CN101127123A - Sign point hole filling method based on neural network in tri-D scanning point cloud - Google Patents

Abstract

The utility model provides a point cloud hole filling method that is based on three-dimensional scanning of neural networks, ensures the continuity of the filled holes data and the surrounding data, and has better point cloud characteristic. The utility model has the advantage of simple method. The utility model is mainly applied to the application situations of filling holes with various complicated curved surface shapes produced by marking points in three-dimensional scanning system. With the neural network method of the utility model, a network with filled holes is obtained, sampling points in the hole regions are sampled subsequently based on the density of the hole boundary points, further the points of filling holes are adjusted according to the curvature of the points to achieve smooth filling of holes.

Description

In the 3-D scanning point cloud based on the complementing method of the monumented point hole of neural network

Technical field

The present invention relates to a kind of method for repairing and mending, relate in particular in a kind of 3-D scanning point cloud complementing method based on the monumented point hole of neural network to three-dimensional picture.

Background technology

Reverse-engineering (Reverse Engineering, RE) technology is the later stage eighties 20th century to appear at the new technology in the advanced manufacturing field, it generally comprises four basic links: three-dimensional body detects and conversion (acquisition of physical data), data pre-service (put cloud processing, identification, look splicing more), the foundation of cad model (surface reconstruction), the moulding of CAM product, its basic flow sheet as shown in Figure 1.In the process that three-dimensional body detects and changes, by the three-dimensional digital scanner three-dimensional scanning survey is fast carried out on the mock-up surface, under the prerequisite that satisfies the discrete sampling speed and the quality of data, obtain the 3 d-dem data of product, in looking splicing more, a lot of algorithms are used the splicing based on monumented point, and at the object exterior pasting monumented point, the losing of data that must cause the zone that is capped, produce hole, this type of hole has the characteristics of himself: certainly lead to, the position size can be determined, quantity is more, the appearance of dot cloud hole has caused the imperfect of these data, therefore hole is compensated that promptly data to be carried out pre-service be a ring of forming a connecting link very crucial in the reverse-engineering, directly influence the quality of reconstruct success or not and cad model, its follow-up link is played very crucial restrictive function.The present invention mainly relates in the digitized process of reverse-engineering, obtains a kind of automatic complementing method of the monumented point hole of product point cloud model appearance with 3 D scanning system (Fig. 2).

In recent years, the algorithm of filling up of dot cloud hole has all been obtained very big progress at home and abroad, delivered a considerable amount of documents, the some of them algorithm has obtained comparatively widely to use, as the algorithm based on the various hole repairings of energy-optimised and segmentation, surface fitting, triangle grid model.The point cloud model that these algorithms need earlier 3-D scanning directly to be obtained is done certain processing in early stage or dot cloud hole is carried out Boundary Recognition, and real-time is not strong, and complexity is also than higher.In application of practical project, should be able in time solve the problem that occurs.

Document " Minimal energy surfaces using parametric splines. " (Gregory E.Fasshauer, LarryL.Schumarker.ComputerAided Geometric Design, 1996,13:45～79) by finding the solution optimization aim function based on " strain energy of distortion function ", realization is filled up hole, guaranteed certain fairness, yet these class methods need subdivision curved surface mostly, and a plurality of subsurface sheets are spliced, therefore the continuity of surface boundary is had relatively high expectations, be difficult to reach the Second Order Continuous between curved surface usually.Because the related algorithm comparative maturity of existing triangle grid model, existing algorithm all is based on this model mostly, document " Filling holes in meshes using a mechanical model to simulate thecurvature variation minimization. " (Jean-Philippe Pemot for example, George Moraru, Philippe Veron.Computer﹠amp; Graphics, 2006,30:892～902) hole repairing is summed up as the triangulation problem of a space polygon.Behind trigonometric ratio, seek the border of triangular network, automatically detect hole, before perforations adding, at first the border triangle is handled, remove the triangle that direction is bad and degenerate, use the triangular network perforations adding at last, but the method can only be repaired closed hole, and this class algorithm for the cloud data of complexity programming to set up the triangle grid model operand bigger, in filling up the process of hole, also need to change the topological structure of triangle gridding, therefore have triangle grid model revised and the shortcoming of designed capacity deficiency again, complexity height, these shortcomings have all limited its application in practice.Filling algorithm to a class polygon hole is not suitable for general point cloud model.

Filling algorithm for general surface mesh all needs to obtain the hole boundary information accurately, and difficulty is big, is difficult for realizing.Compensate in the algorithmic procedure of hole by scattered point set around the dot cloud hole is carried out surface fitting, the parametrization of point at random is essential.Even parametrization, entad parametrization and accumulation Chord Length Parameterization method are arranged usually, and these methods are primarily aimed at the data point that is topological rectangular array, to the cloud data of random distribution, need sort to it, difficulty is also bigger.

Algorithm based on the neural network perforations adding has also obtained using comparatively widely now, has also obtained effect preferably simultaneously.Document " based on the some cloud incomplete data method for repairing and mending of neural network " (Liu Jun for example, He Jianying. the mechanical engineer, 2007 (2): the BP data network method for repairing and mending that has proposed a kind of three-dimensional incomplete data 32～34), frame is selected the sample point set of incomplete data boundary vicinity, and trained, BP network after the training is used for the prediction at incomplete number of regions strong point, finishes the repairing of data.This class methods accommodation is wider, for nonlinear data, noisy data, the indefinite data of pattern feature reasonable effect is arranged all, but not enough place is arranged: operation is many, each hole, and all the very important person selects the boundary sample point set for frame; The density of repairing area point is with inconsistent on every side; For the big characteristic area of some curved transition, it is good inadequately that feature shows.

In the application process of reality, in 3 D scanning system, introduce monumented point for the looking splicing an of cloud more, just inevitably there is the hole phenomenon this moment, it is crucial that these holes are filled up timely.For this reason, use a kind of new algorithm that the monumented point hole is filled up among the present invention.How to guarantee that the hole data filled up and ambient data are continuously and the difficult point that is more preferably this type of algorithm of character representation.Cloud data density is selected the method for resampling density and can be reached smooth filling to hole for the some cloud of repairing by curvature adjustment around the basis that proposes among the present invention.

Summary of the invention

The invention provides a kind of can guarantee the hole data filled up and ambient data continuously and some cloud character representation better based on the method for filling dot cloud hole of the 3-D scanning of neural network, the present invention has the simple advantage of method.

The present invention adopts following technical scheme:

In a kind of 3-D scanning point cloud based on the complementing method of the monumented point hole of neural network:

The first step: the hole that forms for the monumented point of object exterior pasting, in 3-D scanning point cloud, when obtaining on every side sample point set of hole, according to the three-dimensional coordinate (x of monumented point _b, y _b, z _b), will be that data point in the cube at center is as the sample point P of perforations adding with the monumented point _s(s=0,1 ..., t), cubical 8 summits are respectively (x _b-r, y _b-r, z _b-r), (x _b-r, y _b-r, z _b+ r), (x _b-r, y _b+ r, z _b-r), (x _b-r, y _b+ r, z _b+ r), (x _b+ r, y _b-r, z _b-r), (x _b+ r, y _b-r, z _b+ r), (x _b+ r, y _b+ r, z _b-r), (x _b+ r, y _b+ r, z _b+ r), wherein r is the thresholding of being got, and in order to obtain suitable sample point set, gets 1.2～1.5 times of radius of monumented point;

Second the step: according to sample points according to P _sThe neural network of training perforations adding, the neural network that is used to fill up hole is three layers a reverse transmittance nerve network, its input layer is 2 neurons, the x coordinate figure and the y coordinate figure of the corresponding sample data of difference, hidden layer is a m neuron, and output layer is the z coordinate figure of 1 neuron and corresponding sample data, and the excitation function of hidden layer and output layer is respectively the tanh sigmoid function $f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$ With linear function g (x)=x, and use Levenberg-Marquardt algorithm training counterpropagation network, hidden neuron number m chooses 20, obtains the neural network net of perforations adding by training;

The 3rd step: after obtaining neural network net, choose resample points,, calculate all sample point P for density and its ambient data that incomplete zone is heavily adopted is consistent along step-lengths such as x and y directions _s, s=0 wherein, 1 ..., t, as the step-length L that resamples, the scope of resampling is interval (x at the mean distance of the point of xoy plane projection and on every side minor increment _Min, x _Max), (y _Min, y _Max), x wherein _Min, x _MaxBe minimum value and the maximal value of the coordinate x of sample point, y _Min, y _MaxBe minimum value and the maximal value of the coordinate y of sample point, heavily adopt ading up to a little

Wherein

Expression rounds downwards, and the x coordinate figure of resample points is from x _MinBeginning with step-length L equidistant increase until

The y coordinate figure of resample points is from y _MinBeginning with step-length L equidistant increase until

The x coordinate figure of the point that resamples and y coordinate figure as the input of neural network, just can be obtained resample points z coordinate figure, and then obtain all resample points;

The 4th step: the curvature according to resample points is added sampled point, when asking for discrete space point curvature, all resample points are projected to the xoy plane, by the 3rd step as can be known, all subpoints, equidistantly distribute with spacing L along x and y direction, get ask point in 5 * 5 neighborhoods on the xoy plane a little as the neighborhood point, remove totally 24 point itself,, obtain the curvature ρ of all resample points according to existing method as a little k neighborhood of ask _s(s=0,1 ..., n-1), according to given curvature thresholding ρ _τ, ρ _τGet 0.1, get curvature greater than ρ _τResample points, obtain the deep camber point, around deep camber point, increase sampled point then, the strategy of taking is, all resample points are projected on the xoy plane, supposes that it is (x that the coordinate of deep camber point subpoint on the xoy plane is arranged _ρ, y _ρ), check that it is along x four the subpoint (x adjacent with the y direction _ρ-L, y _ρ), (x _ρ+ L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L), if the deep camber point is wherein also arranged, then two resample points of equidistant increase between deep camber point are without loss of generality, and suppose to project to (x _ρ+ L, y _ρ) point of coordinate is the deep camber point, then increase

2 points, all the other are not on the direction of deep camber point, increase a bit, (the x that projects to as described above _ρ-L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L) 3 are not the deep camber points, then increase

3 points, adopt above strategy to mend point at all deep camber points, at last among the neural network net that the x of the point of all benefits and second step of y coordinate figure input are obtained, the z coordinate figure of asking, all resample points have so just been obtained according to the curvature increase, resample points together with obtaining in the 3rd step obtains final resample points;

The 5th step: with in the final resample points with 3-D scanning point cloud in the distance of closest approach less than 0.5d _3DResample points give up, remaining sampled point is added in the 3-D scanning point cloud the filling up of complement mark point hole, wherein d _3DBe the density of 3-D scanning point cloud, d _3DBe set to the mean value of the each point of a cloud and minor increment on every side.

The present invention is mainly used in the application scenario that the hole of the various complex-curved shapes that produced by monumented point in the 3 D scanning system is filled up.Utilize the neural net method among the present invention, can obtain filling up the network of hole, according to the density of hole frontier point, get sampled point subsequently, and further adjust the point of perforations adding, reach smooth filling hole according to the curvature of point in the hole zone.This method mainly contains following advantage:

(1) gridding of comparing in the But most of algorithms is handled, and the present invention is primarily aimed at the original point cloud model that 3 D scanning system obtains, without any need for gridding handle, applicability is wide, speed is fast.

(2) because monumented point is attached to each different position of object, therefore the hole that forms also is various shape, therefore in the parametric surface fitting complementing method, just can't select to be fit to the parametric surface of all holes, and neural network has very strong functional approximation capability, three-layer network can with arbitrary accuracy approach any continuous function with and all-order derivative, therefore compare with method based on various parametric surface fittings, based on the hole of neural network method, can obtain and better fill up precision the various complex-curved shapes of monumented point formation.

(3) during the triangular network perforations adding, it is closed that the hole of being mended necessarily requires, and this method can both be filled up wide adaptability for the hole of the arbitrary shape that monumented point forms.

(4) hole that forms for monumented point, because the D coordinates value and the size of monumented point are known, therefore hole is discerned automatically according to the monumented point position, obtaining also of sample point finished automatically around the hole, need not any manual operation, after obtaining 3-D scanning point cloud, just can finish filling up of all monumented point holes automatically.

(5) can be according to actual needs, the artificially number of times of Control Training and precision in network training have been avoided unnecessary calculating redundancy.

When (6) getting on the whole, this method resamples to the hole zone according to the density of hole frontier point, has guaranteed the continuity of hole repairing zone with the border.

(7) this method is to the preliminary resample points of hole, curvature according to resample points, near the big point of curvature, add sampled point, make hole in the big place of curved transition, the point of filling up is more, the general algorithm of evenly filling up of comparing more can show the feature of a cloud, and the hole of being filled up is more smooth.

Description of drawings

Fig. 1 is the reverse-engineering process flow diagram.

Fig. 2 is a grating style three-dimension scanning system composition diagram.

Fig. 3 is algorithm overall flow figure.

Fig. 4 is the pictorial diagram that contains the monumented point hole.

Fig. 5 chooses the monumented point cube synoptic diagram of sample point set on every side.

Fig. 6 chooses k neighborhood point set synoptic diagram.

Fig. 7 is that curvature function is found the solution figure.

Fig. 8 increases the sampled point synoptic diagram according to curvature.

Fig. 9 is the algorithm flow chart that increases sampled point according to curvature.

Figure 10 is the positive effect figure of final perforations adding.

Figure 11 is the design sketch of another angle of final perforations adding.

Embodiment

In a kind of 3-D scanning point cloud based on the complementing method of the monumented point hole of neural network:

The first step: the hole that the monumented point of pasting for Venus's head portrait surface forms, in 3-D scanning point cloud, when obtaining on every side sample point set of hole, according to the three-dimensional coordinate (x of monumented point _b, y _b, z _b), will be that data point in the cube at center is as the sample point P of perforations adding with the monumented point _s(s=0,1 ..., t), cubical 8 summits are respectively (x _b-r, y _b-r, z _b-r), (x _b-r, y _b-r, z _b+ r), (x _b-r, y _b+ r, z _b-r), (x _b-r, y _b+ r, z _b+ r), (x _b+ r, y _b-r, z _b-r), (x _b+ r, y _b-r, z _b+ r), (x _b+ r, y _b+ r, z _b-r), (x _b+ r, y _b+ r, z _b+ r), wherein r is the thresholding of being got, and in order to obtain suitable sample point set, gets 1.2～1.5 times of radius of monumented point;

Second the step: according to sample points according to P _sThe neural network of training perforations adding, the neural network that is used to fill up hole is three layers a reverse transmittance nerve network, its input layer is 2 neurons, the x coordinate figure and the y coordinate figure of the corresponding sample data of difference, hidden layer is a m neuron, and output layer is the z coordinate figure of 1 neuron and corresponding sample data, and the excitation function of hidden layer and output layer is respectively the tanh sigmoid function $f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$ With linear function g (x)=x, and use Levenberg-Marquardt algorithm training counterpropagation network, hidden neuron number m chooses 20, and the step-length of network training gets 2000, and training precision gets 0.001, obtains the neural network net of perforations adding by training;

The 3rd step: after obtaining neural network net, choose resample points along step-lengths such as x and y directions, for density and its ambient data that incomplete zone is heavily adopted is consistent, calculates each sample point earlier and put minor increment d at the point of xoy plane projection in the point of xoy plane projection and all the other _i, again with the average of all minimum distances as heavily adopting step-length L, $L=\frac{1}{t+1}\underset{i=0}{\overset{t}{\mathrm{\Σ}}}{d}_i,$ The scope that resamples is interval (x _Min, x _Max), (y _Min, y _Max), x wherein _Min, x _MaxBe minimum value and the maximal value of the coordinate x of sample point, y _Min, y _MaxBe minimum value and the maximal value of the coordinate y of sample point, heavily adopt ading up to a little Wherein

Expression rounds downwards, and the x coordinate figure of resample points is from x _MinBeginning with step-length L equidistant increase until

The y coordinate figure of resample points is from y _MinBeginning with step-length L equidistant increase until

The x coordinate figure of the point that resamples and y coordinate figure as the input of neural network, just can be obtained resample points z coordinate figure, and then obtain all resample points Q _s(s=0,1 ... n-1);

The 4th step: according to resample points Q _s(s=0,1, n-1) curvature is added sampled point, when asking for discrete space point curvature, all resample points are projected to the xoy plane, go on foot as can be known by the 3rd, all subpoints equidistantly distribute with spacing L along x and y direction, get ask point in 5 * 5 neighborhoods on the xoy plane a little as the neighborhood point, remove totally 24 point itself, as ask a little k neighborhood K (Q _s), S (Q _s) be resample points Q _sK neighbor point least square fitting plane, the order Be resample points Q _sK neighbor point set K (Q _s) the centre of form, be called resample points Q _sCentral point, this central point is:

$\stackrel{\‾}{Q}=\frac{1}{(k+1)}\underset{Q_s\∈k\left(Q_s\right)}{\mathrm{\Σ}}{Q}_s$

If resample points Q _sJ neighbor point to least square plane S (Q _s) distance be d _j, arrive

Distance be λ _j, so to resample points Q _sJ point have a function f _j(Q _s), this function f _j(Q _s) be:

$f_j\left(Q_s\right)=\frac{d_j}{{\mathrm{\λ}}_j}$

Resample points Q so _sCurvature function can be expressed as:

$f\left(Q_s\right)=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{f}_j\left(Q_s\right)$

According to curvature function f (Q _s) obtain the curvature ρ of all resample points _s(s=0,1 ..., n-1), according to given curvature thresholding ρ _τ, ρ _τGet 0.1, get curvature greater than ρ _τResample points, obtain the deep camber point, around deep camber point, increase sampled point then, the strategy of taking is, all resample points are projected on the xoy plane, supposes that it is (x that the coordinate of deep camber point subpoint on the xoy plane is arranged _ρ, y _ρ), check that it is along x four the subpoint (x adjacent with the y direction _ρ-L, y _ρ), (x _ρ+ L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L), if the deep camber point is wherein also arranged, then two resample points of equidistant increase between deep camber point are without loss of generality, and suppose to project to (x _ρ+ L, y _ρ) point of coordinate is the deep camber point, then increase

2 points, all the other are not on the direction of deep camber point, increase a bit, (the x that projects to as described above _ρ-L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L) 3 are not the deep camber points, then increase

3 points, adopt above strategy to mend point at all deep camber points, at last among the neural network net that the x of the point of all benefits and second step of y coordinate figure input are obtained, the z coordinate figure of asking, all resample points have so just been obtained according to the curvature increase, resample points together with obtaining in the 3rd step obtains final resample points Q _s' (s=0,1 ..., n-1 ..., n+l-1);

The 5th step: with in the final resample points with 3-D scanning point cloud in the distance of closest approach less than 0.5d _3DResample points give up, remaining sampled point is added in the 3-D scanning point cloud the filling up of complement mark point hole, wherein d _3DBe the density of 3-D scanning point cloud, d _3DBe set to the mean value of the each point of a cloud and minor increment on every side.

With reference to the accompanying drawings, the present invention is described in detail:

In reverse-engineering, consider that the measurement data amount is big and at random, contain non-linear strong characteristics of noise and nerual network technique to the advantage on this class data processing, the hole of the various complex-curved shapes that form for monumented point, we introduce neural network (neural network) technology the hole in measuring are repaired, neural network has very strong functional approximation capability, and three layers of feedforward network can approach any continuous function and all-order derivative thereof with arbitrary accuracy.Adopt neural network to replace traditional curved surface fitting method based on spline base function, the mapping relations of expression curved surface are stored in the connection weights and threshold value of neural network, the information stores pattern of this holographic formula, make model have stronger fault freedom and associative ability, can not have a strong impact on its overall performance because partial nerve unit is impaired; Can seriously not distort output yet, have robustness because input signal is subjected to the pollution of noise to a certain degree.Overall algorithm flow is seen Fig. 3.

The present invention relates generally to the content of following four aspects:

1) the sample point set around the monumented point hole obtains and the training of neural network

The hole that the monumented point of pasting for Venus's head portrait surface forms is seen Fig. 4, in 3-D scanning point cloud, when obtaining on every side sample point set of hole, according to the three-dimensional coordinate (x of monumented point _b, y _b, z _b), will be that data point in the cube at center is as the sample point P of perforations adding with the monumented point _s, s=0 wherein, 1 ..., t, the three-dimensional coordinate on cubical 8 summits is respectively (x _b-r, y _b-r, z _b-r), (x _b-r, y _b-r, z _b+ r), (x _b-r, y _b+ r, z _b-r), (x _b-r, y _b+ r, z _b+ r), (x _b+ r, y _b-r, z _b-r), (x _b+ r, y _b-r, z _b+ r), (x _b+ r, y _b+ r, z _b-r), (x _b+ r, y _b+ r, z _b+ r), wherein r is the thresholding of being got, and gets 1.2～1.5 times of radius of monumented point, sees Fig. 5.

According to sample point set P _sThe neural network of training perforations adding, the Neural Network Toolbox of utilization matlab is set up the BP neural network, network is three layers, and input layer is 2 neurons, respectively the x coordinate figure and the y coordinate figure of corresponding sample data, hidden layer is a m neuron, output layer is 1 neuron, and the z coordinate figure of corresponding sample data, the excitation function of hidden layer and output layer are respectively the tanh sigmoid functions $f\left(x\right)\frac{e^x-e^{-x}}{e^x+e^{-x}}$ With linear function g (x)=x, and use Levenberg-Marquardt algorithm BP network, hidden neuron number m gets 20, and the step-length of network training gets 2000, and training precision gets 0.001, obtains the neural network net of perforations adding by training.

2) resampling of perforations adding data and by the curvature adjustment

The strategy that resamples is after obtaining neural network net, choose resample points along step-lengths such as x and y directions, for density and its ambient data that makes incomplete area resample is consistent, calculate the resampling step-length in accordance with the following methods: calculate earlier each sample point and put minor increment d at the point of xoy plane projection in the point of xoy plane projection and all the other _i, again with the average of all minimum distances as heavily adopting step-length L, $L=\frac{1}{t+1}\underset{i=0}{\overset{t}{\mathrm{\Σ}}}{d}_i.$

The scope that resamples is interval (x _Min, x _Max), (y _Min, y _Max), x wherein _Min, x _MaxBe minimum value and the maximal value of the coordinate x of sample point, y _Min, y _MaxBe minimum value and the maximal value of the coordinate y of sample point,, get ading up to of resample points according to resampling scope and resampling step-length

Wherein

Expression rounds downwards, and the x coordinate figure of resample points is from x _MinBeginning with step-length L equidistant increase until

The y coordinate figure of resample points is from y _MinBeginning with step-length L equidistant increase until

The x coordinate figure of the point that resamples and y coordinate figure as the input of neural network, just can be obtained resample points z coordinate figure, and then obtain all resample points Q _s(s=0,1 ... n-1).

To preliminary resample points, adjust according to curvature, in the bigger place of curved transition, increase some sampled points, the more smooth hole of filling up.

Resample points Q _sThe method of asking spatial discrete points curvature is used in the calculating of curvature, asks k neighborhood K (Q a little earlier _s), owing to equidistantly sample along x and y direction when resampling, therefore the subpoint of all resample points equidistantly distributes with spacing L along x and y direction, get ask point in 5 * 5 neighborhoods on the xoy plane a little as the neighborhood point, see Fig. 6, remove self totally 24 point, gather K (Q as ask k neighbor point a little _s), the Q of match resample points again _sK neighborhood least square fitting planar S (Q _s), order

Be resample points Q _sK neighbor point set K (Q _s) the centre of form, be called resample points Q _sCentral point, this central point is:

$\stackrel{\‾}{Q}=\frac{1}{(k+1)}\underset{Q_s\∈k\left(Q_s\right)}{\mathrm{\Σ}}{Q}_s$

If resample points Q _sJ neighbor point to least square plane S (Q _s) distance be d _j, arrive Distance be λ _j, see Fig. 7, so to resample points Q _sJ point have a function f _j(Q _s), this function f _j(Q _s) be:

$f_j\left(Q_s\right)=\frac{d_j}{{\mathrm{\λ}}_j}$

Resample points Q so _sCurvature function can be expressed as:

$f\left(Q_s\right)=\frac{1}{k}\underset{j=1}{\overset{k}{\mathrm{\Σ}}}{f}_j\left(Q_s\right)$

According to curvature function f (Q _s) obtain the curvature ρ of all resample points _s(s=0,1 ..., n-1).

After obtaining the curvature of all sampled points, according to given curvature thresholding ρ _τ, ρ _τGet 0.1, get curvature greater than ρ _τResample points, obtain the deep camber point, around deep camber point, increase sampled point then, the strategy of taking is, all resample points are projected on the xoy plane, supposes that it is (x that the coordinate of deep camber point subpoint on the xoy plane is arranged _ρ, y _ρ), check that it is along x four the subpoint (x adjacent with the y direction _ρ-L, y _ρ), (x _ρ+ L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L), if the deep camber point is wherein also arranged, then two resample points of equidistant increase between deep camber point are without loss of generality, and suppose to project to (x _ρ+ L, y _ρ) point of coordinate is the deep camber point, then increase

2 points, all the other are not on the direction of deep camber point, increase a bit, (the x that projects to as described above _ρ-L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L) 3 are not the deep camber points, then increase

3 points, see Fig. 8, adopt above strategy to mend point at all deep camber points, at last among the neural network net that the x and the input of y coordinate figure of the point of all benefits trained, try to achieve the z coordinate figure, so just obtained all resample points, added the resampling point set Q that begins to try to achieve according to the curvature increase _s, obtain final resampling point set Q _s' (s=0,1 ..., n-1 ..., n+l-1), (n point is preliminary resampling point set Q wherein _s, the sampled point of back l point for increasing according to the curvature self-adaptation).The algorithm flow chart that increases point according to curvature is seen Fig. 9.

3) finally fill up determining of hole point

Because the scope that resamples is determined by the sample point set, if therefore directly final resample points is added original some cloud, borderline region at hole has the overlay region, for fear of adding redundant points, with in the final resample points with 3-D scanning point cloud in the distance of closest approach less than 0.5d _3DResample points give up, remaining resample points is added in the 3-D scanning point cloud the filling up of complement mark point hole, wherein d _3DBe the density of 3-D scanning point cloud, d _3DBe set to the mean value of the each point of a cloud and minor increment on every side.Some cloud behind the final perforations adding, Figure 10 is positive effect figure, and Figure 11 is the design sketch of another angle, and the rectangle frame zone among the two secondary figure is the zone of perforations adding.

Claims (1)

1. In the 3-D scanning point cloud based on the complementing method of the monumented point hole of neural network, it is characterized in that:

   The first step: the hole that forms for the monumented point of object exterior pasting, in 3-D scanning point cloud, when obtaining on every side sample point set of hole, according to the three-dimensional coordinate (x of monumented point _b, y _b, z _b), will be that data point in the cube at center is as the sample point P of perforations adding with the monumented point _s, s=0 wherein, 1 ..., t, the three-dimensional coordinate on cubical 8 summits is respectively (x _b-r, y _b-r, z _b-r), (x _b-r, y _b-r, z _b+ r), (x _b-r, y _b+ r, z _b-r), (x _b-r, y _b+ r, z _b+ r), (x _b+ r, y _b-r, z _b-r), (x _b+ r, y _b-r, z _b+ r), (x _b+ r, y _b+ r, z _b-r), (x _b+ r, y _b+ r, z _b+ r), wherein r is the thresholding of being got, and gets 1.2～1.5 times of radius of monumented point:

   Second the step: according to sample points according to P _sThe neural network of training perforations adding, the neural network that is used to fill up hole is three layers a reverse transmittance nerve network, its input layer is 2 neurons, the x coordinate figure and the y coordinate figure of the corresponding sample data of difference, hidden layer is a m neuron, and output layer is the z coordinate figure of 1 neuron and corresponding sample data, and the excitation function of hidden layer and output layer is respectively the tanh sigmoid function $f\left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$ With linear function g (x)=x, and use Levenberg-Marquardt algorithm training counterpropagation network, hidden neuron number m chooses 20, obtains the neural network net of perforations adding by training;

   The 3rd step: after obtaining neural network net, choose resample points,, calculate all sample point P for density and its ambient data that incomplete zone is heavily adopted is consistent along step-lengths such as x and y directions _s, s=0 wherein, 1 ..., t, as the step-length L that resamples, the scope of resampling is interval (x at the mean distance of the point of xoy plane projection and on every side minor increment _Min, x _Max), (y _Min, y _Max), x wherein _Min, x _MaxBe minimum value and the maximal value of the coordinate x of sample point, y _Min, y _MaxBe minimum value and the maximal value of the coordinate y of sample point, heavily adopt ading up to a little

   

   Wherein "

   

   Expression rounds downwards, and the x coordinate figure of resample points is from x _MinBeginning with step-length L equidistant increase until

   

   The y coordinate figure of resample points is from y _MinBeginning with step-length L equidistant increase until

   

   With the x coordinate figure and the y of the point that resamples, coordinate figure just can obtain resample points z coordinate figure as the input of neural network, and then obtains all resample points;

   The 4th step: the curvature according to resample points is added sampled point, when asking for discrete space point curvature, all resample points are projected to the xoy plane, by the 3rd step as can be known, all subpoints, equidistantly distribute with spacing L along x and y direction, get ask point in 5 * 5 neighborhoods on the xoy plane a little as the neighborhood point, remove totally 24 point itself,, obtain the curvature ρ of all resample points according to existing method as a little k neighborhood of ask _s, s=0 wherein, 1 ..., n-1 is according to given curvature thresholding ρ _τ, ρ _τGet 0.1, get curvature greater than ρ _τResample points, obtain the deep camber point, around deep camber point, increase sampled point then, the strategy of taking is, all resample points are projected on the xoy plane, supposes that it is (x that the coordinate of deep camber point subpoint on the xoy plane is arranged _ρ, y _ρ), check that it is along x four the subpoint (xs adjacent with the y direction _ρ-L, y _ρ), (x _ρ+ L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L), if the deep camber point is wherein also arranged, then two resample points of equidistant increase between deep camber point are without loss of generality, and suppose to project to (x _ρ+ L, y _ρ) point of coordinate is the deep camber point, then increase

   

   

   2 points, all the other are not on the direction of deep camber point, increase a bit, (the x that projects to as described above _ρ-L, y _ρ), (x _ρ, y _ρ-L), (x _ρ, y _ρ+ L) 3 are not the deep camber points, then increase

   

   

   

   3 points, adopt above strategy to mend point at all deep camber points, at last among the neural network net that the x of the point of all benefits and second step of y coordinate figure input are obtained, the z coordinate figure of asking, all resample points have so just been obtained according to the curvature increase, resample points together with obtaining in the 3rd step obtains final resample points;

   The 5th step: with in the final resample points with 3-D scanning point cloud in the distance of closest approach less than 0.5d _3DResample points give up, remaining resample points is added in the 3-D scanning point cloud the filling up of complement mark point hole, wherein d _3DBe the density of 3-D scanning point cloud, d _3DBe set to the mean value of the each point of a cloud and minor increment on every side.

Images