CN104463826A - Novel point cloud parallel Softassign registering algorithm - Google Patents

Abstract

The invention discloses a novel point cloud parallel Softassign registering algorithm. On the basis that CUDA performs parallel acceleration on a Softassign algorithm, by use of a method of combining three-dimensional point cloud discrete curvature estimation and three-dimensional Kd-tree, point-cloud simplification is performed on a three-dimensional object point cloud to enable the simplified three-dimensional object point cloud to maintain sufficient geometric characteristics, and then Softassign registering is performed on the simplified object point cloud, such that the registering precision of the Softassign algorithm in three-dimensional object point cloud registering is improved. The algorithm provided by the invention has the following advantages: first of all, the point-cloud simplification is performed on the three-dimensional object point cloud, such that the information content of the Softassign registering is reduced; and secondly, through the improved Softassign registering algorithm, the registering precision in object cloud registering is improved, and through a parallel acceleration technology, the operation speed of the Softassign registering method is enhanced.

Description

A kind of some cloud newly walks abreast Softassign registration Algorithm

Technical field

The present invention relates to a kind of image processing techniques, particularly a kind of simplification of three-dimensional body point cloud information and registration technology.

Background technology

Along with the full-fledged of 3-D scanning technology and application popularization, identify at meter Cheng Wei in recent years the research direction that field is new based on the identification of object three-dimensional shape information and coupling.Point cloud matching is one of key issue of the three-dimensional data coupling of object scan, first due to the restriction of scanning angle, complete subject depth image cannot be obtained by a 3-D scanning, the cloud data that each scanning obtains is the surface data of part, need the data fusion of Multiple-Scan to be spliced into a width Complete three-dimensional cloud data, and the core technology of anastomosing and splicing and three-dimensional registration; Secondly object model to be identified and the object model in reference library are carried out shape facility when comparing, also need to mate between character pair, judge matching precision by comparison match error rice.

Along with the precision of three-dimensional scanning device improves, data scale and registration accuracy also improve thereupon, and traditional serial registration Algorithm efficiency reduces, and cannot meet the demand of real-time; High performance Graphics Processing Unit (Graphic Processing Unit, and unified calculation framework (Compute Unified Device Architecture GPU), CUDA) providing high performance parallel computation environment for addressing this problem, making the real-time solution of the problems such as data scale is huge, point cloud registering precision is higher become possibility.

Point cloud registering adopts ICP (Iterative Closest Point) algorithm usually, namely makes the distance mean square deviation on two subject to registration some clouds between corresponding point obtain minimum value by iterative computation, thus realizes the exact matching between subject to registration some cloud.But the deficiency of ICP algorithm is to need initial alignment, so usually adopt the methods such as PCA (Principal Component Analysis) to carry out initial registration, i.e. thick alignment, then with initial registration result for condition carries out ICP registration, i.e. Accurate align.Rangarajan in 1997 etc. propose Softassign algorithm for all kinds of matching problem, and optimize the solution procedure of corresponding relation between points, improve optimum solution, this algorithm also has certain robustness to abnormity point problem.No matter ICP algorithm or Softassign algorithm are all containing extensive matrix operation, these computings possess the feasibility of parallel accelerate, significantly can improve the deficiency that serial registration efficiency is lower, but due to the restriction in video memory space, this work has been carried out simple stochastic sampling to larger point cloud model and has been simplified, and causes final registration accuracy to there is clear and definite loss.

Summary of the invention

Huge in order to overcome data scale, the problem that point cloud registering precision is not high, of the present inventionly proposes a kind of some cloud newly and to walk abreast Softassign registration Algorithm, avoids the problem of local registration, improves the Efficiency and accuracy of object point cloud registering.

This algorithm solves the technical scheme that its technical matters adopts:

Utilize the method that three-dimensional point cloud Discrete Curvature Estimation and three dimensions kd-tree combine, carry out point cloud simplification to three-dimensional body point cloud, after enabling simplification, three dimensional object point cloud retains enough geometric properties.

Softassign registration is carried out to simplification object point cloud, improves the registration accuracy of Softassign algorithm in three dimensional object point cloud registering.

CUBLAS (CUDA Basic Linear Algebra Subprograms) is utilized to carry out computing between acceleration vector and matrix.Utilize the computing between CUDAkernel function acceleration matrix element.

The beneficial effect of this algorithm is: carry out point cloud simplification to three-dimensional body point cloud, decrease the quantity of information of Softassjgn registration; And the Softassign registration Algorithm by improving, improve the registration accuracy in object cloud registration; Parallel accelerate technology is adopted to improve the arithmetic speed of Softassign registration Algorithm.

Accompanying drawing explanation

Fig. 1 is point cloud simplification details of the present invention signal

Embodiment

Below in conjunction with the drawings and specific embodiments, illustrate the present invention further.

1. utilize kd-tree to carry out point cloud simplification

If the original point cloud data position X of three-dimensional body, first set up the kd-tree (K-dimension tree) of original point, for fast search neighborhood and closest approach.Kd-tree is the sharp data structure divided data at k dimension space.Three dimensional object point cloud can set up kd-tree, and main application is data point search, and it is neighborhood search that its way of search comprises two kinds: one, can search the neighborhood point of specifying in the radius of neighbourhood; Two is k neighborhood search, can search the k strong point that distance objective point is nearest, search for closest approach exactly as k=1.Because kd-tree is the data structure based on spatial division, when carrying out data search, search for from the little space of bottom, can search efficiency be improved.

In the process simplifying some cloud, a cloud is divided into key point and non-key point.First the Gaussian curvature of each point in solution point cloud, and compare with appointment threshold value, if Gaussian curvature is greater than specify threshold value, judge that this point is as key point, directly copy to and simplify in some cloud; If be less than appointment threshold value, judge that this point is non-key point, examination ball is set up by k neighborhood search, dot density in examination ball, if dot density is greater than appointment threshold value, then in mark examination ball equalization point be a little key point, otherwise be labeled as key point a little by ball, realize the simplification to non-key point accordingly.

The object-point cloud shortcut calculation process according to Gaussian curvature and kd-tree that the present invention realizes is as shown in algorithm 1.

In algorithm 1, the detailed description of Step6 and Step7 as shown in Figure 1, the x that sets up an office is the random point that Step4 chooses, as central point, with the detailed description of Step6 and Step7 in an x algorithm 1 as shown in Figure 1, the x that sets up an office is the random point that Step4 chooses, and as central point, utilizes kd-tree to carry out k neighborhood search centered by an x, x is followed successively by from an x distance incremental order in k=3, Fig. 1 ₁, x ₂, x ₃, x ₄.As Fig. 1 a: if the key point wherein do not marked, be then the centre of sphere with x, with r ₃=|| x ₃-x|| is that radius sets up examination ball, and in figure, circle represents examination ball; As Fig. 1 b: if x ₃for marking key point, be the centre of sphere with x, with r ₂=|| x ₂-x|| is that radius sets up examination ball, and Fig. 1 a and 1b is sufficient in the step a of Step6 in algorithm 1 and the signal of step b respectively.Fig. 1 c is the final result simplified, with x, x ₁, x ₂the equalization point x ' of 3 replaces original 3 points, and marks.

2. adopt the Softassign registration Algorithm improved

For two object data point set X and Y to be matched, utilize point cloud simplification algorithm to be simplified a cloud and be designated as X _rand Y _r.Traditional ICP algorithm is at point set Y _rin find and point set X _rin certain some x _ricorresponding some y _rj, make objective function minimum.And for Softassign algorithm, point set Y _rin each point can with x _rihave corresponding relation, but the alignment probability of correspondence is not identical.Softassign algorithm calculates R and t by minimizing registration error E (such as formula 1).

$E=\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}\left|\right|x_{\mathrm{Ri}}-(R^{k-1}{y}_{\mathrm{Rj}}+t^{k-1})\left|\right|-\mathrm{\α}\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}-\frac{1}{\mathrm{\β}}\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}(\log m_{\mathrm{ij}}-1)---\left(1\right)$

Wherein m _ijrepresent some x _iwith y _jthe probability of alignment, makes M=(m _ij) represent alignment weight matrix, notice that M contains extra row and processes the abnormal conditions in aliging with row.N _x, n _yrepresent some cloud X _rand Y _rdata scale after simplification, R ^k-1, t ^k-1represent this registration R, the initial value of t, α, β are constant, specify when algorithm runs.

3.Softassign algorithm parallel accelerate

M in the Step4 of algorithm 2 _ijrepresent and simplify some cloud X _rmiddle x _riand Y _rmiddle y _rjregistration probability, then solve m _ijtime, simplify some cloud point X _rand Y _rhelp the relation of mapping, note matrix M=(m _ij), its scale is two simplification point cloud X _rand Y _rthe product of vertex number, that is: | M|=n _x× n _y, when matrix multiplication is carried out to matrix M, the computing scale of more than 1,000,000 times can be reached.

The key of algorithm parallel accelerate is the calculating process of former algorithm to be divided into computing and the computing of matrix interior element between vector and matrix, thus utilizing CUBLAS (CUDABasic Linear Algebra Subprograms) to carry out computing between acceleration vector and matrix on the one hand, CUBLAS is the Parallel Implementation of BLAS based on CUDA; Utilize the computing between CUDAkernel function acceleration matrix element on the other hand.Below elaborate and how to utilize CUBLAS and CUDA kernel function to carry out parallel accelerate to the whole process of Softassign algorithm.

Claims (5)

1. the invention discloses a kind of some cloud newly to walk abreast Softassign registration Algorithm, it is characterized in that, kd-tree is utilized to carry out point cloud simplification, by point cloud simplification algorithm improvement Softassign registration Algorithm, improve the registration accuracy in object cloud registration, and adopt parallel accelerate technology to improve arithmetic speed.

2. method according to claim 1, is characterized in that, three dimensional object point cloud can set up kd-tree, and main application is data point search, and it is neighborhood search that its way of search comprises two kinds: one; Two is k neighborhood search.

3. method according to claim 1, is characterized in that, for two three object data point set X and Y to be matched, utilizes point cloud simplification algorithm to be simplified a cloud and is designated as X _rand Y _rthe Softassign algorithm improved, point set Y _rin each point can with x _rihave corresponding relation, but the alignment probability of correspondence being not identical, calculating R and t by minimizing registration error E.

$E=\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}\left|\right|x_{\mathrm{Ri}}-(R^{k-1}{y}_{\mathrm{Rj}}+t^{k-1})\left|\right|-\mathrm{\α}\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}-\frac{1}{\mathrm{\β}}\underset{i=1}{\overset{n_x}{\mathrm{\Σ}}}\underset{j=1}{\overset{n_y}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}(\log m_{\mathrm{ij}}-1)$

4. method according to claim 1, the calculating process of former algorithm is it is characterized in that to be divided into computing and the computing of matrix interior element between vector and matrix, thus utilizing CUBLAS to carry out computing between acceleration vector and matrix on the one hand, CUBLAS is the Parallel Implementation of BLAS based on CUDA; Utilize the computing between CUDAkernel function acceleration matrix element on the other hand.

5. method according to claim 2, is characterized in that, in the process simplifying some cloud, a cloud is divided into key point and non-key point.First the Gaussian curvature of each point in solution point cloud, and compare with appointment threshold value, if Gaussian curvature is greater than specify threshold value, judge that this point is as key point, directly copy to and simplify in some cloud; If be less than appointment threshold value, judge that this point is non-key point, examination ball is set up by k neighborhood search, and the dot density examined or check in ball, if dot density is greater than appointment threshold value, then in mark examination ball equalization point be a little key point, otherwise be labeled as key point a little by ball, realize the simplification to non-key point accordingly.