CN106384386B - Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model - Google Patents
Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model Download PDFInfo
- Publication number
- CN106384386B CN106384386B CN201610882819.4A CN201610882819A CN106384386B CN 106384386 B CN106384386 B CN 106384386B CN 201610882819 A CN201610882819 A CN 201610882819A CN 106384386 B CN106384386 B CN 106384386B
- Authority
- CN
- China
- Prior art keywords
- grid
- lod
- hole
- mesh
- combined
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/30—Polynomial surface description
Abstract
The invention discloses grid processing method, system and 3D method for reconstructing and systems in a kind of generation of LOD model.Wherein, the method for grid being handled in LOD model generating process includes: acquisition initial mesh;The initial mesh is sheared to obtain multiple LOD subregions;The LOD subregion of current level is sutured with combined grid by mesh update;It is lattice simplified to combined grid application to obtain the LOD subregion of next level;The mesh update is executed to the LOD subregion iteration of next level until obtaining final combined grid.For by the 3D reconstruction technique of rapid growth come the growing demand of band for be robust and expansible.It optimizes three main grid operations, and the experiment on large data collection shows it with robustness and scalability.
Description
Technical field
Embodiment of the present invention is related to grid processing side in the generation of LOD (Level-of-detail, level of detail) model
Method, system and 3D (three-dimensional) method for reconstructing and system.
Background technique
Past 10 years, by the universalness that improved computing hardware and unmanned plane are photographed large-scale 3D reconstruction is become could
Energy.These 3D models rebuild are typically used as to such as machine with protecting from Buildings Modeling, urban planning, measurement, traces
The extensive use of the application of the relatively frontier of obstacle avoidance and 3D printing in people's technology etc..Commercial unmanned plane photography is especially
It is fast-developing in recent years.Unquestionably, following to may require that even more large-scale and higher-quality 3D reconstruction model.
The possible mode for handling this large data is divide and conquer (divide-and-conquer).Identical technology is
It is the height for the object or scene for generating capture applied to multi-angle of view stereo reconstruction (multi view stereo, MVS) step
The required step of density points cloud and be the prior step that large-scale LOD model generates.Hereafter, net is extracted in each cluster of Cong Dianyun
Lattice face.This means that generated grid surface is usually segment and discontinuous.Moreover, the ruler in usually these initial mesh faces
It is very little too big so that it cannot by the Internet transmission, and limit its usability.Ideal solution is initial according to these
Grid surface generates multi-resolution models or level of detail (LOD) grid, to meet the accuracy requirements of the different levels of model.Example
Such as, transmission application requires most detailed model, and the model simplified is best suited for on-line normalization and quickly previewing scene.Only pass through
The grid for generating LOD form may make that large-scale 3D reconstruction model continues to develop and more practical.
However, developed most of grid operations technologies are conceived to the single object of processing, hardly it is applied to handle this
Kind large data.When being used for this large data, some in these technologies start to expose some problems, and these are asked
The past is inscribed almost without causing to pay close attention to.Moreover, segment and highdensity initial mesh is converted into continuous LOD form grid
Need to mix several kinds of grid operations, which again increases one layer of difficulty.
Summary of the invention
In computer graphics, level of detail (LOD) is to choice (trade between model fidelity and rendering speed
Off research).Basic thought is using simplified model to indicate inessential or the object further away from viewpoint is to reduce rendering
Load.For many years many researchers have proposed different methods and generate LOD model and put into application.Works " 3D figure
Level of detail (Level of Detail For 3D Graphics) " be extraordinary general survey and its intermediary in terms of LOD
The classical technology to continue is still related so far.Except all LOD frameworks, the present invention is made on discrete LOD.Discrete LOD
The same model for referring to different level of detail is that discrete and each level generates offline and first based on observer and Mo
The distance between type selects suitably to render at runtime horizontal.Advantage is that each level of model is static and thus
Graphic hardware can be had an impact to increase its rendering speed.Relative to other frameworks, such as continuous LOD or view correlation LOD, from
Scattered LOD is more easier to program, because it is decoupled lattice simplified with rendering.This decoupling also makes offline lattice simplified parallelization
And it is thus more expansible.Fig. 5 shows the typical example of the discrete LOD of single object.
Different from simple, the single object LOD example of the foregoing description, the project that the present invention faces is that scale is much bigger simultaneously
And constantly increase, in LOD process must also use divide and rule technology.The present invention provides nets in a kind of generation of LOD model as a result,
Lattice processing method, system and 3D method for reconstructing and system.
It on the one hand, can be with according to a kind of method for handling grid in LOD model generating process of embodiment of the present invention
It include: acquisition initial mesh;The initial mesh is sheared to obtain multiple LOD subregions;By mesh update by current level
LOD subregion is sutured with combined grid;It is lattice simplified to the combined grid application to be divided with the LOD for obtaining next level
Area;The mesh update is executed to the LOD subregion iteration of next level until obtaining final combined grid.
In another embodiment of the present invention, the method also includes carrying out hole filling to hole existing for combined grid
So that adjacent mesh surface is adapted.
Optionally, the hole filling may include the first filling processing: the grid of topological structure of the building with known hole
Patch;The border vertices in the hole on target gridding are mapped to the border vertices of the grid patch;Use the position of the border vertices
It sets and internal vertex is carried out smoothly as constraint condition.In addition, the hole filling may also include the second filling processing: being mended to grid
Fourth carries out Initial Triangulation Approach;Trigonometric ratio is refined by increasing internal vertex;Increased internal vertex, which is placed on, to be made to mend
The position that fourth surface and adjacent surface are adapted.
Wherein, carrying out Initial Triangulation Approach to grid patch may include: two error functions for defining triangle: WAreaWith
WDihedral angle-angle, wherein WAreaRefer to the area of triangle, WDihedral angle-angleRefer to the maximum dihedral angle angle with adjacent triangle;Definition
Grid patch error function W is WAreaTotal minimal error or WDihedral angle-angleMaximum value minimal error;By keeping W minimum
Change to complete Initial Triangulation Approach.
In addition, or alternatively, second filling processing further include: the surface in the hole smoothly newly filled minimizes side
Normal variations between boundary vertex and increased internal vertex.
In an alternative embodiment of the invention, the first filling processing is used for the hole more than or equal to predetermined threshold,
The second filling processing is for the hole less than the predetermined threshold.
Optionally, the shearing may include: definition secant plane;All and cutting is searched in the initial mesh
The triangle of level-crossing;The triangle of the intersection is separated into small triangle along shear line;By using described small by three
Angular geometric center as test point by all small triangle marks the secant plane either side, and remove should
Unwanted triangle in side.
Optionally, the mesh update may include: to cut away predetermined offset along the combined side of grid to be combined to build
Stand the merging gap between grid to be combined;Along the combined side of grid to be combined by matching between two merging grids most
Short distance border vertices establish the bridge between grid to be combined;Hole filling is carried out to complete to close to hole between the net to be combined
And.In addition, or optionally, the mesh update further include: along bridge described in Boundary Traversal, and record the sequence of bridge, check
Discontinuous sequence, the bridge intersected by the sequential search of bridge, and abandon the bridge of the intersection.
Optionally, the various different embodiments of the present invention are based on QSlim algorithm and carry out institute using following Quadratic Surface Equations
It states lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates v's
Support plane set, KpIndicate that the squared-distance matrix for supporting plane, α ∈ [0,1] has determined detected on vertex in grid
The significance level of plane domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)Indicate net
In lattice vertex to affiliated characteristic plane squared-distance matrix.
It on the other hand, may include: by multi-angle of view solid weight according to a kind of 3D method for reconstructing of embodiment of the present invention
Build the point cloud that MVS obtains object;Surface mesh is obtained from described cloud;Based on the surface mesh of the acquisition, using above-mentioned each
Grid processing method described in kind different embodiments generates LOD grid;The 3D of the object is completed using the LOD grid of generation
It rebuilds.
In another aspect, a kind of system for handling grid in LOD model generating process according to embodiment of the present invention can
To include:
Module is obtained, for obtaining initial mesh;
Shear module obtains multiple LOD subregions for shearing the initial mesh;
Merging module, for being sutured the LOD subregion of current level with combined grid by mesh update;
Simplify module, for lattice simplified to obtain the LOD subregion of next level to the combined grid application;
Wherein, the mesh update is executed to the LOD subregion iteration of next level until obtaining final merging
Grid.
Preferably, the system also includes carry out hole filling so that adjacent mesh surface phase to hole existing for combined grid
The filling module of adaptation.Optionally, the filling module can be used for executing the first filling processing: building is opened up with known hole
Flutter the grid patch of structure;The border vertices in the hole on target gridding are mapped to the border vertices of the grid patch;Use this
The position of border vertices carries out internal vertex as constraint condition smooth.In addition, the filling module can also be used in second and fill out
It fills processing: Initial Triangulation Approach is carried out to grid patch;Trigonometric ratio is refined by increasing internal vertex;By increased inside
Vertex is placed on the position for being adapted patch surface and adjacent surface.Wherein, carrying out Initial Triangulation Approach to grid patch can wrap
It includes: defining two error functions of triangle: WAreaAnd WDihedral angle-angle, wherein WAreaRefer to the area of triangle, WDihedral angle-angleRefer to
With the maximum dihedral angle angle of adjacent triangle;Definition grid patch error function W is WAreaTotal minimal error or
WDihedral angle-angleMaximum value minimal error;Initial Triangulation Approach is completed by minimizing W.In optional embodiment party of the invention
In formula, the second filling processing may also include that the surface in the hole smoothly newly filled, and minimize border vertices and increased inside
Normal variations between vertex.
In selective embodiment of the invention, the first filling processing is for more than or equal to predetermined threshold
Hole, the second filling processing is for the hole less than the predetermined threshold.
Preferably, the shearing can include: define secant plane;It is searched in the initial mesh all flat with the cutting
The triangle that face intersects;The triangle of the intersection is separated into small triangle along shear line;By using the small triangle
The geometric center of shape as test point by all small triangle marks in the either side of the secant plane, and remove the side
In unwanted triangle.
Preferably, the mesh update can include: cut away predetermined offset along the combined side of grid to be combined to establish
Merging gap between grid to be combined;Pass through along the combined side of grid to be combined nearest between two merging grids of matching
The bridge between grid to be combined is established apart from border vertices;Hole filling is carried out to complete to merge to hole between the net to be combined.
Optionally, the mesh update may also include that along bridge described in Boundary Traversal, and record the sequence of bridge, check discontinuous
Sequentially, the bridge intersected by the sequential search of bridge, and abandon the bridge of the intersection.
Preferably, the system of embodiment of the present invention is based on QSlim algorithm using described in the progress of following Quadratic Surface Equations
It is lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates v's
Support plane set, KpIndicate that the squared-distance matrix for supporting plane, α ∈ [0,1] has determined detected on vertex in grid
The significance level of plane domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)Indicate net
In lattice vertex to affiliated characteristic plane squared-distance matrix.
Another aspect may include: according to a kind of 3D reconstructing system of embodiment of the present invention
First acquisition unit obtains the point cloud of object by multi-angle of view stereo reconstruction MVS;
Second acquisition unit obtains surface mesh from described cloud;
Generation unit, based on the surface mesh of the acquisition, at grid described in above-mentioned various different embodiments
Reason method generates LOD grid;
Reconstruction unit is rebuild using the 3D that the LOD grid of generation completes the object.
According to the various embodiments of the present invention, for the 3D reconstruction technique by rapid growth come the growing need of band
It is robust for asking and expansible.It optimizes three main grid operations, the grid operations be mainly used for process,
The fixed challenging grid flaw generated by aforementioned reconstruction procedures merges grid and declines without overlay requirements or do not generate
Subtract triangle, be suitable for City scenarios grid by general simplified algorithm optimization.Experiment on large data collection shows that it has
Robustness and scalability.
Detailed description of the invention
Fig. 1 shows the example flow of the grid processing method according to embodiment of the present invention.
Fig. 2 shows the example flows of the 3D method for reconstructing according to embodiment of the present invention.
Fig. 3 shows the exemplary architecture of the lattice processing system according to embodiment of the present invention.
Fig. 4 shows the exemplary architecture of the 3D reconstructing system according to embodiment of the present invention.
Fig. 5 shows the typical example of the discrete LOD of single object.
Fig. 6 shows a kind of original LOD design.
Fig. 7 a shows the 3D reconstruction image example obtained using original LOD design.
Fig. 7 b shows restrained boundary during simplification and generates undesirable star network.
Fig. 8 shows a kind of LOD design according to embodiment of the present invention.
Fig. 9 shows the definition in hole.
Figure 10 is shown across the detection of the hole of subregion.
Figure 11 a shows the typical input of hole filling algorithm.
Figure 11 b shows the hole filling algorithm for not carrying out refinement and shaping.
Figure 11 c shows the hole filling algorithm for having carried out refinement and shaping.
Figure 11 d shows hole filling algorithm compared with true ground.
Figure 12 shows typical input data scale.
Figure 13 shows the effect of α in archioporus filling algorithm.
Figure 14 shows new hole fill method.
Figure 15 shows the sparse linear system used in new method.
Figure 16 shows the visual comparison (line architecture (wireframe)) of archioporus filling algorithm and new method.
Figure 17 shows the visual comparisons (surface) of archioporus filling algorithm and new method.
Figure 18 shows the limitation of new method.
Figure 19, which is shown, there is strange Self-crossover after the region for simplifying new method filling.
Figure 20 shows a case that possible in shearing.
Figure 21 shows the input for merging algorithm.
Figure 22 a shows the step 1 for merging algorithm.
Figure 22 b shows the step 2 for merging algorithm.
Figure 22 c shows the step 3 for merging algorithm.
Figure 23 shows and improves the final result of algorithm.
Figure 24 shows processing bridge cross-cutting issue.
Figure 25 shows the purpose of QSlim geometric error company.
Figure 26 shows the comparison of the simplification method of QSlim and embodiment of the present invention, wherein black region is in input
The plane domain detected.
Figure 27 shows the comparison of the simplification method of QSlim and embodiment of the present invention.
Figure 28 shows the large data collection for experiment.
Figure 29 shows the 5th grade of LOD grid of Toulouse.
Figure 30 shows the 4th grade of LOD grid of Toulouse.
Figure 31 shows the 3rd level LOD grid of Toulouse.
Figure 32 shows the 2nd grade of LOD grid of Toulouse.
Figure 33 shows the 1st grade of LOD grid of Toulouse.
Figure 34 shows the 0th grade of LOD grid of Toulouse.
Specific embodiment
The various aspects of technical solution, feature and advantage to facilitate the understanding of the present invention, with reference to the accompanying drawing to this hair
It is bright to be specifically described.It should be appreciated that following various embodiments are served only for for example, and is not intended to limit the present invention
Protection scope.
This is described below with reference to the example block diagram of method, system, unit and programming and computer program product
Invention.It should be understood that each of example block diagram frame and combinations thereof can be realized by programming instruction, the programming refers to
Enabling includes computer program instructions.These computer program instructions can be loaded to computer or other programmable datas processing dress
It sets, the instruction executed on computer or other programmable data processing units in this way generates for realizing in block diagram or flow chart
The instruction of specified function.These computer program instructions are also storable in computer-readable memory, can indicate computer
Or other programmable data processing units are run in a specific way, in this way, the instruction being stored in computer-readable memory produces
Raw includes the manufacture for realizing the instruction for the function of specifying in block diagram or flow chart.Computer program instructions can also be loaded to calculating
Machine or other programmable data processing units are executed with generating a series of in the computer or other programmable devices to produce
The operating procedure of raw executive program, in this way, the instruction executed on the computer or other programmable devices provides
The step of realizing the function of specifying in block diagram or flow chart.Programming instruction also can be stored in electronic circuit and/or pass through electricity
Implement to realize each function of the invention, step on sub-line road.
Grid processing method
Fig. 1 shows a kind of method that grid is handled in LOD model generating process according to embodiment of the present invention.Such as
It is described to may include shown in Fig. 1, but be not limited to:
110. obtaining initial mesh;
120. shearing the initial mesh to obtain multiple LOD subregions;
130. the LOD subregion of current level is sutured with combined grid by mesh update;
140. lattice simplified to the combined grid application to obtain the LOD subregion of next level;
The LOD subregion iteration of 150. pairs of next levels executes the mesh update until obtaining final combined net
Lattice.
In selective embodiment of the invention, the method can also include carrying out to hole existing for combined grid
Hole is filled so that adjacent mesh surface is adapted.Optionally, the hole filling may include the first filling processing, as follows: building band
There is the grid patch of the topological structure in known hole;The border vertices in the hole on target gridding are mapped to the side of the grid patch
Boundary vertex;The position of the border vertices is used to carry out smoothly as constraint condition to internal vertex.In addition, the hole filling may be used also
It is handled including the second filling, it is as follows: Initial Triangulation Approach is carried out to grid patch;Trigonometric ratio is carried out carefully by increasing internal vertex
Change;Increased internal vertex is placed on to the position for being adapted patch surface and adjacent surface.
Wherein, carrying out Initial Triangulation Approach to grid patch may include: two error functions for defining triangle: WAreaWith
WDihedral angle-angle, wherein WAreaRefer to the area of triangle, WDihedral angle-angleRefer to the maximum dihedral angle angle with adjacent triangle;Definition
Grid patch error function W is WAreaTotal minimal error or WDihedral angle-angleMaximum value minimal error;By keeping W minimum
Change to complete Initial Triangulation Approach.
In addition, or alternatively, the second filling processing can also include: the surface in the hole smoothly newly filled, minimum
Change the Normal variations between border vertices and increased internal vertex.
In an alternative embodiment of the invention, the first filling processing is used for the hole more than or equal to predetermined threshold,
That is macropore, the second filling processing is for the hole less than the predetermined threshold, i.e. aperture.The predetermined threshold can be according to reality
Border needs to be determined, for example, 600 vertex.
Optionally, in one embodiment of the invention, the shearing may include, but be not limited to: definition cutting is flat
Face;All triangles intersected with the secant plane are searched in the initial mesh;By the triangle of the intersection along cutting
Tangent line is separated into small triangle;By using the geometric center of the small triangle as test point by all small triangle marks
Remember the either side in the secant plane, and removes unwanted triangle in the side.
Optionally, in one embodiment of the invention, the mesh update may include, but be not limited to: along to
Merge the merging gap that the combined side of grid cuts away predetermined offset to establish between grid to be combined;Along grid to be combined
Combined side merges the minimum distance border vertices between grid by matching two and establishes the bridge between grid to be combined;To described
Hole carries out hole filling to complete to merge between net to be combined.In addition, the mesh update can also include: along Boundary Traversal institute
Bridge is stated, and records the sequence of bridge, checks discontinuous sequence, the bridge intersected by the sequential search of bridge, and abandon the friendship
The bridge of fork.
Optionally, in the various different embodiments of the present invention, based on QSlim algorithm using following Quadratic Surface Equations into
Row is described lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates v's
Support plane set, KpIndicate that the squared-distance matrix for supporting plane, α ∈ [0,1] has determined detected on vertex in grid
The significance level of plane domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)Indicate net
In lattice vertex to affiliated characteristic plane squared-distance matrix.
According to the various embodiments of the present invention, for the 3D reconstruction technique by rapid growth come the growing need of band
It is robust for asking and expansible.It optimizes three main grid operations, the grid operations be mainly used for process,
The fixed challenging grid flaw generated by aforementioned reconstruction procedures merges grid and declines without overlay requirements or do not generate
Subtract triangle, be suitable for City scenarios grid by general simplified algorithm optimization.Experiment on large data collection shows that it has
Robustness and scalability.
3D method for reconstructing
Fig. 2 shows a kind of 3D method for reconstructing according to embodiment of the present invention.As shown in Fig. 2, the method for reconstructing can
To include, but are not limited to:
210. obtain the point cloud of object by multi-angle of view stereo reconstruction MVS;
220. obtain surface mesh from described cloud;
230. surface meshes based on the acquisition, using grid processing method described in above-mentioned various different embodiments
Generate LOD grid;
240. are rebuild using the 3D that the LOD grid generated completes the object.
According to the various embodiments of the present invention, for the 3D reconstruction technique by rapid growth come the growing need of band
It is robust for asking and expansible.It optimizes three main grid operations, the grid operations be mainly used for process,
The fixed challenging grid flaw generated by aforementioned reconstruction procedures merges grid and declines without overlay requirements or do not generate
Subtract triangle, be suitable for City scenarios grid by general simplified algorithm optimization.Experiment on large data collection shows that it has
Robustness and scalability.
Lattice processing system
Fig. 3 shows a kind of system that grid is handled in LOD model generating process according to embodiment of the present invention.Such as
Shown in Fig. 3, lattice processing system may include, but be not limited to obtain module 310, shear module 320, merging module 330 and
Simplify module 330.Wherein, obtaining module 310 can be used for obtaining initial mesh, and shear module 320 can be used for shearing by described
The initial mesh of the acquisition of module 310 is obtained to obtain multiple LOD subregions, merging module 330 can be used for work as by mesh update
The LOD subregion suture of preceding level is with combined grid, and simplifies module 340 and can be used for the combined grid application
It is lattice simplified to obtain the LOD subregion of next level, wherein the grid is executed to the LOD subregion iteration of next level
Merge until obtaining final combined grid.
In an alternative embodiment of the invention, the system, which may also include, fills out the progress hole of hole existing for combined grid
It is filled with the filling module for keeping adjacent mesh surface adaptable.Optionally, the filling module can be used for executing the first filling processing:
The grid patch of topological structure of the building with known hole;The border vertices in the hole on target gridding are mapped to the grid to mend
The border vertices of fourth;The position of the border vertices is used to carry out smoothly as constraint condition to internal vertex.In addition, the filling
Module can also be used in the second filling processing: carry out Initial Triangulation Approach to grid patch;By increase internal vertex to trigonometric ratio into
Row refinement;Increased internal vertex is placed on to the position for being adapted patch surface and adjacent surface.Wherein, to grid patch
Carry out Initial Triangulation Approach can include: define two error functions of triangle: WAreaAnd WDihedral angle-angle, wherein WAreaRefer to triangle
Area, WDihedral angle-angleRefer to the maximum dihedral angle angle with adjacent triangle;Definition grid patch error function W is WAreaIt is total
The minimal error or W of meterDihedral angle-angleMaximum value minimal error;Initial Triangulation Approach is completed by minimizing W.In this hair
In bright optional embodiment, the second filling processing may also include that the surface in the hole smoothly newly filled, and minimize boundary top
Normal variations between point and increased internal vertex.
In selective embodiment of the invention, the first filling processing is for more than or equal to predetermined threshold
Hole, the second filling processing is for the hole less than the predetermined threshold.
In the various embodiments of the present invention, the shearing can include: define secant plane;In the initial mesh
Search all triangles intersected with the secant plane;The triangle of the intersection is separated into small triangle along shear line;
By using the geometric center of the small triangle as test point by all small triangle marks in the secant plane
Either side, and remove unwanted triangle in the side.Preferably, the mesh update can include: along grid to be combined
The combined side merging gap of cutting away predetermined offset to establish between grid to be combined;It is logical along the combined side of grid to be combined
The overmatching two minimum distance border vertices merged between grid establish the bridge between grid to be combined;To the net to be combined
Between hole carry out hole filling to complete to merge.Optionally, the mesh update may also include that along bridge described in Boundary Traversal, and
And the sequence of bridge is recorded, and check discontinuous sequence, the bridge intersected by the sequential search of bridge, and abandon the bridge of the intersection.
In the various embodiments of the present invention, the lattice processing system is based on QSlim algorithm and uses following secondary songs
Face equation carries out described lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates v's
Support plane set, KpIndicate that the squared-distance matrix for supporting plane, α ∈ [0,1] has determined detected on vertex in grid
The significance level of plane domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)Indicate net
In lattice vertex to affiliated characteristic plane squared-distance matrix.
According to the various embodiments of the present invention, for the 3D reconstruction technique by rapid growth come the growing need of band
It is robust for asking and expansible.It optimizes three main grid operations, the grid operations be mainly used for process,
The fixed challenging grid flaw generated by aforementioned reconstruction procedures merges grid and declines without overlay requirements or do not generate
Subtract triangle, be suitable for City scenarios grid by general simplified algorithm optimization.Experiment on large data collection shows that it has
Robustness and scalability.
3D reconstructing system
Fig. 4 shows a kind of 3D reconstructing system according to embodiment of the present invention.In embodiments of the present invention, the 3D
Reconstructing system may include:
First acquisition unit 410 obtains the point cloud of object by multi-angle of view stereo reconstruction MVS;
Second acquisition unit 420 obtains surface mesh from described cloud;
Generation unit 430, based on the surface mesh of the acquisition, using grid described in the above various different embodiments
Processing method generates LOD grid;
Reconstruction unit 440 is rebuild using the 3D that the LOD grid of generation completes the object.
According to the various embodiments of the present invention, for the 3D reconstruction technique by rapid growth come the growing need of band
It is robust for asking and expansible.It optimizes three main grid operations, the grid operations be mainly used for process,
The fixed challenging grid flaw generated by aforementioned reconstruction procedures merges grid and declines without overlay requirements or do not generate
Subtract triangle, be suitable for City scenarios grid by general simplified algorithm optimization.Experiment on large data collection shows that it has
Robustness and scalability.
It is described the invention in detail above based on method, the various different embodiments of system, below with specific
Application the present invention is illustrated.
Using example
As previously mentioned, simple, the single object LOD example from the foregoing description is different, the project that the present invention faces is rule
Mould is much bigger and constantly increases, in LOD process must also use divide and rule technology.The step of before LOD process is MVS,
Generate the discontinuous surface mesh of high density.Input of these initial mesh as LOD process.It should be noted that at the beginning of these
Beginning grid is in shared cooperative system, and therefore, they are overlapped around in its borderline region.
A kind of original design is to handle each initial mesh as single object and independently generate LOD grid, referring to
Fig. 6.Although this design has extraordinary scalability, major defect is there is boundary difference between adjacent grid
It is different.A kind of means to save the situation is to constrain or limit these boundaries during lattice simplified, but this is lattice simplified by big havoc
Validity because the ratio regular meeting on fixed vertex quicklys increase and generates the grid of star-shaped over-simplification, referring to Fig. 7 a and
7b。
Therefore, embodiment of the present invention is designed using quaternary tree (quad-tree).Referring to Fig. 8.First by all input nets
It is overall situation 2D grid (global 2D grid) that lattice, which shear (trim), and lesser rectangular mesh is known as subregion (tiles).Each side
The number of upward subregion is 2n, wherein n is natural number.The input level is known as n-th grade of LOD.Generating each next stage
During LOD, the LOD subregion of 2 × 2 current levels is sutured by mesh update first.Then, by all grid flaws
(mesh artifacts) is fixed in each combined grid.Finally, lattice simplified to combined grid application to obtain down
The LOD of level-one.The three step process is sequentially repeated until obtaining the grid (that is, the 0th grade LOD) that only one finally merges.
The number of level is then n+1, also, at n-th grade, the sum of subregion is 2n×2n=22n.Therefore, by moving to next stage,
The number of subregion is reduced to a quarter of current level.
Quaternary tree design can solve the problems, such as border-differential and over-simplification problem in grid.The design also reduces text
Number of packages mesh and network I/O expense (overhead), this is because reducing total grid number.For example, if obtaining 24×24=16 ×
16=256 subregion is as input, then will generate 2 in total8+26+24+22+20=341 discrete grids, in comparison,
In the original design of the LOD with the level quantity, grid sum is 28× 5=1280.As it can be seen that total grid number is able to significantly
It reduces.The design of embodiment of the present invention is expansible, because being only used only Local Phase neighbour's when generating next stage LOD
Grid can carry out in a distributed manner.A data aggregate is only needed in each level.The granularity of subregion, the number of LOD level and
Simplifying rate depends on hardware limitation and application purpose.In general, the granularity of subregion is more fine and LOD level it is deeper become
In taking longer time, but need less strength machine (powerful machines) and higher parallelization
(parallelization)。
Three kinds of main grid operations can be used to implement this design, that is, fill (hole filling), shearing and close in hole
And (trimming and merging), simplified (simplification).The purpose of " hole filling " is fixed initial mesh
Existing flaw (artifacts), the flaw can be due to removing non-manifold (non-manifold) grid and flowing
What journey internal dynamic generated." shearing " removes overlapping region applied to initial mesh and " merging " is used to remove border-differential simultaneously
Move up quaternary tree." simplification " is the core that LOD is generated, wherein simplification can subtract while maintaining grid geometric fidelity
Few size of mesh opening.Each operation is thoroughly discussed below.
[hole filling]
Theoretically, two kinds of hole involved in grid, that is, boundary hole and topological hole (topological hole), ginseng
See Fig. 9.It is discussed for the sake of boundary hole and simplicity herein referring briefly to boundary hole as hole.Actual hole filling algorithm is presented first
And its limitation and problem are discussed.Then introduce the method for the processing limitation and problem.
Hole is topological characteristic, can easily be detected by using half of (Half-Edge) data structure, the Half-
Edge data structure is to indicate the most common mode of grid.Each non-boundary edge can and boundary adjacent with two gores
Edge is only adjacent with a face, and it is open to make the other side, and thus becomes a part in hole.Then by passing through all boundaries
Half edge can detect all holes and each hole be expressed as to the ordered list at half edge of boundary (or vertex).Of the invention real
It applies in mode, it is more using CGAL (Computational Geometry Algorithms Library, computational geometry algorithms library)
Face volume data structure is substantially the half of data structure for representing grid.
In one embodiment of the invention, LOD designs all input grids and is cut into 2D lattice structure and each
Rectangular mesh is referred to as subregion.How to detect so big hole existing for the subregion, as illustrated in Figure 10 (left side) one
A hole.
The detection in hole can be by identifying and selecting first when being projected in X/Y plane with all of range are not completely covered
Subregion.The coverage area of subregion may be defined as its projected area relative to rectangular area (that is, area of rectangular sub-area completely)
Ratio.Its projected area can amount to the delaunay trigonometric ratio of its 2D limitation by using boundary hole one side of something as the edge of limitation
Triangle in (delaunay triangulation) and find.Another quick estimation to projected area is to calculate its projection
The area of the convex surface the 2D shell on vertex.For example, CGAL calculates the software package (packages) of two values with providing convenience.Then
The selection for all that not being completely covered subregion expands outwardly the subregion of an adjacent ring.Finally, merging selected
Subregion and carry out hole detection.Referring to Figure 10.
In one embodiment of the invention, the hole filling algorithm can be analyzed to three key steps:
1. Initial Triangulation Approach (Initial triangulation).Two error functions of triangle, W are defined firstArea
And WDihedral angle-angle, WAreaRefer to the area of triangle, WDihedral angle-angleRefer to the maximum dihedral angle angle with adjacent triangle.Then definition
Grid patch (mesh patch) error function W is WAreaTotal minimal error or its trigonometric ratio WDihedral angle-angleMaximum value
Minimal error.Therefore, the Initial Triangulation Approach of grid patch be make W minimize Initial Triangulation Approach and this can by dynamic compile
Journey is searched.
2. refining (Refinement).Initial Triangulation Approach have base closed hole, but trigonometric ratio it is too coarse so that
The patch (patch) of nature complexion can not be shown, because it lacks internal vertex.This step is based on some heuristic rules
(heuristic criteria) iteratively increases more internal vertexs on trigonometric ratio.The standard by referred to as α parameter control
System.In short, higher α causes degree of refinement higher and takes longer time operation.
3. shaping (Fairing).It is inserted into refinement step after new vertex, shaping will place them in so that mending
Fourth surface and adjacent surface seem natural position.This is pushed up by surface, the minimum boundary in the hole for repeating smoothly newly to fill
Point and Normal variations between the new vertex being inserted into refinement step are completed.
These three steps are illustrated in Figure 11 a, 11b, 11c, 11d.Wherein, without refinement step, Initial Triangulation Approach
Strange shape can be generated.It is refining with after shaping step, the shape and adjacent surface of the grid patch of filling are adapted.
Because the thorough dynamic programming of the step 1 in algorithm is retrieved, the step 1 in algorithm is in O (n3) time operation.In step
Runing time complexity in rapid 2 is more difficult to accurately define, because of the number on vertex of insertion and the surface area of patch and parameter alpha
Proportional raising, but surface area can not be defined on hole well in the 3 d space.In practice, it can produce good-looking net
Lattice patch is adapted from different adjacent shapes but is for fine tuning parameter and using given input prediction total run time
It is highly desirable to skill.
In another embodiment of the invention, improved hole filling algorithm is provided.In embodiments of the present invention,
The appearance in hole is from three sources.
1) they are the flaws for the initial mesh that MVS step generates.It is known as the water surface, the reflecting surface of the window etc. in building
It can not rebuild and leave macropore in surface mesh.
2) in following shearings and merging process, in order to be filled deliberately to generate hole as combined purpose.
3) hole filling algorithm generates non-manifold grid sometimes.By removing non-manifold surface, new hole is generated.
Most latter two source in hole can be effectively treated since its size is limited by the algorithm.However, in initial mesh
Flaw brings very big challenge to algorithm, because they have very big input size and cover very big area.Greatly
Input size will slow down step 1, due to its O (n3) runing time essence, and biggish area will also slow down step 2 and
Step 3.
The present inventor is tested on the machine with intel i7-4770K CPU, which has four
Kernel operation device simultaneously supports 8 threads and 32GB memory.Test result is as shown in Figure 12 and table 1.
Table 1: the assessment of hole filling algorithm
Exponentially form gradually rises runing time when can be seen that from table above when input size increase.In order to
It solves the problems, such as, in another embodiment of the invention, algorithm can be extended to branch by the dynamic programming in decomposition step 1
Hold multithreading.Although this can successfully improve runing time several times (several folds), this acceleration nevertheless suffers from machine
The limitation of device core number and scale can not be expanded, and the size of macropore is practically without the upper limit.
Moreover, the inventors found that parameter alpha default value (that is,) (that define the degree of refinement)
It can not generate and seem satisfactory grid patch.The inventor is by testing adjusting parameter repeatedly and finally setting α's
Default value about 1.8.Referring to Figure 13.
Finally, the present inventor has handled the macropore of these types by another entirely different method.This method
Inspiration by following the description: these most of macropores are substantially flat.Initial Triangulation Approach step be usually it is extra, because
It is for minimum error result it is contemplated that it is only plane.Actually nor is it necessary that will iteratively tessellated mesh patch
To adapt to neighbouring grid.In another embodiment of the invention, the basic thought of the new method is building first with
Know the grid patch of topological structure and the border vertices on target gridding are mapped to the border vertices of the grid patch.Then
Border vertices position is used to carry out smoothly as constraint condition to internal vertex.Referring to Figure 14.
Smoothing problasm is represented by sparse linear system (sparse linear system) Ax=B.For example, with reference to figure
Example shown in 15.
All internal vertexs are located at the mean place of its adjacent ring and all border vertices are fixed.Due to matrix
Sparse essence, linear system can be used the linear solution device as Biconjugate gradient method (linear solver) big
Cause O (n2) rapid solving in the time.Obtained grid patch is the smooth surface with rough isodensity vertex.In order to further
It using the usually simple fact of the shape on macropore boundary, is sampled by reducing, each k-th of the vertex of extraction is come approximate as input
Input boundary.Setting input threshold value first, such as 500 or 1000, then reduce sampling input vertex, in this way its equivalent inpnt ruler
It is very little to be lower than threshold value.In this way, the runing time of filling macropore can finally be limited.
When hole is larger, almost plane and when convex surface, this method operation is best.The grid that the speed of service is exceedingly fast and obtains
The quality of patch can match in excellence or beauty with the quality of the grid patch from primal algorithm, referring to Figure 16,17.
In addition, because this method does not consider any aesthetics as minimum with the angle of the dihedral angle of adjacent triangle
Standard, obtained grid patch seem unnatural.This method also creates Self-crossover.Referring to Figure 18.
Fortunately, these macropores generally fall into plane domain, and as the lake or river in natural phenomena, and grid patch can
It is thoroughly mixed with adjacent surface.Therefore, the LOD of embodiment of the present invention is designed to handle aperture and macropore respectively, i.e. aperture
Using primal algorithm, the method based on new linear system is used to macropore.In one embodiment of the invention, by macropore
Threshold value is set as 600 vertex, and realizes balance result well.
One final temporary problem be when the grid patch to these types carries out lattice simplified, find it is another with
The related problem of this method.It mentions before, initial mesh is obtained from the 3D of Nature condition reconstruction, the Nature condition tool
There are natural geometry feature or variation, and this method will generate artificial smooth surface.It is lattice simplified to tend to over-simplification these are smooth
Region generates aciculiform and overturning triangle.Referring to Figure 19.In order to solve this problem, artificial random noise is introduced finally
The vertex position obtained by linear system.Select the amplitude of this noise as the sub-fraction of average edge length to maintain it
It influences to minimize and prevent patch over-simplification.
In general, the hole filling algorithm of embodiment according to the present invention, is extended to support multithreading, is run
Time improves several times (several folds).For actual use angle, macropore occurs really and the property of primal algorithm
It can be due to its O (n3) essence and start fast decay.Therefore, the present invention implement it is another based on the algorithm of linear system with special
Handle this macropore.Embodiment of the present invention applies reduction sampling also to limit runing time.Although the new method is logical
With in case have its limitation, but its in the case that these macropores be usually plane and convex surface 3D rebuild run it is good
It is good.Runing time is improved hundreds times by this method and obtained mesh quality can match in excellence or beauty with primal algorithm.
[shear and merge]
Due to the essence of dividing and ruling of LOD process design, grid using shearing and is merged, this is not in other LOD design
Common.
<shearing>
Shearing is quite direct grid operations.In embodiments of the present invention, secant plane is defined first, is then looked up
All triangles with the level-crossing.Then, those triangles intersected are separated into smaller triangle along shear line.
Finally, by using their geometric center as test point by all triangle marks in the either side of plane, and remove
The triangle for wanting to remove is removed in the side.There are three kinds of crossing instances, and can according to cross edge and the number on vertex to this three
Kind crossing instances classification.It is as follows:
0 side of (1) 1 vertex: without operation, label triangle to the +/- side of plane;
1 side of (2) 1 vertex: triangle is separated into two triangles, increases a new summit;
2 side of (3) 0 vertex: triangle is separated into three triangles, increases by two new summits;
Although the triangle decayed for many grid operations be it is undesirable, it is tolerable in shearing
's.It is fully located at the triangle of the decaying of shear plane if there is three vertex, can simply directly assign them to any
Side and this will not influence result.Otherwise, them are handled as conventional triangle.It is fortunately that shearing manipulation both will not
The triangle of decaying will not be generated by generating non-manifold grid.Referring to fig. 20.
Wherein, runing time is O (n), wherein n is face number.
<merging>
Existing multiple developed mesh update algorithms.One type depends on to surface resampling and from adopting again
The surface mesh of merging is extracted in the point cloud of sample.Although they work under general case, they are undesirable, because
Resampling can destroy original mesh topological structure and cannot keep enough geometric fidelities, rebuild this violates 3D
Primary objective.Some other methods dependent on the overlapping region found between combineds grid and carefully increase new vertex with
Face is to connect.However, this method can not handle non-overlap input grid.
In fact, the LOD design of embodiment of the present invention shows that merging is special circumstances, because merging grid (that is, dividing
Area) it is rectangular by pre-processing, and known combined position is without detecting overlapping region.Figure 21 shows typical input
Example.
In embodiments of the present invention, implement new merging algorithm by reseting hole filling algorithm.Merging algorithm tool
There are three key steps:
1) cuts away minor shifts amount (offset) along combined side to establish the merging gap merged between grid.
2) is established along the minimum distance border vertices that their combined side is merged between grid by matching two and is merged
Bridge or connection between grid.
3) filling of applicable holes is to complete to merge.
This three step is shown in Figure 22 a, 22b, 22c and Figure 23.Wherein, Figure 22 a shows foundation merging gap, figure
22b, which is shown, establishes connecting bridge, and Figure 22 c shows filling hole, and Figure 23 shows merging algorithm in the form of illustrating only boundary
Final result.
Most delicate part is step 2.First choice needs set distance threshold value to prevent from matching vertex far apart, phase
It is usually mistake away from too far vertex.The threshold value can pass through the width of zoom factor and bridge joint gap (bridging gap)
Be multiplied setting.Secondly, the number across the merging boundary of two grids, which may not be 1 pair 1, matches and may occur bridge intersection
(bridge crossing) problem.This will lead to the topological structure or non-manifold grid of mistake.In fact it can detect bridge intersection to ask
Topic, for example, along Boundary Traversal (traverse) and record the sequence of bridge and check discontinuous sequence, it then can be in step
The bridge of those intersections is simply discarded before 3.Referring to fig. 24, left-hand side shows the accurate row from B1 to B5 and from C1 to C5
Sequence;Right-hand side shows bridge cross-cutting issue when in grid Y across boundary C.After depositing B9, it is envisaged that B10 is arrived,
But A4 is obtained instead of it, after depositing A5, it is envisaged that arrive A6, but obtain B8 instead of it.
Embodiment according to the present invention, the merging algorithm get rid of the requirement of overlapping, and overlapping is other mesh update algorithms
In necessary to.Merge gap due to establishing in step 1, triangle of the grid finally merged without containing decaying.
[simplification]
It is lattice simplified to be core algorithm used in discrete LOD is generated and many algorithms are suggested in the past.Some calculations
Method is to start specification error threshold value based on fidelity, then simplifies grid as far as possible within the scope of the bouds on error.
The error is typically measured as inputting and simplifying the geometric error between grid, can be global error or local error, this takes
Certainly in practical operation.Some algorithms are based on budget, with the algorithm based on fidelity on the contrary, instead of fixed error
The target numbers for the triangle realized are wanted in boundary, the method setting based on budget after simplifying.Then apply a series of offices
For portion's grid operations with simplified model, this minimizes error.
Although none size is fitted all simplification algorithms, QSlim proposes a classical way, will be simple
Change is formulated as greedy algorithm (greedy algorithm) to minimize the local geometric in each grid operations step process
Error.Its maximum contribution is to invent using matrix, thus compactly and easily indicate error matrix, the error matrix
For calculating the optimal layout for collapsing vertex after an opposite vertexes collapse.This method has proved to be successfully, because it is strong
And effectively and for completely new a kind of algorithm that simplifies provide enlightenment.
In embodiments of the present invention, inventor is based on QSlim algorithm and provides a kind of lattice simplified method, helps to protect
It deposits and reinforced structure feature.
The general thought of QSlim algorithm is iteratively to select and destroy an opposite vertexes for new vertex and by new vertex
It is placed on optimum position, in this way, the approximate geometry of obtained grid minimizes the error.As shown in figure 25, the mesh of each iteration
Be to minimize geometric error, the geometric error be defined as from vertex to its support plane squared-distance.
Assuming that plane p=[a b c d] indicates the plane defined by equation ax+by+cz+d=0, wherein a2+b2+c2=1
And vertex v=[v1 v2 v31] some arbitrary points in representation space.The then squared-distance expression from v to p are as follows:
squareddist(v,p)=(pTv)2=(vTp)(pTV)=vTKpv
Wherein:
The squared-distance information coding of the plane is symmetrical 4 × 4 matrix data packet by Kp, with 10 uniquenesses in total
Floating number.From point v to one, the summation of the squared-distance of group plane (planes) can be calculated effectively are as follows:
∑p∈planes(v)squareddist(v,p)=∑p∈planes(v)(vTKpV)=vT(∑p∈planes(v)Kp)v,
It is the simple summation of matrix K p.In the pre-treatment step of algorithm, by the corresponding Kp of each triangle with
Each vertex QvIt is associated, QvIt is represented simply as quadratic surface, wherein Qv=∑p∈planes(v)Kp, wherein planes (v) is v
Support plane set.
Then, each feasible vertex is calculated to Vi,Vj, be they collapse vertex v ' optimum position, in this way,
ErrorVi,Vj=v 'T(QVi+QVj)v′It is minimized.V' can be by Errorvi,vjPart derivation simultaneously sets it to 0 and obtains
Linear system solve and quickly found, by all errors insertion pile structure (heap structure) and in each iteration
Middle extraction minimal error vertex collapses (collapse) to execute.After collapsing every time, new vertex v ' will be from Qi Muding
It is Q that quadratic surface is inherited in pointv=(Qvi+Qvj), and accumulation tree is updated, it all is affected for what is be connected with the new summit
Vertex pair.If the v' can not be solved, simply return to be assigned therein as viOr vjOr their intermediate point.Repeating should
Process is until reach the triangle of target numbers.
Two kinds of conclusions are obtained accordingly.Firstly, the formula of QSlim algorithm shows that minimization problem is local.The top collapsed
The optimum position of point is determined by the support planar set of first vertex set, from the support plane inherit to obtain described in collapse
Vertex.The support planar set grows out near plane and the step that extends outwardly.The considerations of lacking to global characteristics, this makes
It is sensitive to local grid noise to obtain QSlim.Secondly, input grid is always subjected to the influence of inevitable noise in MVS step.
Even in input grid, the completely flat surface in actual scene may include noise.
This generates undesirable as a result, the plane domain of grid energetically can not sufficiently simplify and other valuable richnesses
Having the region of feature, there is relatively high priority to be simplified.This is in 3D large-scale reconstruction and scene comprising a large amount of
It is especially prominent when flat surface.
By the enlightenment of above content, in embodiments of the present invention, inventor improves quadratic surface formula so that its
Be conducive to the vertex in simplification and the secant plane region in plane domain to form plane belonging to vertex, thus it is smooth that
Noise in a little flat surfaces.For doing so, the plane domain and its plane equation in grid are detected first.The present invention is implemented
Existing method is optimized in mode, exactly matches it with application scenarios of the present invention, and has obtained the two of following improvement
Secondary surface equation:
α ∈ [0,1] has determined the significance level of plane domain detected and corresponding vertex is forced to be cut to its inspection
The plane measured.β ∈ [0,1] has determined the simplification priority of these plane domains in all feasible vertex pair.In this hair
In the experiment of bright embodiment, α is set as 0.9, β 0.5.The new quadratic surface is initialized when algorithm starts and other
Step does not change.Figure 26 and 27 shows the difference between standard QSlim and embodiment of the present invention modified version.Input net
Lattice have the vertex 1.9M and 3.9M triangle.The two is reduced to the 1% of input according to the number of triangle.
Above-mentioned experiment confirm embodiment of the present invention proposes newly design and show QSlim really and lead to vertex more
Add and be uniformly distributed, and the improved method of embodiment of the present invention can more energetically simplify plane domain, rich in feature
Region (for example, vehicle, tree, elbow and edge) leaves more details.
Experiment
Previously described improved grid operations used below, it is enterprising by large-scale Grid dataset on a single machine
Row experiment designs to assess the LOD of embodiment of the present invention.
<input data>
Large data collection is named as Toulouse, is the city in the French west and south.The data set passes through 3D reconstructing system
It generates.It has been cut into 128 × 64 subregion grid in advance, is made of 8044 available partitions.It has the top 407M in total
Point and the face 798M and the size for constituting 31GB in total in memory.They generate the 7th grade by LOD process is supplied to
The grid of LOD grid to the 0th grade.It has 4 cross partitions hole and their size is very big.Figure 28 is shown the 5th
Data set in grade LOD, be not filled by cross partitions hole so that reader to understand what input data set appears to such general
It reads.6th grade or the 7th grade too greatly so that it cannot load is in grid browser.
<starting>
A part that the LOD process of embodiment of the present invention is implemented in C++ and is integrated into 3D reconstructing system.Process
Flowing by python Script controlling.When generating next stage LOD, 3 steps for each 2 × 2 grid are handled (that is, merging, filling out
Fill hole, simplification) operation parallel.In order to fill the hole of cross partitions, be added to after generating the 7th grade of LOD additional step with
Detect its filling.
The purpose of embodiment of the present invention is that the LOD grid that mean size is 200KB is generated in all LOD levels.Cause
This, in each level from the 7th grade to the 0th grade setting simplify than for 0.25 to compensate quaternary tree design.Because of input file
Mean size be 4MB, so simplifying the 7th grade of input setting than being 0.05.
In embodiments of the present invention, the machine of operation test is equipped with i7-470K CPU, band there are four processing core and
Support 8 threads and 32GB RMA.
<result>
Total runing time of process is 21,681 seconds or about 6 hours.Here is experimental summary.
Table 2: the experimental summary of large data collection
Figure 29 shows the 5th grade of LOD grid of Toulouse, and Figure 30 shows the 4th grade of LOD grid of Toulouse, figure
31 show the 3rd level LOD grid of Toulouse, and Figure 32 shows the 2nd grade of LOD grid of Toulouse, and Figure 33 is shown
The 1st grade of LOD grid of Toulouse, Figure 34 show the 0th grade of LOD grid of Toulouse.
The experimental results showed that triangle, non-manifold edge and the top do not decayed by all grids that LOD process generates
Point.
In conclusion the invention proposes LOD grids to generate process, for the 3D reconstruction technique by rapid growth come band
Growing demand for be robust (robust) and expansible.It optimizes three main grid operations, institute
Grid operations are stated to be mainly used for process, the fixed challenging grid flaw generated by aforementioned reconstruction procedures, merge grid
It does not generate without overlay requirements or decaying triangle, be suitable for City scenarios grid by general simplified algorithm optimization.In large size
Experiment on data set shows it with robustness and scalability.
Through the above description of the embodiments, those skilled in the art can be understood that the present invention can be by
The mode of software combination hardware platform is realized.Based on this understanding, technical solution of the present invention makes tribute to background technique
That offers can be embodied in the form of software products in whole or in part, which can store is situated between in storage
In matter, such as ROM/RAM, magnetic disk, CD, including some instructions use is so that a computer equipment (can be individual calculus
Machine, server or network equipment etc.) execute method described in certain parts of each embodiment of the present invention or embodiment.
Skilled person should be appreciated that disclosed above is only embodiments of the present invention, certainly cannot be with
This interest field to limit the present invention still belongs to the claims in the present invention institute according to equivalent variations made by embodiment of the present invention
The range covered.
Claims (20)
1. a kind of method for handling grid in LOD model generating process, which is characterized in that the described method includes:
Obtain initial mesh;
The initial mesh is sheared to obtain multiple LOD subregions;
The LOD subregion of current level is sutured with combined grid by mesh update;
It is lattice simplified to the combined grid application to obtain the LOD subregion of next level;
The mesh update is executed to the LOD subregion iteration of next level until obtaining final combined grid;
The method also includes carrying out hole filling to hole existing for combined grid so that adjacent mesh surface is adapted;
Wherein, the hole filling includes that the first filling is handled:
The grid patch of topological structure of the building with known hole;
The border vertices in the hole on target gridding are mapped to the border vertices of the grid patch;
The position of the border vertices is used to carry out smoothly as constraint condition to internal vertex.
2. the method as described in claim 1, which is characterized in that the hole filling further includes the second filling processing:
Initial Triangulation Approach is carried out to grid patch;
Trigonometric ratio is refined by increasing internal vertex;
Increased internal vertex is placed on to the position for being adapted patch surface and adjacent surface.
3. method according to claim 2, which is characterized in that carrying out Initial Triangulation Approach to grid patch includes:
Define two error functions of triangle: WAreaAnd WDihedral angle-angle, wherein WAreaRefer to the area of triangle, WDihedral angle-angleRefer to
With the maximum dihedral angle angle of adjacent triangle;
Definition grid patch error function W is WAreaTotal minimal error or WDihedral angle-angleMaximum value minimal error;
Initial Triangulation Approach is completed by minimizing W.
4. method according to claim 2, which is characterized in that the second filling processing further include:
The surface in the hole smoothly newly filled,
Minimize the Normal variations between border vertices and increased internal vertex.
5. method according to claim 2, which is characterized in that the first filling processing is for being greater than or equal to predetermined threshold
Hole, it is described second filling processing for less than the predetermined threshold hole.
6. the method as described in claim 1, which is characterized in that the shearing includes:
Define secant plane;
All triangles intersected with the secant plane are searched in the initial mesh;
The triangle of the intersection is separated into small triangle along shear line;
It is as test point that all small triangle marks is flat in the cutting by using the geometric center of the small triangle
The either side in face, and remove unwanted triangle in the side.
7. the method as described in claim 1, which is characterized in that the mesh update includes:
The merging gap for cutting away predetermined offset along the combined side of grid to be combined to establish between grid to be combined;
Merge the minimum distance border vertices between grid by matching two along the combined side of grid to be combined to establish wait close
And the bridge between grid;
Hole is carried out to the hole between the grid to be combined to fill to complete to merge.
8. the method for claim 7, which is characterized in that the mesh update further include:
Along bridge described in Boundary Traversal, and the sequence of bridge is recorded, checks discontinuous sequence,
The bridge intersected by the sequential search of bridge, and abandon the bridge of the intersection.
9. the method as described in claim 1, which is characterized in that carried out based on QSlim algorithm using following Quadratic Surface Equations
It is described lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates the Zhi Chiping of v
Face set, KpIndicate that vertex to the squared-distance matrix for supporting plane, α ∈ [0,1] has determined plane area detected in grid
The significance level in domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)It indicates to push up in grid
It puts to the squared-distance matrix of affiliated characteristic plane.
10. a kind of 3D method for reconstructing characterized by comprising
The point cloud of object is obtained by multi-angle of view stereo reconstruction MVS;
Surface mesh is obtained from described cloud;
Based on the surface mesh of the acquisition, LOD grid is generated using method described in claim 1 to 9 any one;
It is rebuild using the 3D that the LOD grid of generation completes the object.
11. a kind of system for handling grid in LOD model generating process, which is characterized in that the system comprises:
Module is obtained, for obtaining initial mesh;
Shear module obtains multiple LOD subregions for shearing the initial mesh;
Merging module, for being sutured the LOD subregion of current level with combined grid by mesh update;
Simplify module, for lattice simplified to obtain the LOD subregion of next level to the combined grid application;
Wherein, the mesh update is executed to the LOD subregion iteration of next level until obtaining final combined grid;
The system also includes carry out hole filling to hole existing for combined grid so that the filling that adjacent mesh surface is adapted
Module;
The filling module is for executing the first filling processing:
The grid patch of topological structure of the building with known hole;
The border vertices in the hole on target gridding are mapped to the border vertices of the grid patch;
The position of the border vertices is used to carry out smoothly as constraint condition to internal vertex.
12. system as claimed in claim 11, which is characterized in that the filling module is also used to the second filling processing:
Initial Triangulation Approach is carried out to grid patch;
Trigonometric ratio is refined by increasing internal vertex;
Increased internal vertex is placed on to the position for being adapted patch surface and adjacent surface.
13. system as claimed in claim 12, which is characterized in that carrying out Initial Triangulation Approach to grid patch includes:
Define two error functions of triangle: WAreaAnd WDihedral angle-angle, wherein WAreaRefer to the area of triangle, WDihedral angle-angleRefer to
With the maximum dihedral angle angle of adjacent triangle;
Definition grid patch error function W is WAreaTotal minimal error or WDihedral angle-angleMaximum value minimal error;
Initial Triangulation Approach is completed by minimizing W.
14. system as claimed in claim 12, which is characterized in that the second filling processing further include:
The surface in the hole smoothly newly filled,
Minimize the Normal variations between border vertices and increased internal vertex.
15. system as claimed in claim 12, which is characterized in that the first filling processing is for being greater than or equal to predetermined threshold
The hole of value, the second filling processing is for the hole less than the predetermined threshold.
16. system as claimed in claim 11, which is characterized in that the shearing includes:
Define secant plane;
All triangles intersected with the secant plane are searched in the initial mesh;
The triangle of the intersection is separated into small triangle along shear line;
It is as test point that all small triangle marks is flat in the cutting by using the geometric center of the small triangle
The either side in face, and remove unwanted triangle in the side.
17. system as claimed in claim 11, which is characterized in that the mesh update includes:
The merging gap for cutting away predetermined offset along the combined side of grid to be combined to establish between grid to be combined;
Merge the minimum distance border vertices between grid by matching two along the combined side of grid to be combined to establish wait close
And the bridge between grid;
Hole is carried out to the hole between the grid to be combined to fill to complete to merge.
18. system as claimed in claim 17, which is characterized in that the mesh update further include:
Along bridge described in Boundary Traversal, and the sequence of bridge is recorded, checks discontinuous sequence,
The bridge intersected by the sequential search of bridge, and abandon the bridge of the intersection.
19. system as claimed in claim 11, which is characterized in that based on QSlim algorithm using following Quadratic Surface Equations into
Row is described lattice simplified:
Wherein, QvIt is expressed as quadratic surface, v indicates vertex, and p indicates the support plane on vertex, and planes (v) indicates the Zhi Chiping of v
Face set, KpIndicate that vertex to the squared-distance matrix for supporting plane, α ∈ [0,1] has determined plane area detected in grid
The significance level in domain, β ∈ [0,1] have determined the simplification priority of plane domain, Kplanar region plane(v)It indicates to push up in grid
It puts to the squared-distance matrix of affiliated characteristic plane.
20. a kind of 3D reconstructing system characterized by comprising
First acquisition unit obtains the point cloud of object by multi-angle of view stereo reconstruction MVS;
Second acquisition unit obtains surface mesh from described cloud;
Generation unit generates LOD using method described in claim 1 to 9 any one based on the surface mesh of the acquisition
Grid;
Reconstruction unit is rebuild using the 3D that the LOD grid of generation completes the object.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610882819.4A CN106384386B (en) | 2016-10-08 | 2016-10-08 | Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610882819.4A CN106384386B (en) | 2016-10-08 | 2016-10-08 | Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106384386A CN106384386A (en) | 2017-02-08 |
CN106384386B true CN106384386B (en) | 2019-05-03 |
Family
ID=57937158
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610882819.4A Active CN106384386B (en) | 2016-10-08 | 2016-10-08 | Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106384386B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP3467783A1 (en) | 2017-10-06 | 2019-04-10 | Thomson Licensing | Method and device for hole filling of a point cloud |
CN107908880A (en) * | 2017-11-17 | 2018-04-13 | 浙江文瑞科技发展有限公司 | Face abstracting method in the complex thin-walled member model of guarantor's topology based on virtual dividing |
CN108877476B (en) * | 2018-05-31 | 2020-05-22 | 北京金阳普泰石油技术股份有限公司 | Hierarchical virtual grid acquisition method, device and system |
CN112188201A (en) * | 2019-07-03 | 2021-01-05 | 腾讯美国有限责任公司 | Method and device for encoding video stream, electronic equipment and storage medium |
CN111753451A (en) * | 2020-06-23 | 2020-10-09 | 中国水利水电科学研究院 | Unstructured grid splitting and merging method suitable for water conservancy relevant numerical simulation |
CN112370777A (en) * | 2020-11-12 | 2021-02-19 | 网易(杭州)网络有限公司 | Image rendering method and device and electronic equipment |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102044088B (en) * | 2010-11-09 | 2012-09-26 | 广州市城市规划勘测设计研究院 | LOD (level of detail) model quick constructing method for scanning mass scattered point cloud by ground laser in single station |
CN102663801B (en) * | 2012-04-19 | 2015-07-01 | 北京天下图数据技术有限公司 | Method for improving three-dimensional model rendering performance |
CN103886635B (en) * | 2014-04-18 | 2017-02-15 | 重庆市勘测院 | Self-adaption LOD model establishing method based on face clustering |
CN104063903B (en) * | 2014-07-08 | 2016-09-14 | 清华大学 | The tetrahedral grid of three-dimensional entity model generates method and apparatus |
-
2016
- 2016-10-08 CN CN201610882819.4A patent/CN106384386B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN106384386A (en) | 2017-02-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106384386B (en) | Grid processing method, system and 3D method for reconstructing and system in the generation of LOD model | |
Aliaga et al. | Interactive example-based urban layout synthesis | |
JP2013507679A (en) | Method and system capable of 3D printing of 3D object model | |
Lee et al. | Perceptual organization of 3D surface points | |
JP2002501639A (en) | Adaptive mesh refinement method and apparatus | |
JP2002501640A (en) | Adaptive mesh refinement method and apparatus | |
CN107248142B (en) | A kind of cultural relic fragments method for automatically split-jointing | |
CN110533778B (en) | Large-scale image point cloud parallel distributed gridding reconstruction method, system and device | |
Mencl | A graph–based approach to surface reconstruction | |
CN102306397A (en) | Method for meshing point cloud data | |
CN110084894A (en) | Partial enlargement methods of exhibiting, device and the electronic equipment of threedimensional model | |
Garcia-Cantero et al. | NeurotessMesh: a tool for the generation and visualization of neuron meshes and adaptive on-the-fly refinement | |
Zhao et al. | Automatic repair of CityGML LOD2 buildings using shrink-wrapping | |
Wang | Computing on rays: A parallel approach for surface mesh modeling from multi-material volumetric data | |
Liu et al. | Automatic sizing functions for unstructured mesh generation revisited | |
Miranda et al. | Surface mesh regeneration considering curvatures | |
JP5320576B1 (en) | Highland leveling program, dynamic link library and landscape examination device | |
Olsen et al. | Hinged, pseudo-grid triangulation method for long, near-linear cliff analyses | |
Zhao et al. | HSW: Heuristic Shrink-wrapping for automatically repairing solid-based CityGML LOD2 building models | |
Weier et al. | Generating and rendering large scale tiled plant populations | |
Angjeliu et al. | Challenges in modelling complex geometry in historical buildings for numerical simulations | |
KR101751687B1 (en) | Interpretation technique of 3-dimensional sequential joint inversion based on model fusion cube | |
KR20100099903A (en) | Method and apparatus for providing 3d model cutting simulation | |
Lizier et al. | Comparing techniques for tetrahedral mesh generation | |
Bærentzen et al. | Reconstruction of a Botanical Tree from a 3D Point Cloud |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |