CN106997591A - A kind of super voxel dividing method of RGB D image mutative scales - Google Patents

Abstract

The invention discloses the super voxel dividing method of mutative scale in one towards RGB D view data, seed point is selected in data using the sampling of Poisson dish, then cluster is iterated according to the color distance and space length between data point and each seed point, obtain initial super voxel segmentation result, the super voxel obtained using initial segmentation is summit, the syntople of super voxel is that non-directed graph is set up on side, super voxel is carried out using the method based on graph theory to merge, the super voxel segmentation result that scale size differs is obtained in same RGB D images, the super voxel segmentation of mutative scale is realized.The inventive method belongs to data prediction, splits obtained super voxel and obtains the super voxel of large scale in the high region of data consistency, obtains the super voxel of small yardstick in the place of data consistency difference, more meets human vision cognitive features.

Description

A kind of super voxel dividing method of RGB-D images mutative scale

Technical field

The present invention relates to a kind of super voxel dividing method, particularly a kind of super voxel of mutative scale suitable for RGB-D images Dividing method.

Background technology

With the progress of sensor technology, the procurement cost of RGB-D images is more and more lower, how more effectively to RGB-D Image carries out the important research content that pretreatment is computer vision in recent years.In order to make full use of in RGB-D images Three-dimensional geometric information, similar with the concept of two dimensional image super-pixel over-segmentation, it is a kind of that RGB-D images are too segmented into super voxel Effective pretreatment mode, can effectively reduce the data volume of subsequent algorithm processing.

Obtained each super voxel size uniformity of the super voxel partitioning algorithm conventional at present after scale parameter determination compared with Height, if follow-up need to carry out multiscale analysis, the scale factor different by setting controls the size of super voxel, so existed The super voxel segmentation result for obtaining a kind of yardstick will be calculated when analyzing each yardstick, so as to add amount of calculation.Base of the present invention The thought analyzed in mutative scale, when carrying out super voxel and splitting, the high region segmentation of data consistency into large scale super body Element, the region segmentation of data consistency difference is into the super voxel of small yardstick, so that according to data local distribution, it is adaptive general RGB-D data are divided into the super voxel of different scale, subsequent algorithm is carried out mutative scale meter in an over-segmentation data Calculate, reduce amount of calculation.

The content of the invention

, can be preferably right it is an object of the invention to provide a kind of super voxel dividing method of RGB-D images mutative scale RGB-D images carry out over-segmentation pretreatment.

The technical solution for realizing the object of the invention is：A kind of super voxel dividing method of RGB-D images mutative scale, including Following steps：

Step 1, seed point are chosen, if a frame RGB-D view data is P, resolution ratio is arranged for m rows n, and each data point is included 3 Color Channels (r, g, b) and 1 depth channel (d).A point p is randomly selected from P₀As initial seed point, half is set Footpath threshold value R is sampled in P using Poisson dish sampling algorithm as the minimum range between seed point to be generated and obtains seed point set Close；Seed point is chosen and specifically includes following steps：

Step 1-1, in a frame RGB-D view data P randomly select a point p₀As initial seed point, will actively it adopt Sampling point queue L₁Sky is initialized as, p₀Add L₁, by inactive sampled point queue L₂It is initialized as sky；

Step 1-2, judgement enliven sampled point queue L₁Whether it is empty, if L₁It is not sky, then from L₁In go out one point of team p_i, with p_iFor the center of circle, R and 2R are random selection candidate's sampled point in the concentric circular regions of radius, if candidate's sampled point is with having planted The distance of son point is then added into L more than R₁；If the attempt to K times still without qualified candidate's sampled point, then by p_iFrom L₁In Delete, and add L₂；Wherein R uses pixel coordinate unit, and K is default numerical value；

Step 1-3, L₂In point be selected seed point.

Step 2, super voxel pre-segmentation, are converted to Lab, depth value d is converted to by the color space of each point in data from RGB Three dimensional space coordinate (x, y, z), using each seed point as cluster centre, Color distance and three dimensions distance iterative calculation are non- Distance between seed point and seed point, gathers for a class apart from the point of seed point recently and the seed point, obtains initial super voxel and divide Cut result；Following steps are specifically included to super voxel pre-segmentation：

Step 2-1, the depth value at RGB-D images midpoint is converted into three dimensional space coordinate, if the point that the i-th row jth is arranged in P For p_ij, its depth value is d_ij, depth value is converted into three dimensional space coordinate using formula (1)：

Wherein f is the focal length of video camera, (c_x,c_y) be image centre coordinate, will with the color space conversion formula of standard The RGB color of each point is converted to Lab colors in P, and the point in P is expressed as into 6 dimensional vectors [l, a, b, x, y, z]；

Step 2-2, the individual seed point obtained using step 1 carry out range searching cluster as initial cluster center, calculate each poly- The distance of point and the cluster centre in the 2R*2R contiguous ranges of class center, each non-cluster central point is ranged special with it The minimum cluster centre of distance is levied to complete two point p in first time cluster process, P_iAnd p_jBetween characteristic distance measure formulas such as Under：

In formula (2), d_labFor color distance, d_xyzFor space length, λ is for determining colouring information and space length information Weight, λ more large space distance it is more important, the smaller color distances of λ are more important；

Step 2-3, iteration cluster process, the result clustered for the first time according to step 2-2, recalculate the cluster of each class Center；New cluster centre characteristic value is the average value of all point features of each class, is then found in a class and new cluster The immediate point of central feature value calculates according to characteristic distance in 2-2 as new cluster centre point and classifying method is counted again Calculate each non-cluster central point generic；K end of iteration；After iterative calculation terminates, the point in each class is to be formed at the beginning of one Begin super voxel.

Step 3, the fusion of super voxel, using initially super voxel as summit, set up side construction non-directed graph G=between adjacent super voxel (V, E), the difference between adjacent super voxel is measured using the Lab color spaces distance and normal vector angular separation of each initial super voxel It is different, minimized using difference in class and class inherited maximizes the initial super voxel of thought fusion, obtain the super voxel of mutative scale.Super body Element fusion, specifically includes following steps：

Step 3-1, the obtained super set of voxels of initial segmentation is set as C, each super voxel is c_i, with super voxel c_iFor top Point v_iVertex set V is set up, the adjacent super voxel v for having public boundary_i, v_jBetween set up side e_ijSide collection E is formed, nothing is built with V and E To figure G；According to the three dimensional space coordinate at P midpoints, k neighbour's point coordinates covariance matrix principal component decomposition method meters of standard are used The normal vector of each point is calculated, then each edge with formula (5) to G is assigned to weight w (v_i,v_j)

Wherein,l_maxRespectively super voxel c_iMiddle mean flow rate and high-high brightness a little, θ_ijFor super voxel c_iAnd c_j Normal vector angle, the normal vector of super voxel be taken as in the super voxel normal vector average a little, α is weight factor；

Step 3-2, initial super voxel is merged, preceding each super voxel will be merged and be initialized as a region A_i, v_kAnd v_l It is region A_iIn two super voxels, the internal dissmilarity degree of definition for the region minimum spanning tree maximum side, as shown in formula (7)

Int(A_i)=maxw (v_k,v_l),v_k,v_l∈A_i,(v_k,v_l)∈E (7)

Two region A are defined with formula (8)_iAnd A_jOutside dissmilarity degree, wherein v_mIt is region A_iIn super voxel, v_nIt is area Domain A_jIn super voxel

Dif(A_i,A_j)=minw (v_m,v_n),v_m∈A_i,v_n∈A_j,(v_m,v_n)∈E (8)

MInt(A_i,A_j) it is A_iAnd A_jThe internal dissmilarity degree of minimum in the two regions, is calculated with formula (9) afterwards

MInt(A_i,A_j)=min ((Int (A_i)+τ(A_i)),(Int(A_j+τ(A_j))) (9)

Wherein τ (A_i)=e/ | A_i|, | A_i| it is region A_iThe number of point is included, e is the constant of setting；

The outside dissmilarity degree and internal dissmilarity degree in two regions are compared afterwards, closed if the formula that meets (10) And the two regions, otherwise nonjoinder

MInt(A_i,A_j)<Dif(A_i,A_j) (10)

The region that region merging technique process is performed until in P obtains the super voxel of mutative scale without untill can merging.

Compared with prior art, its remarkable advantage is the present invention：(1) present invention is initial using the selection of the Poisson dish method of sampling Seed point, compared with the uniform grid method of sampling, this method more meets the distribution character of cone cell on the retina, contributes to Uniform sampling is concentrated in uneven three-dimensional data points；(2) present invention merges depth value and normal vector etc. on the basis of color Geological information, is more suitable for the data for handling RGB-D types；(3) the inventive method is by once calculating in same frame RGB-D data It is upper split obtain the super voxel of different scale, hence it is evident that be different from and once calculate that can only to obtain yardstick on same frame data more consistent Super voxel undue segmentation method, be more beneficial for continuing after the decrease the amount of calculation and multiscale analysis of algorithm.

The present invention will be further described below in conjunction with the accompanying drawings.

Brief description of the drawings

Fig. 1 for the present invention algorithm pre-segmentation with merge after the super voxel segmentation result of mutative scale.

Fig. 2 is the super voxel dividing method flow chart of RGB-D image mutative scales of the invention.

Embodiment

A kind of super voxel dividing method of RGB-D images mutative scale of the present invention, comprises the following steps：

Step 1, seed point are chosen, if a frame RGB-D view data is P, resolution ratio arranges for m rows n, the letter of each data point Breath includes 3 Color Channels (r, g, b) and 1 depth channel (d).A point p is randomly selected from P₀As initial seed point, Set radius threshold R as the minimum range between seed point to be generated, sampled using Poisson dish sampling algorithm in P and obtain seed Point set.Specifically include following steps：

Step 1-1, a point p is randomly selected in P₀As initial seed point, sampled point queue L will be enlivened₁It is initialized as Sky, p₀Add L₁, by inactive sampled point queue L₂It is initialized as sky；

If step 1-2, L₁It is not sky, then from L₁In go out one point p of team_i, with p_iFor the center of circle, with R and 2R, (R is used respectively Pixel coordinate unit) to randomly choose candidate's sampled point in the concentric circular regions of radius, if candidate's sampled point and existing seed point Distance be then added into L more than R₁.If the attempt to K times still without qualified candidate's sampled point, then by p_iFrom L₁In delete Remove, and add L₂；

Step 1-3, L₂In point be selected seed point.

Step 2, super voxel pre-segmentation, are turned the color space of each point in P from RGB with the color space conversion method of standard Be changed to Lab, depth value d is converted to three dimensional space coordinate (x, y, z), using each seed point as cluster centre, Color distance and Three dimensions distance iterates to calculate the distance between non-seed point and seed point, obtains initial super voxel segmentation result.Specifically include Following steps：

Step 2-1, the depth value at RGB-D images midpoint is converted into three dimensional space coordinate, if the point that the i-th row jth is arranged in P For p_ij, its depth value is d_ij, depth value is converted into three dimensional space coordinate using formula (1)：

Wherein f is the focal length of video camera, (c_x,c_y) be image centre coordinate, will with the color space conversion formula of standard The RGB color of each point is converted to Lab colors in P.Changed by color space and coordinate space, the point in P be expressed as 6 dimensions to Measure [l, a, b, x, y, z]；

Step 2-2, each seed point obtained using step 1 carry out range searching cluster as initial cluster center, calculate each poly- The distance of point and the cluster centre in the 2R*2R contiguous ranges of class center, each non-cluster central point is ranged special with it The minimum cluster centre of distance is levied to complete two point p in first time cluster process, P_iAnd p_jBetween characteristic distance measure formulas such as Under：

In formula (2), d_labFor color distance, d_xyzFor space length, λ is for determining colouring information and space length information Weight, λ more large space distance it is more important, the smaller color distances of λ are more important, with specific reference to concrete application demand set.

Step 2-3, iteration cluster process, the result clustered for the first time according to step 2-2, recalculate the cluster of each class Center.New cluster centre characteristic value is the average value of such all point feature, then in a class in searching and new cluster The immediate point of heart characteristic value calculates according to characteristic distance in 2-2 as new cluster centre point and classifying method is recalculated Each non-cluster central point generic.K end of iteration.After iterative calculation terminates, the point in each class is to form one initially Super voxel.

Step 3, the fusion of super voxel, using initially super voxel as summit, syntople is side construction non-directed graph G between each super voxel =(V, E), the difference between adjacent super voxel is measured using the Lab color spaces distance and normal vector angular separation of each initial super voxel It is different, minimized using difference in class and class inherited maximizes the initial super voxel of thought fusion, obtain the super voxel of mutative scale.Specifically Comprise the following steps：

Step 3-1, the obtained super set of voxels of initial segmentation is set as C, each super voxel is c_i, with super voxel c_iFor top Point v_iVertex set V is set up, the adjacent super voxel v for having public boundary_i, v_jBetween set up side e_ijSide collection E is formed, nothing is built with V and E To figure G.According to the three dimensional space coordinate at P midpoints, k neighbour's point coordinates covariance matrix principal component decomposition method meters of standard are used The normal vector of each point is calculated, then each edge with formula (5) to G is assigned to weight w (v_i,v_j)

Wherein,l_maxRespectively super voxel c_iMiddle mean flow rate and high-high brightness a little, θ_ijFor super voxel c_iAnd c_j Normal vector angle, the normal vector of super voxel be taken as in the super voxel normal vector average a little, α is weight factor.

Step 3-2, initial super voxel is merged, merge preceding each super voxel and oneself be initialized as a region A_i, v_kWith v_lIt is region A_iIn two super voxels, the internal dissmilarity degree of definition for the region minimum spanning tree maximum side, such as formula (7) institute Show

Int(A_i)=maxw (v_k,v_l),v_k,v_l∈A_i,(v_k,v_l)∈E (7)

The outside dissmilarity degree in two regions is defined with formula (8)

Dif(A_i,A_j)=minw (v_m,v_n),v_m∈A_i,v_n∈A_j,(v_m,v_n)∈E (8)

MInt (Ai, Aj) is A_iAnd A_jThe minimum internal diversity in the two regions, is calculated with formula (9) afterwards

MInt(A_i,A_j)=min ((Int (A_i)+τ(A_i)),(Int(A_j+τ(A_j))) (9)

Wherein τ (A_i)=e/ | A_i|, | A_i| it is region A_iThe number of point is included, e is the constant of setting；

The outside dissmilarity degree and internal dissmilarity degree in two regions are compared afterwards, closed if the formula that meets (10) And the two regions, otherwise nonjoinder

MInt(A_i,A_j)<Dif(A_i,A_j) (10)

The region that region merging technique process is performed until in P obtains the super voxel of mutative scale without untill can merging.

Invention is using Poisson dish method of sampling selection initial seed point, and compared with the uniform grid method of sampling, this method is more Meet the distribution character of cone cell on the retina, help to concentrate uniform sampling in uneven three-dimensional data points.

With reference to specific embodiment, the present invention will be further described.

Embodiment 1

Fig. 1 carries out pre-segmentation and mutative scale point for the present invention's to the RGB-D images that a frame Kinect depth cameras are gathered The process and its result schematic diagram cut.

Split a frame RGB-D images underneath with the undue segmentation method of the super voxel of mutative scale of the present invention, comprise the following steps：

Step 1, seed point are chosen, and a frame RGB-D view data is P, and resolution ratio is that 480 rows 640 are arranged, each data point Packet contains 3 Color Channels (r, g, b) and 1 depth channel (d).A point p is randomly selected from P₀It is used as initial seed Point, is set radius threshold R=20 as the minimum range between seed point to be generated, is sampled using Poisson dish sampling algorithm in P Obtain seed point set.Specifically include following steps：

Step 1-1, a point p is randomly selected in P₀As initial seed point, sampled point queue L will be enlivened₁It is initialized as Sky, p₀Add L₁, by inactive sampled point queue L₂It is initialized as sky；

If step 1-2, L₁It is not sky, then from L₁In go out one point p of team_i, with p_iFor the center of circle, R and 2R are the concentric of radius Circle randomly chooses candidate's sampled point in region, and L is added into if the distance of candidate's sampled point and existing seed point is more than R₁.Such as Fruit attempts K=30 times still without qualified candidate's sampled point, then by p_iFrom L₁It is middle to delete, and add L₂；

Step 1-3, L₂In point be selected seed point

Step 2, super voxel pre-segmentation, are turned the color space of each point in P from RGB with the color space conversion formula of standard Be changed to Lab, depth value d is converted to three dimensional space coordinate (x, y, z), using each seed point as cluster centre, Color distance and Three dimensions distance iterates to calculate the distance between non-seed point and seed point, obtains initial super voxel segmentation result.Specifically include Following steps：

Step 2-1, the depth value at RGB-D images midpoint is converted into three dimensional space coordinate, if the point that the i-th row jth is arranged in P For p_ij, its depth value is d_ij, depth value is converted into three dimensional space coordinate using formula (1)：

According to the exploitation handbook of Kinect depth cameras, f=570.3, c_x=320, c_y=240.It is empty with the color of standard Between conversion formula the RGB color of each point in P is converted into Lab colors.Changed by color space and coordinate space, the point in P It is expressed as 6 dimensional vectors [l, a, b, x, y, z]；

Step 2-2, the individual seed point obtained using step 1 carry out range searching cluster as initial cluster center, calculate each poly- The distance of point and the cluster centre in the 2R*2R contiguous ranges of class center, each non-cluster central point is ranged special with it The minimum cluster centre of distance is levied to complete two point p in the first cluster process, P_iAnd p_jBetween characteristic distance measure formulas such as Under：

In formula (2), d_labFor color distance, d_xyzFor space length, λ=1.

Step 2-3, iteration cluster process, the result clustered for the first time according to step 2-2, recalculate the cluster of each class Center.New cluster centre characteristic value is the average value of all point features of each class, is then found in a class and new cluster The immediate point of central feature value calculates according to characteristic distance in 2-2 as new cluster centre point and classifying method is counted again Calculate each non-cluster central point generic.K=10 end of iteration.After iterative calculation terminates, the point in each class is to form one Individual initial super voxel, as shown in pre-segmentation result in Fig. 1.

Step 3, the fusion of super voxel, using initially super voxel as summit, syntople is side construction non-directed graph G between each super voxel =(V, E), the difference between adjacent super voxel is measured using the Lab color spaces distance and normal vector angular separation of each initial super voxel It is different, minimized using difference in class and class inherited maximizes the initial super voxel of thought fusion, obtain the super voxel of mutative scale.Specifically Comprise the following steps：

Step 3-1, the obtained super set of voxels of initial segmentation is set as C, each super voxel is c_i, with super voxel c_iFor top Point v_iVertex set V is set up, the adjacent super voxel v for having public boundary_i, v_jBetween set up side e_ijSide collection E is formed, nothing is built with V and E To figure G.According to the three dimensional space coordinate at P midpoints, k neighbour's point coordinates covariance matrix principal component decomposition method meters of standard are used The normal vector of each point is calculated, then each edge with formula (5) to G is assigned to weight w (v_i,v_j)

Wherein,l_maxRespectively super voxel c_iMiddle mean flow rate and high-high brightness a little, θ_ijFor super voxel c_iAnd c_j Normal vector angle, the normal vector of super voxel be taken as in the super voxel normal vector average a little, α=0.5.

Step 3-2, initial super voxel merge, and each super voxel oneself is initialized as a region A before merging_i, v_kAnd v_lIt is area Domain A_iIn two super voxels, the internal dissmilarity degree of definition for the region minimum spanning tree maximum side, as shown in formula (7)

Int(A_i)=maxw (v_k,v_l),v_k,v_l∈A_i,(v_k,v_l)∈E (7)

Two region A are defined with formula (8)_iAnd A_jOutside dissmilarity degree, wherein v_mIt is region A_iIn super voxel, v_nIt is area Domain A_jIn super voxel

Dif(A_i,A_j)=minw (v_m,v_n),v_m∈A_i,v_n∈A_j,(v_m,v_n)∈E (8)

MInt(A_i,A_j) it is A_iAnd A_jThe internal dissmilarity degree of minimum in the two regions, is calculated with formula (9) afterwards

MInt(A_i,A_j)=min ((Int (A_i)+τ(A_i)),(Int(A_j+τ(A_j))) (9)

Wherein τ (A_i)=e/ | A_i|, | A_i| it is region A_iInclude the number of point, e=100.

The outside dissmilarity degree and internal dissmilarity degree in two regions are compared afterwards, closed if the formula that meets (10) And the two regions, otherwise nonjoinder

MInt(A_i,A_j)<Dif(A_i,A_j) (10)

Region that region merging technique process is performed until in P is not untill can merging, finally giving after merging Each region is the super voxel of mutative scale that segmentation is obtained, as shown in the super voxel segmentation result of mutative scale in Fig. 1.

The inventive method obtains the super voxel of different scale by once calculating the segmentation in same frame RGB-D data, bright Show to be different from once to calculate and the undue segmentation method of the more consistent super voxel of yardstick can only be obtained on same frame data, be more beneficial for Continue the amount of calculation and multiscale analysis of algorithm after the decrease.

Claims (4)

1. a kind of super voxel dividing method of RGB-D images mutative scale, it is characterised in that comprise the following steps：

Step 1, seed point are chosen, if a frame RGB-D view data is P, resolution ratio arranges for m rows n, the packet of each data point Containing 3 Color Channels (r, g, b) and 1 depth channel (d)；A point p is randomly selected from P₀(numeral 0 is subscript) is as just Beginning seed point, is set radius threshold R as the minimum range between seed point to be generated, is adopted using Poisson dish sampling algorithm in P Sample obtains seed point set；

Step 2, pre-segmentation is carried out to super voxel, the color space of each point in data is converted into Lab from RGB, depth value d is turned Three dimensional space coordinate (x, y, z) is changed to, afterwards using each seed point as cluster centre, Color distance and three dimensions distance change In generation, calculates the distance between non-seed point and seed point, obtains initial super voxel segmentation result；

Step 3, super voxel is merged, using initially super voxel as summit, to be that side is constructed undirected for syntople between each super voxel G=(V, E) is schemed, using between the Lab color spaces distance and the adjacent super voxel of normal vector angular separation measurement of each initial super voxel Difference, obtain the super voxel of mutative scale by merging initial super voxel, complete and the super voxel of RGB-D image mutative scales is divided Cut.

2. the super voxel dividing method of RGB-D images mutative scale according to claim 1, it is characterised in that seed in step 1 Point selection specifically includes following steps：

Step 1-1, in a frame RGB-D view data P randomly select a point p₀As initial seed point, sampled point will be enlivened Queue L₁Sky is initialized as, p₀Add L₁, by inactive sampled point queue L₂It is initialized as sky；

Step 1-2, judgement enliven sampled point queue L₁Whether it is empty, if L₁It is not sky, then from L₁In go out one point p of team_i, with p_i For the center of circle, R and 2R are random selection candidate's sampled point in the concentric circular regions of radius, if candidate's sampled point and existing seed point Distance is then added into L more than R₁；If the attempt to K times still without qualified candidate's sampled point, then by p_iFrom L₁It is middle to delete, And add L₂；Wherein R uses pixel coordinate unit, and K is default numerical value；

Step 1-3, L₂In point be selected seed point.

3. the super voxel dividing method of RGB-D images mutative scale according to claim 1 or 2, it is characterised in that in step 2 Following steps are specifically included to super voxel pre-segmentation：

Step 2-1, the depth value at RGB-D images midpoint is converted into three dimensional space coordinate, if the point that the i-th row jth is arranged in P is p_ij, its depth value is d_ij, depth value is converted into three dimensional space coordinate using formula (1)：

$p_{ij}{(x,y,z)}^T=d_{ij}{(\frac{i-c_x}{f},\frac{j-c_y}{f},1)}^T---\left(1\right)$

Wherein f is the focal length of video camera, (c_x,c_y) be image centre coordinate, with the color space conversion formula of standard by P The RGB color of each point is converted to Lab colors, and the point in P is expressed as into 6 dimensional vectors [l, a, b, x, y, z]；

Step 2-2, the individual seed point obtained using step 1 are carried out range searching cluster, calculated in each cluster as initial cluster center The distance of point in heart 2R*2R contiguous ranges and the cluster centre, by each non-cluster central point range with its feature away from From minimum cluster centre to complete two point p in first time cluster process, P_iAnd p_jBetween characteristic distance measure formulas it is as follows：

$D(p_i,p_j)=d_{lab}+\mathrm{\&lambda;}\frac{d_{xyz}}{R}---\left(2\right)$

$d_{lab}=\sqrt{{(l_i-l_j)}^2+{(a_i-a_j)}^2+{(b_i-b_j)}^2}---\left(3\right)$

$d_{xyz}=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2+{(z_i-z_j)}^2}---\left(4\right)$

In formula (2), d_labFor color distance, d_xyzFor space length, λ is the power for determining colouring information and space length information Weight, λ more large space distances are more important, and the smaller color distances of λ are more important；

In step 2-3, iteration cluster process, the result clustered for the first time according to step 2-2, the cluster for recalculating each class The heart；New cluster centre characteristic value is the average value of all point features of each class, then in a class in searching and new cluster The immediate point of heart characteristic value calculates according to characteristic distance in 2-2 as new cluster centre point and classifying method is recalculated Each non-cluster central point generic；K end of iteration；After iterative calculation terminates, the point in each class is to form one initially Super voxel.

4. three-dimensional point cloud point normal estimation method according to claim 1, it is characterised in that super voxel melts in step 3 Close, specifically include following steps：

Step 3-1, the obtained super set of voxels of initial segmentation is set as C, each super voxel is c_i, with super voxel c_iFor vertex v_i Vertex set V is set up, the adjacent super voxel v for having public boundary_i, v_jBetween set up side e_ijSide collection E is formed, non-directed graph is built with V and E G；According to the three dimensional space coordinate at P midpoints, calculate each using k neighbour's point coordinates covariance matrix principal component decompositions method of standard Point normal vector, then each edge with formula (5) to G be assigned to weight w (v_i,v_j)

$w(v_i,v_j)=\mathrm{\&alpha;}\frac{|{\stackrel{\&OverBar;}{l}}_i-{\stackrel{\&OverBar;}{l}}_j|}{l_{max}}+(1-\mathrm{\&alpha;})\frac{{\mathrm{\&theta;}}_{ij}}{\mathrm{\&pi;}}---\left(5\right)$

${\mathrm{\&theta;}}_{ij}={\cos }^{-1}\left(\frac{{\stackrel{\&RightArrow;}{n}}_i\&CenterDot;{\stackrel{\&RightArrow;}{n}}_j}{\left|{\stackrel{\&RightArrow;}{n}}_i\right|\left|{\stackrel{\&RightArrow;}{n}}_j\right|}\right)---\left(6\right)$

Wherein,l_maxRespectively super voxel c_iMiddle mean flow rate and high-high brightness a little, θ_ijFor super voxel c_iAnd c_jMethod Vector angle, the normal vector of super voxel be taken as in the super voxel normal vector average a little, α is weight factor；

Step 3-2, initial super voxel is merged, preceding each super voxel will be merged and be initialized as a region A_i, v_kAnd v_lIt is area Domain A_iIn two super voxels, the internal dissmilarity degree of definition for the region minimum spanning tree maximum side, as shown in formula (7)

Int(A_i)=max w (v_k,v_l),v_k,v_l∈A_i,(v_k,v_l)∈E (7)

Two region A are defined with formula (8)_iAnd A_jOutside dissmilarity degree, wherein v_mIt is region A_iIn super voxel, v_nIt is region A_j In super voxel

Dif(A_i,A_j)=min w (v_m,v_n),v_m∈A_i,v_n∈A_j,(v_m,v_n)∈E (8)

MInt(A_i,A_j) it is A_iAnd A_jThe internal dissmilarity degree of minimum in the two regions, is calculated with formula (9) afterwards

MInt(A_i,A_j)=min ((Int (A_i)+τ(A_i)),(Int(A_j+τ(A_j))) (9)

Wherein τ (A_i)=e/ | A_i|, | A_i| it is region A_iThe number of point is included, e is the constant of setting；

The outside dissmilarity degree and internal dissmilarity degree in two regions are compared afterwards, merge this if the formula that meets (10) Two regions, otherwise nonjoinder

MInt(A_i,A_j)<Dif(A_i,A_j) (10)

The region that region merging technique process is performed until in P obtains the super voxel of mutative scale without untill can merging.