CN106652048A - Three-dimensional model point-of-interest extraction method based on 3D-SUSAN (Small Univalue Segment Assimilating Nucleus) operator - Google Patents

Abstract

The invention puts forward a three-dimensional model point-of-interest extraction method based on a 3D-SUSAN (Small Univalue Segment Assimilating Nucleus) operator. The method comprises the following steps that: firstly, reading the vertex of a three-dimensional model, simplifying the vertex of the model, carrying out model meshing, and calculating the curvature of each vertex; then, through Gaussian filtering, processing the curvature of each vertex, and carrying out denoising processing; thirdly, determining a central point, taking 36 neighbor points around the central point, if the curvature difference of the central vertex and the neighbor point is smaller than a similarity threshold value, entering a next step, and if the USAN value of each vertex is smaller than a geometrical threshold value, proving that the central vertex is a point of interest; and finally, evaluating the three-dimensional model point of interest which is extracted on the basis of the 3D-SUSAN operator and is put forward by the invention from the aspects of fault rate, miss rate and error rate. By use of the method, the three-dimensional model point of interest can be extracted, and the advantages of high anti-noise capability, small calculation amount, high efficiency and stable performance of the SUSAN still can be kept.

Description

Threedimensional model point-of-interest extracting method based on 3D-SUSAN operators

Technical field

The present invention relates to threedimensional model point-of-interest is extracted, specifically be based on 3D-SUSAN angle point operators with regard to a kind of Threedimensional model point-of-interest extracting method.

Background technology

3D applications in recent years have obtained extensive concern, and during threedimensional model develops, it is very that point-of-interest is extracted The essence of many 3D application technologies, for example, the matching retrieval of lattice simplified, mesh segmentation, viewpoint selection and threedimensional model.Use feeling is emerging Interesting Point matching threedimensional model, it is desirable to provide the local feature region of model.This method is also applied for tri-dimensional picture identification, matching.

In the field that point-of-interest is extracted, either in theory or in actual 3D applications, we can mesh See the huge advance of this technology.These progress can be emerged from from model analysis, model transmission and model rendering.Sense is emerging Interest point extractive technique is usually used mathematical measure and extracts feature, such as the curvature in geometry teaching, curvature is more big more notable.

Threedimensional model point-of-interest is extracted and is also faced with some challenges and difficulty.First, do not exist till now bright The true definition with regard to point-of-interest.Show that point-of-interest is to give directions to there is larger difference with other points based on some subjectivity researchs Different point.Then, because the topological direction of 3D grids is arbitrary, model vertices just have arbitrary neighbours summit.Finally, except The positional information on summit, without other about summit information, it is therefore desirable to calculate other characteristic informations on summit.

Two-dimentional corner detection operator can broadly be summarized as 3 classes to be respectively based on the Corner Detection of gray level image, is based on The Corner Detection of bianry image and the Corner Detection based on contour curve.Common angle point operator includes Moravec Corner Detections Operator, Harris corner detection operators and SUSAN corner detection operators etc..Wherein SUSAN corner detection operators according to it is each The related regional area of picture point has identical pixel value.If each pixel value in a certain window area and the window The pixel value at center is same or similar, and the region is referred to as USAN regions by us.If USAN regions are less than threshold value, we will recognize Central point for the region is angle point.

Bibliography

[1]HelinDutagaci,Chun Pan Cheung,AfzalGodil,“Evaluation of 3D interest point detection techniquesvia human-generated ground truth”, Springer-Verlag,2012.

[2]C.H.Lee,A.Varshney,and D.W.Jacobs,"Mesh Saliency,"ACM SIGGRAPH, vol.174,pp.659-666,2005.

[3]A Benchmark for 3D Interest Points Marked by Human Subjects, [Online].

Available:http://www.itl.nist.gov/iad/vug/sharp/benchmark/ 3DInterestPoint.

[4]Castellani,U.,Cristani,M.,Fantoni,S.,Murino,V.:Sparse pointsmatching by combining 3Dmesh saliency with statistical descriptors.Comput.Graph.Forum27(2),643–652(2008).

The content of the invention

The purpose of the present invention is in view of threedimensional model point-of-interest is to lattice simplified, mesh segmentation, viewpoint selection and 3D moulds The importance of type matching retrieval, proposes a kind of threedimensional model point-of-interest extracting method based on 3D-SUSAN angle point operators, should The SUSAN corner detection operators of two dimensional image feature point extraction are introduced field of three dimension by method, it is proposed that a kind of 3D-SUSAN angles Point detective operators, and effectively extract threedimensional model point-of-interest.Anti-noise ability that the method inherits SUSAN operators is strong, calculate Measure the features such as little, efficiency high and stable performance.

To achieve these goals, the present invention is achieved by the following technical programs, specifically includes following steps：：

Step (1)：Algorithm is read using grid and calculate the vertex position coordinate of threedimensional model to be extracted, and utilize grid Simplify operator to be simplified on threedimensional model summit to be extracted.

Step (2)：The curvature on each summit is calculated using Curvature Operator.

Step (3)：Gauss process is carried out to curvature value using Gaussian function, noise is removed, the curvature value after denoising is obtained.

Step (4)：Summit cluster is defined, according to the distance of neighbours summit and current vertex, selected distance is most short by front 36 Individual neighbours summit is used as the summit cluster compared with current vertex curvature value.

Step (5)：According to the summit cluster of step (4) definition, if current vertex and other neighbours summits in the cluster of summit Curvature difference less than definition similarity threshold, then using the current vertex as candidate point.Then the USAN of candidate point is compared Value and the size of geometry threshold value, if USAN values are less than geometry threshold size, return the candidate point as point-of-interest.

Step (6)：Define evaluation operators FPE, FNE and WME^[1], according to algorithm miss rate, perissology, error rate to step (5) point-of-interest for obtaining is estimated.

Utilization Curvature Operator described in step (2) calculates the curvature on each summit, and wherein curvature estimation formula is as follows：

If three dimensional curve equations are：X=x (t), y=y (t), z=z (t)；

1) derivation respectively, obtains x ' (t), y ' (t), z ' (t)；

2) 2 ranks are asked to lead respectively, (t), z is " (t) to obtain x " (t), y "；

3) three single orders are led to be combined and regards a trivector as：

R '=(x ' (t), y ' (t), z ' (t))；

4) three second orders are led to be combined and regards a trivector as：

R "=(x " (t), (t), y " z " is (t))；

5) curvature estimation formula is：

Gaussian function process is carried out to vertex curvature in step (2) using Gauss formula in described step (3), σ is obtained The Gauss weighted average of average curvature in region, described Gaussian function formula is as follows：

G is the Gauss weighted average of average curvature in σ regions,It is the curvature value on summit, x is currently counted in σ regions Calculate the neighbours summit of vertex v.

Summit cluster is chosen described in step (4) and refers to that centered on currently calculating summit selected distance is most short by front 36 Individual neighbours summit is used as the summit contrasted with current vertex.

Threedimensional model vertex distance computing formula is：

(x₁,y₁,z₁),(x₂,y₂,z₂) be respectively two summits coordinate.

The step (5) to implement process as follows：

5-1. needs the affiliated summit cluster of each culminating point of dynamic calculation first of all for threshold adaptive ability is increased Incurvature difference similarity threshold.

Wherein, culminating point refers to the current summit for calculating in the cluster of summit；A represents the mean value of curvature difference.c(v₀) table Show culminating point curvature, c (v) represents other neighbours' vertex curvatures in the cluster of summit.

T represents curvature similarity threshold, and a represents the mean value of curvature difference, and c represents the curvature value of cluster inner vertex, and N is represented Cluster point number, takes here 36.

If current vertex is less than curvature similarity threshold with the curvature absolute value differences on its neighbours summit, by current vertex As candidate point.

5-2. calculates the USAN values of each candidate point：

T represents curvature similarity threshold, c (v₀) culminating point curvature is represented, c (v) represents other neighbours in the cluster of summit Vertex curvature.D represents the USAN values of certain candidate point, and its initial value is 0.

With reference to the SUSAN operators applied in the extraction of two dimensional image focus, typically when angle point is extracted, geometry threshold value is equal to The 1/2 of USAN regions, therefore in 3D-SUSAN operators, we equally choose the 1/2 of USAN regions as geometry threshold value：

g(v₀) represent v₀Geometry threshold value, d (v₀) represent v₀USAN regional values.

If the USAN values of candidate point are less than geometry threshold value, then it is assumed that the candidate point is angle point.

5-3. does non-maxima suppression to the angle point for obtaining, and finally gives point-of-interest.

Assessment algorithm described in step (6) is specific as follows：Assessment algorithm is specifically divided into two steps：One is to set up the true field of model Scape, two is using the real scene assessment point-of-interest extraction algorithm set up.

(1) evaluation criteria of model real scene interest point identification is set up, the part refer to website A Benchmark for 3D Interest Points Marked by Human Subjects^[2]Database, including threedimensional model and every The artificial extraction point-of-interest coordinate set of individual model.

(2) using the real scene assessment interest point extraction algorithm set up, using FPE (false positive Errors), the interest that FNE (false negative errors) and WME (weighted miss error) are extracted to algorithm Point is estimated.

G_r=(p ∈ M | d (g, p)≤r)

G_rModel real scene interest point set is represented, M represents the point of interest of algorithm extraction model, and d represents summit and summit Distance, r represents the region of radius.

N_CRepresent the correct point quantity that algorithm is extracted, N_GRepresent point of interest number total in real scene.

N_ARepresent the point of interest number that algorithm is extracted, N_CRepresent that algorithm extracts the correct point quantity for obtaining.

n_iRepresent n_iIndividual volunteer identifies i points, g_iReal scene 3D-SUSAN assessment results when representing that radius is r.

Beneficial effects of the present invention are as follows：

The present invention proposes a kind of threedimensional model point-of-interest extracting method based on 3D-SUSAN angle point operators, and the method will SUSAN corner detection operators for two dimensional image feature point extraction introduce field of three dimension, it is proposed that a kind of 3D-SUSAN angle points Detective operators, and efficiently extract the point-of-interest of threedimensional model.

3D-SUSAN corner detection operators not only can extract threedimensional model point of interest and still remain the anti-noise of SUSAN Ability is strong, the little efficiency high of amount of calculation, the advantage of stable performance.

Description of the drawings

Fig. 1 is threedimensional model interest point extraction flow chart of the present invention based on 3D-SUSAN algorithms.

Fig. 2 is summit cluster diagram.

Fig. 3 is non-maxima suppression flow chart.

Fig. 4 is 3D-SUSAN algorithm evaluation result figures.

Fig. 5 is interest point identification.

Specific embodiment

Below in conjunction with the accompanying drawings the invention will be further described.

As shown in figure 1, a kind of threedimensional model point-of-interest extracting method based on 3D-SUSAN angle point operators, specifically includes Following steps：

Step (1)：Algorithm is read using grid and calculate the vertex position coordinate of threedimensional model to be extracted, and utilize grid Simplify operator to be simplified on threedimensional model summit to be extracted.

Step (2)：The curvature on each summit is calculated using Curvature Operator.

Step (3)：Gauss process is carried out to curvature value using Gaussian function, noise is removed, the curvature value after denoising is obtained.

Step (4)：Summit cluster is defined, according to the distance of neighbours summit and current vertex, selected distance is most short by front 36 Used as the summit cluster compared with current vertex curvature value, accompanying drawing 2 is summit cluster diagram on individual neighbours summit.

Step (5)：According to the summit cluster of step (4) definition, if current vertex and other neighbours summits in the cluster of summit Curvature difference less than definition similarity threshold, then using the current vertex as candidate point.Then the USAN of candidate point is compared Value and the size of geometry threshold value, if USAN values are less than geometry threshold size, return the candidate point as point-of-interest.

Step (6)：Define evaluation operators FPE, FNE and WME^[1], according to algorithm miss rate, perissology, error rate to step (5) point-of-interest for obtaining is estimated.

Utilization Curvature Operator described in step (2) calculates the curvature on each summit, and wherein curvature estimation formula is as follows：

If three dimensional curve equations are：X=x (t), y=y (t), z=z (t)；

6) derivation respectively, obtains x ' (t), y ' (t), z ' (t)；

7) 2 ranks are asked to lead respectively, (t), z is " (t) to obtain x " (t), y "；

8) three single orders are led to be combined and regards a trivector as：

R '=(x ' (t), y ' (t), z ' (t))；

9) three second orders are led to be combined and regards a trivector as：

R "=(x " (t), (t), y " z " is (t))；

10) curvature estimation formula is：

Gaussian function process is carried out to vertex curvature in step (2) using Gauss formula in described step (3), σ is obtained The Gauss weighted average of average curvature in region, described Gaussian function formula is as follows：

G is the Gauss weighted average of average curvature in σ regions,It is the curvature value on summit, x is currently counted in σ regions Calculate the neighbours summit of vertex v.

Summit cluster is chosen described in step (4) and refers to that centered on currently calculating summit selected distance is most short by front 36 Individual neighbours summit is used as the summit contrasted with current vertex.

Threedimensional model vertex distance computing formula is：

(x₁,y₁,z₁),(x₂,y₂,z₂) be respectively two summits coordinate.

The step (5) to implement process as follows：

5-1. needs the affiliated summit cluster of each culminating point of dynamic calculation first of all for threshold adaptive ability is increased Incurvature difference similarity threshold.

Wherein, culminating point refers to the current summit for calculating in the cluster of summit；A represents the mean value of curvature difference.c(v₀) table Show culminating point curvature, c (v) represents other neighbours' vertex curvatures in the cluster of summit.

T represents curvature similarity threshold, and a represents the mean value of curvature difference, and c represents the curvature value of cluster inner vertex, and N is represented Cluster point number, takes here 36.

If current vertex is less than curvature similarity threshold with the curvature absolute value differences on its neighbours summit, by current vertex As candidate point.

5-2. calculates the USAN values of each candidate point：

T represents curvature similarity threshold, c (v₀) culminating point curvature is represented, c (v) represents other neighbours in the cluster of summit Vertex curvature.D represents the USAN values of certain candidate point, and its initial value is 0.

With reference to the SUSAN operators applied in the extraction of two dimensional image focus, typically when angle point is extracted, geometry threshold value is equal to The 1/2 of USAN regions, therefore in 3D-SUSAN operators, we equally choose the 1/2 of USAN regions as geometry threshold value：

g(v₀) represent v₀Geometry threshold value, d (v₀) represent v₀USAN regional values.

If the USAN values of candidate point are less than geometry threshold value, then it is assumed that the candidate point is angle point.

5-3. does non-maxima suppression to the angle point for obtaining, and finally gives point-of-interest, and accompanying drawing 3 is non-maxima suppression stream Cheng Tu.

Assessment algorithm described in step (6) is specific as follows：Assessment algorithm is specifically divided into two steps：One is to set up the true field of model Scape, two is using the real scene assessment point-of-interest extraction algorithm set up.

(1) evaluation criteria of model real scene interest point identification is set up, the part refer to website A Benchmark for 3D Interest Points Marked by Human Subjects^[2]Database, including threedimensional model and every The artificial extraction point-of-interest coordinate set of individual model.

(2) using the real scene assessment interest point extraction algorithm set up, using FPE (false positive Errors), the interest that FNE (false negative errors) and WME (weighted miss error) are extracted to algorithm Point is estimated.

G_r=(p ∈ M | d (g, p)≤r)

G_rModel real scene interest point set is represented, M represents the point of interest of algorithm extraction model, and d represents summit and summit Distance, r represents the region of radius.

N_CRepresent the correct point quantity that algorithm is extracted, N_GRepresent point of interest number total in real scene.

N_ARepresent the point of interest number that algorithm is extracted, N_CRepresent that algorithm extracts the correct point quantity for obtaining.

n_iRepresent n_iIndividual volunteer identifies i points, g_iReal scene 3D-SUSAN assessment results when representing that radius is r are shown in Accompanying drawing 43D-SUSAN algorithm evaluation result figures.

FNE represents that miss rate, FPE represent that perissology, WME represent error rate, when they reduce speed it is faster, it was demonstrated that the calculation It is better that method extracts point of interest result.

The present invention extracts the result of threedimensional model point of interest for algorithm, is commented from error rate, miss rate, error rate Estimate, and the point of interest extracted with Meshsaliency is made comparisons.Just as can be seen that and Mesh saliency from accompanying drawing 5^[3]Phase Than the point of interest that 3D-SUSAN is extracted is substantially few than the point of interest that Mesh saliency are extracted, and can be seen that from accompanying drawing 5 The point of interest of the mark result of 3D-SUSAN and mankind's mark is about the same.Therefore deduce that the invention is carried in reduction algorithm Take and increased in threedimensional model point of interest quantity.

Claims (6)

1. the threedimensional model point-of-interest extracting method of 3D-SUSAN operators is based on, it is characterised in that comprised the steps：：

Step (1)：The vertex position coordinate that algorithm calculates threedimensional model to be extracted is read using grid, and using lattice simplified Operator is simplified on threedimensional model summit to be extracted；

Step (2)：The curvature on each summit is calculated using Curvature Operator；

Step (3)：Gauss process is carried out to curvature value using Gaussian function, noise is removed, the curvature value after denoising is obtained；

Step (4)：Summit cluster is defined, according to neighbours summit and the distance of current vertex, most short front 36 neighbours of selected distance Summit is occupied as the summit cluster compared with current vertex curvature value；

Step (5)：According to the summit cluster of step (4) definition, if the song on current vertex and interior other neighbours summits of summit cluster Rate difference is less than the similarity threshold of definition, then using the current vertex as candidate point；Then compare the USAN values of candidate point with The size of geometry threshold value, if USAN values are less than geometry threshold size, returns the candidate point as point-of-interest；

Step (6)：Evaluation operators FPE, FNE and WME are defined, step (5) is obtained according to algorithm miss rate, perissology, error rate To point-of-interest be estimated.

2. the threedimensional model point-of-interest extracting method based on 3D-SUSAN operators according to claim 1, its feature exists The curvature on each summit is calculated in the utilization Curvature Operator described in step (2), wherein curvature estimation formula is as follows：

If three dimensional curve equations are：X=x (t), y=y (t), z=z (t)；

1) derivation respectively, obtains x ' (t), y ' (t), z ' (t)；

2) 2 ranks are asked to lead respectively, (t), z is " (t) to obtain x " (t), y "；

3) three single orders are led to be combined and regards a trivector as：

R '=(x ' (t), y ' (t), z ' (t))；

4) three second orders are led to be combined and regards a trivector as：

R "=(x " (t), (t), y " z " is (t))；

5) curvature estimation formula is：

3. the threedimensional model point-of-interest extracting method based on 3D-SUSAN operators according to claim 2, its feature exists Vertex curvature carries out Gaussian function process during Gauss formula is utilized in described step (3) to step (2), obtains in σ regions The Gauss weighted average of average curvature, described Gaussian function formula is as follows：

G is the Gauss weighted average of average curvature in σ regions,It is the curvature value on summit, x is that current calculating is pushed up in σ regions The neighbours summit of point v.

4. the threedimensional model point-of-interest extracting method based on 3D-SUSAN operators according to claim 3, its feature exists Choose summit cluster described in step (4) to refer to centered on currently calculating summit, most short front 36 neighbours of selected distance Summit is used as the summit contrasted with current vertex；

Threedimensional model vertex distance computing formula is：

$d=\sqrt{{(x_1-x_2)}^2+{(y_1-y_2)}^2+{(z_1-z_2)}^2}$

(x₁,y₁,z₁),(x₂,y₂,z₂) be respectively two summits coordinate.

5. the threedimensional model point-of-interest extracting method based on 3D-SUSAN operators according to claim 1, its feature exists In the step (5) to implement process as follows：

5-1. needs the affiliated summit cluster introversion of each culminating point of dynamic calculation first of all for threshold adaptive ability is increased Rate similarity threshold；

$a=\frac{1}{36}\mathrm{\Σ}|c\left(v\right)-c\left(v_0\right)|$

Wherein, culminating point refers to the current summit for calculating in the cluster of summit；A represents the mean value of curvature difference；c(v₀) in expression Heart vertex curvature, c (v) represents other neighbours' vertex curvatures in the cluster of summit；

$t=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\Σ}}{(a-c)}^2}$

T represents curvature similarity threshold, and a represents the mean value of curvature difference, and c represents the curvature value of cluster inner vertex, and N represents cluster Point number, takes here 36；

If the curvature absolute value differences on current vertex and its neighbours summit are less than curvature similarity threshold, using current vertex as Candidate point；

5-2. calculates the USAN values of each candidate point：

$d\left(v_0\right)=d+\exp \{-\[\frac{c\left(v\right)-c\left(v_0\right)}{t}\]\}$

T represents curvature similarity threshold, c (v₀) culminating point curvature is represented, c (v) represents that other neighbours summits are bent in the cluster of summit Rate；D represents the USAN values of certain candidate point, and its initial value is 0；

With reference to the SUSAN operators applied in the extraction of two dimensional image focus, typically when angle point is extracted, geometry threshold value is equal to USAN areas The 1/2 of domain, therefore in 3D-SUSAN operators, we equally choose the 1/2 of USAN regions as geometry threshold value：

$g\left(v_0\right)=\frac{1}{2}\×d\left(v_0\right)$

g(v₀) represent v₀Geometry threshold value, d (v₀) represent v₀USAN regional values；

If the USAN values of candidate point are less than geometry threshold value, then it is assumed that the candidate point is angle point；

$R\left(v_0\right)=\left\{\begin{array}{cc}g-d\left(v_0\right)& d\left(v_0\right)<g\\ 0& d\left(v_0\right)\&GreaterEqual;g\end{array}\right.$

5-3. does non-maxima suppression to the angle point for obtaining, and finally gives point-of-interest.

6. the threedimensional model point-of-interest extracting method based on 3D-SUSAN operators according to claim 1, its feature exists It is specific as follows in the assessment algorithm described in step (6)：Assessment algorithm is specifically divided into two steps：One is to set up model real scene, two It is using the real scene assessment point-of-interest extraction algorithm set up；

(1) evaluation criteria of model real scene interest point identification is set up, reference data includes threedimensional model and each model Artificial extraction point-of-interest coordinate set；

(2) using the real scene assessment interest point extraction algorithm set up, the interest extracted to algorithm using FPE, FNE and WME Point is estimated；

G_r=(p ∈ M | d (g, p)≤r)

G_rRepresent model real scene interest point set, M represents the point of interest of algorithm extraction model, d represent summit and summit away from From r represents the region of radius；

$FNE\left(r\right)=1-\frac{N_C}{N_G}$

N_CRepresent the correct point quantity that algorithm is extracted, N_GRepresent point of interest number total in real scene；

N_ARepresent the point of interest number that algorithm is extracted, N_CRepresent that algorithm extracts the correct point quantity for obtaining；

$WME=1-\frac{1}{{\mathrm{\&Sigma;}}_{i=1}^{N_G}{n}_i}\underset{i=1}{\overset{N_G}{\&Sigma;}}{n}_i{\mathrm{\&delta;}}_i$

n_iRepresent that ni volunteer identifies i points, g_iReal scene 3D-SUSAN assessment results when representing that radius is r.