A 3D real scene modeling and dynamic updating method based on multi-source data fusion
By combining tilt photogrammetry and lidar scanning technology, multi-source data fusion and point cloud registration are carried out, and remote sensing change detection and close-up photography technology are used to solve the problems of bottom adhesion and distortion and blind spots on the top of the building in three-dimensional real-life modeling. , realizes three-dimensional real scene refined modeling and dynamic updates, and improves the efficiency and accuracy of data acquisition.
Patent Information
- Application Number
- CN202411178355.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-27
AI Technical Summary
The existing three-dimensional real-life modeling technology has the tendency of bottom adhesion and distortion of tilt photogrammetry models, and the laser scanning technology has problems with blind spots and texture distortion at the top of buildings, and it is difficult to quickly and efficiently obtain geographic information data, and urban expansion requires irregular periods. Updated the three-dimensional model.
Combining tilt photogrammetry technology and lidar scanning technology, multi-source data fusion method is adopted to obtain tilt images and remote sensing images through drones, and point cloud data is obtained using ground lidar scanning, point cloud registration and fusion modeling, and Remote sensing change detection and close-up photography technology are used to reshoot the changing areas to achieve refined modeling and dynamic updates in three-dimensional real scenes.
The accuracy and details of three-dimensional modeling are improved, and the problems of adhesion and distortion at the bottom of the model and blind spots at the top of the building are solved, which realizes dynamic updates of the three-dimensional model and efficient acquisition of geographic information data.
Smart Images

Figure CN118691776B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of three-dimensional real scene modeling, and in particular relates to a three-dimensional real scene modeling and dynamic updating method based on multi-source data fusion. Background Art
[0002] With the continuous development of China's economy and technology, urban construction is moving towards digitalization and intelligence. In the rapid development of "digital cities" and "smart cities", people's demand for high-quality spatial geographic information data has gradually shifted from traditional two-dimensional planes to more realistic and accurate three-dimensional real-scene modeling. Traditional drone photogrammetry and post-manual modeling methods can no longer meet the needs of fast and efficient acquisition of geographic information data, and their costs are high and efficiency is low.
[0003] At present, the main methods for constructing real-life 3D models are oblique photogrammetry and laser scanning technology. Among them, oblique photogrammetry uses drones to collect oblique images and then generate 3D models. However, oblique image measurement is affected by the flight altitude of the aerial photography platform, the camera angle of view and the occlusion of buildings, and the model is prone to bottom adhesion and distortion. Laser scanning technology can quickly move and measure, and obtain a large amount of high-density point cloud data in a short time. However, this method also has certain limitations. Due to the ground installation of the instrument, the top of the building often becomes a scanning blind spot, and the 3D modeling effect will be distorted or even "void" due to texture information. Therefore, how to integrate other data sources and technologies to improve the accuracy and details of 3D modeling has extremely high theoretical and practical value. In addition, with the continuous acceleration of urban expansion, it is necessary to detect changing areas from time to time and update the 3D real-life model. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a three-dimensional real scene modeling and dynamic updating method which complements the advantages of oblique photogrammetry technology and lidar scanning technology, combines remote sensing image change detection technology and uses close-range photogrammetry to perform close-up reshoots of changed areas, and can realize three-dimensional real scene refined modeling and dynamic updating based on multi-source data fusion.
[0005] The technical solution of the present invention is as follows:
[0006] A three-dimensional real scene modeling and dynamic updating method based on multi-source data fusion includes the following steps:
[0007] S1: Carry out aerial photography preparation according to the building height, vegetation and road distribution of the test area, use drones to obtain oblique images and remote sensing image data of the entire test area, incorporate the oblique images and remote sensing image data into a unified object coordinate system through aerial triangulation, obtain sparse point clouds based on the oblique images and remote sensing image data, perform closed-loop scanning of the test area through ground-based laser radar, obtain point cloud data while moving, and pre-process the point cloud data to obtain radar point clouds;
[0008] S2: Based on the point cloud registration of sparse point cloud and radar point cloud, the real scene 3D model is generated by fusion modeling;
[0009] S3: Re-aerial triangulation of the area to be measured and quality inspection of the real-scene 3D model are performed. If qualified, an encrypted point cloud is generated through aerial triangulation, and a 3D model is further generated based on the real-scene 3D model. If it does not meet the requirements, the UAV route is re-planned and the oblique photogrammetry of the area to be measured is re-performed;
[0010] S4: Shoot the later images of the area to be measured according to the shooting cycle, and perform remote sensing change detection on the remote sensing period images at different times in the later images. If a change is detected, obtain the changed area and the corresponding coordinate information, and further use the drone close-up photography technology to locally reshoot the changed area. If no change is detected, there is no need to update the three-dimensional model;
[0011] S5: Generate an encrypted point cloud by performing aerial triangulation on the close-up images of the changed area and update the real-scene 3D model.
[0012] Furthermore, the real-scene 3D model quality detection in step S3 performs a quantitative quality evaluation on the real-scene 3D modeling based on three dimensions: boundary point accuracy, side length accuracy, and area accuracy.
[0013] Further, the remote sensing change detection in step S4 includes the following steps:
[0014] S41. Remote sensing images from different periods , are sent to the Transformer model respectively, where X (t1) and X (t2) are remote sensing images of two periods respectively, H and W represent the height and width of the image respectively, and C represents the number of channels of the image. Represents dimension information;
[0015] S42. The Transformer model converts each pixel or pixel block in the input image into a feature vector sequence, and uses the self-attention mechanism of the Transformer model to calculate the global correlation of the feature vector sequence to obtain the attention matrix. The process can be expressed as:
[0016] ,
[0017] in, represents the self-attention mechanism, Q, K, and V represent the query, key, and value vectors obtained by linearly changing the triplet of the feature vector sequence. is the scaling factor, represents the transposition operation, softmax represents the softmax function, and the self-attention mechanism is expanded into the form of multi-head self-attention MultiHead. The expression is as follows:
[0018] ,
[0019] ,
[0020] in, represents the number of attention heads, , , , Represents the weight parameter of linear transformation, Concat represents the feature concatenation operation, represents the i-th attention head;
[0021] S43. After obtaining the attention matrix, the output SA of the self-attention mechanism is obtained by performing a weighted sum operation on the input feature vector sequence. The expression is:
[0022] ;
[0023] Where X is the remote sensing image input X at different times (t1) or X (t2) ;
[0024] S44, the output SA of the self-attention mechanism is further processed by the feed-forward neural network FFN. The calculation process of the FFN network is:
[0025] ,
[0026] in, , represents the weight matrix of the linear transformation, , is the corresponding bias, x represents any input variable;
[0027] S45. Obtain the final output feature sequence of the Transformer model through residual connection , the expression is:
[0028] ,
[0029] Therefore, the two remote sensing images input from different periods are , , the output feature sequences obtained after the Transformer model are , The Euclidean distance is further used to calculate the distance D between two feature sequences, which is expressed as: ,
[0030] Among them, yi represents the constituent elements of the feature sequence, i represents the index of the sequence length, and N represents the length of the feature sequence. The distance between the two feature sequences is obtained and compared with the preset threshold to obtain the result of change detection, thereby detecting the changes of remote sensing image objects.
[0031] Furthermore, the point cloud registration in step S2 is achieved by combining manual coarse registration with automatic fine registration.
[0032] Furthermore, the flight altitude of the UAV is set to 120m, the lateral overlap of the oblique photography is set to 75%, the heading overlap is set to 80%, and multiple sorties are used to complete the aerial photography work.
[0033] Furthermore, the sparse point cloud uses Context Capture to process the data images obtained by oblique photography and imports remote sensing image data, and generates a high-precision sparse point cloud through image matching and point cloud processing.
[0034] Furthermore, during close-range photogrammetry, the UAV's fine three-dimensional route planning, flight mode setting, shooting position, flight attitude and camera viewing angle parameters are set according to the initial three-dimensional information of the area to be measured.
[0035] Furthermore, the close-range image and the oblique image are subjected to feature point extraction and matching through a scale-invariant feature conversion algorithm, so as to perform point cloud registration and fusion modeling with the sparse point cloud.
[0036] Compared with the prior art, the present invention has the following beneficial effects:
[0037] The present invention complements the advantages of oblique photogrammetry technology and lidar scanning technology, combines remote sensing image change detection technology and uses close-range photogrammetry to take close-up photos of the changed area, thereby realizing refined modeling and dynamic updating of three-dimensional real scenes based on multi-source data fusion. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the process of the present invention; DETAILED DESCRIPTION
[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0040] like Figure 1 As shown, a 3D real scene modeling and dynamic updating method based on multi-source data fusion includes the following steps:
[0041] S1: Carry out aerial photography preparation according to the building height, vegetation and road distribution of the test area, use drones to obtain oblique images and remote sensing image data of the entire test area, incorporate the oblique images and remote sensing image data into a unified object coordinate system through aerial triangulation, obtain sparse point clouds based on the oblique images and remote sensing image data, perform closed-loop scanning of the test area through ground-based laser radar, obtain point cloud data while moving, and pre-process the point cloud data to obtain radar point clouds;
[0042] The generation process of sparse point cloud includes the following steps:
[0043] S101. According to the building height, vegetation and road distribution in the test area, a five-lens rotary wing drone is used to obtain an oblique image of the entire test area. In this embodiment, a CW-10 Dapeng electric vertical take-off and landing fixed-wing drone is used as a flight device, and an S5260 five-lens aerial survey camera is used to collect ground object information in real time.
[0044] S102. When performing aerial photography, according to the "Low-altitude Digital Aerial Photography Specifications", the flight altitude is set to 120m, the lateral overlap of oblique photography is set to 75%, the heading overlap is set to 80%, and multiple flights are used to complete the aerial photography work;
[0045] S103. In order to make up for the defect that oblique photogrammetry cannot obtain ground images, a backpack-type mobile laser scanner is used to perform closed-loop scanning of the area to be measured, and high-precision point cloud data is obtained during movement;
[0046] In step S101, the UAV oblique photogrammetry operation process includes:
[0047] S1011, Field Survey: Before field operations, conduct a field survey of the target area. This step helps to assess site conditions, identify potential flight obstacles and determine suitable take-off and landing points;
[0048] S1012, Airspace Application: Apply for a legal flight airspace permit based on the specific requirements of the flight mission, which involves complying with local and national aviation regulations and ensuring that the flight is conducted within legal airspace;
[0049] S1013, Plan Design: Design a flight plan based on the survey results and measurement requirements, including determining the flight altitude, flight path, shooting parameters, etc., to maximize the efficiency and quality of data collection;
[0050] S1014, Route planning: Use drone route planning software to design flight routes. These routes are usually based on the predetermined flight altitude and shooting requirements, and the optimal route is planned on the map software;
[0051] S1015, Aerial photography preparation: Prepare drones and oblique photography equipment, including installation of oblique cameras and pre-flight checks to ensure that the equipment is functioning properly and calibrated to ensure accuracy during data collection.
[0052] S1016, data collection: actual drone flight and oblique photography data collection, the drone will fly according to the predetermined route and shooting parameters, and the oblique photography equipment will continuously shoot oblique images of the ground.
[0053] S1017, Image preprocessing: Perform color uniformity processing on the acquired oblique images, including adjustment of hue, brightness and contrast, to improve the visual consistency and quality of the images.
[0054] S1018, model output. Use Context Capture to process the data images obtained by field oblique photography, import remote sensing image data, and generate a high-precision 3D model through image matching and point cloud processing;
[0055] Steps S1011-S1015 are the aerial photography preparation part;
[0056] The process of 3D modeling using Context Capture software includes:
[0057] S1019, Aerial triangulation: The main task of aerial triangulation is to extract and match feature points from aerial images, calculate the exterior orientation elements of the images by connecting these feature points, and finally incorporate all images into a unified object coordinate system. Multiple sets of oblique photography data and their POS (latitude and longitude) information are imported into the Context Capture software, and a pyramid matching strategy is adopted to match the same-name points on images of different levels, restore the relative position relationship between images, and calculate the exterior orientation elements of each image. High-density point cloud data (sparse point cloud) can be generated through dense matching of multi-view images.
[0058] S1020, 3D reconstruction: 3D reconstruction is performed using high-density point cloud data generated by aerial triangulation. The point cloud is kept dense, the 3D reconstruction range is determined, the block area is planned and the block center points are marked. The high-density point cloud data is then cut and divided into blocks to ensure that the vertices of the 3D mesh model are in the best position to improve the geometric accuracy and reduce model defects.
[0059] S1021. Construct TIN network: TIN network is the full name of irregular triangulated network. It divides the point cloud data of each block according to the pre-set priority order to form an irregular triangulated network. Use software optimization algorithms to delete or repair incorrectly matched triangulated network vertices to obtain ideal three-dimensional expression effects.
[0060] S1022, model generation: Based on the spatial position information of the TIN network vertices, texture is given to the model and the image with the best viewing angle is matched, and the image is pasted on the model surface to generate a high-precision real-scene three-dimensional model.
[0061] S1023, Model repair and editing: There may be reflection holes and other problems in the real-life 3D model, which may cause the model to be inconsistent with the actual situation of the area to be measured. Surveying and mapping personnel need to check the quality of the model, find out the holes, and use professional software such as MeshLab to repair and edit the model until all the holes are repaired;
[0062] The generation process of radar point cloud includes the following steps:
[0063] The equipment used to acquire ground-based laser radar data is the Leica BLK360 laser scanner. When acquiring field laser data, the first step is to determine the specific area to be measured that needs to be laser scanned. Then, based on the size and shape of the area to be measured and the technical requirements of the laser scanner, the appropriate measuring station location is selected for the layout of the measuring station.
[0064] The layout of the measuring stations should be able to effectively cover the entire area to be measured to ensure that the laser scanner can obtain complete ground data from different angles and positions. After the measuring stations are laid out, the laser scanner is used to collect data. After completing the data collection of all measuring stations, the collected radar point cloud data is processed and registered (preprocessing). The processing and registration include denoising, filtering and registration of the point cloud to ensure the accuracy and integrity of the final generated point cloud model;
[0065] S2: Perform point cloud registration on sparse point cloud and radar point cloud, and generate a real-scene 3D model by fusion modeling;
[0066] The point cloud registration and fusion modeling includes the following contents: making full use of the advantages and disadvantages of the two data sources of oblique image and ground radar, and performing point cloud registration on the processed sparse point cloud and the laser radar point cloud;
[0067] Point cloud registration is the process of converting point cloud data from different coordinate systems into a unified coordinate system, that is, determining the rotation, translation and scale transformation relationship between them. In the process of point cloud registration, in order to improve the registration accuracy, a combination of manual coarse registration and automatic fine registration ICP algorithm is generally used. Manual coarse registration usually estimates the initial transformation parameters manually or with prior knowledge to roughly align the point cloud data set. This rough alignment can help avoid the automatic fine registration ICP algorithm from falling into a local optimal solution due to initial position deviation. The automatic fine registration ICP algorithm can accurately calculate the optimal transformation between point cloud data through iterative optimization.
[0068] The basic principle of the automatic fine registration ICP algorithm is: first obtain the feature points of the two point sets, match the data according to the feature points, assume that these matching points are corresponding points, and solve the motion parameters according to the corresponding relationship, and finally use these parameters to convert the data. The objective function of the automatic fine registration ICP algorithm is The expression is as follows:
[0069] ,
[0070] Where B is the number of iterations, is a point in the reference point cloud, is the corresponding point in the target point cloud, R and T are the translation and rotation matrices. The iterative algorithm is used to solve the minimum error data and achieve the optimal parameter configuration of the algorithm;
[0071] S3: Re-aerial triangulation of the area to be measured and quality inspection of the real-scene 3D model are performed. If qualified, an encrypted point cloud is generated through aerial triangulation, and a 3D model is further generated based on the real-scene 3D model. If it does not meet the requirements, the UAV route is re-planned and the oblique photogrammetry of the area to be measured is re-performed;
[0072] The quality inspection of the real-life 3D model can conduct a quantitative quality evaluation of the real-life 3D modeling from three dimensions: boundary point accuracy, side length accuracy, and area accuracy. The 3D model effect can also be evaluated through visual judgment.
[0073] Boundary point accuracy refers to the accuracy and consistency of the boundary points of buildings or plots drawn during the modeling process. It is necessary to accurately measure and record the coordinates of key locations such as the walls and corners of the building. These coordinates are reference points for subsequent verification of the accuracy of the floor plan; measure the position of each boundary point in the floor plan, use RTK technology to obtain the actual measurement value, calculate the error of the boundary point position, and ensure that the error rate is less than 3% and the error is less than 0.03m.
[0074] Edge length accuracy refers to the difference between the length of the line segment or edge drawn in the modeling and the length of the corresponding line segment in the actual geographic space. It can be achieved by measuring the length of each segment of the building or plot boundary in the model and comparing it with the corresponding length in the field measurement. In the plane analysis stage, a certain number of edges can be randomly selected from the plan view of the area to be measured, and the corresponding photogrammetry results can be used as the basis for inspection. Generally, the error rate of edge length accuracy does not exceed 1%.
[0075] Area accuracy refers to the difference between the area of the building or plot drawn in the modeling and the area of the corresponding area in the actual geographic space. A certain number of buildings can be randomly selected from the floor plan of the area to be measured and the areas of the buildings can be measured on site to serve as the basic information for calculating the area accuracy.
[0076] Visual inspection: It is an intuitive and effective method to help evaluate the accuracy and integrity of real-life 3D models, especially in checking surface distortion, holes, texture splay, distortion and blur;
[0077] S4: Shoot the later images of the area to be measured according to the shooting cycle, and perform remote sensing change detection on the remote sensing period images at different times in the later images. If a change is detected, obtain the changed area and the corresponding coordinate information, and further use the drone close-up photography technology to locally reshoot the changed area. If no change is detected, there is no need to update the three-dimensional model;
[0078] The remote sensing change detection in step S4 comprises the following steps:
[0079] S41. Remote sensing images from different periods , are sent to the Transformer model respectively, where X (t1) and X (t2) are remote sensing images of two periods respectively, H and W represent the height and width of the image respectively, and C represents the number of channels of the image. Represents dimension information;
[0080] S42. The Transformer model converts each pixel or pixel block in the image into a feature vector sequence, and uses the self-attention mechanism of the Transformer model to calculate the global correlation of the feature vector sequence to obtain the attention matrix. The process can be expressed as:
[0081] ,
[0082] in, represents the self-attention mechanism, Q, K, and V represent the query, key, and value vectors obtained by linearly changing the triplet of the feature vector sequence. is the scaling factor, represents the transposition operation, and ssoftmax represents the softmax function. Here, the self-attention mechanism is expanded into the form of multi-head self-attention MultiHead, and the expression is as follows:
[0083] ,
[0084] ,
[0085] in, represents the number of attention heads, , , , Represents the weight parameter of linear transformation, Concat represents the feature concatenation operation, represents the i-th attention head.
[0086] S43. After obtaining the attention matrix, the output SA of the self-attention mechanism is obtained by performing a weighted sum operation on the input feature vector sequence. The expression is: ,
[0087] Where X is the remote sensing image input X at different times (t1) or X (t2) .
[0088] S44, the output SA of the self-attention mechanism is further processed by the feed-forward neural network FFN. The calculation process of the FFN network is: ,
[0089] in, , represents the weight matrix of the linear transformation, , is the corresponding bias, x represents any input variable;
[0090] S45. Obtain the final output feature sequence of the Transformer model through residual connection , the expression is: ,
[0091] Therefore, the two remote sensing images input from different periods are , , the output feature sequences obtained after the Transformer model are , , and further use the Euclidean distance to calculate the distance D between two feature sequences, which is expressed as:
[0092] ,
[0093] Among them, yi represents the constituent elements of the feature sequence, i represents the index of the number of iterations, and N represents the length of the feature sequence. According to the distance between the two feature sequences obtained, the change detection result is obtained by comparing it with the preset threshold, and then the change of the remote sensing object is detected.
[0094] During close-range photogrammetry, the drone's fine three-dimensional route planning, flight mode setting, shooting position, flight attitude, and camera viewing angle parameters are set according to the initial three-dimensional information of the area to be measured. The present invention uses a DJI Phantom 4 equipped with a Canon SLR camera for close-range photogrammetry;
[0095] The feature points of the drone close-up image and the early drone oblique image are extracted and matched for point cloud registration and fusion modeling with the sparse point cloud. The SIFT (Scale Invariant Feature Transform) algorithm is used here. The process is as follows: ① construct scale space; ② detect spatial extreme points; ③ accurately locate key points; ④ remove low-contrast and edge points; ⑤ determine the direction of key points; ⑥ generate key point descriptors to form a 128-dimensional feature vector; ⑦ calculate the Euclidean distance of feature points for feature matching;
[0096] S5: Generate an encrypted point cloud by performing aerial triangulation on the close-up images of the changed area and update the real-scene 3D model;
[0097] Further use of close-range images for aerial triangulation to regenerate encrypted point clouds, perform point cloud registration and fusion modeling with early radar point clouds and sparse point clouds, and update the real-scene 3D model to achieve dynamic updating of the real-scene 3D model.
[0098] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A three-dimensional real scene modeling and dynamic updating method based on multi-source data fusion, characterized in that: The steps include: S1: Carry out aerial photography preparation according to the building height, vegetation and road distribution of the test area, use drones to obtain oblique images and remote sensing image data of the entire test area, incorporate the oblique images and remote sensing image data into a unified object coordinate system through aerial triangulation, obtain sparse point clouds based on the oblique images and remote sensing image data, perform closed-loop scanning of the test area through ground-based laser radar, obtain point cloud data while moving, and pre-process the point cloud data to obtain radar point clouds; S2: Based on the point cloud registration of sparse point cloud and radar point cloud, the real scene 3D model is generated by fusion modeling; S3: Re-aerial triangulation of the area to be measured and quality inspection of the real-scene 3D model are performed. If qualified, an encrypted point cloud is generated through aerial triangulation, and a 3D model is further generated based on the real-scene 3D model. If it does not meet the requirements, the UAV route is re-planned and the oblique photogrammetry of the area to be measured is re-performed; S4: Shoot the later images of the area to be measured according to the shooting cycle, and perform remote sensing change detection on the remote sensing period images at different times in the later images. If a change is detected, obtain the changed area and the corresponding coordinate information, and further use the drone close-up photography technology to locally reshoot the changed area. If no change is detected, there is no need to update the three-dimensional model; The remote sensing change detection in step S4 comprises the following steps: S41. Remote sensing images from different periods , are sent to the Transformer model respectively, where and are remote sensing images of two periods respectively, H and W represent the height and width of the image respectively, and C represents the number of channels of the image. Represents dimension information; S42. The Transformer model converts each pixel or pixel block in the input image into a feature vector sequence, and uses the self-attention mechanism of the Transformer model to calculate the global correlation of the feature vector sequence to obtain the attention matrix. The process can be expressed as: , in, represents the self-attention mechanism, Q, K, and V represent the query, key, and value vectors obtained by linearly changing the triplet of the feature vector sequence. is the scaling factor, represents the transposition operation, softmax represents the softmax function, and the self-attention mechanism is expanded into the form of multi-head self-attention MultiHead. The expression is as follows: , , in, represents the number of attention heads, , , , Represents the weight parameter of linear transformation, Concat represents the feature concatenation operation, represents the i-th attention head; S43. After obtaining the attention matrix, the output SA of the self-attention mechanism is obtained by performing a weighted sum operation on the input feature vector sequence. The expression is: ; Where X is the remote sensing image input at different times or ; S44, the output SA of the self-attention mechanism is further processed by the feed-forward neural network FFN. The calculation process of the FFN network is: ; in, , represents the weight matrix of the linear transformation, , is the corresponding bias, x represents any input variable; S45. Obtain the final output feature sequence of the Transformer model through residual connection , the expression is: ; Therefore, the two remote sensing images input from different periods are , , the output feature sequences obtained after the Transformer model are , , and further use the Euclidean distance to calculate the distance D between two feature sequences, which is expressed as: , Among them, yi represents the constituent elements of the feature sequence, i represents the index of the sequence length, and N represents the length of the feature sequence. The distance between the two feature sequences is obtained and compared with the preset threshold to obtain the result of change detection, thereby detecting the change of remote sensing image objects; S5: Generate an encrypted point cloud by performing aerial triangulation on the close-up images of the changed area and update the real-scene 3D model.
2. The three-dimensional real scene modeling and dynamic updating method based on multi-source data fusion according to claim 1 is characterized in that: The real-scene 3D model quality detection in step S3 performs a quantitative quality evaluation on the real-scene 3D modeling based on three dimensions: boundary point accuracy, side length accuracy, and area accuracy.
3. The method for three-dimensional real scene modeling and dynamic updating based on multi-source data fusion according to claim 1, characterized in that: The point cloud registration in step S2 is achieved by combining manual coarse registration and automatic fine registration.
4. The method for 3D real scene modeling and dynamic updating based on multi-source data fusion according to claim 1, characterized in that: The UAV flight altitude is set to 120m, the lateral overlap of oblique photography is set to 75%, the heading overlap is set to 80%, and multiple sorties are used to complete the aerial photography work.
5. The three-dimensional real scene modeling and dynamic updating method based on multi-source data fusion according to claim 1 is characterized in that: The sparse point cloud uses Context Capture to process the data images obtained by oblique photography and imports remote sensing image data, and generates a high-precision sparse point cloud through image matching and point cloud processing.
6. The method for 3D real scene modeling and dynamic updating based on multi-source data fusion according to claim 1, characterized in that: During close-range photogrammetry, the UAV's detailed three-dimensional route planning, flight mode setting, shooting position, flight attitude and camera viewing angle parameters are set according to the initial three-dimensional information of the area to be measured.
7. The method for 3D real scene modeling and dynamic updating based on multi-source data fusion according to claim 1, characterized in that: The close-range image and the oblique image are subjected to feature point extraction and matching through a scale-invariant feature conversion algorithm, so as to perform point cloud registration and fusion modeling with the sparse point cloud.
