding encoding layers. Finally, the output convolutional layer compresses the high-dimensional feature map into 3 channels and obtains the repaired image through the sigmoid function and inverse normalization. Each decoding layer includes an upsampling module, a stitching module, and two gated convolutional modules.

Images