CN115719407A - Distributed multi-view stereo reconstruction method for large-scale aerial images - Google Patents
Distributed multi-view stereo reconstruction method for large-scale aerial images Download PDFInfo
- Publication number
- CN115719407A CN115719407A CN202310011438.9A CN202310011438A CN115719407A CN 115719407 A CN115719407 A CN 115719407A CN 202310011438 A CN202310011438 A CN 202310011438A CN 115719407 A CN115719407 A CN 115719407A
- Authority
- CN
- China
- Prior art keywords
- images
- point cloud
- calculating
- image
- points
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Image Processing (AREA)
Abstract
The invention discloses a distributed multi-view stereo reconstruction method facing a large-scale aerial image. The method makes full use of the regionality among large-scale aerial images, converts the multi-view stereo reconstruction problem of a large-scale scene into a small-scale multi-view stereo reconstruction problem which can be solved on a low-performance computer, improves the time efficiency of three-dimensional reconstruction, and reduces the cost of three-dimensional reconstruction.
Description
Technical Field
The invention relates to computer vision and image processing technology, in particular to a distributed multi-view stereo reconstruction method for large-scale aerial images.
Background
Multi-view Stereo (MVS) is a technique for computing a scene dense point cloud model from image data, usually with the output information of a Structure from Motion (SfM), i.e., a sparse point cloud model and camera parameters, as input information to the MVS. Currently, great research progress is made on the multi-view stereo reconstruction problem of small-scale image data (e.g., small-scale scene image data acquired by a handheld camera), however, for a large-scale outdoor scene, the existing multi-view stereo reconstruction method needs to be further improved. In addition, with the popularization of consumer-grade unmanned aerial vehicle equipment, large-scale data of outdoor scenes can be acquired easily. The existing multi-view stereo reconstruction method mainly has the following challenges when processing large-scale aerial image data: (a) The method is very time-consuming, the multi-view stereo reconstruction process is very time-consuming, and particularly when outdoor large-scale aerial image data are processed, the existing multi-view stereo reconstruction method cannot calculate a dense point cloud model within limited time and is difficult to meet the timeliness requirement of a high-level application system; (b) The storage overflows, the multi-view stereo reconstruction method has a large demand on a computer memory, and particularly when the data volume of an aerial image is large, the single-version multi-view stereo reconstruction method has the problem of memory overflow, so that the three-dimensional reconstruction process fails.
The above problems seriously hinder the development and application of the multi-view stereo reconstruction technology, and expose the shortcomings of the single-machine version multi-view stereo reconstruction method in processing large-scale aerial image data. Therefore, a distributed multi-view stereo reconstruction method oriented to large-scale aerial images is needed, so that dense point cloud models of scenes can be rapidly calculated from the large-scale aerial images.
At present, the classical papers on multi-view stereo research are mainly: [1] accurate, dense, and Robust Multi-View Steropsis, [2] Pixel View Selection for Unstructured Multi-View Stereo, [3] BlendedMVS: A large-scale dataset for generated Multi-View Stereo networks. A paper [1] published in 2007 on a CVPR conference is a multi-view stereo reconstruction method based on seed point diffusion; paper [2] published in the ECCV conference in 2016, is a multi-view stereo reconstruction method based on image block matching; paper [3] was published in CVPR in 2020, and is a multi-view stereo reconstruction method based on deep learning, and the depth map of each image is estimated mainly by using a deep learning technique. The important point of the multi-view stereo reconstruction methods is how to improve the precision of a multi-view stereo reconstruction model (three-dimensional dense point cloud), and the processed targets are image data of small-scale scenes.
Therefore, when the existing multi-view stereo reconstruction method is applied to large-scale aerial image data, the following challenges still face: (1) When the existing single-machine version multi-view stereo reconstruction method is used for processing large-scale aerial image data, a larger memory space is needed, for example, the content space needed for processing 1000 pieces of image data is 64 Gb, and even exceeds the maximum memory space supported by the existing hardware equipment, for example, when the image data reaches 1500 pieces, the memory space of 128Gb is needed, and the maximum memory space range supported by a single computer is far exceeded; (2) The existing multi-view stereo reconstruction method has too low operating efficiency, and is difficult to meet the time efficiency requirement of large-scale three-dimensional reconstruction based on aerial image data, for example, the time for processing 1000 aerial images takes 10 days.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to solve the defects in the prior art, provides a distributed multi-view stereo reconstruction method for large-scale aerial images,
the dense point cloud model of the scene is rapidly calculated from the large-scale aerial image data on the basis of the distributed operation environment, the progress of the multi-view three-dimensional reconstruction technology for the large-scale aerial image is promoted, and the purpose of rapidly calculating the high-quality dense point cloud of the large-scale scene is achieved.
The technical scheme is as follows: the invention discloses a distributed multi-view stereo reconstruction method for large-scale aerial images, which comprises the following steps of:
s1, for a given large-scale aerial image data set,wherein, in the step (A),representing the number of aerial images, and calculating a sparse point cloud model and camera parameters of a corresponding scene:
the sparse point cloud model is SWhereinThe number of three-dimensional points in the sparse point cloud representing the entire scene,denotes the firstiA three-dimensional pointThe position in the world coordinate system is,a serial number representing a three-dimensional point;
camera parameters CWherein, in the step (A),indicating the number of aerial images to be taken,denotes the firstThe internal parameter matrix of the individual cameras,denotes the firstThe rotation matrix of the individual cameras is,is shown asThe translation vector of each of the cameras is,a serial number representing a camera;
s2, dividing the large-scale sparse point cloud model S into different sub-regions to obtain SIn whichThe number of sub-areas is indicated,a serial number representing a three-dimensional point,is shown asA sub-region;
dividing camera parameters C intoObtaining the camera parameters C of each subarea corresponding to the sparse point cloud model S;
S3, calculating a depth map of the image in each area, and recording the areaTherein is provided withTaking aerial images, areasThe corresponding depth image is,(ii) a Wherein the content of the first and second substances,is shown asA number of the amplitude depth map;
s4, for each regionOf the depth image dataSelecting two optimal initial fusion views, and recording asIn whichIndicating areaNumber of aerial images inThe amount of the (B) component (A),andis a subscript, used to distinguish different depth images;andrepresenting an optimal initial fusion view;
s5, fusion regionAll depth images in (1)I.e. obtaining the regionCorresponding dense point cloud model, noteWherein, in the process,representing the number of three-dimensional points in the dense cloud after the depth map in the sub-region is fused,indicating pointsA position in a world coordinate system;
s6, modeling the dense point cloud of each areaAre combined into a wholeThe dense point cloud model of the complete scene can be obtained and recorded asWherein, in the process,indicating the number of sub-regions.
Further, in step S1, a hybrid motion inference structure method is used to extract the aerial image data set from the aerial image data setCalculating sparse point cloud model and camera parameters of a scene, and specifically comprising the following steps:
step S1.1, aerial image matching
Firstly, detecting feature points and calculating feature descriptors by using a local feature detection method based on deep learning, then calculating a matching relation between the feature descriptors by using a local perception hash method, and finally eliminating wrong matching points according to geometric consistency between images to obtain correct feature matching points;
step S1.2, calculating camera parameters
Calculating the parameters of the camera by using an incremental motion inference structure method according to the feature matching points obtained in the S1.1; firstly, calculating relative attitude information of a camera by using a five-point algorithm, then calculating absolute attitude information of the camera by using a three-point method, and finally calculating focal length information of each image by using a camera self-calibration method;
s1.3, calculating a sparse point cloud model of the regional image
And calculating a sparse point cloud model of the area scene by using a global motion inference structure method according to the feature matching points obtained in the step S1.1 and the camera parameters obtained in the step S1.2, so that the time efficiency of three-dimensional reconstruction is improved. First, the region is dividedThe three-dimensional points corresponding to all the images are registered in a world coordinate system, and then Bundle adjustment (Bundle A) is useddjustment) method simultaneously optimizes the camera parameters and three-dimensional points in the world coordinate system until convergence, and then an accurate sparse point cloud model can be obtained; wherein, the first and the second end of the pipe are connected with each other,is shown asThe serial number of each region.
Further, in step S2, a dominant set clustering method is used to divide the large-scale sparse point cloud into a plurality of sub-region scenes, and the specific method is as follows:
note the bookRepresent an inclusionAerial image and sparse point cloud modelThe set of (a) and (b),indicates a device havingRows and columnsA square matrix of columns for recording similarity between images;andrespectively representing imagesAnd imageSet of all three-dimensional points that can be observed, imageAnd an imageThe similarity between them is defined as:
wherein, the first and the second end of the pipe are connected with each other,representing a vectorSum vectorThe angle between the two (C) and the angle between the two (C) are determined,the calculation method of (2) is as follows:
wherein the content of the first and second substances,andare all intermediate calculation variables;
to this end, a similarity matrix between the images is calculatedAccording toConstruct a graph structure from the values ofWhereinThe position of the vertex is represented and,representing an edge;
note the bookIs represented by containingA vector of elements, then arbitraryTime of day, vectorThe values of each element in (a) are as follows:
wherein the content of the first and second substances,subscripts of (2)To representTime of day vectorTo (1) aThe number of the components is such that,to representTransposing;
therefore, the temperature of the molten metal is controlled,time of day andtime of day, vectorThe error between is:
if it is usedAnd terminating the iterative calculation process, namely dividing the large-scale sparse point cloud model into a plurality of sub-area sparse point cloud models.
Further, in the step S3, a stereo matching method based on image blocks is used to calculate a corresponding high-quality depth map for each image, so as to lay a foundation for reconstructing a high-quality dense point cloud model. The detailed calculation steps are as follows:
s3.1, performing parallax propagation by using an odd-even iteration test, wherein the strategy comprises space propagation, visual angle propagation and time propagation;
s3.2, when the space transmission is respectively iterated for odd and even times, the transmission directions respectively start from the upper left corner and the lower right corner, cost comparison is carried out on each pixel point and the parallax planes of the left pixel and the right pixel, and matching cost of the parallax of the upper left corner and the lower right corner and the current parallax value is calculated;
s3.3, finally, taking the parallax value with the minimum substitution value as the optimal parallax value;
and step S3.4, iterating and calculating the steps S3.1 to S3.4 until the optimal parallax values of all pixels are calculated, namely obtaining the optimal depth image.
Further, in the step S4, a multiple constraint method is used to select two optimal initial fusion images for each sub-region, so as to ensure the geometric consistency between the generated dense point cloud and the real scene; the detailed calculation steps are as follows:
s4.1, calculating the characteristic matching points meeting the homography matrix constraint between any two images in the input images by using a homography matrix constraint method, and recording the characteristic matching points as;
S4.2, calculating the characteristic matching points meeting the basic matrix constraint relation between any two images in the input images according to the basic matrix constraint relation, and recording the characteristic matching points as;
S4.3, calculating characteristic matching points meeting the intrinsic matrix constraint relation between any two images in the input images according to the intrinsic matrix constraint relation, and recording the characteristic matching points as;
Step S4.4, matching point is takenMatching pointAnd matching pointThe intersection between them to obtain the matching point;
Step S4.5, matching points are satisfiedAnd (5) taking the two images with the least error points as initial fusion images.
Further, in the step S5, the depth images in the region are fused together by fully utilizing the normal vector information of the depth map, so as to obtain a dense point cloud model corresponding to the image in the region, and the detailed steps are as follows:
s5.1, calculating the confidence coefficient of each vertex of the depth image to be fused;
s5.2, deleting some redundant overlapped points from the depth image to be fused according to the confidence coefficient of each vertex in the depth image to obtain the topological information of each area image in each depth image;
s5.3, carrying out weighting operation on the vertexes on the depth image according to the topological information to obtain the geometric information of the image;
and S5.4, stitching the area images according to the topological information and the geometric information so as to obtain a dense point cloud model of the corresponding area. Further, in step S6, a global iterative nearest neighbor method is used to merge a plurality of dense point cloud models with feature overlap into a dense point cloud model of a complete scene, and the detailed steps are as follows:
Step S6.2, finding out origin cloudCorresponding point set in (1)And is andso thatAndthe distance betweenMinimum;
Step S6.4, use of rotation matrixAnd translation vectorTo pairThe points are subjected to rotation transformation and translation transformation, and a new point set is calculated;
Step S6.5, calculating point setAnd point setAverage distance therebetween,Representing the number of three-dimensional points in the point set;
step S6.6, ifIf the number of iterations is less than the preset threshold or greater than the preset number of iterations, the calculation process is terminated, otherwise, the step S6.2 is returned until the calculation process converges.
Has the beneficial effects that: compared with the prior art, the invention has the following advantages:
(1) According to the method, the sparse point cloud model of the large-scale scene is divided into different subset areas, so that the problems of memory overflow and system crash of a single-version multi-view stereo reconstruction method are avoided, and the large-scale multi-view stereo reconstruction is possible.
(2) According to the method, on different nodes of a distributed system, sparse point cloud models and camera parameters (including camera focal length, rotation matrix and translation vector) of different sub-areas are independently processed, the dense point cloud models are rapidly calculated, and the time efficiency of large-scale multi-view three-dimensional reconstruction is improved.
(3) The method not only can solve the problem of memory space overflow of a single-version multi-view stereo reconstruction method, but also can improve the time efficiency of large-scale multi-view stereo reconstruction, and lays an important foundation for the application of unmanned aerial vehicle aerial images in the field of three-dimensional reconstruction and the development and application of three-dimensional reconstruction technology.
Drawings
FIG. 1 is a schematic overall flow diagram of the present invention;
FIG. 2 is an aerial image dataset in an embodiment;
FIG. 3 is a large-scale sparse point cloud model in an embodiment;
FIG. 4 is a sparse point cloud model and camera pose information for a subregion in an embodiment;
FIG. 5 is an original image (non-depth image) of a primary fused view of a sub-region in an embodiment;
FIG. 6 is a dense point cloud model of a sub-region in an embodiment;
FIG. 7 is a further prior art dense point cloud model;
FIG. 8 is a block diagram of a complete dense point cloud model according to an embodiment of the present invention.
Detailed Description
The technical solution of the present invention is described in detail below, but the scope of the present invention is not limited to the embodiments.
The invention discloses a distributed multi-view stereo reconstruction method for large-scale aerial images, which aims to reconstruct scene dense point cloud from the large-scale aerial images, and has the following core idea: firstly, dividing a sparse point cloud model corresponding to a large-scale aerial image into a plurality of sub-regions; secondly, distributing the aerial images, the sparse point cloud models and the camera parameters corresponding to the sub-regions on different nodes of a distributed environment for processing, and calculating a depth image of each image; then, fusing depth images in the sub-regions on different nodes of the distributed environment, so as to obtain dense point cloud models of the sub-regions; finally, merging the dense point clouds on the sub-nodes on a main control machine in the distributed environment, so as to obtain a dense point cloud model of a complete scene; finally, the method makes it possible to quickly calculate a high-quality dense point cloud model from large-scale aerial image data.
As shown in fig. 1, the distributed multi-view stereo reconstruction method for large-scale aerial images of the present invention includes the following steps:
s1, for a given large-scale aerial image data set,wherein, in the step (A),representing the number of aerial images, and calculating a sparse point cloud model and camera parameters of a corresponding scene:
the sparse point cloud model is SWhereinThe number of three-dimensional points is represented,indicating pointsThe position in the world coordinate system is,inSubscriptA serial number representing a three-dimensional point;
camera parameters CWherein, in the process,indicating the number of aerial images to be taken,denotes the firstThe internal parameter matrix of the individual cameras,is shown asThe rotation matrix of the individual cameras is,is shown asThe translation vector of the individual cameras is,a serial number indicating a camera;
s1.1, detecting feature points and calculating feature descriptors by using a SuperPoint method based on deep learning, then calculating a matching relation between the feature descriptors by using a local perception Hash method, and finally eliminating wrong matching points according to geometric consistency between images to obtain correct feature matching points;
s1.2, calculating camera parameters by using an incremental motion inference structure method according to the feature matching points obtained in the S1.1; firstly, calculating relative attitude information of a camera by using a five-point algorithm, then calculating absolute attitude information of the camera by using a three-point method, and finally calculating focal length information of each image by using a camera self-calibration method;
and S1.3, calculating a sparse point cloud model of the area scene by using a global motion inference structure method according to the feature matching points obtained in the S1.1 and the camera parameters obtained in the S1.2, and improving the time efficiency of three-dimensional reconstruction. First, the region is dividedRegistering three-dimensional points corresponding to all the images in a world coordinate system, and then optimizing camera parameters and the three-dimensional points in the world coordinate system by using a Bundle Adjustment (Bundle Adjustment) method until convergence to obtain an accurate sparse point cloud model; wherein the content of the first and second substances,is shown asThe serial number of each region;
s2, dividing the large-scale sparse point cloud model S into different areas to obtain SWherein, in the process,representing the number of three-dimensional points in the sparse point cloud throughout the scene,indicating pointsThe position in the world coordinate system is such that,indicating the number of sub-regions,A number representing a three-dimensional point is indicated,is shown asA sub-region;
dividing camera parameters C intoDifferent regions, obtaining camera parameters C of each regionWherein, in the step (A),indicating the number of aerial images to be taken,is shown asThe internal parameter matrix of the individual cameras,is shown asThe rotation matrix of the individual cameras is,is shown asThe translation vector of the individual cameras is,representing the number of sub-regions;
dividing a large-scale sparse point cloud into a plurality of sub-region scenes by using a dominance set clustering method, wherein the specific method comprises the following steps:
note the bookShow a package ofAerial image and sparse point cloud modelThe set of (a) and (b),indicates a device havingRows and columnsA square matrix of columns for recording similarity between images;andrespectively representing imagesAnd imageSet of all three-dimensional points that can be observed, imageAnd imageIn betweenThe similarity is defined as:
wherein the content of the first and second substances,representing a vectorSum vectorThe angle between the two (C) and the angle between the two (C) are determined,the calculation method of (2) is as follows:
wherein the content of the first and second substances,andare all intermediate calculation variables;
to this end, a similarity matrix between the images is calculatedAccording toValue of (a) to construct a graph structureWhereinThe position of the vertex is represented and,representing an edge;
note bookIs represented by containingA vector of elements, then arbitraryTime of day, vectorThe values of each element in (a) are as follows:
wherein the content of the first and second substances,subscripts of (2)To representTime of day vectorTo (1)The number of the components is such that,to representTransposing;
therefore, the temperature of the molten metal is controlled,time of day andtime of day, vectorThe error between is:
wherein, the first and the second end of the pipe are connected with each other,representTransposing;
if it is usedIf the iterative computation process is ended, the large-scale sparse point cloud model can be divided into sparse point cloud models of a plurality of sub-areas;
s3, calculating a depth map of the image in each area, and recording the areaIs provided with thereinTaking aerial images, areasThe corresponding depth image is,(ii) a Calculating a corresponding high-quality depth map for each image by using a stereo matching method based on image blocks, and laying a foundation for reconstructing a high-quality dense point cloud model; the specific process is as follows:
s3.1, performing parallax propagation by using an odd-even iteration test, wherein the strategy comprises space propagation, visual angle propagation and time propagation;
s3.2, when the space transmission is respectively iterated for odd and even times, the transmission directions respectively start from the upper left corner and the lower right corner, cost comparison is carried out on each pixel point and the parallax planes of the left pixel and the right pixel, and the matching cost of the parallax of the upper left corner and the lower right corner and the current parallax value is calculated;
s3.3, finally, taking the parallax value with the minimum substitution value as the optimal parallax value;
step S3.4, iterative computation step S3.1 to step S3.4, until the optimal parallax value of all pixels is computed, the optimal depth image can be obtained
S4, for each regionDepth image data ofSelecting two optimal initial fusion views, and recording asWhereinIndicating areaThe number of aerial images in the interior,andis a subscript, used to distinguish different depth images; selecting two optimal initial fusion images for each sub-region by using a multiple constraint method, and ensuring the geometric consistency of the generated dense point cloud and a real scene; the specific process is as follows:
s4.1, calculating the characteristic matching points meeting the homography matrix constraint between any two images in the input images by using a homography matrix constraint method, and recording the characteristic matching points as;
S4.2, calculating the characteristic matching points meeting the basic matrix constraint relation between any two images in the input images according to the basic matrix constraint relation, and recording the characteristic matching points as;
S4.3, calculating characteristic matching points meeting the intrinsic matrix constraint relation between any two images in the input images according to the intrinsic matrix constraint relation, and recording the characteristic matching points as;
Step S4.4, matching point is takenMatching pointAnd matching pointThe intersection between them to obtain the matching point;
Step S4.5, matching points are satisfiedTwo images which are constrained and have the least error points are used as initial fusion images;
s5, fusion regionAll depth images in (1)To obtain the regionCorresponding dense point cloud model, noteWherein, in the step (A),representing the number of three-dimensional points in the dense cloud generated by the depth map fusion,indicating pointsA location in a world coordinate system; the method comprises the following steps of fully utilizing normal vector information of a depth map to fuse depth images in a region into a whole to obtain a dense point cloud model corresponding to the images in the region, wherein the specific process comprises the following steps:
s5.1, calculating the confidence coefficient of each vertex of the depth image to be fused;
s5.2, deleting some redundant overlapped points from the depth image to be fused according to the confidence of each vertex in the depth image to obtain the topological information of each region image in each depth image;
s5.3, carrying out weighting operation on the vertex on the depth image according to the topological information to obtain the geometric information of the image;
s5.4, stitching the area images according to the topological information and the geometric information so as to obtain dense point cloud models of corresponding areas;
s6, modeling the dense point cloud of each areaCombining them into a whole body to obtain complete dense point cloud model, and marking it asWherein, in the step (A),representing the number of sub-regions; combining a plurality of dense point cloud models with feature overlapping into a dense point cloud model of a complete scene by using a global iterative nearest neighbor method, wherein the specific process is as follows:
step S6.2, finding out origin cloudCorresponding point set in (1)And is andso thatAndthe distance betweenMinimum;
Step S6.4, using the rotation matrixAnd translation vectorFor is toThe points are subjected to rotation transformation and translation transformation, and a new point set is calculated;
Step S6.5, calculating point setAnd point setAverage distance between,Representing the number of three-dimensional points in the point set;
step S6.6, ifIf the value is less than the preset threshold value or greater than the preset iteration number, the calculation process is terminated, otherwise, the step S6.2 is returned until the calculation process converges.
Example 1:
the original aerial image of the embodiment is shown in fig. 2, the final reconstruction result of the embodiment is shown in fig. 8, and the dense point cloud model reconstructed from the large-scale aerial image has high geometric consistency with the real scene.
As can be seen from the above embodiments, the sparse point cloud model of the complete scene is calculated firstly (as shown in fig. 3), then the large-scale sparse point cloud model is divided into sparse point cloud models of a plurality of sub-regions (as shown in fig. 4), then the depth map of the image in each sub-region is calculated, and an initial fusion view is selected for each sub-region (as shown in fig. 5, the initial fusion view is an original image (non-depth image)) of the initial fusion view), and fourth, the depth images in the sub-regions are fused to obtain dense point cloud models corresponding to the sub-regions (as shown in fig. 6); and finally, combining the dense point clouds of the sub-regions into a whole to obtain a dense point cloud model of the complete scene (as shown in fig. 8). In addition, as can be seen from fig. 7, the dense point cloud model calculated by other prior art schemes has a large difference from the real scene in terms of geometric consistency.
In addition, the technical scheme of the embodiment only needs to use 1000 aerial images for 8 hours and only occupies 24Gb memory space, that is, the time efficiency of large-scale multi-view stereo reconstruction can be improved, and the problem of memory overflow can be avoided.
The invention can also be applied to other fields, such as the Yuan universe, digital Chinese construction, digital country construction, digital city construction, military simulation, unmanned driving, automatic navigation under the condition of no satellite, digital protection of cultural heritage, three-dimensional scene monitoring, shooting and manufacturing of large-scale movies and televisions, three-dimensional surveying of natural disaster sites, three-dimensional visual science popularization creation, virtual reality and augmented reality.
Claims (7)
1. A distributed multi-view stereo reconstruction method for large-scale aerial images is characterized by comprising the following steps: the method comprises the following steps:
s1, for a given large-scale aerial image data set,wherein, in the step (A),representing the number of aerial images, and calculating a sparse point cloud model and camera parameters of a corresponding scene:
the sparse point cloud model is SIn whichThe number of three-dimensional points in the sparse point cloud representing the entire scene,is shown asiA three-dimensional pointThe position in the world coordinate system is,a serial number representing a three-dimensional point;
camera parameters CWhereinIndicating the number of aerial images to be taken,is shown asThe internal parameter matrix of the individual cameras,is shown asThe rotation matrix of the individual cameras is,is shown asThe translation vector of each of the cameras is,a serial number indicating a camera;
s2, dividing the large-scale sparse point cloud model S into different sub-regions to obtain SWherein, in the process,the number of sub-regions is indicated,representing three-dimensional pointsThe serial number of (a) is included,denotes the firstA sub-region;
dividing camera parameters C intoSub-regions corresponding to the sparse point cloud model S, and obtaining camera parameters C of each sub-region;
S3, calculating a depth map of the image in each area, and recording the areaTherein is provided withTaking aerial images, i.e. areasThe corresponding depth image is,(ii) a Wherein, the first and the second end of the pipe are connected with each other,denotes the firstA number of the depth map;
s4, for each regionDepth image data ofSelecting two optimal initial fusion views, and recording asWhereinIndicating areaThe number of aerial images in the interior,andis a subscript, used to distinguish different depth images;andrepresenting an optimal initial fused view;
s5, fusion regionAll depth images in (1)To obtain the regionCorresponding dense point cloud model, noteWherein, in the step (A),representing the number of three-dimensional points in the dense cloud after the depth map in the sub-region is fused,indicating pointsA position in a world coordinate system;
2. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in step S1, a hybrid motion inference structure method is used to extract a set of aerial image data from a set of aerial image dataCalculating sparse point cloud model and camera parameters of a scene, and specifically comprising the following steps:
step S1.1, aerial image matching
Firstly, detecting feature points and calculating feature descriptors by using a local feature method based on deep learning, then calculating a matching relation between the feature descriptors by using a local perception hash method, and finally eliminating wrong matching points according to geometric consistency between images to obtain correct feature matching points;
step S1.2, calculating camera parameters
Calculating camera parameters by using an incremental motion inference structure method according to the feature matching points obtained in the S1.1; firstly, calculating relative attitude information of a camera by using a five-point algorithm, then calculating absolute attitude information of the camera by using a three-point method, and finally calculating focal length information of each image by using a camera self-calibration method;
s1.3, calculating a sparse point cloud model of the regional image
And (3) according to the feature matching points obtained in the step (S1.1) and the camera parameters obtained in the step (S1.2), calculating a sparse point cloud model of the area scene by using a global motion inference structure method:
first, the region is dividedRegistering three-dimensional points corresponding to all the images in a world coordinate system, and then optimizing camera parameters and the three-dimensional points in the world coordinate system by using a bundling adjustment method until convergence, so as to obtain an accurate sparse point cloud model; wherein, the first and the second end of the pipe are connected with each other,is shown asThe serial number of each region.
3. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in the step S2, the dominant set clustering method is used to divide the sub-regions, and the specific method is as follows:
note the bookShow a package ofAerial image and sparse point cloud modelThe set of (a) and (b),indicates a device havingRows and columnsA square matrix of columns for recording similarity between images;andrespectively representing imagesAnd imageSet of all three-dimensional points that can be observed, then imageAnd an imageThe similarity between them is defined as:
wherein, the first and the second end of the pipe are connected with each other,representing a vectorSum vectorThe angle between the two (C) and the angle between the two (C) are determined,the calculation method of (2) is as follows:
wherein the content of the first and second substances,andare all intermediate calculation variables;
to this end, a similarity matrix between the images is calculatedAccording toValue of (a) to construct a graph structureWhereinThe position of the vertex is represented and,representing an edge;
note bookIs expressed as containingA vector of elements, then arbitraryTime of day, vectorThe values of each element in (a) are as follows:
wherein the content of the first and second substances,subscript ofRepresentTime of day vectorTo (1)The number of the components is such that,to representTransposing;
therefore, the temperature of the molten metal is controlled,time of day andtime and vectorThe error between is:
4. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in step S3, a stereo matching method based on image blocks is used to calculate a corresponding high-quality depth map for each image, and the detailed calculation steps are as follows:
s3.1, performing parallax propagation by using an odd-even iteration test, wherein the strategy comprises space propagation, visual angle propagation and time propagation;
s3.2, when the space transmission is respectively iterated for odd and even times, the transmission directions respectively start from the upper left corner and the lower right corner, cost comparison is carried out on each pixel point and the parallax planes of the left pixel and the right pixel, and the matching cost of the parallax of the upper left corner and the lower right corner and the current parallax value is calculated;
s3.3, finally, taking the parallax value with the minimum substitution value as the optimal parallax value;
and step S3.4, iterating and calculating the steps S3.1 to S3.4 until the optimal parallax values of all pixels are calculated, namely obtaining the optimal depth image.
5. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in the step S4, two optimal initial fusion images are selected for each sub-region by using a multiple constraint method, and the detailed calculation steps are as follows:
s4.1, calculating the characteristic matching points meeting the homography matrix constraint between any two images in the input images by using a homography matrix constraint method, and recording the characteristic matching points as;
S4.2, calculating the characteristic matching points meeting the basic matrix constraint relation between any two images in the input images according to the basic matrix constraint relation, and recording the characteristic matching points as;
S4.3, calculating the feature matching points meeting the intrinsic matrix constraint relation between any two images in the input images according to the intrinsic matrix constraint relation, and recording the feature matching points as;
Step S4.4, matching point is takenMatching pointsAnd matching pointThe intersection between them to obtain the matching point;
6. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in the step S5, the depth images in the region are fused into a whole by fully utilizing the normal vector information of the depth image, and a dense point cloud model corresponding to the images in the region is obtained; the detailed steps are as follows:
s5.1, calculating the confidence coefficient of each vertex of the depth image to be fused;
s5.2, deleting some redundant overlapped points from the depth image to be fused according to the confidence of each vertex in the depth image to obtain the topological information of each region image in each depth image;
s5.3, carrying out weighting operation on the vertexes on the depth image according to the topological information to obtain the geometric information of the image;
and S5.4, stitching the area images according to the topological information and the geometric information so as to obtain a dense point cloud model of the corresponding area.
7. The distributed multi-view stereo reconstruction method for large-scale aerial images according to claim 1, characterized in that: in the step S6, combining a plurality of dense point cloud models with feature overlapping into a dense point cloud model of a complete scene by using a global iterative nearest neighbor method; the detailed steps are as follows:
step S6.2, finding out origin cloudCorresponding point set in (1)And is andso thatAnda distance therebetweenMinimum;
Step S6.4, using the rotation matrixAnd translation vectorTo pairThe points are subjected to rotation transformation and translation transformation, and a new point set is calculated;
Step S6.5, calculating point setAnd point setAverage distance between, Representing the number of three-dimensional points in the point set;
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310011438.9A CN115719407B (en) | 2023-01-05 | 2023-01-05 | Large-scale aerial image-oriented distributed multi-view three-dimensional reconstruction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310011438.9A CN115719407B (en) | 2023-01-05 | 2023-01-05 | Large-scale aerial image-oriented distributed multi-view three-dimensional reconstruction method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN115719407A true CN115719407A (en) | 2023-02-28 |
CN115719407B CN115719407B (en) | 2023-06-27 |
Family
ID=85257835
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310011438.9A Active CN115719407B (en) | 2023-01-05 | 2023-01-05 | Large-scale aerial image-oriented distributed multi-view three-dimensional reconstruction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115719407B (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116071504A (en) * | 2023-03-06 | 2023-05-05 | 安徽大学 | Multi-view three-dimensional reconstruction method for high-resolution image |
CN116805355A (en) * | 2023-08-25 | 2023-09-26 | 安徽大学 | Multi-view three-dimensional reconstruction method for resisting scene shielding |
CN116993925A (en) * | 2023-09-25 | 2023-11-03 | 安徽大学 | Distributed bundling adjustment method for large-scale three-dimensional reconstruction |
CN117408999A (en) * | 2023-12-13 | 2024-01-16 | 安格利(成都)仪器设备有限公司 | Method for automatically detecting corrosion pits of containers and pipelines by utilizing point cloud complement |
CN117437363A (en) * | 2023-12-20 | 2024-01-23 | 安徽大学 | Large-scale multi-view stereoscopic method based on depth perception iterator |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108416840A (en) * | 2018-03-14 | 2018-08-17 | 大连理工大学 | A kind of dense method for reconstructing of three-dimensional scenic based on monocular camera |
CN111968218A (en) * | 2020-07-21 | 2020-11-20 | 电子科技大学 | Three-dimensional reconstruction algorithm parallelization method based on GPU cluster |
CN112085845A (en) * | 2020-09-11 | 2020-12-15 | 中国人民解放军军事科学院国防科技创新研究院 | Outdoor scene rapid three-dimensional reconstruction device based on unmanned aerial vehicle image |
US20210183080A1 (en) * | 2019-12-13 | 2021-06-17 | Reconstruct Inc. | Interior photographic documentation of architectural and industrial environments using 360 panoramic videos |
CN113284227A (en) * | 2021-05-14 | 2021-08-20 | 安徽大学 | Distributed motion inference structure method for large-scale aerial images |
WO2021185322A1 (en) * | 2020-03-18 | 2021-09-23 | 广州极飞科技有限公司 | Image processing method and related device |
US20210358206A1 (en) * | 2020-05-14 | 2021-11-18 | Star Institute Of Intelligent Systems | Unmanned aerial vehicle navigation map construction system and method based on three-dimensional image reconstruction technology |
CN115205489A (en) * | 2022-06-06 | 2022-10-18 | 广州中思人工智能科技有限公司 | Three-dimensional reconstruction method, system and device in large scene |
-
2023
- 2023-01-05 CN CN202310011438.9A patent/CN115719407B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108416840A (en) * | 2018-03-14 | 2018-08-17 | 大连理工大学 | A kind of dense method for reconstructing of three-dimensional scenic based on monocular camera |
US20210183080A1 (en) * | 2019-12-13 | 2021-06-17 | Reconstruct Inc. | Interior photographic documentation of architectural and industrial environments using 360 panoramic videos |
WO2021185322A1 (en) * | 2020-03-18 | 2021-09-23 | 广州极飞科技有限公司 | Image processing method and related device |
US20210358206A1 (en) * | 2020-05-14 | 2021-11-18 | Star Institute Of Intelligent Systems | Unmanned aerial vehicle navigation map construction system and method based on three-dimensional image reconstruction technology |
CN111968218A (en) * | 2020-07-21 | 2020-11-20 | 电子科技大学 | Three-dimensional reconstruction algorithm parallelization method based on GPU cluster |
CN112085845A (en) * | 2020-09-11 | 2020-12-15 | 中国人民解放军军事科学院国防科技创新研究院 | Outdoor scene rapid three-dimensional reconstruction device based on unmanned aerial vehicle image |
CN113284227A (en) * | 2021-05-14 | 2021-08-20 | 安徽大学 | Distributed motion inference structure method for large-scale aerial images |
CN115205489A (en) * | 2022-06-06 | 2022-10-18 | 广州中思人工智能科技有限公司 | Three-dimensional reconstruction method, system and device in large scene |
Non-Patent Citations (5)
Title |
---|
BARBARA ROESSLE,ET AL.: "Dense Depth Priors for Neural Radiance Fields from Sparse Input Views", 《IEEE XPLORE》 * |
MINGWEI CAO,ET AL.: "Parallel surface reconstruction for large-scale scenes in the wild", 《WILEY》 * |
姜翰青;赵长飞;章国锋;王慧燕;鲍虎军;: "基于多视图深度采样的自然场景三维重建", 计算机辅助设计与图形学学报, no. 10 * |
张琮毅 等: "尺度可变的快速全局点云配准方法", 《计算机学报》 * |
颜深 等: "大规模室外图像3维重建技术研究进展", 《中国图象图形学报》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116071504A (en) * | 2023-03-06 | 2023-05-05 | 安徽大学 | Multi-view three-dimensional reconstruction method for high-resolution image |
CN116805355A (en) * | 2023-08-25 | 2023-09-26 | 安徽大学 | Multi-view three-dimensional reconstruction method for resisting scene shielding |
CN116805355B (en) * | 2023-08-25 | 2023-11-21 | 安徽大学 | Multi-view three-dimensional reconstruction method for resisting scene shielding |
CN116993925A (en) * | 2023-09-25 | 2023-11-03 | 安徽大学 | Distributed bundling adjustment method for large-scale three-dimensional reconstruction |
CN116993925B (en) * | 2023-09-25 | 2023-12-01 | 安徽大学 | Distributed bundling adjustment method for large-scale three-dimensional reconstruction |
CN117408999A (en) * | 2023-12-13 | 2024-01-16 | 安格利(成都)仪器设备有限公司 | Method for automatically detecting corrosion pits of containers and pipelines by utilizing point cloud complement |
CN117408999B (en) * | 2023-12-13 | 2024-02-20 | 安格利(成都)仪器设备有限公司 | Method for automatically detecting corrosion pits of containers and pipelines by utilizing point cloud complement |
CN117437363A (en) * | 2023-12-20 | 2024-01-23 | 安徽大学 | Large-scale multi-view stereoscopic method based on depth perception iterator |
CN117437363B (en) * | 2023-12-20 | 2024-03-22 | 安徽大学 | Large-scale multi-view stereoscopic method based on depth perception iterator |
Also Published As
Publication number | Publication date |
---|---|
CN115719407B (en) | 2023-06-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN115719407A (en) | Distributed multi-view stereo reconstruction method for large-scale aerial images | |
CN112765095B (en) | Method and system for filing image data of stereo mapping satellite | |
CN113516769B (en) | Virtual reality three-dimensional scene loading and rendering method and device and terminal equipment | |
CN113096250A (en) | Three-dimensional building model library system construction method based on unmanned aerial vehicle aerial image sequence | |
CN113808261B (en) | Panorama-based self-supervised learning scene point cloud completion data set generation method | |
CN108648264A (en) | Underwater scene method for reconstructing based on exercise recovery and storage medium | |
CN113284227B (en) | Distributed motion inference structure method for large-scale aerial images | |
CN106408531A (en) | GPU acceleration-based hierarchical adaptive three-dimensional reconstruction method | |
CN113192200A (en) | Method for constructing urban real scene three-dimensional model based on space-three parallel computing algorithm | |
CN114202632A (en) | Grid linear structure recovery method and device, electronic equipment and storage medium | |
Wang et al. | Fast 3D reconstruction method based on UAV photography | |
Lai et al. | 2D3D-MVPNet: Learning cross-domain feature descriptors for 2D-3D matching based on multi-view projections of point clouds | |
CN113886510A (en) | Terminal interaction method, device, equipment and storage medium | |
CN117456136A (en) | Digital twin scene intelligent generation method based on multi-mode visual recognition | |
US20080111814A1 (en) | Geometric tagging | |
CN111260712B (en) | Depth estimation method and device based on refocusing polar line graph neighborhood distribution | |
CN116152442A (en) | Three-dimensional point cloud model generation method and device | |
CN116612235A (en) | Multi-view geometric unmanned aerial vehicle image three-dimensional reconstruction method and storage medium | |
CN115578260A (en) | Attention method and system for direction decoupling for image super-resolution | |
CN113487713B (en) | Point cloud feature extraction method and device and electronic equipment | |
CN113850293A (en) | Positioning method based on multi-source data and direction prior joint optimization | |
Li et al. | BDLoc: Global localization from 2.5 D building map | |
CN116805355B (en) | Multi-view three-dimensional reconstruction method for resisting scene shielding | |
CN117078982B (en) | Deep learning-based large-dip-angle stereoscopic image alignment dense feature matching method | |
CN117437363B (en) | Large-scale multi-view stereoscopic method based on depth perception iterator |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |