A three-dimensional thermal environment neural field reconstruction method and system for building facade defect diagnosis

Abstract

This invention discloses a three-dimensional thermal environment neural field reconstruction method and system for diagnosing building facade defects, belonging to the field of building energy conservation technology. It includes: simultaneously acquiring multi-view image data using an integrated thermal infrared and visible light dual-modal device; achieving cross-modal camera pose calibration using an improved Structure for Motion Restoration (SfM) algorithm; constructing an implicit thermal radiation model using neural radiation fields (NeRF) to introduce temperature information and enhance thermal environment characterization; and generating an explicit triangular mesh model with both geometric details and thermal properties using the signed distance function (SDF) and Marching Cubes algorithm. The invention's multi-modal neural radiation field fusion framework complements thermal infrared and visible light data to achieve three-dimensional thermal environment reconstruction. This method can be widely applied to building facade safety diagnosis, energy-saving renovation, and low-carbon operation and maintenance.

Description

Technical Field

[0001] This invention relates to a three-dimensional thermal environment neural field reconstruction method and system for diagnosing defects in building facades, belonging to the field of building energy conservation technology. Background Technology

[0002] Globally, building energy consumption has become a significant component of total social energy consumption, especially with the accelerating pace of urbanization, where its proportion continues to rise. Simultaneously, safety accidents caused by facade defects are frequent, making improved diagnosis of these defects crucial not only for environmental protection and energy conservation but also for urban public safety. Thermal environment information of building surfaces, particularly the temperature distribution of building facades, is vital for assessing building energy consumption levels, identifying potential building hazards, and developing low-carbon retrofit strategies. Therefore, accurately acquiring and analyzing changes in the thermal environment of building surfaces is crucial for improving building energy efficiency and implementing energy-saving retrofits.

[0003] Traditional 3D reconstruction techniques have been widely applied in the field of RGB imaging, especially in urban planning, architectural design, and post-disaster reconstruction. However, in the field of thermal infrared imaging, traditional 3D reconstruction methods still face many challenges. While thermal infrared imaging can provide detailed information about the temperature distribution on building surfaces, helping to reveal heat loss and energy efficiency issues, it has proven effective in various ways. For example, in 2017, a research team in Paris, France, collected nighttime building thermal data using a handheld thermal imager and combined it with structure-of-motion (SfM) and multi-view stereo matching (MVS) to generate 3D models, successfully locating 15% of the exterior facade heat leakage areas, resulting in a 12% efficiency improvement after energy-saving renovations. In 2019, a team from Columbia University in the United States used drone thermal infrared aerial photography and LiDAR point cloud registration, combined with machine learning, to identify 23% of the exterior wall thermal bridging effects in old buildings in New York, prompting revisions to local energy efficiency laws. In 2021, the Fraunhofer Institute in Germany innovatively integrated visible light and thermal infrared imaging, combined with BIM dynamic simulation, to reduce heating energy consumption in industrial buildings by 18%. However, despite the valuable practical experience provided by these studies, the accuracy is still limited by the low resolution of thermal infrared images, their susceptibility to environmental interference, and the insufficient adaptability of traditional algorithms. In the future, we need to further improve their practicality through deep learning and multimodal data fusion.

[0004] Currently, the application of traditional 3D reconstruction technology to thermal infrared imagery faces several challenges. Firstly, the quality of thermal infrared images is affected by numerous factors, such as the temperature difference between the target and background, signal noise, and the resolution of the imaging equipment. These factors limit the accuracy of 3D reconstruction. Furthermore, due to the non-line-of-sight nature of thermal infrared imaging, the spatial resolution of building surface temperature information is typically low, further increasing the difficulty of the reconstruction process. For energy efficiency analysis of large-scale buildings or urban areas, how to efficiently and accurately process and reconstruct thermal infrared data remains a pressing issue. These problems restrict the comprehensive application of thermal infrared imagery in building energy efficiency assessment. Future improvements in technological innovation and algorithm optimization are urgently needed to achieve more accurate reconstruction of the building thermal environment. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a three-dimensional thermal environment neural field reconstruction method and system for diagnosing defects in building facades, so as to achieve high-precision three-dimensional reconstruction of the thermal environment of building surfaces and support thermal analysis and diagnosis from any viewpoint.

[0006] To achieve the above objectives / to solve the above technical problems, the present invention is implemented using the following technical solution:

[0007] First aspect: A three-dimensional thermal environment neural field reconstruction method for diagnosing building facade defects, the method comprising:

[0008] By using thermal infrared cameras and visible light cameras, the building surface is simultaneously captured from multiple perspectives according to a preset path to obtain thermal infrared images and visible light images, thus constructing a dual-modal image dataset.

[0009] Based on the dual-modal image dataset, the pose of the visible light camera is recovered using the structure-of-motion-recovery algorithm, and the relative pose relationship between the thermal infrared camera and the visible light camera is determined through joint calibration experiments, thereby achieving spatial alignment of cross-modal data.

[0010] By utilizing neural radiation fields and fusing temperature information from thermal infrared images and visible light texture features from a dual-modal image dataset, an implicit three-dimensional model of the building's thermal radiation field is constructed; the implicit three-dimensional model generates a thermal environment representation from any viewpoint through volumetric rendering.

[0011] Based on the dual-modal image dataset and the poses of the thermal infrared camera and the visible light camera, a symbolic distance function and the Marching Cubes algorithm are introduced to extract the explicit triangular mesh surface from the implicit 3D model. The temperature information of the thermal infrared image and the visible light texture features are then mapped onto the triangular mesh surface to reconstruct the texture-mapped triangular mesh model.

[0012] Optionally, the preset path is dynamically planned based on the building's geometric features, which can ensure that the image coverage and overlap rate meet the preset thresholds. A camera parameter distortion correction algorithm is used to eliminate lens distortion of thermal infrared and visible light cameras.

[0013] Optionally, the motion recovery structure algorithm optimizes the objective function of bundle adjustment by introducing sparse feature point matching of thermal infrared images.

[0014] Optionally, the joint calibration experiment establishes the conversion relationship between the intrinsic and extrinsic parameters of the thermal infrared and visible light cameras using the checkerboard calibration method.

[0015] Optionally, the neural radiation field adopts a dual-branch network architecture to learn visible light texture and thermal infrared temperature respectively.

[0016] Optionally, the volume rendering formula is:

[0017] ;

[0018] in, These are image pixel values. This represents the small interval represented by the i-th sampling point. It is cumulative transparency. Indicates the opacity of the i-th sampling point. The color of the sampling point is represented by N, which is the total number of sampling points on the light ray.

[0019] Optionally, the neural radiation field simultaneously monitors the temperature prediction information of the sampling points in the dual-modal image dataset.

[0020] The second aspect: a three-dimensional thermal environment neural field reconstruction system for diagnosing defects in building facades, the system comprising:

[0021] The multimodal data acquisition module is configured to simultaneously capture images of the building surface from multiple perspectives using a thermal infrared camera and a visible light camera, following a preset path, to obtain thermal infrared images and visible light images, and to construct a dual-modal image dataset.

[0022] The cross-modal calibration module is configured to: recover the pose of the visible light camera based on the structure of motion recovery algorithm according to the dual-modal image dataset, and determine the relative pose relationship between the thermal infrared camera and the visible light camera through joint calibration experiments, thereby achieving spatial alignment of cross-modal data;

[0023] An implicit reconstruction module is configured to: construct an implicit 3D model of the building's thermal radiation field based on the neural radiation field, by fusing temperature information from thermal infrared images and visible light texture features from a dual-modal image dataset; the implicit 3D model generates a thermal environment representation from any viewpoint through volumetric rendering;

[0024] Explicit Surface Generation Module: Configured to: extract explicit triangular mesh surfaces from implicit 3D models based on the dual-modal image dataset and the poses of thermal infrared and visible light cameras, introduce a signed distance function and the Marching Cubes algorithm, map the temperature information of thermal infrared images and visible light texture features onto the triangular mesh surfaces, and reconstruct the texture-mapped triangular mesh model.

[0025] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0026] This invention integrates multimodal devices and data-driven cross-modal calibration, accurately aligning thermal infrared and visible light data, and supporting common-path simulation sampling;

[0027] This invention presents a multimodal neural radiation field fusion framework that complements thermal infrared and visible light data to achieve three-dimensional reconstruction of the thermal environment. This method can be widely applied to building facade safety diagnosis, energy-saving renovation, and low-carbon operation and maintenance. Attached Figure Description

[0028] Figure 1 The diagram shown is a flowchart of the three-dimensional thermal environment neural field reconstruction method provided in an embodiment of the present invention.

[0029] Figure 2 The diagram shown is a flowchart of the triangular mesh model construction process of the three-dimensional thermal environment neural field reconstruction method provided in this embodiment of the invention.

[0030] Figure 3 The image shown is a schematic diagram of thermal infrared and visible light images of a building surface provided in an embodiment of the present invention;

[0031] Figure 4 The diagram shown is a schematic diagram of thermal infrared image distortion correction provided in an embodiment of the present invention;

[0032] Figure 5 The figure shown is a schematic diagram of the explicit surface reconstruction results based on SDF and Marching Cube algorithms provided in an embodiment of the present invention;

[0033] Figure 6 The diagram shown is a schematic diagram of the thermal environment analysis results of building surfaces provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0035] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0036] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0037] like Figures 1-2 As shown: A three-dimensional thermal environment neural field reconstruction method for diagnosing building facade defects, the method includes:

[0038] Step 1: Simultaneously capture multi-view images of the building surface using a thermal infrared camera and a visible light camera, following a preset path, to obtain thermal infrared and visible light images and construct a dual-modal image dataset;

[0039] Step 2: Based on the dual-modal image dataset, the pose of the visible light camera is recovered using the structure-of-motion-recovery algorithm, and the relative pose relationship between the thermal infrared camera and the visible light camera is determined through joint calibration experiments to achieve spatial alignment of cross-modal data;

[0040] Step 3: Based on the neural radiation field, temperature information from thermal infrared images and visible light texture features from the dual-modal image dataset are fused to construct an implicit 3D model of the building's thermal radiation field; the implicit 3D model is then used to generate a thermal environment representation from any viewpoint through volumetric rendering.

[0041] Step 4: Based on the dual-modal image dataset and the poses of the thermal infrared camera and the visible light camera, the symbolic distance function and the Marching Cubes algorithm are introduced to extract the explicit triangular mesh surface from the implicit 3D model. The temperature information of the thermal infrared image and the visible light texture features are then mapped onto the triangular mesh surface to reconstruct the texture-mapped triangular mesh model.

[0042] Step 5: Combine the semantic database of building materials with the thermodynamic simulation engine to analyze the temperature field distribution in the reconstructed model, supporting the identification of defective areas such as hollow areas and leaks on the building facade.

[0043] In this embodiment, in step 1, the preset path is dynamically planned based on the building's geometric features, which enables the image coverage and overlap rate to meet the preset threshold. The camera parameter distortion correction algorithm is used to eliminate lens distortion of thermal infrared and visible light cameras and improve data spatial consistency.

[0044] In the specific implementation process: High-precision thermal infrared and visible light cameras are used on drones to simultaneously capture images of the building surface from multiple different perspectives, acquiring thermal and geometric information of the building surface to form thermal infrared and visible light image datasets. During the acquisition process, attention is paid to the coverage and overlapping areas of the images to ensure that the images fully cover the main parts of the building surface and have sufficient overlap for subsequent image registration and 3D reconstruction.

[0045] In this embodiment, in step 2, the objective function of bundle adjustment (BA) is optimized by introducing sparse feature point matching of thermal infrared images based on the structure of motion recovery (SfM) algorithm; the joint calibration experiment establishes the intrinsic and extrinsic parameter conversion relationship between thermal infrared and visible light cameras through the checkerboard calibration method.

[0046] In the specific implementation process: Camera pose recovery is the core link in 3D reconstruction and image distortion correction, including two steps: extrinsic parameter calculation of visible light images and intrinsic parameter calibration of thermal infrared images. First, the intrinsic parameters of the thermal infrared camera are calibrated using the checkerboard method. This involves capturing a checkerboard image with known dimensions and calibration points, and calculating the camera's intrinsic parameters, such as focal length, principal point position, and distortion coefficients. The checkerboard method relies on the accurate identification and matching of the calibration board corner points in the image to ensure accurate camera intrinsic parameters. Next, the SfM algorithm is used to recover the camera's extrinsic parameters from the visible light image. SfM extracts and matches feature points from the image, using these matching points to estimate the camera's position and pose. Bundle adjustment (BA) is then used to globally optimize the pose and 3D point coordinates of all cameras, thereby improving the accuracy of pose estimation. Finally, by fusing the extrinsic parameters of the visible light camera and the intrinsic parameters of the thermal infrared camera, effective distortion correction can be achieved for both images, realizing the alignment and fusion of thermal infrared and visible light images in 3D space, providing high-precision image data for subsequent thermal environment reconstruction and analysis. See attached document for details. Figure 4 : Schematic diagram of thermal infrared image distortion correction.

[0047] In this embodiment, for step 3, the NeRF model (Neural Radiation Field) adopts a dual-branch network architecture to learn visible light texture and thermal infrared temperature respectively, and enhances the sensitivity to thermal radiation changes through a temperature loss function; in the volume rendering formula, the rendering weight of thermal radiation distribution is dynamically adjusted.

[0048] In practical implementation: Based on Neural Radiation Field (NeRF) technology, the 3D reconstruction process involves fusing thermal infrared images with traditional visual information to more accurately capture and reconstruct the building's thermal environment. NeRF learns the density, color, and temperature of each point in 3D space through an implicit representation based on neural networks, transforming the input 2D images (thermal infrared and visible light images) into 3D data in space. The network learns the visual characteristics of each point and the temperature information of its location.

[0049] Specifically, such as Figure 3 As shown, in the 3D reconstruction process based on Neural Radiation Field (NeRF) technology, the temperature value of the thermal infrared image is used as an additional input channel, along with spatial coordinates and viewpoint information, and fed into the neural network. NeRF generates 3D thermal environment images through volumetric rendering technology. This technology outputs the color and density of each sampling point from the network and uses the following volumetric rendering formula to generate the final image:

[0050] （1）；

[0051] in, These are image pixel values. This represents the small interval represented by the i-th sampling point. It is cumulative transparency. Indicates the opacity of the i-th sampling point. The color of the sampling point is represented by N, which is the total number of sampling points on the light ray.

[0052] Furthermore, temperature information is incorporated into the volumetric rendering process to enhance the ability to capture changes in thermal radiation, thereby improving sensitivity to the thermal environment of the building's exterior surface. Therefore, the network simultaneously monitors the temperature prediction information at the sampling points. By analyzing the temperature at each sampling point, NeRF (Neural Radiation Field) can more accurately reconstruct the thermal environment image. Finally, by combining this temperature information with spatial coordinates and viewpoint information, the network can generate a detailed 3D thermal environment reconstruction that accurately reflects the building's thermal radiation characteristics.

[0053] Introducing temperature information during volumetric rendering helps the network better capture changes in thermal radiation. This means the network can accurately simulate temperature distribution, heat propagation, and radiation characteristics at different 3D spatial locations, thereby improving reconstruction accuracy, especially for building thermal energy assessment. The introduction of temperature gradient information enhances the model's sensitivity to changes in thermal radiation, making the final generated 3D thermal environment image more realistic and accurately reflecting key information such as heat leakage and energy efficiency weaknesses in buildings.

[0054] In this embodiment, in step 4, explicit surface reconstruction is achieved by obtaining the zero isosurface in the SDF; the Laplacian smoothing and edge length optimization algorithm is used to post-process the triangular mesh to improve the surface smoothness and geometric consistency of the model.

[0055] In practical implementation: Based on Neural Radiation Field (NeRF) technology, explicit surface reconstruction extends the NeRF network structure to simultaneously output the Signed Distance Function (SDF) value for each sampling point. The SDF is a function describing the orthogonal distance from a point in 3D space to an object's surface. Obtaining its zero isosurface allows for the construction of a 3D model of the object's surface, optimization of network parameters, and improvement of the accuracy of 3D surface reconstruction.

[0056] The specific process is as follows: The network estimates the distance to the object's surface by learning the SDF value of each sampling point. After the network obtains the SDF value of each sampling point in space through optimization, the Marching Cube algorithm is used to extract the explicit triangular mesh model of the building surface from the SDF representation. The Marching Cube algorithm finds zero isosurfaces (i.e., points with an SDF value of 0) in the SDF field and gradually connects adjacent points to form a triangular mesh, thereby forming the object's surface.

[0057] This algorithm takes the SDF value as input and generates an object surface composed of a triangular mesh. After generating the initial triangular mesh, mesh quality needs to be improved through mesh optimization techniques, including mesh smoothing, reducing redundant faces, and increasing mesh detail. The mesh optimization methods we use include Laplacian smoothing and edge length optimization, resulting in a finer and smoother mesh, ensuring reconstruction quality. Through this process, NeRF can not only generate high-quality 3D reconstructions of the thermal environment but also provide accurate explicit 3D surface representations, offering more detailed spatial data for building energy efficiency assessments and energy-saving retrofits. See the appendix for details. Figure 5 : Schematic diagram of explicit surface reconstruction results based on SDF and Marching Cube algorithms.

[0058] In this embodiment, step 5 involves in-depth analysis of the reconstructed 3D thermal environment model, incorporating building material and semantic information such as window frames and marble walls. By analyzing the temperature field distribution of the building envelope, defects in the building facade, such as hollow areas and leaks, can be detected and diagnosed. This stage of work not only demonstrates the practical application value of 3D reconstruction technology but also provides strong technical support for building energy consumption assessment and low-carbon retrofitting.

[0059] In the specific implementation process: After completing the explicit surface reconstruction, thermal environment analysis combines building material characteristics and semantic information to conduct in-depth analysis of the reconstructed 3D thermal environment model. By analyzing the temperature distribution in different areas of the building surface, utilizing temperature values ​​from thermal infrared images and the building's material characteristics, different building surfaces can be effectively identified and distinguished. This is because some materials exhibit different thermal responses compared to their surrounding areas due to variations in thermal conductivity. We can then further use material classification and temperature analysis to accurately detect defects in the building facade, such as hollow areas and leaks. Generally, abnormal areas in the temperature field usually correspond to potential heat loss problems in the building structure or facade. In the model, based on the behavior of these abnormal temperature areas and combined with semantic information, we can make accurate judgments and provide a theoretical basis for energy-saving renovations of buildings. See the appendix for details. Figure 6 : Schematic diagram of the thermal environment analysis results of building surface, showing the temperature field distribution and defect detection.

[0060] Example 2: A three-dimensional thermal environment neural field reconstruction system for diagnosing building facade defects, the system comprising:

[0061] The multimodal data acquisition module is configured to simultaneously capture images of the building surface from multiple perspectives using a thermal infrared camera and a visible light camera, following a preset path, to obtain thermal infrared images and visible light images, and to construct a dual-modal image dataset.

[0062] The cross-modal calibration module is configured to: recover the pose of the visible light camera based on the structure of motion recovery algorithm according to the dual-modal image dataset, and determine the relative pose relationship between the thermal infrared camera and the visible light camera through joint calibration experiments, thereby achieving spatial alignment of cross-modal data;

[0063] An implicit reconstruction module is configured to: construct an implicit 3D model of the building's thermal radiation field based on the neural radiation field, by fusing temperature gradient information from thermal infrared images and visible light texture features from a dual-modal image dataset; the implicit 3D model generates a thermal environment representation from any viewpoint through volumetric rendering.

[0064] Explicit Surface Generation Module: Configured to: extract explicit triangular mesh surfaces from implicit 3D models based on the dual-modal image dataset and the poses of thermal infrared and visible light cameras, by introducing a signed distance function and the Marching Cubes algorithm, and map the temperature gradient information of thermal infrared images and visible light texture features onto the triangular mesh surfaces to reconstruct the texture-mapped triangular mesh model.

[0065] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional thermal environment neural field reconstruction for building facade defect diagnosis, characterized in that, The method comprises: Through a thermal infrared camera and a visible light camera, multi-view synchronous shooting is performed on a building surface according to a preset path, thermal infrared images and visible light images are acquired, and a dual-mode image dataset is constructed; According to the dual-mode image dataset, the pose of the visible light camera is recovered based on a motion recovery structure algorithm, and the relative pose relationship between the thermal infrared camera and the visible light camera is determined through a joint calibration experiment, so that spatial alignment of cross-modal data is realized, thereby realizing alignment and fusion of thermal infrared and visible light images in three-dimensional space; By extending the neural radiation field network structure, the neural radiation field network structure can output the signed distance function of each sampling point simultaneously, a double-branch network architecture is adopted to learn the visible light texture and the thermal infrared temperature respectively, the sensitivity of thermal radiation change is enhanced through a temperature loss function, the temperature information of the thermal infrared images and the visible light texture features in the dual-mode image dataset are fused, an implicit three-dimensional model of the building thermal radiation field is constructed, and the temperature information is introduced in the volume rendering process, and the sensitivity of thermal radiation change is enhanced through the temperature loss function; the implicit three-dimensional model can accurately simulate temperature distribution, heat propagation and radiation characteristics at different three-dimensional space positions through volume rendering; According to the dual-mode image dataset and the poses of the thermal infrared camera and the visible light camera, the signed distance function and the Marching Cubes algorithm are introduced, an explicit triangular mesh surface is extracted from the obtained implicit three-dimensional model, and the temperature information of the thermal infrared images and the visible light texture features are mapped to the triangular mesh surface, so that a triangular mesh model after texture mapping is reconstructed.

2. The three-dimensional thermal environment neurofield reconstruction method for building facade defect diagnosis according to claim 1, characterized in that, The preset path is dynamically planned based on building geometric characteristics, so that the image coverage range and the overlap rate meet preset threshold values, and a camera parameter de-distortion algorithm is adopted to eliminate lens distortion of the thermal infrared camera and the visible light camera. 3.The method of claim 1, wherein, The motion recovery structure algorithm optimizes the objective function of the bundle adjustment by introducing sparse feature point matching of the thermal infrared images. 4.The method of claim 1, wherein, The joint calibration experiment establishes the internal parameter matrix and the external parameter conversion relationship of the thermal infrared camera and the visible light camera through a checkerboard calibration method, and constructs a dual-mode image pair directly used for sampling in the same light path.

5. The method of claim 1, wherein the method is a method of building facade defect diagnosis-oriented three-dimensional thermal environment neurofield reconstruction. The volume rendering formula is: ； wherein, is the image pixel value, denotes a micro-interval represented by the i-th sample point, is the accumulated transparency, denotes the opacity of the i-th sample point, denotes the color of the sample point, and N is the total number of sample points on the ray.

6. The three-dimensional thermal environment neural field reconstruction method for building facade defect diagnosis according to claim 1, wherein the neural radiation field introduces temperature information in the volume rendering process to enhance the ability to capture thermal radiation change, and realizes supervision of temperature prediction information of sampling points of the dual-mode image dataset.

7. A three-dimensional thermal environment neural field reconstruction system for building facade defect diagnosis, characterized in that, The system comprises: A multi-modal data acquisition module configured to perform multi-view synchronous shooting on a building surface through a thermal infrared camera and a visible light camera according to a preset path, acquire thermal infrared images and visible light images, and construct a dual-mode image dataset; The cross-modal calibration module is configured to: according to the dual-modal image dataset, recover the pose of the visible light camera based on a motion structure recovery algorithm, and determine the relative pose relationship between the thermal infrared camera and the visible light camera through a joint calibration experiment, and realize spatial alignment of the cross-modal data, wherein the spatial alignment of the cross-modal data is realized, so as to realize alignment and fusion of the thermal infrared and visible light images in a three-dimensional space; The implicit reconstruction module is configured to: by extending the neural radiance field network structure, make it capable of simultaneously outputting the signed distance function of each sampling point, adopt a double-branch network architecture to learn visible light texture and thermal infrared temperature respectively, and through a temperature loss function, enhance the sensitivity of thermal radiation change, fuse the temperature information of the thermal infrared image and the visible light texture features in the dual-modal image dataset, construct an implicit three-dimensional model of the building thermal radiation field, and in the volume rendering process, introduce temperature information, and combine the temperature loss function to enhance the sensitivity of thermal radiation change; the implicit three-dimensional model can accurately simulate temperature distribution, heat propagation and radiation characteristics at different three-dimensional space positions through volume rendering; The explicit surface generation module is configured to: according to the dual-modal image dataset and the pose of the thermal infrared camera and the visible light camera, introduce the signed distance function and combine the Marching Cubes algorithm, extract an explicit triangular mesh surface from the obtained implicit three-dimensional model, and map the temperature information of the thermal infrared image and the visible light texture features to the triangular mesh surface, and reconstruct a triangular net model after texture mapping.

Images