CN102708589B - Three-dimensional target multi-viewpoint view modeling method on basis of feature clustering - Google Patents

Abstract

The invention relates to a three-dimensional target multi-viewpoint view modeling method on the basis of the feature clustering, which comprises four steps of: 1, acquiring an all attitude image of a test target; 2, extracting a target feature vector set <xi|i=1,2,...N> from an acquired all attitude image set and adopting a 7-dimensional invariant moment formed by integrating an affine invariant moment with a geometric invariant moment and further carrying out regularization to describe an appearance feature of the all attitude image; 3, determining an optimal cluster number Cop; and 4, clustering target feature vectors into the Cop type by adopting a k-means clustering method and using a small amount of finally obtained clustering centers as modeling results. Aiming at the defect that under different viewpoints, due to difference of target images, the single target view description cannot identify the target, the invention establishes the three-dimensional target multi-viewpoint view modeling method which has a small description number and little redundant information and can be used for well describing the target all-attitude feature vector set. The three-dimensional target multi-viewpoint view modeling method has high practical value and wide application prospect in the field of the pattern recognition.

Description

A kind of many viewpoints of objective View Modeling Method based on feature clustering

Technical field

The present invention relates to a kind of many viewpoints of objective View Modeling Method based on feature clustering, belong to area of pattern recognition, be specifically related to the aspects such as target identification, Target Modeling and Data Reduction.For the modeling of many viewpoints of objective, be applicable to the problem that simple target view description that under different points of view, target image difference causes can not be identified target.

Background technology

Objective identification is an important research direction in computer vision field.The three-dimensional information that obtains in actual applications at present target is often very difficult, and the image that identification objective still mainly forms by identification target two-dimensional projection completes.Target two-dimensional imaging (projection) process has caused partial information loss, and under different points of view, the appearance difference of complex target is obvious, makes the objective identification difficulty of fast and stable very large.

By objective being carried out to modeling to form the comprehensive description to target shape, be one of means that address this problem.This just need to study many viewpoints View Modeling Method of objective.Idea is under a plurality of viewpoints, the description of target image to be integrated to the description as target intuitively.Yet for same target, its attitude number cannot be exhaustive, between different description, a lot of redundant informations have also been comprised.Therefore, need to when describing, yojan redundancy retain as far as possible important description.

The viewpoint spherical surface partitioning algorithm of classical many viewpoints of objective View Modeling Method based target geometry and visible relation, the geometry topological characteristic of main based target image, the modeling result finally obtaining is the multi-pose characteristic view collection about target.But these class methods depend on geometry and the feature of target to the structure of target signature view collection, generally can only, for a certain class certain objects of certain complexity, as solid of revolution, quadric surface body etc., in reality, be difficult to application.

Summary of the invention

The object of the invention is to: a kind of many viewpoints of objective View Modeling Method based on feature clustering is provided, the simple target view description that it causes for target image difference under different points of view can not be identified the shortcoming of target, has built a kind of many viewpoints of objective View Modeling Method of describing that quantity is little, redundant information is few, can better describing the full posture feature vector set of target.

Its technical scheme is as follows:

A kind of many viewpoints of objective View Modeling Method based on feature clustering of the present invention, it comprises the following steps:

Step 1: the full attitude image that obtains test target.Target is placed in to viewpoint ball center, and target carriage change is equivalent to the difference object observing of video camera in viewpoint spherical surface.Uniform sampling in viewpoint spherical surface, obtaining quantity is the full attitude image set of the objective of N.

Step 2: extract target feature vector set { x from the full attitude image set obtaining _i| i==1,2 ... N}.Employing combine affine not bending moment and geometric invariant moment and further form after regularization 7 tie up not bending moment and describe the full attitude image of target resemblance.

Step 3: determine optimum clustering number C _op.By analyzing cluster total error quadratic sum J _eand the relation curve (J between cluster numbers C _e-c curve) and J _e-c second derivative curve zero crossing, determines optimum clustering number C _op.

Step 4: adopt k-means Method that target feature vector is gathered for C _opclass, using a small amount of cluster centre finally obtaining as modeling result.

Wherein, the full attitude image that obtains test target described in step 1, specific implementation process is as follows:

First create the target three-dimensional model that need to carry out the modeling of many viewpoints view.Then object module is placed in to the centre of sphere of imaginary viewpoint ball, because target carriage change is equivalent to the difference object observing of video camera in viewpoint spherical surface, therefore obtains video camera and observe the process that the process of viewpoint is division viewpoint spherical surface.Adopt the mode of similar division earth longitude and latitude, with same intervals, evenly divide viewpoint spherical surface, and at corresponding division points place to target imaging, can obtain the full attitude image set of target.

Wherein, the affine not bending moment described in step 2 is obtained by (1) formula, wherein μ _pqpresentation video p+qJie center square.

$\left\{\begin{array}{c}{I}_1=\frac{({\mathrm{\&mu;}}_{20}{\mathrm{\&mu;}}_{02}-{\mathrm{\&mu;}}_{11}^2)}{{\mathrm{\&mu;}}_{00}^4}\\ I_2=\frac{{\mathrm{\&mu;}}_{30}^2{\mathrm{\&mu;}}_{03}^2-6{\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{03}{\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{12}+4{\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{12}^3+4{\mathrm{\&mu;}}_{21}^3{\mathrm{\&mu;}}_{03}-3{\mathrm{\&mu;}}_{12}^2{\mathrm{\&mu;}}_{21}^2}{{\mathrm{\&mu;}}_{00}^{10}}\\ I_3=\frac{{\mathrm{\&mu;}}_{20}({\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{03}-{\mathrm{\&mu;}}_{12}^2)-{\mathrm{\&mu;}}_{11}({\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{03}-{\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{12})+{\mathrm{\&mu;}}_{02}({\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{12}-{\mathrm{\&mu;}}_{21}^2)}{{\mathrm{\&mu;}}_{00}^7}\end{array}\right.---\left(1\right)$

For carrying out front four components of comprehensive geometric invariant moment with above-mentioned affine not bending moment, by (2) formula, obtained

$\left\{\begin{array}{c}{\mathrm{\&Phi;}}_1={\mathrm{\&eta;}}_{20}+{\mathrm{\&eta;}}_{02}\\ {\mathrm{\&Phi;}}_2={({\mathrm{\&eta;}}_{20}-{\mathrm{\&eta;}}_{02})}^2+{4\mathrm{\&eta;}}_{11}^2\\ {\mathrm{\&Phi;}}_3={({\mathrm{\&eta;}}_{30}-3{\mathrm{\&eta;}}_{12})}^2+{(3{\mathrm{\&eta;}}_{21}-{\mathrm{\&eta;}}_{03})}^2\\ {\mathrm{\&Phi;}}_4={({\mathrm{\&eta;}}_{30}+{\mathrm{\&eta;}}_{12})}^2+{({\mathrm{\&eta;}}_{21}+{\mathrm{\&eta;}}_{03})}^2\end{array}\right.---\left(2\right)$

For solving the above-mentioned not difference of the order of magnitude between bending moment, need further above-mentioned not bending moment to be carried out to regularization computing

$\stackrel{\&OverBar;}{I_k}=\mathrm{sign}\left(I_k\right){\left|I_k\right|}^{2/S}---\left(3\right)$

I in formula _kthe combination that represents not regularization is bending moment not, not bending moment is combined in expression regularization, and Sign represents Heaviside function, and S represents the summation of square factor exponent number in each square component.Finally obtain describing 7 of target shape information and tie up not bending moment of regularization.

Wherein, the definite optimum clustering number C described in step 3 _op, specific implementation process is as follows:

Adopt k-means Method to carry out cluster to the full posture feature vector set of above-mentioned target, making classification count c progressively increases, c=2, and c=3 ..., record the total error quadratic sum J of each cluster _e, final according to obtaining J _e-c curve, further calculates J _e-c second derivative curve, in the situation that considering certain numerical value fluctuation, selects J _eon-c second derivative curve first close to zero point and relatively the mid point between flat zone as optimum cluster numbers c _opestimated value.

The advantage that the present invention has is: by feature clustering, with minority cluster centre, represent the distribution of the full posture feature vector of target in feature space, realized Data Reduction, method highly versatile, can conveniently determine optimum clustering number.

Accompanying drawing explanation

Fig. 1 is processing flow chart of the present invention;

Fig. 2 is the viewpoint ball schematic diagram relating in the present invention;

Fig. 3 is J in the present invention _e-c curve synoptic diagram;

Fig. 4 is that the present invention determines optimum clustering number C _opschematic diagram.

Embodiment

The present invention is directed to many viewpoints of objective view modeling problem, proposed a kind of many viewpoints of objective View Modeling Method based on Support Vector data description.

See Fig. 1, a kind of many viewpoints of objective View Modeling Method based on feature clustering of the present invention, it comprises the following steps:

Step 1: the full attitude image that obtains test target.Target is placed in to viewpoint ball center, and target carriage change is equivalent to the difference object observing of video camera in viewpoint spherical surface.Uniform sampling in viewpoint spherical surface, obtaining quantity is the full attitude image set of the objective of N.

Step 2: extract target feature vector set { x from the full attitude image set obtaining _i| i=1,2 ... N}.Employing combine affine not bending moment and geometric invariant moment and further form after regularization 7 tie up not bending moment and describe the full attitude image of target resemblance.

Step 3: determine optimum clustering number C _op.By analyzing cluster total error quadratic sum J _eand the relation curve (J between cluster numbers C _e-c curve) and J _e-c second derivative curve zero crossing, determines optimum clustering number C _op.

Step 4: adopt k-means Method that target feature vector is gathered for C _opclass, using a small amount of cluster centre finally obtaining as modeling result.

First, obtain the full attitude image of target and need to create the target three-dimensional model for the treatment of modeling.See Fig. 2, object module is placed in to the centre of sphere of viewpoint ball.Viewpoint ball is the imaginary unit's ball being defined in three dimensional euclidean space, under different points of view, the imaging process of target can be regarded target as and take the centre of sphere as the process of starting point to different directions projection, in viewpoint spherical surface, the corresponding projecting direction vector of every bit P P, sees target along the rectangular projection of projecting direction P corresponding to the point of P viewpoint spherical surface.Therefore obtain video camera and observe the process that the process of viewpoint is division viewpoint spherical surface.Adopt the mode of similar division earth longitude and latitude, with same intervals, evenly divide viewpoint spherical surface, and at corresponding division points place to target imaging, can obtain element number is the full attitude image set of target of N.

Attitude image set is extracted to target signature, obtain the proper vector set { x that describes the full attitude appearance information of this target _i| i=1,2 ... N}.Adopt 7 of comprehensive affine not bending moment and geometric invariant moment formation to tie up not bending moment and described target shape.Affine not bending moment is obtained by (1) formula, wherein μ _pqpresentation video p+qJie center square.

$\left\{\begin{array}{c}{I}_1=\frac{({\mathrm{\&mu;}}_{20}{\mathrm{\&mu;}}_{02}-{\mathrm{\&mu;}}_{11}^2)}{{\mathrm{\&mu;}}_{00}^4}\\ I_2=\frac{{\mathrm{\&mu;}}_{30}^2{\mathrm{\&mu;}}_{03}^2-6{\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{03}{\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{12}+4{\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{12}^3+4{\mathrm{\&mu;}}_{21}^3{\mathrm{\&mu;}}_{03}-3{\mathrm{\&mu;}}_{12}^2{\mathrm{\&mu;}}_{21}^2}{{\mathrm{\&mu;}}_{00}^{10}}\\ I_3=\frac{{\mathrm{\&mu;}}_{20}({\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{03}-{\mathrm{\&mu;}}_{12}^2)-{\mathrm{\&mu;}}_{11}({\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{03}-{\mathrm{\&mu;}}_{21}{\mathrm{\&mu;}}_{12})+{\mathrm{\&mu;}}_{02}({\mathrm{\&mu;}}_{30}{\mathrm{\&mu;}}_{12}-{\mathrm{\&mu;}}_{21}^2)}{{\mathrm{\&mu;}}_{00}^7}\end{array}\right.$

For carrying out front four components of comprehensive geometric invariant moment with above-mentioned affine not bending moment, by (2) formula, obtained

$\left\{\begin{array}{c}{\mathrm{\&Phi;}}_1={\mathrm{\&eta;}}_{20}+{\mathrm{\&eta;}}_{02}\\ {\mathrm{\&Phi;}}_2={({\mathrm{\&eta;}}_{20}-{\mathrm{\&eta;}}_{02})}^2+{4\mathrm{\&eta;}}_{11}^2\\ {\mathrm{\&Phi;}}_3={({\mathrm{\&eta;}}_{30}-3{\mathrm{\&eta;}}_{12})}^2+{(3{\mathrm{\&eta;}}_{21}-{\mathrm{\&eta;}}_{03})}^2\\ {\mathrm{\&Phi;}}_4={({\mathrm{\&eta;}}_{30}+{\mathrm{\&eta;}}_{12})}^2+{({\mathrm{\&eta;}}_{21}+{\mathrm{\&eta;}}_{03})}^2\end{array}\right.$

For solving the above-mentioned not difference of the order of magnitude between bending moment, need further above-mentioned not bending moment to be carried out to regularization computing

$\stackrel{\&OverBar;}{I_k}=\mathrm{sign}\left(I_k\right){\left|I_k\right|}^{2/S}$

The combination that wherein Ik represents not regularization is bending moment not, not bending moment is combined in expression regularization, and Sign represents Heaviside function, and S represents the summation of square factor exponent number in each square component.For example, for second component Ф of geometric invariant moment ₂, S=(2+0) * 2=4, wherein (2+0) represents square factor η ₂₀exponent number and, * 2 represent square factor η ₂₀power operation.

Finally obtain describing 7 of target shape information and tieed up not bending moment of regularization.

The full posture feature vector set of above-mentioned target is carried out to cluster, using a small amount of cluster centre finally obtaining as modeling result.Adopted k-means Method.The basis of k-average is error sum of squares criterion.If N _ii cluster Γ _iin number of samples, m _iit is the average of these samples.Γ _iin each sample y and average m _ibetween error sum of squares after all classes are added be:

$J_e=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{y\&Element;{\mathrm{\&Gamma;}}_i}{\mathrm{\&Sigma;}}{\left|\right|y-m_i\left|\right|}^2---\left(4\right)$

J _emeasured with c cluster centre m ₁, m ₂..., m _cthe total square-error producing while representing c sample set.To different cluster J _evalue be different, make J _eminimum cluster is the optimal result under error sum of squares criterion.

Adopt following methods to determine optimum clustering number c _op: as shown in Figure 3, making classification count c progressively increases, to c=2, and c=3 ... use respectively k-mean algorithm, error sum of squares will reduce along with the increase of c in theory.There is certain critical point OP, when cluster numbers is c _opwhen individual, J _ealong with cluster numbers, c is increased to c from 1 _opand reduce rapidly J when c increases again _ethough reduce to some extent, the speed reducing slows down.As shown in Figure 4, according to calculating J _e-c curve, further calculates J _e-c second derivative curve, in the situation that considering certain numerical value fluctuation, selects J _eon-c second derivative curve first close to zero point and relatively the mid point between flat zone as optimum cluster numbers c _opestimated value.

Finally, adopting k-means clustering algorithm, is c by the full posture feature vector set of target cluster _opindividual classification, using the cluster centre of each subclass as many viewpoints of objective view modeling result.

Claims (1)

1. many viewpoints of the objective View Modeling Method based on feature clustering, is characterized in that: it comprises the following steps:

Step 1: the full attitude image that obtains test target; Target is placed in to viewpoint ball center, and target carriage change is equivalent to the difference object observing of video camera in viewpoint spherical surface, uniform sampling in viewpoint spherical surface, and obtaining quantity is the full attitude image set of the objective of N;

Step 2: extract the full posture feature vector of target set { x from the full attitude image set obtaining _i| i=1,2 ... N}, adopt combine affine not bending moment and geometric invariant moment and further form after regularization 7 tie up not bending moment and describe the full attitude image of target resemblance;

Step 3: determine optimum clustering number C _op; By analyzing cluster total error quadratic sum J _eand the relation curve between cluster numbers C is J _e-c curve and J _e-c second derivative curve zero crossing, determines optimum clustering number C _op;

Step 4: adopt k-means Method that the set of the full posture feature vector of target is gathered for C _opclass, using a small amount of cluster centre finally obtaining as modeling result;

Wherein, the full attitude image that obtains test target described in step 1, specific implementation process is as follows: first create the target three-dimensional model that need to carry out the modeling of many viewpoints view, then object module is placed in to the centre of sphere of imaginary viewpoint ball, because target carriage change is equivalent to the difference object observing of video camera in viewpoint spherical surface, therefore obtain video camera and observe the process that the process of viewpoint is division viewpoint spherical surface; Adopt the mode of similar division earth longitude and latitude, with same intervals, evenly divide viewpoint spherical surface, and at corresponding division points place to target imaging, obtain the full attitude image set of target;

Wherein, the affine not bending moment described in step 2 is obtained by (1) formula, wherein μ _pqpresentation video p+qJie center square;

For carrying out front four components of comprehensive geometric invariant moment with above-mentioned affine not bending moment, by (2) formula, obtained

For solving the difference of the order of magnitude between affine not bending moment and geometric invariant moment, need further above-mentioned not bending moment to be carried out to regularization computing

In formula, I _kthe combination that represents not regularization is bending moment not, not bending moment is combined in expression regularization, and sign represents Heaviside function, and S represents the summation of square factor exponent number in each square component, finally obtains describing 7 of target shape information and ties up not bending moment of regularization;

Wherein, the definite optimum clustering number C described in step 3 _op, specific implementation process is as follows:

Adopt k-means Method to carry out cluster to the full posture feature vector set of above-mentioned target, making classification count c progressively increases, c=2, and c=3 ..., record the total error quadratic sum J of each cluster _e, final according to obtaining J _e-c curve, further calculates J _e-c second derivative curve, in the situation that considering certain numerical value fluctuation, selects J _eon-c second derivative curve first close to zero point and relatively the mid point between flat zone as optimum cluster numbers c _opestimated value.