CN107123164B

CN107123164B - Three-dimensional reconstruction method and system for keeping sharp features

Info

Publication number: CN107123164B
Application number: CN201710151230.1A
Authority: CN
Inventors: 刘琼; 吴铭荃
Original assignee: South China University of Technology SCUT
Current assignee: South China University of Technology SCUT
Priority date: 2017-03-14
Filing date: 2017-03-14
Publication date: 2020-04-28
Anticipated expiration: 2037-03-14
Also published as: CN107123164A

Abstract

The invention discloses a three-dimensional reconstruction method and a three-dimensional reconstruction system for keeping sharp features. The method comprises the following steps: 1) performing smooth denoising on the point cloud which is input and subjected to rough registration by adopting a two-dimensional image filtering method expanded to a three-dimensional space; 2) removing outliers of the smoothed rough registration point cloud by adopting an improved region growing method; 3) performing fine registration on the coarse registration point cloud after the outlier is removed by adopting an ICP algorithm based on kd-tree acceleration; 4) fusing the fine registration point cloud by adopting a fusion method based on neighborhood searching and boundary point detection; 5) and performing surface reconstruction on the fused fine registration point cloud by adopting a method based on feature point detection and self-adaptive step length updating. The three-dimensional reconstruction system realized based on the method can keep sharp characteristics at the edge of the surface of the reconstruction model and also has reconstruction speed on the premise of ensuring precision.

Description

Three-dimensional reconstruction method and system for keeping sharp features

Technical Field

The invention belongs to the field of computer graphics, computer vision, reverse engineering and computer aided geometric design, and particularly relates to a three-dimensional reconstruction method and a three-dimensional reconstruction system for keeping sharp features.

Background

The three-dimensional reconstruction technology is a technology for establishing a proper mathematical representation and processing model on a computer for an object and objectively reproducing a real object in the computer.

According to the difference of the three-dimensional point cloud data acquisition modes, the three-dimensional reconstruction technology is roughly divided into the following two types: three-dimensional reconstruction is performed based on a plurality of images and three-dimensional reconstruction is performed based on point cloud data with depth information. The method comprises the steps of carrying out three-dimensional reconstruction based on a plurality of images, carrying out image acquisition through a camera to obtain sequence images of an object to be reconstructed at different angles, carrying out point cloud estimation through characteristic relation and position relation among the images, and then reconstructing point cloud. The method has the advantages that the used equipment is simple and the cost is low; however, the method has the disadvantages that the point cloud position information can be obtained through a series of operations, the process is complicated, the quality of the reconstruction result is greatly influenced by the angle and the number of the images, the method is sensitive to the shooting environment, and the three-dimensional reconstruction with high precision requirement is difficult to complete. The method comprises the steps of carrying out three-dimensional reconstruction on the basis of point cloud data with depth information, obtaining three-dimensional point cloud with the depth information in the modes of raster scanning, laser scanning and the like, and reconstructing the point cloud data with the depth information through the spatial relationship of the point cloud. The method has the advantages that the point cloud information with higher accuracy can be directly acquired, but expensive and professional equipment is needed, and calibration is needed during data acquisition to meet the requirement of acquiring the accuracy of the point cloud data.

In the process of three-dimensional reconstruction through point cloud, in order to enable measurement and splicing work of a reconstructed model to be carried out smoothly, how to keep sharp features of an object on the surface edge is a key link. Zhang Xin et al (Zhang Xin, King Zhang, Van Gengqi, et al. three-dimensional model of feature preserving trilateral filtering denoising algorithm [ J ]. computer aided design and graphics, 2009,21(7): 936) 942.) maintain the features of the surface of the point cloud by filtering normal vectors in the point cloud using trilateral filtering functions. The main idea of this method is to fit a model of the surface approximation by each point and the points in its neighborhood. The method can generate a better reconstruction effect on a discrete curve fitting model. However, in order to pursue accuracy, the method needs to be iterated continuously, and the fitting process is time-consuming. Sahay P et al (Sahay P, Rajagopalan A N. geometrical input of3D structures [ C ]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition works 2015:1-7.) propose a "hole filling" method for a three-dimensional model of a defect, the region after hole filling can keep the characteristics of the whole model. Such methods deal directly with three-dimensional models while preserving the surface features of the object but they focus primarily on the conformity of the features of the object in the missing regions with the overall model, while not much discussion is given of the features of the object at the edges. Qian J et al (Qian J, Zhang Y. Sharp Feature prediction in Octree-Based Hexaregenerative Mesh Generation for CAD Assembly Models [ M ]// Proceedings of the 19th International Meshing roundable.2009: 243-262.) use an octree-Based approach to preserve features of CAD Models at boundaries. It acts directly on the CAD model, i.e. by changing the shape of the model to create a different model. Although not operating directly in the point cloud, the above-described method provides a reference to the present invention, i.e., the operation of holding features can be performed directly on the model without the need to perform such operation indirectly in the point cloud. Wang J et al (Wang J, Yu Z, Zhu W, et al. feature-Preserving surface reconstruction From Unoriented, noise Point Data [ J ]. Computer Graphics Forum,2013,32(1): 164-: finding out outliers in the point cloud by calculating the most appropriate tangent plane of each point, calculating characteristic points and non-characteristic points of an object, dividing the point cloud into a plurality of parts, and finally reconstructing the point cloud by a method for keeping the characteristics. The method is applied to the surface reconstruction of an object to obtain good effect, however, the characteristic point detection method proposed by the method and the final surface reconstruction have no correlation, and the prior knowledge of the previous processing is not utilized in the surface reconstruction, so that the processing process is more complicated.

In summary, although a three-dimensional reconstruction method for maintaining sharp features based on point cloud has achieved certain results, in order to meet the requirements of practical applications, further improvements in reconstruction time and reconstruction accuracy are urgently needed.

Disclosure of Invention

The embodiment of the invention aims to provide a three-dimensional reconstruction method and a three-dimensional reconstruction system for keeping sharp features, and aims to solve the problems that the sharp features near the edge cannot be completely kept, the reconstruction time needs to be improved, the robustness is not strong and the like in the conventional three-dimensional reconstruction method.

The three-dimensional reconstruction method for maintaining sharp features completes the process of reconstructing a three-dimensional model from point cloud and maintains the sharp features near the surface edges of the model in the reconstruction process through the optimization of the processing of each stage of the point cloud and the improvement of the surface reconstruction process. The method specifically comprises the following steps:

(1) performing smooth denoising on the point cloud which is input and subjected to rough registration by adopting a two-dimensional image filtering method expanded to a three-dimensional space;

(2) removing outliers of the smoothed rough registration point cloud by adopting an improved region growing method;

(3) performing fine registration on the coarse registration point cloud after the outlier is removed by adopting an ICP (Iterative closest point) algorithm accelerated based on a kd-tree (k-dimensional tree);

(4) fusing the fine registration point cloud by adopting a fusion method based on neighborhood searching and boundary point detection;

(5) and performing surface reconstruction on the fused fine registration point cloud by adopting a method based on feature point detection and self-adaptive step length updating.

Further, the step (1) of performing smooth denoising on the point cloud which is input and subjected to coarse registration refers to: the method comprises the steps of applying the smooth filtering method of a two-dimensional image to smooth denoising of three-dimensional point cloud in an expanding mode, taking x and y of point cloud data points (x, y and z) as plane positioning marks, taking z as pixel values in the two-dimensional image, and conducting smooth denoising of the z values on point cloud blocks point by point, wherein the point cloud blocks refer to point clouds collected under the same visual angle; the roughly registered point cloud refers to a point cloud obtained by performing coordinate transformation on a series of point cloud blocks PQ at different shooting angles by adopting a classical coordinate transformation method;

further, the performing z-value smoothing and denoising on the point cloud blocks point by point refers to: and acquiring a z value in a neighborhood with the current (x, y) as a center, and selecting a two-dimensional image smoothing filtering method to smooth the z value according to experience, wherein the neighborhood with the (x, y) as the center is a square, and the side length is selected within 3-15.

Further, the three-dimensional reconstruction method for maintaining sharp features is characterized in that, in the step (2), performing outlier removal on the smoothed coarse registration point cloud by using an improved region growing method refers to: counting the number of points in a point cloud area divided by a kd-tree by adopting a KNN (K-nearest neighbor) clustering mode, and removing outliers by judging the number of points in the current area; further, the outlier refers to an isolated point cloud different from a main point cloud, the main point cloud is a point cloud belonging to the surface of the reconstruction object, the isolated point cloud includes isolated points and a sparse point cloud, the sparse point cloud is a local point cloud having a density less than 40% of the density of the whole point cloud or a preset value, and the sparse point cloud may be closer to the main point cloud; the specific algorithm is as follows:

1) for point cloud P ═ P_i,i＝1,2,3,…,N},p_iAs points in the point cloud, N (N)>1) Calculating P density rho for the point number of the point cloud; p is a radical of_iThe flag value of the flag bit flag is flag num, the initial value flag num is 0, the flag is flag num, the point cloud P is divided by using kd-tree, and i is 0;

2) increasing i by 1;

3) if p is_iIf Flag of (1) is 0, then p is defined_iflagNum +1, Flag ═ flagNum and p_iPush stack S, otherwise return 2);

4) pop up S Stack Top, denoted as p_sUsing a radius search algorithm to find p_sPoint cloud P in a region centered at r and a radius_r(p_j,j＝1,2,…,N_r)，N_rIs P_rThe number of points in; calculating P_rDensity p_r(ii) a If ρ_rP is eliminated if p is less than or equal to 0.4 rho_sI.e. removing sparse points if rho_rJump 6, 0);

5) to P_rPoint P where middle Flag is 0_zero(p_k′,k＝1,2,…,N_zero)，N_zeroIs P_zeroNumber of points in, if N_zeroIs greater than or equal to 1, executing p point by point_k' Flag ═ flagNum, and S was pressed;

6) if S is not empty, returning to 4);

7) if i is less than or equal to N, returning to 2);

8) counting the number Num of point clouds clustered one by one_l＝{l_jJ ═ 1,2, …, flagNum }, when l_j<Th (Th is a preset threshold), and removing the sparse point cloud.

Further, the three-dimensional reconstruction method for keeping sharp features is characterized in that in the step (3), a kd-tree-based accelerated ICP algorithm is adopted to perform fine registration on coarse registration point clouds after outliers are removed, and series point clouds PQ subjected to coarse registration are subjected to fine registration, wherein the specific algorithm is as follows:

1) establishing an RT transformation matrix of the point cloud Q, setting the initial value of the transformation matrix to be 0, and setting i to be 0;

2) increment i by 1, find pairs of neighboring PQs: traversal point cloud P { P_i,i＝1,2,3,…,N_pPoint p in_iPoint-to-point cloud Q { Q_j,j＝1,2,3,…,N_qBuilding kd-Tree, searching off p_iMost recent q_j；

3) Passing point pair series p_iq_jMapping to obtain a rotation matrix R and a translation matrix T;

4) updating an RT transformation matrix of the point cloud Q;

5) calculating the mean square error sigma between PQ after the point cloud Q is subjected to RT conversion, if the sigma is more than or equal to Th (Th is a preset threshold) or the iteration number N is less than or equal to N_max(N_maxPreset maximum number of iterations), return to 2);

6) and finishing point cloud fine registration.

Further, the fusing the point cloud of the fine registration by using the fusion method based on the neighborhood searching and the boundary point detection in the step (4) comprises: the domain searching method comprises the steps of partitioning point cloud through a space grid, and finely partitioning the current neighborhood by adopting a kd-tree; further, the space grid segmentation method is to calculate the minimum external cuboid (namely bounding box) of the current point cloud area to be searched, and then divide the current bounding box into a plurality of small cubes according to the set side length; the set side length is 5% of the shortest side length of the current bounding box;

further, the boundary point detection method projects a neighborhood point set of the current point to a micro-tangent plane of the current point, and judges the maximum value theta of an included angle between the current point and the neighborhood point set_maxWhether or not: theta_max>θ_th(θ_thTo set a threshold), if yes, the current point is considered as a boundary point and marked.

Further, the fusing the fine registration point cloud means removing redundant point cloud in the fine registration point cloud; the redundant point cloud means that the current point cloud P is judged firstly_i(i-1, 2, … …, N-1) whether or not there is an adjacent point cloud P in the vicinity thereof_i+1If the point exists, further judging whether the point belongs to a boundary point, if so, keeping the point to maintain the consistency of the point cloud, otherwise, removing the point as a redundant point; the specific algorithm is as follows:

1) let i equal to 0, j equal to 0, P_i(i＝1,2,……,N-1)；

2) Increase i by 1, P_iRecording as the point cloud P which is spliced at present_cCalculate P_cSurround theA box, and the bounding box is divided into a plurality of small cubes by using a space grid;

3) increment j by 1, traverse P_iPoint V in_j(j＝1,2,……,M)；

4) If V_jIs not at P_cIn a bounding box, V_jIncorporation of P_cAnd if j is less than or equal to M, returning to 3);

5) judgment V_jSet of points V' (not containing V) in the small cube and the adjacent small cubes (27 in total)_j) Whether or not at P_cIn the bounding box, if not, V_jIncorporation of P_cAnd if j is less than or equal to M, returning to 3);

6) to P_cCreate kd-Tree with V_jSearching membership P in neighborhood with circle center as r as radius_cSet of points of V' (without V)_j) Counting the number l, if l is less than or equal to Th (Th is a preset threshold), V_jIncorporation of P_cAnd if j is less than or equal to M, returning to 3);

7) judgment V_jWhether or not there is a set of boundary points V' (without V) in the neighborhood_j) If present, V_jIncorporation of P_cOtherwise remove V_jAnd if j is less than or equal to M, returning to 3);

8) i < N, return 2);

9) and finishing point cloud fusion.

Further, the step (5) of performing surface reconstruction on the fused fine registration point cloud by using a feature point detection and adaptive step size updating method includes: 1) an improved poisson surface reconstruction algorithm; 2) a sharp feature point detection algorithm; 3) and updating the coordinates of the sharp peaks of the triangular meshes.

Further, the improved poisson surface reconstruction algorithm refers to: and removing the triangular patch of the expansion area on the premise of keeping the hole filling effect of the classical Poisson surface reconstruction algorithm. The specific algorithm is as follows:

1) aiming at the point cloud P, a triangular mesh (V) is obtained by adopting a classic Poisson three-dimensional surface reconstruction algorithm_i&T_j,i＝1,3,……,N_v,j＝1,2,……,N_t}，V_iRepresenting the vertices of a triangular mesh, T_jFor patches in a triangular mesh, N_vFor vertices in a triangular meshNumber, N_tThe number of patches in the triangular mesh;

2) setting V point by point_iThe flag removeFlag is false, and j is 0;

3) increasing j by 1, performing grid division on P, and calculating T_j＝(V_j1,V_j2,V_j3) The grid location;

4) finding T_jGrid and T_jWhether a point in P exists in the adjacent grids (totally 27 small cubes) or not is judged, and if not, (V) is carried out one by one_j1,V_j2,V_j3) The flag bit removeFlag is true, and if j is less than or equal to N_tReturn to 3);

5) clustering triangular grid points with removeFlag as true by using a region growing method, and counting the number of areas by regions_numIf there is a region: area thereof_num≤Th_rec(Th_recIs a preset threshold value), the current point of true is reset to false;

6) filtering T to mesh one by one_jIf there is a T containing true_j(V_j1,V_j2,V_j3) Vertex, then remove T_j。

Further, the sharp feature point detection algorithm refers to: detecting and marking sharp edge points of the current point cloud P, wherein the normal vector distribution difference of the micro-plane where the middle points of the sharp edge point set P' are located is obvious; the specific algorithm is as follows:

1) setting the P topology structure by using kd-tree;

2) increase i by 1, and add P to any point in P_iAs the center of a circle, r is the radius search point set P_nei(not containing p)_i)，P_neiMiddle arbitrary point p_j(j＝1,2,…,N_neiIn which N is_neiIs P_neiNumber of medium normal vectors) of normal vectors η_j，P_neiThe normal vector set of all points is denoted as gamma_neiI.e. η_j∈Γ_nei，p_iNormal vector gamma of_iAnd η_jIs theta_tj(ii) a The radius r is set according to 1.5 times of the density P;

3) calculating all included angles theta_tj(j-1, 2, …)Standard deviation sigma_tIf σ is_t≤σ_hold(σ_holdPreset threshold), return 2), otherwise, consider p_iPossibly a sharp edge point;

4) calculating theta_tjMaximum included angle θ of_maxAt a minimum angle theta_minIf (theta)_max-θ_min)≤Δθ_hold(Δθ_holdPreset threshold), return 2), otherwise, consider p_iPossibly a sharp edge point;

5) will theta_tjMapping to unit spherical surface and carrying out Gaussian clustering, if the category is 1, the current region belongs to a flat region, otherwise, p_iAt sharp edges or sharp corner points and marking.

Further, a kd-tree is established on the P ', the triangular mesh is searched point by point to find out the vertex near the P', and the vertex is marked as a sharp edge point in the triangular mesh.

Further, the algorithm for updating the coordinates of the sharp vertices of the triangular mesh refers to: aiming at the obtained triangular mesh sharp points, quickly and accurately adjusting the coordinates of the marked triangular mesh sharp points through a normal vector bilateral filtering algorithm and self-adaptive step iteration, and recovering the surface of an object with sharp features; the specific algorithm is as follows:

1) mesh { V } for the reconstruction surface triangular mesh_i&T_j,i＝1,3,……,N_v,j＝1,2,……,N_tTriangular patch T with sharp point marked_j′(j＝1,2,……,N_loc)，N_locUsing bilateral filtering algorithm (1) to process T by T for the number of triangular patches containing sharp points_j′(V_j1′,V_j2′,V_j3') obtaining N_locA normal vector N_j′(j＝1,2,……,N_loc)：

Wherein i is 1,2, …, c_jAnd c_iRespectively representing the barycentric position coordinates of the current and neighborhood triangular patches,

is a standard Gaussian function of (| cj-ci |), σ_cIs the standard deviation of (| cj-ci |),

is (N)_j·(N_j-N_i) Standard Gaussian function of), N_j·(N_j-N_i) Representing the projection, σ, of the difference between the normal vectors of adjacent triangular meshes on the current triangular mesh normal vector_dIs (N)_j·(N_j-N_i) Standard deviation of);

2) defining a matching function P (T) where l is 0 and loop times_l) To obtain the matching degree of the updated point and the current normal vector, the formula (2) is shown, wherein, the subscript F of sigma represents any triangle patch in the mesh, sigma represents the sum of F in the traversal mesh and the formula in the formula (2), (v) represents the sum of F in the traversal mesh and the formula in the formula (2)_i′，v_j′，v′_k) Setting an initial value P (T) of P (T) for three points (i, j, k represent three different vertexes of any triangular patch and have no numerical significance) in the triangular patch₀) For larger values, the step length λ is λ₀(λ₀Is a preset step length);

P(T_l)＝∑_F((N_j′·(v_i′-v_j′))²+(N_j′·(v_i′-v′_k))²+(N_j′·(v′_k-v_j′))²) (2)

3) increasing l by 1, calculating P (T) from equation (2)_l)；

4) Updating the coordinates of the marked sharp points in the mesh by the formula (3);

v_j′＝v_j+λ∑_j∈fN_f′.N_f′·(v_i-v_j)+N_f′.N_f′·(v_k-v_j) (3)

wherein f is a point v_jAssociated triangular patch, (v)_i，v_j，v_k) Three points in the associated triangular patch. I.e. point v_jNormal vector is N_f'stretching in the direction of the triangular patch'

5) If P (T)_l)≤P(T_l-1) And then returns to 3),

6) l-1, if λ <0.04, making λ - λ/2, otherwise, λ -0.02, and returning to 4);

7)ΔP(V)≥P_thand l is less than or equal to l_max(P_thThreshold value representing the difference between the current value and the last value determined by equation (2) < CHEM >, < CHEM >_maxRepresenting the maximum number of cycles), return to 3);

8) and (6) ending.

The invention also provides a system for realizing the three-dimensional reconstruction method for keeping sharp features, which comprises the following steps: the system comprises a point cloud preprocessing module, a point cloud merging module and a surface reconstruction module;

the point cloud preprocessing module removes noise of different point clouds by adopting a smooth denoising method, clusters the point clouds by using an improved region growing method, and removes outliers in the point clouds according to a clustering result; the point cloud preprocessing module corresponds to the step (1) and the step (2) in the three-dimensional reconstruction method;

the point cloud merging module is used for registering and fusing the preprocessed plurality of point clouds, and comprises a method based on kd-tree acceleration for registering the point clouds and a boundary point detection method used in point cloud fusion; the point cloud merging module corresponds to the step (3) and the step (4) in the three-dimensional reconstruction method;

the surface reconstruction module carries out surface reconstruction on the fused whole point cloud, including using the characteristic point to detect

The method for measuring and self-adapting step length updating accelerates the process of keeping the sharp characteristics of the surface of the model; the point cloud is combined

And the module corresponds to step (5) in the three-dimensional reconstruction method.

Compared with the existing three-dimensional reconstruction technology, the three-dimensional reconstruction method for keeping the sharp features provided by the invention has the following advantages and effects:

the outlier of the point cloud is removed in the preprocessing process through an improved region growing method, so that point cloud data which are more in line with the real shape of an original object can be generated, and the influence caused by noise generated by the outlier in the subsequent point cloud registration process can be reduced; the registered point clouds are fused based on a neighborhood searching and boundary point detecting method, so that redundant point clouds can be removed more accurately, and the point cloud density and the calculation cost consumed by subsequent surface reconstruction are reduced; the fused point cloud is subjected to surface reconstruction with sharp edge characteristics by using a method based on characteristic point detection and self-adaptive step length updating, higher reconstruction precision can be achieved in a shorter time, and simultaneously, the step length can be reduced according to comparison of a matching function under the condition that a larger initial value is taken as the step length, so that the error condition can not occur in the process of iterating point coordinates. The invention is suitable for three-dimensional reconstruction of point clouds with rich sharp features at edges, such as point clouds of industrial metal parts acquired through reverse engineering. In the system, in the process of processing hundreds of thousands of point clouds, the time for removing outliers is about 2.5 seconds, the time for fusing the point clouds is about 3 seconds, and the time for reconstructing the surface is about 8.5 seconds, so that the practical application requirement can be well met.

Drawings

FIG. 1 is a flow chart of a three-dimensional reconstruction method for maintaining sharp features according to an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a three-dimensional reconstruction system with sharp features maintained according to an embodiment of the present invention; in the figure: A. a point cloud preprocessing module; B. a point cloud merging module; C. a surface reconstruction module;

FIG. 3 is a flowchart of a method for removing outliers in a point cloud according to an embodiment of the present invention;

FIG. 4a is a distribution diagram of origin clouds before removing isolated point clouds in an embodiment of the invention;

FIG. 4b is a distribution of the point clouds with the isolated point clouds removed in accordance with an embodiment of the present invention;

FIG. 5 is a graph illustrating the operational efficiency of an outlier algorithm provided by an embodiment of the present invention;

FIG. 6 is a flow chart of a point cloud fusion method provided by an embodiment of the invention;

FIG. 7a is a first block of point cloud distribution map before point cloud fusion in an embodiment of the present invention;

FIG. 7b is a second distribution plot of the point clouds of the embodiment of the present invention before point cloud fusion;

FIG. 7c is a point cloud distribution diagram after point cloud fusion according to an embodiment of the present invention;

FIG. 8 is a diagram illustrating an embodiment of operating efficiency of a point cloud fusion algorithm provided by an embodiment of the present invention;

fig. 9 is a flowchart of a surface reconstruction method for maintaining sharp features according to an embodiment of the present invention.

FIG. 10a is a cloud distribution diagram of an origin before surface reconstruction in an embodiment of the present invention;

FIG. 10b is a triangular mesh pattern diagram after surface reconstruction using a modified Poisson reconstruction algorithm in an embodiment of the present invention;

fig. 10c is a triangular mesh style diagram after surface reconstruction using a reconstruction algorithm that preserves sharp features in an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The application of the principles of the present invention will be further described with reference to the accompanying drawings and specific embodiments.

As shown in fig. 1, the three-dimensional reconstruction method for maintaining sharp features of the embodiment of the present invention includes the following steps:

s101, performing smooth denoising on the point cloud which is input and subjected to rough registration by adopting a two-dimensional image filtering method expanded to a three-dimensional space;

s102, removing outliers of the smoothed rough registration point cloud by adopting an improved region growing method;

s103, performing fine registration on the coarse registration point cloud after the outlier is removed by adopting an ICP (inductively coupled plasma) algorithm based on kd-tree acceleration;

s104, fusing the fine registration point cloud by adopting a fusion method based on neighborhood searching and boundary point detection;

and S105, performing surface reconstruction on the fused fine registration point cloud by adopting a characteristic point detection-based and self-adaptive step length updating method.

Step S101, performing smooth denoising on the point cloud which is input and subjected to rough registration: the method comprises the steps of applying the smooth filtering method of a two-dimensional image to smooth denoising of three-dimensional point cloud in an expanding mode, taking x and y of point cloud data points (x, y and z) as plane positioning marks, taking z as pixel values in the two-dimensional image, and conducting smooth denoising of the z values on point cloud blocks point by point, wherein the point cloud blocks refer to point clouds collected under the same visual angle; the roughly registered point cloud refers to a point cloud obtained by performing coordinate transformation on a series of point cloud blocks PQ at different shooting angles by adopting a classical coordinate transformation method;

step S101, performing z-value smoothing and denoising on the point cloud blocks point by point: and acquiring a z value in a neighborhood with the current (x, y) as a center, and selecting a two-dimensional image smoothing filtering method to smooth the z value according to experience, wherein the neighborhood with the (x, y) as the center is a square, and the side length is selected within 3-15.

Step S102, the step of removing outliers from the smoothed rough registration point cloud by adopting an improved region growing method is as follows: counting the number of points in the point cloud area divided by the kd-tree by adopting a KNN clustering mode, and judging the number of points in the current area to remove outliers; further, the outlier refers to an isolated point cloud different from a main point cloud, the main point cloud is a point cloud belonging to the surface of the reconstruction object, the isolated point cloud includes isolated points and a sparse point cloud, the sparse point cloud is a local point cloud having a density less than 40% of the density of the entire point cloud or a preset value, and the sparse point cloud may be closer to the main point cloud.

And S103, performing fine registration on the coarse registration point cloud after the outlier is removed by adopting a kd-tree-based accelerated ICP algorithm, and performing fine registration on the series of point cloud PQ subjected to coarse registration.

The step S104 of fusing the point cloud of the fine registration by adopting a fusion method based on neighborhood searching and boundary point detection comprises the following steps: the domain searching method comprises the steps of partitioning point cloud through a space grid, and finely partitioning the current neighborhood by adopting a kd-tree; further, the space grid segmentation method is to calculate a minimum external cuboid (namely a bounding box) of a current point cloud area to be searched, and then divide the current bounding box into a plurality of small cubes according to a set side length; the set side length is 5% of the shortest side length of the current bounding box.

Step S104, the boundary point detection method projects a neighborhood point set of the current point to a micro-tangent plane of the current point, and judges the maximum value theta of an included angle between the current point and the neighborhood point set_maxWhether or not: theta_max>θ_th(θ_thTo set a threshold), if yes, the current point is considered as a boundary point and marked.

The step S104 of fusing the fine registration point cloud means removing redundant point cloud in the fine registration point cloud; the redundant point cloud is to firstly judge whether a point adjacent to the current point cloud exists or not, if so, further judge whether the point belongs to a boundary point or not, if so, keep the point to maintain the consistency of the point cloud, otherwise, reject the point as a redundant point.

Step S105, the surface reconstruction of the fused fine registration point cloud by using the method based on feature point detection and adaptive step length updating comprises the following steps: 1) an improved poisson surface reconstruction algorithm; 2) a sharp feature point detection algorithm; 3) and updating the coordinates of the sharp peaks of the triangular meshes.

The improved poisson surface reconstruction algorithm in step S105 refers to: removing the triangular patch of the expansion area on the premise of keeping the hole filling effect of the classic Poisson surface reconstruction algorithm; the sharp feature point detection algorithm is as follows: detecting and marking sharp edge points of the current point cloud according to the significance of the normal vector distribution difference of the micro plane where the sharp edge point concentration points are located; the algorithm for updating the sharp vertex coordinates of the triangular mesh refers to: aiming at the sharp points of the triangular grids obtained before, the coordinates of the marked sharp points of the triangular grids are quickly and accurately adjusted through a normal vector bilateral filtering algorithm and self-adaptive step iteration, and the surface of an object keeping sharp features is recovered.

As shown in fig. 2, the three-dimensional reconstruction system maintaining sharp features according to the embodiment of the present invention mainly includes a point cloud preprocessing module a, a point cloud merging module B, and a surface reconstruction module C.

The point cloud preprocessing module A removes noise of different point clouds by adopting a smooth denoising method, clusters the point clouds by using an improved region growing method, and removes outliers in the point clouds according to a clustering result. This module corresponds to step S101 and step S102 in the three-dimensional reconstruction method.

And the point cloud merging module B is used for registering and fusing the preprocessed plurality of point clouds, and comprises a method for registering the point clouds by using a kd-tree acceleration-based method and a boundary point detection method used in the point cloud fusion. This module corresponds to step S103 and step S104 in the three-dimensional reconstruction method.

And the surface reconstruction module C is used for performing surface reconstruction on the integrated point cloud after fusion, and comprises a process of accelerating the maintenance of sharp characteristics of the surface of the model by using a characteristic point detection and self-adaptive step length updating method. This block corresponds to step S105 in the three-dimensional reconstruction method.

The specific embodiment of the invention:

the overall flow of the method of the invention is shown in figure 1, and the main body of the method of the invention comprises three parts: 1. removing point cloud outliers based on an improved region growing method in point cloud preprocessing; 2. a method for fusing point clouds based on neighborhood searching and boundary point detection in point cloud merging; 3. the method for maintaining sharp edge features of a reconstructed model based on feature point detection and adaptive step size updating in surface reconstruction.

1. Point cloud removal outliers

The operation of removing outliers in the point cloud is used for processing the point cloud which is far away from the main point cloud and the point cloud which is close to the main point cloud but has sparse point cloud density in the point cloud data after the point cloud smoothing filtering. The flow chart of point cloud outlier removal is shown in fig. 3, and mainly includes a kd-tree-based partitioning method, a region growth improvement clustering method, and a point cloud removal method based on statistical analysis.

1.1 splitting method based on kd-Tree

Due to the fact that the number of the point clouds is large, the space of the point clouds needs to be divided before the point clouds are subjected to local neighborhood searching, and searching efficiency is improved. The method for partitioning the data based on the kd-tree comprises the following specific steps: calculating a bounding box of the point cloud of the current area to be searched, and using the side with the longest length of the bounding box as a dividing axis of a dividing space; splitting the splitting axis using a plane perpendicular to the splitting axis and assigning the current point to a left or right sub-tree of the current tree by determining whether the current point is located on the left or right side of the splitting plane; then, the space of the current point is divided in a recursive manner until the space only contains one data point. The method can access according to the rules adopted during the segmentation during the access, and is simple to realize.

1.2 clustering method to improve region growing

The general idea of removing outliers in point clouds is to cluster the point clouds by a region growing clustering method, find out the main point clouds, and then remove the points of the non-main point clouds. However, in the actual removing process, it is found that some sparse point clouds next to the dense point clouds cannot generate a good reconstruction effect in the subsequent reconstruction process, and the sparse point clouds can be used as outliers to be removed. However, when the region growing method is used for clustering, the sparse point clouds are all kept. In order to solve the problems, the invention improves the region growing method, and the specific idea is as follows: in the process of carrying out region growing on each point, searching points near the current point by using a KNN (K nearest neighbor) search method; then calculating the average distance between the adjacent points and the search point, and taking the average distance as the density judgment of the adjacent points of the point; if the density of the points around the point is far less than the density of the whole point cloud, the point is removed and is not used as a point for region growing, and sparse points are marked.

1.3 Point cloud removal method based on statistical analysis

In order to acquire and remove the outlier point clouds, a clustering mode based on radius search is used for searching point cloud areas, then the number of point clouds in each area is counted, and the outlier point clouds are removed through a set threshold value. The specific steps are shown in algorithm 1.

Algorithm 1, outlier point cloud removal algorithm:

1) for point cloud P ═ P_i,i＝1,2,3,…,N},p_iAs points in the point cloud, N (N)>1) Calculating P density rho for the point number of the point cloud; p is a radical of_iThe flag value of flag bit flag is flag num, the initial value flag num is 0, and the flag value flag nu ism, dividing the point cloud P by using a kd-tree, wherein i is 0;

2) increasing i by 1;

6) if S is not empty, returning to 4);

7) if i is less than or equal to N, returning to 2);

According to the method, the operation of removing outliers is carried out on the point clouds with different orders of magnitude, the distribution condition of the original point clouds is shown in fig. 4a, and the distribution of the point clouds after the isolated point clouds are removed is shown in fig. 4b, so that the sparse outliers at the lower half part of the point clouds can be well removed. The operation efficiency of the algorithm is shown in fig. 5, the operation time of the algorithm and the number of the point clouds show a nearly linear relationship, and the average operation time of one hundred thousand point clouds is about 2.5 seconds. Such operating efficiency is within an acceptable range as a part of the pre-treatment stage.

2. Point cloud fusion

The point cloud fusion is performed after point cloud registration, and is used for solving the problem that point cloud density distribution is uneven due to the fact that the registered point cloud has overlapped areas in some areas. The invention uses the thought based on space grid division for point cloud filtering and uses the method based on kd-tree division for fine searching of point cloud neighborhood. The main idea of the algorithm is as follows: firstly, rough searching is used, an interval with larger side length is set for space grid division, and the current point cloud is divided into a plurality of small cube areas. For each new point, if no other point clouds exist in the grid where the new point is located and the surrounding grids, the point is filtered, and the new point is not considered to be the point cloud of the overlapped area; and then, performing fine search by using a smaller radius based on the kd-tree, judging whether a previous point cloud exists around a new point, and if so, considering that the point is in an overlapping area and needs to be removed.

However, in the method, whether the point in the previous point cloud exists in the neighborhood is judged, that is, whether the newly entered point cloud and the previous point cloud have an overlapping area is judged, and the points in the overlapping area are removed. However, the algorithm has a big problem if the new point cloud P is entered_iPrior point cloud P_cIf there are points near the boundary, the points near the boundary are all eliminated in the above determination. This causes a band-shaped hollow region to appear near the boundary line of the previous point cloud during the point cloud fusion. Therefore, the method for point cloud fusion of the present invention needs to go through the above neighborhood searching process of boundary point detection and neighborhood, and the specific flow chart is shown in fig. 6.

2.1 boundary Point detection Algorithm

According to the above discussion, based on the consideration that the boundary point of each point cloud needs to be detected before the point cloud fusion is performed, the point clouds near the boundary point are reserved, and the complete structure of the point clouds is maintained. The main idea of the boundary point detection algorithm is that boundary points may exist at points where the curvature and the distribution of surrounding points vary relatively much. The specific flow of the algorithm is as follows.

Algorithm 2, extracting Point cloud P_cBoundary point algorithm:

1) for the current point cloud P_cConstructing kd-Tree, traversing P_cAll points V in_cDetermining the current point V_cNeighborhood point set V of_t；

2)According to the obtained current point V_cNormal vector of (2), set of neighborhood points V_tProjecting onto a microcut plane;

3) will V_cPoint V on the microcut plane_c' respectively and V_tSet of points V on tangent plane_t' all points are connected to form an edge E;

4) calculating the included angle between two adjacent edges in the E, and recording the maximum value theta of the current included angle_max；

5) According to a set threshold value theta_thresholdWhen theta is_max＞θ_thresholdTime, mark the current point V_cAre boundary points.

2.2 improved Point cloud fusion Algorithm

Obtaining a point cloud P_cAfter marking the boundary point, the newly entered point needs to be judged: and querying a point nearby the point by using a kd-tree, and if a boundary point exists, storing the point. The improved fusion algorithm has the following specific steps.

Algorithm 3, improved point cloud fusion algorithm:

1) let i equal to 0, P_i(i＝1,2,……,N-1)

2) Increase i by 1, P_iRecording as the point cloud P which is spliced at present_cCalculate P_cA bounding box, and the bounding box is divided into a plurality of small cubes by using a space grid;

3) increment j by 1, traverse P_iPoint V in_j(j＝1,2,……,M)；

8) i < N, return 2);

9) and finishing point cloud fusion.

According to the invention, different point clouds are fused by using the point cloud fusion algorithm, fig. 7a and 7b are two point clouds to be fused, and fig. 7c is a fusion result of the two point clouds, so that the basic structural characteristics of a point cloud model can be maintained while redundant point clouds are removed in the process of fusing the point clouds. The operation efficiency of the algorithm is shown in fig. 8, a nearly linear relationship is presented between the time of point cloud fusion operation and the number of point clouds, the fusion time of each twenty thousand point clouds is about 6 seconds, and the fusion speed is high.

3. Surface reconstruction

After the point cloud is preprocessed, registered and fused, the original object is already close to the structure, but a geometric model is still needed to construct the geometric structure of the current object instead of a scattered point cloud. When a common surface reconstruction method such as a poisson surface reconstruction method is used for reconstructing the surface of an object, a reconstructed surface patch is too smooth at a sharp edge or a sharp vertex, and the original style of the object cannot be maintained. The embodiment can maintain the sharp features of the reconstructed model at the surface, as shown in the flowchart of fig. 9, and is implemented in two steps: firstly, the iteration of point cloud coordinates near the feature points is accelerated through a feature point detection method, and the need of iterative updating of all the points is avoided; and secondly, updating the point cloud coordinates by using a self-adaptive step size iterative method so as to obtain a more accurate reconstruction model in a shorter time.

3.1 Sharp feature Point detection

The distribution of the normal vectors of the points near the sharp edge is obvious in difference, and the distribution of the normal vectors of the points near the flat area is uniform. Based on the difference of normal vector distribution difference, the sharp feature point detection algorithm provided by the invention has the main ideas that: firstly, whether a current point is possibly in a region with sharp edges or not is analyzed by calculating the standard deviation and the maximum interval of the included angle between the normal vector of the current point and the normal vector in the neighborhood; secondly, mapping the normal vectors in the neighborhood of the current point into a unit sphere by using a Gaussian mapping mode, namely moving the initial points of all normalized normal vectors to the center of the unit sphere, and then representing the directions of the initial points and the end points of the normalized normal vectors, wherein the distances of points with the same or similar directions on the surface of the sphere are closer, so that the end nodes of all normal vectors mapped to the surface of the sphere are clustered by using a distance-based region growing method, and the change condition of the normal vectors around the current point is judged according to the obtained categories. The specific steps of the algorithm are as follows.

Algorithm 4, feature point detection algorithm:

1) inputting point cloud P to be reconstructed_cUsing kd-Tree to establish the topological structure of current point cloud, for P_cAt any point p_iFinding p using radiusSearch_iNearby point P_neiRecord P_neiNormal vector N of all points in_nei；

2) For all normal vectors N_i∈N_neiAnd point p_iNormal vector p_njAngle of (theta)_ti；

3) Calculating all included angles theta_tiStandard deviation of (a)_tWhen σ is_t＞σ_threshold(σ_thresholdA preset standard deviation threshold), the point is considered to be a sharp point possibly, otherwise, the step 1) is returned to traverse the next point;

4) calculating theta_tiMiddle maximum included angle theta_maxAnd minimum angle theta_minWhen theta is_max-θ_min＞Δθ_threshold(Δθ_thresholdA preset threshold value of the maximum and minimum angle difference), the point is considered to be a sharp point possibly, otherwise, the step 1) is returned to traverse the next point;

5) clustering theta using Gaussian maps_tiMapping to a unit sphere plane for clustering, and considering the region as a flat region when the cluster type is 1, or considering the region as a flat region otherwiseThe points constituting the sharp edges or points on the sharp corner points are marked.

3.2 adaptive step size update preserving Sharp features

The invention provides a surface reconstruction method based on sharp feature point detection, which directly adjusts the positions of points forming a surface triangular grid after Poisson reconstruction, so that the adjusted triangular grid can keep the sharp features of the surface of an original object. The main idea for keeping the sharp features of the model surface is: firstly, updating the normal vector of each grid near a sharp edge point by using a bilateral filtering method, so that the normal vectors at the left side and the right side of the sharp edge are closer to the integral normal value of the side of the sharp edge, rather than the normal value in the middle direction of the two sides independently; and the second method updates the points forming the triangular mesh according to the updated result of the grid normal vector, so that the normal vector of the triangular mesh formed by the updated points is as close to the normal vector value generated by filtering as possible.

In the iterative process, the defined matching function p (t) has smaller and smaller values, and can be closer to the coordinates of the desired point. Setting the convergence condition of iterative update to be DeltaP (T) < P_thNamely, when the difference between the matching functions before and after two updates is smaller than a certain threshold, the point coordinates are stopped to be updated. The value of the iterative step length lambda plays a key role in the process, when the optimization problem is solved, if the step length is too large, the minimum value can be directly skipped during iteration, and if the step length is too small, the optimal solution can be reached in a long time. The invention uses a self-adaptive step length updating method based on the change of the matching function to iteratively update the latest coordinates of the points so as to achieve the aim of keeping the sharp characteristics of the surface of the object. The method comprises the following specific steps.

Algorithm 5, feature preserving surface reconstruction algorithm:

2) j is 0, and a matching function p (t) is defined to obtain the matching degree of the updated point and the current normal vector, as shown in formula (2), where F is the triangular patch in mesh, (v)_i′，v_j′，v′_k) Three points in a triangular patch. Setting an initial value P (T) of P (T)₀) For larger values, the step length λ is λ₀(λ₀Is a preset step length);

PT)＝∑_F((N_j′·(v_i′-v_j′))²+(N_j′·(v_i′-v′_k))²+(N_j′·(v′_k-v_j′))²) (2)

3) increasing j by 1, calculating P (T) from equation (2)_j)；

v_j′＝v_j+λ∑_j∈fN′_f·N_f′·(v_i-v_j)+N′_f·N_f′·(v_k-v_j) (3)

wherein f is a point v_jAssociated triangular patch, (v)_i，v_j，v_k) Three points in a triangular patch. I.e. point v_jVector N of approach_fDirection of 'stretching'

5) If P (T)_j)≤P(T_j-1) Then return to 3), otherwise let j equal j-1 and when λ<When the lambda is 0.04, the lambda is equal to lambda/2, otherwise, the lambda is equal to lambda-0.02, and the return is 4);

6)ΔP(V)≥P_thor Num is not more than Num_maxAnd then return to 3);

7) and (6) ending.

The invention is used for carrying out surface reconstruction of the maintained characteristics on the input point cloud, fig. 10a is the distribution of the original point cloud, fig. 10b is the result of carrying out surface reconstruction by using the improved Poisson reconstruction, it can be seen that the surface at the edge is too smooth at the moment, fig. 10c is the result of carrying out surface reconstruction by using the invention to maintain sharp characteristics, it is obviously seen that the reconstructed model is sharper at the edge, and the distribution condition of the original point cloud can be reflected better. Meanwhile, the comparison between the self-adaptive method and the common fixed step length method used in the invention is shown in table 1, the self-adaptive step length is used to reach the same accurate value than the fixed step length, the number of iterations required to be consumed is less, and the invention can reach higher accuracy in shorter time. In addition, the invention can reduce the step length in a self-adaptive mode under the condition that a larger initial value is taken as the step length, thereby ensuring that the situation of iteration errors can not occur.

TABLE 1 comparison of adaptive iteration method and fixed step method

The present invention is not limited to the above embodiments, and any modifications, equivalents and improvements made within the principle of the present invention are included in the scope of the present invention.

Claims

1. A method of three-dimensional reconstruction while preserving sharp features, comprising the steps of:

specifically, the smoothing and denoising of the point cloud input and subjected to the coarse registration is as follows: the smoothing filtering method of the two-dimensional image is applied to smoothing and denoising of three-dimensional point cloud in an expanding mode, namely x and y of point cloud data points (x, y and z) are used as plane positioning marks, z is used as a pixel value in the two-dimensional image, smoothing and denoising of the z value is carried out on point cloud blocks point by point, and the point cloud blocks refer to point cloud collected under the same visual angle; the roughly registered point cloud refers to a point cloud obtained by performing coordinate transformation on a series of point cloud blocks PQ at different shooting angles by adopting a classical coordinate transformation method;

the step of performing z-value smooth denoising on the point cloud blocks point by point refers to: acquiring a z value in a neighborhood with the current (x, y) as a center, and selecting a two-dimensional image smoothing filtering method to smooth the z value according to experience, wherein the neighborhood with the current (x, y) as the center is a square, and the side length is selected within 3-15;

the step of performing outlier removal on the smoothed rough registration point cloud by adopting an improved region growing method is as follows: counting the number of points in the point cloud area divided by the kd-tree by adopting a KNN clustering mode, and judging the number of points in the current area to remove outliers; the outlier refers to an isolated point cloud different from a main point cloud, the main point cloud is a point cloud belonging to the surface of a reconstruction object, the isolated point cloud comprises isolated points and a sparse point cloud, the sparse point cloud is a local point cloud with the density less than 40% of the density of the whole point cloud or a preset value, and the specific algorithm is as follows:

1) for point cloud P ═ P_i,i＝1,2,3,…,N},p_iIs a point in the point cloud, N is the number of points in the point cloud, N>1, calculating P density rho; p is a radical of_iThe flag value of the flag bit flag is flag num, the initial value flag num is 0, the flag is flag num, the point cloud P is divided by using kd-tree, and i is 0;

2) increasing i by 1;

3) if p is_iIf Flag of (1) is 0, then p is defined_iIncreasing flagNum by 1, Flag ═ flagNum and adding p_iPush stack S, otherwise return 2);

6) if S is not empty, returning to 4);

7) if i is less than or equal to N, returning to 2);

counting the number of point clouds clustered one by one

When in use

Th is a preset threshold value, and sparse point clouds are removed;

(3) performing fine registration on the coarse registration point cloud after the outlier is removed by adopting an ICP algorithm based on kd-tree acceleration;

the method comprises the following steps of performing fine registration on a coarse registration point cloud after outliers are removed by adopting a kd-tree-based accelerated ICP (inductively coupled plasma) algorithm, and performing fine registration on a series of point clouds PQ subjected to coarse registration, wherein the specific algorithm is as follows:

1) establishing an RT transformation matrix of the point cloud Q, setting an initial value of the transformation matrix to be 0, and setting i to be 0;

2) increment i by 1, find the adjacent PQThe point pair of (2): traversal point cloud P { P_i,i＝1,2,3,…,N_pPoint p in_iPoint-to-point cloud Q { Q_j,j＝1,2,3,…,N_qBuilding kd-Tree, searching off p_iMost recent q_j；

4) updating an RT transformation matrix of the point cloud Q;

5) calculating the mean square error sigma between PQ after the point cloud Q is subjected to RT conversion, if the sigma is more than or equal to Th or the iteration number N is less than or equal to N_maxTh is a predetermined threshold, N_maxReturning to 2) for the preset maximum iteration times;

6) finishing point cloud fine registration;

2. The three-dimensional reconstruction method for maintaining sharp features according to claim 1, wherein the step (4) of fusing the fine registration point cloud by using a fusion method based on neighborhood searching and boundary point detection comprises: the neighborhood searching method comprises the steps of partitioning point cloud through a space grid, and finely partitioning the current neighborhood by adopting a kd-tree; the space grid segmentation means that a bounding box which is the minimum external cuboid of a current point cloud area to be searched is calculated, and then the current bounding box is divided into a plurality of small cubes according to a set side length; the set side length is 5% of the shortest side length of the current bounding box;

the boundary point detection means that a neighborhood point set of a current point is projected into a micro-tangent plane of the current point, and the maximum value theta of the included angle between the current point and the neighborhood point set is judged_maxWhether or not: theta_max>θ_th，θ_thIf the threshold value is set, the current point is considered to belong to the boundary point and is marked;

the fusing the fine registration point cloud means removing redundant point cloud in the fine registration point cloud; the redundant point cloud means that the current point cloud P is judged firstly_iWhether there is a neighboring point cloud P in the vicinity_i+1If the point i is 1,2, … …, N-1, further judging whether the point belongs to a boundary point, if so, keeping the point to maintain the consistency of the point cloud, otherwise, removing the point as a redundant point; the specific algorithm is as follows:

1) let i equal to 0, j equal to 0, P_i，i＝1,2,……,N-1；

3) increment j by 1, traverse P_iPoint V in_j，j＝1,2,……,M；

5) judgment V_jWhether the point set V' in the small cube and the adjacent 27 small cubes is in P_cIn a bounding box, wherein V' does not contain V_jIf not, V_jIncorporation of P_cAnd if j is less than or equal to M, returning to 3);

6) to P_cCreate kd-Tree with V_jSearching membership P in neighborhood with circle center as r as radius_cAnd count points

Wherein V' does not contain V_jIf, if

Th is a predetermined threshold value, V_jIncorporation of P_cAnd if j is less than or equal to M, returning to 3);

7) judgment V_jWhether there is a set of boundary points V '"within the neighborhood, where V'" does not contain V_jIf present, V_jIncorporation of P_cOtherwise remove V_jAnd if j is less than or equal to M, returning to 3);

8) i < N, return 2);

9) and finishing point cloud fusion.

3. The three-dimensional reconstruction method for maintaining sharp features according to claim 1, wherein the step (5) of performing surface reconstruction on the fused fine registration point cloud by using a feature point detection and adaptive step size updating method comprises: 1) an improved poisson surface reconstruction algorithm; 2) a sharp feature point detection algorithm; 3) and updating the coordinates of the sharp peaks of the triangular meshes.

4. The three-dimensional reconstruction method for preserving sharp features of claim 3 wherein said modified Poisson surface reconstruction algorithm refers to: removing the triangular patch of the expansion area on the premise of keeping the hole filling effect of the classic Poisson surface reconstruction algorithm; the specific algorithm is as follows:

1) aiming at the point cloud P, a triangular mesh (V) is obtained by adopting a classic Poisson three-dimensional surface reconstruction algorithm_i&T_j,i＝1,3,……,N_v,j＝1,2,……,N_t}，V_iRepresenting the vertices of a triangular mesh, T_jFor patches in a triangular mesh, N_vIs the number of vertices in the triangular mesh, N_tIs the number of patches in the triangular mesh,&represents a logical and;

2) setting V point by point_iThe flag removeFlag is false, which indicates that j is 0;

4) finding T_jGrid and T_jIf the points in P exist in 27 small cubes in the adjacent grids, if not, (V) is carried out_j1,V_j2,V_j3) The flag bit removeFlag is true, true indicates that j is not more than N_tReturn to 3);

5) clustering triangular grid points with removeFlag as true by using a region growing method, and counting the number of areas by regions_numIf there is a region: area thereof_num≤Th_rec，Th_recIf the current value is the preset threshold value, resetting the current true point to false;

5. The three-dimensional reconstruction method for maintaining sharp features according to claim 3, wherein said sharp feature point detection algorithm is characterized by: detecting and marking sharp edge points of the current point cloud P, wherein the normal vector distribution difference of the micro-plane where the middle points of the sharp edge point set P' are located is obvious; the specific algorithm is as follows:

1) setting the P topology structure by using kd-tree;

2) increase i by 1, and add P to any point in P_iAs the center of a circle, r is the radius search point set P_nei，P_neiNot containing pi, P_neiMiddle arbitrary point p_jHas a normal vector of η_j，j＝1,2,…,N_nei，N_neiIs P_neiNumber of medium normal vectors, P_neiThe normal vector set of all points is denoted as gamma_neiI.e. η_j∈Γ_nei，p_iNormal vector gamma of_iAnd η_jIs theta_tj(ii) a The radius r is set according to 1.5 times of the density P;

3) calculating all included angles theta_tjStandard deviation of (a)_tJ is 1,2, …, if σ_t≤σ_hold，σ_holdReturning to 2) for presetting the threshold, otherwise, considering p as_iCandidate sharp edge points;

4) calculating theta_tjMaximum included angle θ of_maxAt a minimum angle theta_minIf (theta)_max-θ_min)≤Δθ_hold(Δθ_holdPreset threshold), return 2), otherwise, consider p_iCandidate sharp edge points;

5) will theta_tjMapping to unit spherical surface and carrying out Gaussian clustering, if the category is 1, the current region belongs to a flat region, otherwise, p_iMarking at sharp edges or sharp corner points;

and establishing a kd-tree on the P ', searching the triangular mesh point by point to find out a vertex near the P', and marking the vertex as a sharp edge point in the triangular mesh.

6. The three-dimensional reconstruction method for maintaining sharp features according to claim 3, wherein the algorithm for updating the sharp vertex coordinates of the triangular mesh refers to: aiming at the sharp edge points in the obtained triangular mesh, quickly and accurately adjusting the coordinates of the sharp points of the marked triangular mesh through a normal vector bilateral filtering algorithm and self-adaptive step iteration, and recovering the surface of an object with sharp features; the algorithm is as follows:

1) mesh { V } for the reconstruction surface triangular mesh_i&T_j,i＝1,3,……,N_v,j＝1,2,……,N_tTriangular patch T with sharp point marked_j′，j＝1,2,……,N_loc，N_locUsing bilateral filtering algorithm (1) to process T by T for the number of triangular patches containing sharp points_j′(V_j1′,V_j2′,V_j3') obtaining N_locA normal vector N_j′，j＝1,2,……,N_loc：

2) number of cycles

Defining a matching function

To obtain the matching degree of the updated point and the current normal vector, as shown in formula (2), where the subscript F of Σ represents any triangle patch in the mesh, Σ represents that F in the mesh is traversed to sum the formula in formula (2),

for three points in a triangular patch, i, j, k representing three different vertices of any triangular patch, the initial value of P (T) is set to P (T)₀) Step λ ═ λ₀，λ₀Is a preset step length;

3) make it

Increased by 1, calculated by equation (2)

wherein f is a point v_jAssociated triangular patch, (v)_i，v_j，v_k) Three points in the associated triangular patch; i.e. point v_jNormal vector is N_f' triangular patch direction "stretch";

5) if it is

Then return to 3),

6)

if λ<When 0.04, let λ be λ/2, otherwise λ be λ -0.02, return to 4);

7)ΔP(V)≥P_thand is

P_thA threshold value representing a difference between the current value and the last value obtained by the expression (2),

if representing the maximum cycle number, returning to 3);

8) and (6) ending.

7. A system for realizing the three-dimensional reconstruction method for maintaining sharp features of any one of claims 1 to 6, which is characterized by comprising: the system comprises a point cloud preprocessing module, a point cloud merging module and a surface reconstruction module;

the surface reconstruction module is used for reconstructing the surface of the integrated point cloud after fusion, and the process of keeping sharp characteristics of the surface of the model is accelerated by using a characteristic point detection and self-adaptive step length updating method; the point cloud merging module corresponds to the step (5) in the three-dimensional reconstruction method.