CN102496184B - Increment three-dimensional reconstruction method based on bayes and facial model - Google Patents

Abstract

The invention discloses an incremental three-dimensional reconstruction method based on Bayesian and panel models, comprising the following steps: Step 1, obtaining a projection matrix corresponding to a two-dimensional image for each viewing angle; Step two, establishing a projection matrix for all two-dimensional images Spherical model, sampling a group of 2D images corresponding to key perspectives; step 3, performing 3D reconstruction based on bins on the 2D images corresponding to the key perspectives to obtain bin cloud; step 4, positioning a new perspective on the spherical model Corresponding 2D image and update the spherical model; step 5, select a subset of bins from the bin cloud; step 6, compare the local 3D surface bin density of the bin subset with the 3D surface bin of the bin cloud Density mean value; Step 7, modeling by Bayesian, so as to realize incremental three-dimensional reconstruction. The invention realizes that incremental reconstruction can be used for real-time three-dimensional reconstruction and multi-resolution reconstruction in the future, and can update existing relevant three-dimensional models at any point in time.

Description

A kind of increment three-dimensional rebuilding method based on Bayes and bin model

Technical field

The present invention relates to Computerized 3 D visual and rebuild field, a kind of increment three-dimensional rebuilding method based on Bayes and bin model specifically, can increment upgrade three-dimensional model with two dimensional image, thereby finally generate three-dimensional model accurately.

Background technology

Along with the development of computer technology, computing machine has related to and mankind's natural interaction and intelligent application, and therefore computer vision should meet and give birth to.The scene perception is a huge challenge to computing machine.The basis of scene perception is the three-dimensional information that obtains scene, and then can be used for realizing identification and analyze.In current scene cognition technology, the technology of getting up early mainly is partial to obtain instrument with three-dimensional information and is directly obtained three-dimensional scenic, but due to the instrument that obtains, as scanner, these can only process small-sized object (corresponding to large scene), and expensive, so can not apply and come widely.In all environment sensing instruments, video camera and camera utilize in a large number with cheap quilt, but how to allow video camera can obtain three-dimensional scene information from perception data as mankind's eyes, are that computing machine moves towards intelligent the only way which must be passed.

At present three-dimensional reconstruction mainly is based upon all pictorial informations is comprehensively realized to three-dimensional information obtains, however these algorithm practical application, because 1) Data Source in reality is all asynchronous, various mostly.The online picture of picture is clapped from different video cameras mostly, without any rule, and illumination condition differs, and picture uploading more new capital is asynchronous, how from new picture mined information to upgrade the three-dimensional model that existed (rebuild in the past or scan) be a urgent problem.2) can not realize real-time application.Under many environment, as real-time monitoring, in the application of robot, need to carry out comprehensive and analysis in time to the real time environment data.

How to realize that it must be the challenge of next three-dimensional reconstruction aspect that increment type is rebuild.

Summary of the invention

Goal of the invention: technical matters to be solved by this invention is for the deficiencies in the prior art, and a kind of increment three-dimensional rebuilding method based on Bayes and bin model is provided.

For making up the deficiencies in the prior art, the invention discloses a kind of three-dimensional increment method for reconstructing based on Bayes and bin model, comprise following steps:

Step 1, carry out the camera parameter demarcation to the two dimensional image under one group of different visual angles of input, obtains the projection matrix of the corresponding two dimensional image in each visual angle;

Step 2, set up spherical model to all two dimensional images, one group of two dimensional image corresponding to crucial visual angle of sampling, and the trigonometric ratio three-dimensional point corresponding to two dimensional image of sampling; Described three-dimensional point i.e. the corresponding point of two dimensional image on spherical model that visual angle is corresponding;

Step 3, the three-dimensional reconstruction that corresponding two dimensional image carries out based on bin to described crucial visual angle obtains bin cloud S _initial;

Step 4, location two dimensional image i corresponding to new visual angle on described spherical model _newand spherical model is upgraded;

Step 5, according to two dimensional image i _newposition on spherical model, from bin cloud S _initialin choose a bin subset P _update;

Step 6, relatively bin subset P _updatemiddle partial 3 d surface bin density and bin cloud S _initialthree-dimensional surface bin density mean value, use synthetic a small amount of samples method expansion bin subset P _update;

Step 7, carry out modeling by Bayes, according to maximum a posteriori to bin subset P _updateupgraded, thereby realized the increment three-dimensional reconstruction.

In step 1 of the present invention, adopt sparse bundle to adjust method two dimensional image carried out to the camera parameter demarcation, obtain projection matrix P corresponding to two dimensional image under each visual angle,

$P=\left[\begin{array}{cccc}{p}_{11}& p_{12}& p_{13}& p_{14}\\ p_{21}& p_{22}& p_{23}& p_{24}\\ p_{31}& p_{32}& p_{33}& p_{34}\end{array}\right],$

Wherein projection matrix P is the real matrix of 3*4.Wherein sparse bundle adjustment method refers to Manolis I.A.Lourakis, Antonis A.Argyros. " SBA:A software package for generic sparse bundle adjustment " .TOMS, vol.36, no.1, pp.1-30,2009.

In step 2 of the present invention, set up spherical model and be: the corresponding two dimensional image for any one visual angle, the normalized vector that the coordinate that makes its corresponding point on spherical model is optical axis vector N, wherein optical axis vector N=(p ₃₁p ₃₂p ₃₃) ^t, p ₃₁, p ₃₂, p ₃₃correspond respectively to the first three columns element of its corresponding projection matrix P the third line; The method of two dimensional image corresponding to one group of crucial visual angle of sampling is: in three values of interval [0,1] stochastic sampling as reference point (v ₁, v ₂, v ₃), find the point nearest with described reference point point Euclidean distance as the three-dimensional point that once sampling obtains on spherical model, X-Y scheme corresponding to described three-dimensional point becomes two dimensional image corresponding to crucial visual angle; Three-dimensional point corresponding to two dimensional image crucial visual angle on spherical model is corresponding by the Delaunay triangulation carried out trigonometric ratio.

In step 3 of the present invention, two-dimension image rebuild corresponding to crucial visual angle that adopts the three-dimensional rebuilding method based on the bin model to obtain step 2 obtains bin cloud S _initial, method for reconstructing is referring to Y.Furukawa and J.Ponce. " Accurate, dense, and robust multiview stereopsis " .PAMI, vol.32, no.8, pp.1362-1376,2010.In the bin model, three-dimensional surface is to be covered by a series of bins, and each local tangential plane is a bin, in model, by regular three-dimensional rectangle, means.To one of them bin p, it includes three attribute: c (p), n (p), R (p).C (p) is bin p centre coordinate; N (p) is normal vector, and direction is pointed to observation point, for weighing surface local curvature; R (p) is the two dimensional image that bin p is corresponding, it has following attribute: two dimensional image R (p) is a two dimensional image in image collection V (p), V (p) is the two dimensional image set that a bin p determines, every two dimensional image in described set can unobstructedly show the projection of bin p fully; The plane of delineation of the correspondence of two dimensional image R (p) is parallel with the section of bin p.The direction on two limits of three-dimensional rectangle accomplishes that wherein the direction on a limit is as far as possible parallel with the x direction of principal axis in camera coordinate system as far as possible, and the rectangle topology size is that its projection in R (p) is no more than the u*u pixel of pressing the axle arrangement, is made as in the present invention 5*5.

Determine the two dimensional image i corresponding to new visual angle of input in step 4 of the present invention by following formula _newthe triangle T of correspondence on spherical model:

$T\←\underset{T}{\arg \max }{\mathrm{\Σ}}_{v\∈T}\left|x_{i_{\mathrm{new}}}^v\right|,$

The summit that wherein v is a triangle T in the middle of spherical model,

two dimensional image i _newthe two dimensional image corresponding with vertex v obtains the match point set by the conversion of yardstick invariant features.That is: T is one and two dimensional image i _newthe triangle that has maximum coupling amount;

Using after the center-of-mass coordinate normalization of triangle T as two dimensional image i _newthree-dimensional point coordinate on spherical model, be designated as point (x, y, z), point (x, y, z) is connected in twos to the spherical model after being upgraded with three summits of triangle T.

In step 5 of the present invention, from initialization bin S set _initialin choose and two dimensional image i _newthe bin subset that correlativity is the highest.This correlativity is embodied in: two dimensional image i _newin can see this bin and bin section and two dimensional image i _newplane of delineation angle less.Bin subset P _updateaccording to following formula, obtain:

$P_{\mathrm{update}}=\underset{v\∈T}{\∪}\left\{p\right|p\∈S_{\mathrm{initial}},\mathrm{visR}\left(p\right)\}.$

In step 6 of the present invention, to bin subset P _updateset is expanded, and spread step takes full advantage of two dimensional image i _newpixel Information and the geological information of three-dimensional model, and can allow the three-dimensional surface bin distribute as far as possible evenly, accomplish that the information that takes full advantage of new input picture goes out some new bins in the three-dimensional surface area extension of low resolution.Spread step is as follows: to bin cloud S _initialin any one bin p calculate local density, by the neighbours' bin quantity D in neighbours' bin set N (p) of bin p _pits local density, the neighbours' bin quantity D of replacing of equal value _paccount form as follows:

N(p)={p′|p′∈S _initial,|(c(p)-c(p'))·n(p)|+|(c(p)-c(p′))·n(p′)|<ρ}，

D _p=|N(p)|，

Wherein ρ is threshold values; ρ corresponds to the number of pixels β in two dimensional image R (p) depth distance by calculating bin p and bin p ′ center automatically determines, is that its depth distance that is 2 pixels during bin p and bin p ′ center correspond to two dimensional image R (p) is multiplied by 2 in the present invention.

By to all at bin cloud S _initialin the D of local density of bin _pask arithmetic mean to calculate bin cloud S _initialthree-dimensional surface bin density mean value D _g; To bin subset P _updatein arbitrary bin p, if the D of local density _pbe less than 1/2nd three-dimensional surface bin density mean value D _g, adopt synthetic minority oversampler method to expand the bin k that makes new advances between bin p and neighbours' bin p.Wherein synthetic minority oversampler method refers to Nitesh V.Chawla, Kevin W.Bowyer, Lawrence O.Hall and W.Philip Kegelmeyer. " SMOTE:Synthetic Minority Over-sampling Technique " .JAIR, vol.16, pp.321-357,2002.

In step 7 of the present invention, by following formula to bin subset P _updaterenewal realizes Bayes's increment three-dimensional modeling:

$p\left(S\right|i_{\mathrm{new}})=\frac{1}{Z}p\left(i_{\mathrm{new}}\right|S)p\left(S\right),S\&Element;\mathrm{\&Omega;},$

Wherein S is real three-dimensional model, i _newthe two dimensional image in step 4, p (S|i _new) be that three-dimensional model S is at two dimensional image i _newunder posterior probability, Z is normaliztion constant, the probability space that Ω is three-dimensional model, by probability space Ω dimensionality reduction to the bin subset P in step 5 _updateon.The level and smooth priori that Probability p (S) is three-dimensional model; p(i _new| S) be two dimensional image i _newlikelihood probability, for weighing three-dimensional model S and two dimensional image i _newthe likelihood degree, be expressed as:

p(i _new|S)∝exp(-ηE _p)，

$E_p=\frac{1}{\left|S\right|}\underset{p\&Element;\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000118628770000014.png,HDA0000118628770000021.png"\; state="\{\{state\}\}">S</figure-callout>}}{\mathrm{\&Sigma;}}\frac{1}{\left|V\left(p\right)\right|-1}\underset{i\&Element;V\left(p\right)/i_{\mathrm{new}}}{\mathrm{\&Sigma;}}h(p,i_{\mathrm{new}},i),$

E wherein _pfor energy function, for weighing the accuracy of the arbitrary bin p of three-dimensional model S in its visible two dimensional image, the variation of the variation of the normal vector of arbitrary bin and bin topology information all can have it indirectly to measure reflection on image.Described accuracy by arbitrary bin p at two dimensional image i _newand the correlativity h between the projection in two dimensional image i (p, i _new, i) determine, wherein i is the two dimensional image in image collection V (p), and η is control variable, and the h calculation procedure is as follows: cover the grid of a u*u on bin p, in the present invention, sizing grid is 5*5; By in the bilinear interpolation computing grid, each is put at two dimensional image i _newwith the projection in two dimensional image i; By the 1 normalization positive correlation amount that deducts grid projection in two width two dimensional images.As can be seen here, during entirely accurate, the h between arbitrary two pictures is 0, last E _pbe also 0 to reach minimum, but because the recovery of measuring error and bin p is a Reverse Problem, so E _palways be greater than 0, work as E _pwhen smaller, illustrate that, under existing measurement environment, bin p is more accurate.η is control variable, in the present invention, is 0.5.

Priori p (S) weighs the level and smooth degree of three-dimensional surface, has showed to a certain extent the geological information of three-dimensional surface.Priori is by energy function E ₁with energy function E ₂be expressed as:

p(S)∝exp(-{λE ₁+ζE ₂})，

Energy function E wherein ₁for weighing the flatness of three-dimensional surface, the curvature with the bin part in the present invention changes to weigh local slickness.Energy function E ₂weigh the divorced degree of bin at whole three-dimensional surface, because at E ₁in only usage vector carry out the level and smooth degree of presentation surface, but for level and smooth between some normal vectors, but coordinate drops on the bin of three-dimensional surface outside, to there is threshold values to control, can play like this purpose of filtering unusual bin, and accomplish truly level and smooth.Energy function E ₁with energy function E ₂computing method are:

${\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA0000118628770000014.png,HDA0000118628770000021.png"\; state="\{\{state\}\}">E</figure-callout>}}_1=\frac{1}{\left|S\right|}\underset{p\&Element;\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000118628770000014.png,HDA0000118628770000021.png"\; state="\{\{state\}\}">S</figure-callout>}}{\mathrm{\&Sigma;}}\frac{1}{\left|N\left(p\right)\right|}\underset{n\&Element;N\left(p\right)}{\mathrm{\&Sigma;}}f(p,v),$

$f(p,v)=\sqrt{{(n\left(p\right)-n\left(v\right))}^T(n\left(p\right)-n\left(v\right))},$

$E_2=\frac{1}{\left|S\right|}\underset{p\&Element;S}{\mathrm{\&Sigma;}}\frac{1}{N\left(p\right)}\underset{v\&Element;N\left(p\right)}{\mathrm{\&Sigma;}}d(p,v),$

$d(p,v)=|n\left(p\right)\&CenterDot;(c\left(v\right)-c\left(p\right))|,$

Wherein f (p, v) be bin p and bin v normal vector between Euclidean distance, d (p, v) is bin p and the absolute value distance of bin v on normal vector n (p), λ, ζ is two control parameters, is respectively 0.3,0.2 in the present invention.

Maximize posteriority p (S|i _new) obtain the maximum likelihood three-dimensional model, for P _updatein bin parameter upgraded; Described parameter is the (P at probability space Ω _update) in three-dimensional coordinate and the normal vector of each bin.Finally obtain the solution of a convergence, be equivalent to:

c(p),n(p)←argmax（exp(-{λE ₁+ζE ₂+ηE _p}),p∈P _update。

Get negative logarithm operation, it is: c (p), n (p) ← argmin (λ E ₁+ ζ E ₂+ η E _p), p ∈ P _update, by the set P of opposite unit _updatethe bin cloud of renewal after finally being upgraded.

In the present invention, adopt conjugate gradient to be optimized problem solving.In order further to reduce dimension, in optimization problem, bin p centre coordinate c (p) only moves on the line of initial point and projection centre, like this three-dimensional coordinate space has been reduced to one dimension, has reduced degree of freedom in the present invention; Normal vector n (p) replaces with two Eulerian angle are approximate simultaneously.So each bin p only uses the three degree of freedom modeling, has increased optimization speed.When inputting new two dimensional image, the present invention turns back to again step 4 and carries out.

Beneficial effect: the present invention has effectively integrated the three-dimensional geometric information reconstructed and the two-dimensional image information newly added by Bayesian frame, break through existing three-dimensional reconstruction algorithm and only focused on image information, and can not realize the leak that increment is rebuild, for the three-dimensional applications in the reality such as in the future real-time, asynchronous reconstruction has been opened a good head, for computer realization intelligence contributes.The present invention can be used as the core technology of a system of exploitation, play the part of the renewal process that transfers data to three-dimensional model from video camera between unmanned car steering, three-dimensional fitting, in Smart Home, large-scale city modeling, so that next step analysis, application intelligence provides reliable data.Due to view data source require low, so increased to a great extent robustness.

The accompanying drawing explanation

Below in conjunction with the drawings and specific embodiments, the present invention is done further and illustrates, above-mentioned and/or otherwise advantage of the present invention will become apparent.

Fig. 1 is spherical model schematic diagram of the present invention.

The schematic diagram that Fig. 2 is bin model modeling three-dimensional surface of the present invention.

Fig. 3 is the schematic diagram that the present invention expands new bin.

Fig. 4 is the present invention and E ₁the schematic diagram that priori is relevant.

Fig. 5 is the present invention and E ₂relevant schematic diagram.

Fig. 6 a1～Fig. 6 c5 is the experimental result schematic diagram under three embodiment data sets of the present invention.

Fig. 7 is that the present invention carries out schematic flow sheet

Embodiment

As shown in Figure 7, the present invention comprises following steps: step 1, the two dimensional image under one group of different visual angles of input is carried out to the camera parameter demarcation, and obtain the projection matrix of the corresponding two dimensional image in each visual angle; Step 2, set up a spherical model to all two dimensional images, one group of two dimensional image corresponding to crucial visual angle of sampling, and the trigonometric ratio three-dimensional point corresponding to two dimensional image of sampling; Described three-dimensional point i.e. the corresponding point of two dimensional image on spherical model that visual angle is corresponding; Step 3, the three-dimensional reconstruction that corresponding two dimensional image carries out based on bin to described crucial visual angle obtains bin cloud S _initial; Step 4, the two dimensional image i that new visual angle is corresponding in location on spherical model _newand spherical model is upgraded; Step 5, according to two dimensional image i _newposition on spherical model, from bin cloud S _initialin choose a bin subset P _update; Step 6, relatively bin subset P _updatemiddle partial 3 d surface bin density and bin cloud S _initialthree-dimensional surface bin density mean value, use synthetic a small amount of samples method expansion bin subset P _update; Step 7, carry out modeling by Bayes, according to maximum a posteriori to bin subset P _updateupgraded, thereby realized the increment three-dimensional reconstruction.

Below in conjunction with accompanying drawing, the present invention is done to detailed introduction.

In step 1, adopt sparse bundle to adjust method two dimensional image carried out to the camera parameter demarcation, obtain projection matrix P corresponding to two dimensional image under each visual angle,

$P=\left[\begin{array}{cccc}{p}_{11}& p_{12}& p_{13}& p_{14}\\ p_{21}& p_{22}& p_{23}& p_{24}\\ p_{31}& p_{32}& p_{33}& p_{34}\end{array}\right],$

Wherein projection matrix P is the real matrix of 3*4.Wherein sparse bundle adjustment method refers to Manolis I.A.Lourakis, Antonis A.Argyros. " SBA:A software package for generic sparse bundle adjustment " .TOMS, vol.36, no.1, pp.1-30,2009.

In step 2, set up spherical model and be: the corresponding two dimensional image for any one visual angle, the normalized vector that the coordinate that makes its corresponding point on spherical model is optical axis vector N, wherein optical axis vector N=(p ₃₁p ₃₂p ₃₃) ^t, p ₃₁, p ₃₂, p ₃₃correspond respectively to the first three columns element of its corresponding projection matrix P the third line; The method of two dimensional image corresponding to one group of crucial visual angle of sampling is: in three values of interval [0,1] stochastic sampling as reference point (v ₁, v ₂, v ₃), find the point nearest with described reference point Euclidean distance as the three-dimensional point that once sampling obtains on spherical model, X-Y scheme corresponding to described three-dimensional point becomes two dimensional image corresponding to crucial visual angle; The quantity of the three-dimensional point of sampling is generally 1/3rd of data set size; Three-dimensional point corresponding to two dimensional image crucial visual angle on spherical model is corresponding by the Delaunay triangulation carried out trigonometric ratio.The spherical model that Fig. 1 is trigonometric ratio, from figure to seeing that a three-dimensional point represents a two dimensional image; Two dimensional image corresponding to crucial visual angle forms triangle by Triangulation Algorithm and covers sphere.

In step 3, the three-dimensional rebuilding method of employing based on the bin model obtains image reconstruction corresponding to crucial visual angle to step 2 and obtains bin cloud S _initial, method for reconstructing is referring to Y.Furukawa and J.Ponce. " Accurate, dense, and robust multiview stereopsis " .PAMI, vol.32, no.8, pp.1362-1376,2010.In the bin model, what three-dimensional surface was approximate covers with local tangential plane, and as shown in Figure 2, each local tangential plane is a bin, in model, by regular three-dimensional rectangle, means.To one of them bin p, it includes three attribute: c (p), n (p), R (p).C (p) is bin p centre coordinate; N (p) is normal vector, and direction is pointed to observation point, for weighing surface local curvature; R (p) is the two dimensional image that bin p is corresponding, it has following attribute: two dimensional image R (p) is a two dimensional image in image collection V (p), V (p) is the visual angle image collection that a bin p determines, every two dimensional image in described set can unobstructedly show the projection of bin p fully; The plane of delineation of the correspondence of two dimensional image R (p) is parallel with the section of bin p.The direction on two limits of three-dimensional rectangle accomplishes that wherein the direction on a limit is as far as possible parallel with the x direction of principal axis in camera coordinate system as far as possible, and the rectangle topology size is that its projection in picture is no more than the u*u pixel of pressing the axle arrangement, is made as in the present invention 5*5.

Step 4 is determined the two dimensional image i corresponding to new visual angle of input by following formula _newthe triangle T of correspondence on spherical model: the summit that wherein v is a triangle T in the middle of spherical model, two dimensional image i _newthe two dimensional image corresponding with vertex v obtains the match point set by the conversion of yardstick invariant features.T is one and two dimensional image i in fact _newthe triangle that has maximum coupling amount; Using after the center-of-mass coordinate normalization of triangle T as two dimensional image i _newthree-dimensional point coordinate on spherical model, be designated as point (x, y, z), point (x, y, z) is connected in twos to the spherical model after being upgraded with three summits of triangle T.

Can see new visual angle i in Fig. 1 _newinterconnect with leg-of-mutton three summits, obtained the spherical model after a renewal.

In step 5, from initialization bin S set _initialin choose and two dimensional image i _newthe bin subset that correlativity is the highest.This correlativity is embodied in: two dimensional image i _newin can see this bin and bin section and two dimensional image i _newplane of delineation angle less.Bin subset P _updateaccording to following formula, obtain:

$P_{\mathrm{update}}=\underset{v\&Element;T}{\&cup;}\left\{p\right|\&Element;S_{\mathrm{initial}},\mathrm{visR}\left(p\right)\}.$

In step 6 of the present invention, to bin subset P _updateset is expanded, and spread step takes full advantage of two dimensional image i _newpixel Information and the geological information of three-dimensional model, and can allow the three-dimensional surface bin distribute as far as possible evenly, accomplish to take full advantage of pictorial information and go out some new bins in the three-dimensional surface area extension of low resolution.Spread step is as follows: to bin cloud S _initialin any one bin p calculate local density, by the neighbours' bin quantity D in neighbours' bin set N (p) of bin p _pits local density, the neighbours' bin quantity D of replacing of equal value _paccount form as follows:

N(p)={p′|p′∈S _initial,|(c(p)-c(p'))·n(p)|+|(c(p)-c(p′))·n(p′)|<ρ}，

D _p=|N(p)|，

The depth distance that wherein ρ corresponds to the number of pixels β in two dimensional image R (p) by calculating bin p and bin p ′ center automatically determines, is that its depth distance that is 2 pixels during bin p and bin p ′ center correspond to two dimensional image R (p) is multiplied by 2 in the present invention; By to all at bin cloud S _initialin the D of local density of bin _pask arithmetic mean to calculate bin cloud S _initialthree-dimensional surface bin density mean value D _g; To bin subset P _updatein arbitrary bin p, if the D of local density _pbe less than 1/2nd three-dimensional surface bin density mean value D _g, adopt synthetic minority oversampler method to expand the bin k that makes new advances between bin p and neighbours' bin p.As can be seen from Figure 3, p ₀and p ₁between a random newly-generated new bin p on the linear space of line _new, its coordinate and normal vector are according to p _newposition in line segment is to p ₀and p ₁consistent attribute weight on average obtain.Wherein synthetic minority oversampler method refers to Nitesh V.Chawla, Kevin W.Bowyer, Lawrence O.Hall and W.Philip Kegelmeyer. " SMOTE:Synthetic Minority Over-sampling Technique " .JAIR, vol.16, pp.321-357,2002..

In step 7 of the present invention, by following formula, replacement problem is carried out to Bayes Modeling:

$p\left(S\right|i_{\mathrm{new}})=\frac{1}{Z}p\left(i_{\mathrm{new}}\right|S)p\left(S\right),S\&Element;\mathrm{\&Omega;},$

Wherein S is real three-dimensional scenic, i _newthe two dimensional image in step 4, p (S|i _new) be that three-dimensional model S is at two dimensional image i _newunder posterior probability, Z is normaliztion constant, the probability space that Ω is three-dimensional model, we are on its dimensionality reduction to one bin subset, i.e. bin subset P in step 5 _update.The level and smooth priori that Probability p (S) is model; p(i _new| S) be two dimensional image i _newlikelihood probability, for weighing three-dimensional model S and two dimensional image i _newthe likelihood degree, be expressed as:

p(i _new|S)∝exp(-ηE _p)，

$E_p=\frac{1}{\left|S\right|}\underset{p\&Element;S}{\mathrm{\&Sigma;}}\frac{1}{V\left(p\right)}\underset{i\&Element;V\left(p\right)}{\mathrm{\&Sigma;}}h(p,i_{\mathrm{new}},i),$

E wherein _pfor energy function, for weighing the accuracy of the arbitrary bin p of three-dimensional model S in its visible two dimensional image, the variation of the variation of the normal vector of arbitrary bin and bin topology information all can have it indirectly to measure reflection on image.Described accuracy by arbitrary bin p at two dimensional image i _newand the correlativity h between the projection in two dimensional image i (p, i _new, i) determine, wherein i is the two dimensional image in image collection V (p), and η is control variable, and the h calculation procedure is as follows: cover the grid of a u*u on bin p, in the present invention, sizing grid is 5*5; By in the bilinear interpolation computing grid, each is put at two dimensional image i _newwith the projection in two dimensional image i; By the 1 normalization positive correlation amount that deducts grid projection in two width two dimensional images.As can be seen here, during entirely accurate, the h between arbitrary two pictures is 0, last E _pbe also 0 to reach minimum, but because the recovery of measuring error and bin p is a Reverse Problem, so E _palways be greater than 0, work as E _pwhen smaller, illustrate that, under existing measurement environment, bin p is more accurate.η is control variable, in the present invention, is 0.5.

Priori P (S) weighs the level and smooth degree on 3D surface, has excavated to a certain extent the geological information of three-dimensional surface.It passes through energy function E priori ₁with energy function E ₂be expressed as:

p(S)∝exp(-{λE ₁+ζE ₂})，

E wherein ₁for weighing the flatness of three-dimensional surface, the curvature with the bin part in the present invention changes to weigh local slickness.E ₁can not accurately weigh its surface smoothness, in the present Fig. 4 of its defect body, be curvature although bin p and neighbours on every side have level and smooth normal vector, because it has broken away from original fit surface, so it is a divorced bin.So propose E in the present invention ₂weigh the divorced degree of bin p at whole three-dimensional surface, the bin that coordinate is dropped on to the three-dimensional surface outside has a threshold values to control, and can play like this purpose of filtering singular point, and accomplishes truly level and smooth.Energy function E ₁with energy function E ₂computing method are:

${\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA0000118628770000014.png,HDA0000118628770000021.png"\; state="\{\{state\}\}">E</figure-callout>}}_1=\frac{1}{\left|S\right|}\underset{p\&Element;\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HDA0000118628770000014.png,HDA0000118628770000021.png"\; state="\{\{state\}\}">S</figure-callout>}}{\mathrm{\&Sigma;}}\frac{1}{\left|N\left(p\right)\right|}\underset{n\&Element;N\left(p\right)}{\mathrm{\&Sigma;}}f(p,v),$

$f(p,v)=\sqrt{{(n\left(p\right)-n\left(v\right))}^T(n\left(p\right)-n\left(v\right))},$

$E_2=\frac{1}{\left|S\right|}\underset{p\&Element;S}{\mathrm{\&Sigma;}}\frac{1}{N\left(p\right)}\underset{v\&Element;N\left(p\right)}{\mathrm{\&Sigma;}}d(p,v),$

d(p,v)=|n(p)·(c(v)-c(p))|，

Wherein f (p, n) be bin p and bin n normal vector between Euclidean distance, mark and can see from Fig. 5, d (p, v) is bin p and the absolute value distance of bin v on normal vector n (p); λ, ζ is two control parameters, is respectively 0.3,0.2 in the present invention.

Maximize posteriority p (S|i _new) obtain the maximum likelihood three-dimensional model, for P _updatemiddle bin parameter is upgraded; Described parameter is at probability space P _updatein three-dimensional coordinate and the normal vector of each bin.Finally obtain the solution of a convergence, be equivalent to:

c(p),n(p)←argmax（exp(-{λE ₁+ζE ₂+ηE _p}),p∈P _update

Get negative logarithm operation, that is:

c(p),n(p)←argmin(λE ₁+ζE ₂+ηE _p),p∈P _update

Like this by the set P of opposite unit _updatethe bin cloud of renewal after finally being upgraded.

In the present invention, adopt conjugate gradient to be optimized problem solving.In order further to reduce dimension, in optimization problem, bin p centre coordinate c (p) only moves on the line of initial point and projection centre, like this three-dimensional coordinate space has been reduced to one dimension, has reduced degree of freedom in the present invention; Normal vector n (p) replaces with two Eulerian angle are approximate simultaneously.So each bin p only uses the three degree of freedom modeling, has increased optimization speed.

When inputting new picture, algorithm turns back to again step 4 and carries out.

Embodiment

As shown in Figure 7, the step of the present embodiment comprises: the two dimensional image to all inputs carries out the camera parameter demarcation by sparse bundle adjustment method, then calculate the optical axis vector of two dimensional image corresponding to each visual angle, and it is carried out to normalization obtain three-dimensional point, set up according to this spherical model, wherein the three-dimensional point on spherical model represents a two dimensional image that visual angle is corresponding; Adopt random algorithm to select at random two dimensional image corresponding to crucial visual angle, the selection repeated is disregarded; Adopt triangle gridding subdivision algorithm to carry out trigonometric ratio to them three-dimensional point corresponding to two dimensional image of selecting to obtain; The two dimensional image corresponding to crucial visual angle carries out the three-dimensional reconstruction based on bin.Initial work completes, and then enters the increment link.Read in two dimensional image corresponding to new visual angle, with yardstick invariant features Transformation Matching algorithm and spherical model, find a triangle, and select according to this bin subset P _update; Spherical model is upgraded; Then with synthetic minority oversampler method expansion P _update, the bin made new advances in sparse area extension; Finally by solving an optimization problem to bin subset P under Bayesian frame _updateupgrade the renewal that reaches the general three model.Along with constantly reading in of new images, the step in the middle of the increment link is constantly carried out.

Fig. 6 a1～Fig. 6 c5 is the experimental result (owing to being picture, can only adopt the performance of gray scale form) listed under three data sets.Fig. 6 a1 is the dinosaur image data set, Fig. 6 a2 uses two dimensional image corresponding to crucial visual angle obtained from dinosaur image data set sampling to carry out the bin cloud design sketch that the dinosaur 3-dimensional reconstruction based on the bin model obtains, and Fig. 6 a3～Fig. 6 a5 is for constantly reading in the bin cloud design sketch of incremental update after two-dimentional dinosaur image corresponding to new visual angle.Fig. 6 b1 is the skull image data set, Fig. 6 b2 uses two dimensional image corresponding to crucial visual angle obtained from skull image data set sampling to carry out the bin cloud design sketch that the skull three-dimensional reconstruction based on the bin model obtains, and Fig. 6 b3～Fig. 6 b5 is for constantly reading in the bin cloud design sketch of incremental update after two-dimentional skull image corresponding to new visual angle.Fig. 6 c1 is the temple image data set, Fig. 6 c2 uses two dimensional image corresponding to crucial visual angle obtained from temple image data set sampling to carry out the bin cloud design sketch that the temple three-dimensional reconstruction based on the bin model obtains, and Fig. 6 c3～Fig. 6 c5 is for constantly reading in the bin cloud design sketch of incremental update after temple two dimensional image corresponding to new visual angle.Can find out, bin milks up, and more and more accurate.In these three embodiment, the present invention has obtained a reasonable result.

The invention provides a kind of increment three-dimensional rebuilding method based on Bayes and bin model; method and the approach of this technical scheme of specific implementation are a lot; the above is only the preferred embodiment of the present invention; should be understood that; for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.In the present embodiment not clear and definite each ingredient all available prior art realized.

Claims (5)

1. A method for incremental three-dimensional reconstruction based on Bayesian and surfel models, characterized in that, comprising the following steps: Step 1, perform camera parameter calibration on a set of input two-dimensional images under different viewing angles, and obtain a projection matrix corresponding to two-dimensional images for each viewing angle; Step 2: Establish a spherical model for all 2D images, sample a set of 2D images corresponding to key perspectives, and triangulate the 3D points corresponding to the sampled 2D images; the 3D points are the 2D points corresponding to a perspective The corresponding point of the image on the spherical model; Step 3, perform bin-based three-dimensional reconstruction on the two-dimensional image corresponding to the key viewing angle to obtain bin cloud S _initial ; Step 4, locate a two-dimensional image i _new corresponding to a new viewing angle on the ball model and update the ball model; Step five, according to the position of the two-dimensional image i _new on the spherical model, select a surfel subset P _update from the bin cloud S _initial ; Step 6, comparing the local three-dimensional surface bin density in the bin subset P _update with the average value of the three-dimensional surface bin density of the bin cloud S _initial , and using a synthetic minority sampling method to expand the bin subset P _update ; Step 7: Modeling is performed by Bayesian, and the surface element subset P _update is updated according to the maximum a posteriori method, so as to realize incremental three-dimensional reconstruction; The expansion steps of the face element subset P _update in step six are as follows: Calculate the local density for any bin p in the bin cloud S _initial , and replace its local density with the number of neighbor bins D p in the neighbor bin set N(p) of the bin _p equivalently, the number of neighbor bins D _p is calculated as follows: N(p)={p′|p′∈S _initial ,|(c(p)-c(p'))·n(p)|+|(c(p)-c(p′))·n (p′)|<ρ}, _Dp = |N(p)|, c(p) is the 3-dimensional geometric center coordinate of surface element p, n(p) is the normal vector of surface element p, wherein the direction of the normal vector points to the direction of the observation point, and ρ is the threshold value; Calculate the three-dimensional surface bin density average value D _g of the bin cloud S _initial by calculating the arithmetic mean of the local density D _p of all bins in the bin cloud S _initial ; For any surface element p in the surface element subset P _update , if the local density D _p is less than half of the three-dimensional surface surface element density average value D _g , use the synthetic minority oversampling method between the surface element p and the neighboring surface elements Extend a new surface element k; In step 7, the Bayesian incremental 3D modeling is realized by updating the panel subset P _update through the following formula: $\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; new</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; Z</span>}}\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; new</span>}}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; \&Omega;</span>}<span\; class="notranslate">\; ,</span>$ Where S is the real 3D model, i _new is the 2D image in step 4, p(S|i _new ) is the posterior probability of the 3D model S under the 2D image i _new , Z is the normalization constant, Ω is the probability space of the 3D model, reduce the dimensionality of the probability space Ω to the surface element subset P _update in step 5, p(S) is the smooth prior probability of the 3D model S; p(i _new |S) is the 2D image The likelihood probability of i _new is used to measure the likelihood of the 3D model S and the 2D image i _new , expressed as: p(i _new |S)∝exp(-ηE _p ), ${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; V</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; new</span>}}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; new</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$ Where E _p is an energy function, which is used to measure the accuracy of any surface element p in the 3D model S in its visible two-dimensional image, and the accuracy is determined by any surface element p in the two-dimensional image i _new and the two-dimensional image i The correlation between the projections of h(p,i _new ,i) is determined, where i is a two-dimensional image in the image set V(p), η is a control variable, and the calculation steps of h(p,i _new ,i) are as follows: Overlay a u*u grid on the surface element p; calculate the projection of each point in the grid in the two-dimensional image i _new and the two-dimensional image i by bilinear interpolation; subtract the grid in the two images by 1 The normalized positive correlation quantity projected in the 2D image; The prior p(S) is expressed by the energy function E ₁ and the energy function E ₂ as: p(S)∝exp(-{λE ₁ +ζE ₂ }), Among them, the energy function _E1 is used to measure the smoothness of the three-dimensional surface, and the energy function _E2 supplements the degree of separation of surface elements on the entire three-dimensional surface. The calculation methods of the energy function _E1 and the energy function _E2 are: ${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$ $\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\sqrt{{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span>$ ${\mathrm{<span\; class="notranslate">\; E.</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; |</span>}\underset{\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$ d(p,v)=|n(p)·(c(v)-c(p))|, Among them, n(p), n(v) correspond to the normal vectors of surface element p and surface element v, c(p), c(v) correspond to the geometric center coordinates of surface element p and surface element v, and f(p, v) is the Euclidean distance between the surface element p and the normal vector of the surface element v, d(p, v) is the absolute value distance between the surface element p and the surface element v on the normal vector n(p), λ, ζ are two Control parameters; Maximize the posterior probability p(S|i _new ) to obtain the maximum likelihood surface model, that is, update P _update to obtain a new three-dimensional model, namely: c(p),n(p)←argmax(exp(-{λE ₁ +ζE ₂ +ηE _p }),p∈P _update , take negative logarithm operation, namely: c(p),n(p)←argmin(λE ₁ +ζE ₂ +ηE _p ),p∈P _update , The updated surface element cloud is finally obtained by updating the surface element set P _update . 2. a kind of incremental three-dimensional reconstruction method based on Bayesian and surfel model according to claim 1, is characterized in that, in step one, adopts sparse beam adjustment method to carry out camera parameter calibration to two-dimensional image, obtains each The projection matrix P corresponding to the two-dimensional image under one viewing angle, $\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{cccc}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 11</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 12</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 13</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 14</span>}}\\ {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; one</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; two</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; three</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; twenty\; four</span>}}\\ {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 31</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 32</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 33</span>}}& {\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; 34</span>}}\end{array}\right]<span\; class="notranslate">\; ,</span>$ Wherein the projection matrix P is a 3*4 real matrix. 3. a kind of incremental three-dimensional reconstruction method based on Bayesian and surfel model according to claim 2, it is characterized in that, in step 2, setting up spherical model is: for the two-dimensional image corresponding to any viewing angle, make The coordinates of its corresponding point on the spherical model are the normalized vectors of the optical axis vector N, where the optical axis vector N=(p ₃₁ p ₃₂ p ₃₃ ) ^T , p ₃₁ , p ₃₂ , and p ₃₃ correspond to It corresponds to the first three columns of elements in the third row of the projection matrix P; The method of sampling two-dimensional images corresponding to a set of key viewing angles is as follows: randomly sample three values in the interval [0,1] as reference points (v ₁ , v ₂ , v ₃ ), and search for points on the spherical model that are consistent with the reference points The point with the closest Euclidean distance is used as a three-dimensional point obtained by one sampling, and the two-dimensional figure corresponding to the three-dimensional point becomes the two-dimensional image corresponding to the key viewing angle; The 3D points corresponding to the 2D images corresponding to the key viewing angles on the spherical model are triangulated by a triangulation algorithm. 4. A kind of incremental three-dimensional reconstruction method based on Bayesian and surfel model according to claim 3, it is characterized in that, in step 4, the two-dimensional image i _new corresponding to the new viewing angle of input is determined by the following formula in The corresponding triangle T on the spherical model: $\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; \&LeftArrow;</span>\underset{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; arg</span>}\mathrm{<span\; class="notranslate">\; max</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{{\mathrm{<span\; class="notranslate">\; i</span>}}_{\mathrm{<span\; class="notranslate">\; new</span>}}}^{\mathrm{<span\; class="notranslate">\; v</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; ,</span>$ Where v is a vertex of a triangle T in the spherical model,

is the two-dimensional image corresponding to the two-dimensional image i _new and the vertex v to obtain a set of matching points through scale-invariant feature transformation; Normalize the centroid coordinates of the triangle T as the three-dimensional point coordinates of the two-dimensional image i _new on the spherical model, which are recorded as points (x, y, z), and the point (x, y, z) and the three-dimensional coordinates of the triangle T Vertices are connected in pairs to obtain an updated ball model. 5. a kind of incremental three-dimensional reconstruction method based on Bayesian and panel model according to claim 4, is characterized in that, the panel subset P _update in step 5 obtains according to following formula: ${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; update</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}}{<span\; class="notranslate">\; \&cup;</span>}<span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; initial</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; visR</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; p</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \}</span><span\; class="notranslate">\; ,</span>$ Among them, R(p) is a two-dimensional image corresponding to surface element p, which has the following properties: two-dimensional image R(p) is a two-dimensional image in the image set V(p), and V(p) is surface element p A set of determined two-dimensional images, each two-dimensional image in the set can fully display the projection of the surface element p without occlusion; the corresponding image plane of the two-dimensional image R(p) is parallel to the tangent plane of the surface element p.

Images