CN104680174A - Mesh animation progressive transmission-orientated frame clustering method - Google Patents

Abstract

The invention discloses a mesh animation progressive transmission-orientated frame clustering method which comprises the following steps: matching all frame models; finding a transformation relation among coordinate matrixes of frame mesh models by an ICP algorithm to obtain a rotating matrix R and a translation matrix T between one frame and another frame, and clustering all frame mesh coordinate data by a class kmeans clustering algorithm; finally obtaining clustering results with continuous frame indexes of all the types, starting to carry out data processing type by type, respectively completing code transmission, and independently transmitting representative frames of all the types. According to the method, a mesh residual value replaces a conventional Euclidean distance during clustering, and a key frame time difference value is added, so that the clustering results are accurate, and the frame indexes included in all the types are continuous; by final type-by-type processing and code transmission, progressive transmission of mesh animations can be realized.

Description

The frame clustering method of Grid-oriented animation progressive transmission

Technical field

The invention belongs to multimedia technology field relevant to portable mobile equipment in wireless network, be specifically related to a kind of frame clustering method of Grid-oriented animation progressive transmission.

Background technology

Under existing network bandwidth conditions, data volume crosses transmission and the application that senior general seriously hinders mesh animation, therefore first performs frame clustering algorithm to mesh animation and is undertaken processing the data-handling efficiency that effectively can improve mesh animation by class again.In addition, select the partial sequence of reconstruct original mesh animation according to the network bandwidth and transmit to client, thus improving grid utilization factor, also meeting the development trend of current grid picture transmission.

But, in current mesh animation Compression Study, owing to reckoning without the progressive transmission of mesh animation, when performing frame clustering algorithm, only consider the similarity of frame coordinate data, thus frame index all kinds of comprised in the cluster result obtained is discrete.Therefore, design one not only ensures similarly between all kinds of frame interior can also make its frame index continuous print frame clustering method, is of great significance the progressive transmission tool realizing mesh animation.

Summary of the invention

The present invention is directed to the deficiency of existing frame clustering method in mesh animation progressive transmission application aspect, provide a kind of frame clustering method of Grid-oriented animation progressive transmission.

The technical solution adopted for the present invention to solve the technical problems is as follows:

Mesh animation sequence A (M is imported at server end ₁, M ₂..., M _f), wherein M _l, l ∈ [1, F] represents the static network lattice model that every frame is corresponding, and its number of vertices is N, and F represents the frame number of this mesh animation.

First, each frame model is mated.Utilize ICP algorithm to find transformation relation between frame grid model coordinates matrix, obtain rotation matrix R and the translation matrix T of frame and interframe.

Secondly, utilize class kmeans clustering algorithm that all frame mesh coordinate data are carried out cluster, its specific algorithm step is as follows:

1) random selecting k frame is as the initial representative frame of k cluster, is kept in array Sframes by its index.

2) select i-th (i=1,2 ...., N) frame data as present frame, calculate present frame to each representative frame j (j=Sframes (1) ..., Sframes (k)) frame pitch d _ij.More all frame pitches, are classified as the i-th frame and representative frame same class when obtaining minimum frame distance.

3) step 2 is repeated) F time.Now all frames are classified all, and this time cluster terminates.

4) for each cluster, the representative frame that such is new is found out.

5) step 2 is repeated) ~ 4), until meet convergence end condition.

6) final cluster result and all final representative frame is preserved.

Finally, obtain all kinds of interior frame index continuous print cluster result, start carry out data processing by class and complete coding transmission respectively, individual transmission is carried out to all kinds of representative frame simultaneously.

Beneficial effect of the present invention: the present invention is a kind of mesh animation frame clustering method for progressive transmission.Substitute traditional Euclidean distance with grid residual error during cluster, and then add key frame time difference, make cluster result not only accurate, can also ensure that all kinds of comprised frame index is continuous.Also separately carry out coding transmission finally by by class process, the progressive transmission of mesh animation can be realized.Individual transmission is carried out to all kinds of representative frame simultaneously, the animation effect that of this mesh animation is basic can be obtained.

Embodiment

Mesh animation sequence A (M is imported at server end ₁, M ₂..., M _f), wherein M _l, l ∈ [1, F] represents the static network lattice model that every frame is corresponding, and its number of vertices is N, and F represents the frame number of this mesh animation.

First, each frame model is mated.Utilize ICP algorithm to find transformation relation between frame grid model coordinates matrix, obtain rotation matrix R and the translation matrix T of frame and interframe.Its specific algorithm is as follows:

If V _i, V _jrepresent the coordinates matrix of the i-th frame and jth frame static network lattice model in mesh animation sequence respectively, R _ij, t _ijrepresent respectively from V _itransform to V _jtime corresponding rotation matrix and translation vector, i.e. R _ijand t _ijmeet the error function value E (R in following formula _ij, t _ij) minimum.

$E(R_{\mathrm{ij}},t_{\mathrm{ij}})=\underset{R,t}{\arg \min }\left|\right|V_j-(V_i\·R_{\mathrm{ij}}+t_{\mathrm{ij}})\left|\right|$

Secondly, utilize class kmeans clustering algorithm that all frame mesh coordinate data are carried out cluster, its specific algorithm step is as follows:

1) random selecting K frame is as the initial representative frame of K cluster, is kept in array Sframes by its index.

2) select i-th (i=1,2 ...., N) frame data as present frame, calculate present frame to each representative frame j (j=Sframes (1) ..., Sframes (K)) frame pitch d _ij, the i-th frame is classified as the C when obtaining minimum frame distance _jin.

Wherein when definition frame distance, classic method only using the geometric distance between two frame coordinate matrixes as module, i.e. d _ij=|| V _i-V _j||.And we consider that the compressed object of this compression algorithm is key-frame animation, also there is the problem at a different time interval in frame and interframe, therefore we are when defining frame pitch, not only replace traditional Euclidean distance with grid residual values, also will consider key frame time difference.With the frame pitch d of the i-th frame to jth frame _ijfor example, its final expression formula is:

d _ij＝||V _j-(V _i·R _ij+T _ij)||+λ|t _i-t _j|

Wherein R _ijwith T _ijmatrix is described above, represents that the jth frame transform that utilizes ICP algorithm to calculate is to the rotation matrix of the i-th frame and translation matrix respectively, and t _iwith t _irepresent the i-th frame and the time coordinate of jth frame in unit grids animation time (1s) respectively.

When λ=0, show that frame pitch does not consider key frame time difference.Now, due to the ambiguity of mesh animation change, therefore after execution clustering algorithm, all kinds of comprised frame sequence not necessarily keeps continuous, and when λ value increases gradually, two interframe the time interval, proportion increased to a certain value time, all kinds of comprised frame frame index value starts continuously, namely to carry out contiguous segmentation to original mesh animation sequence.But when λ value is excessive, frame cluster will be determined primarily of time difference.In order to ensure that cluster is accurate, we get and make frame index start continuous print critical value as final λ value.

3) step 2 is repeated) F time.Now all frames are classified all, and this time cluster terminates.

4) for each cluster, the representative frame that such is new is found out.

Wherein representative frame is defined as: set and perform cluster result that frame clustering algorithm obtains as A (C ₁, C ₂..., C _k), wherein C _k=(V _k1, V _k2..., V _km), k ∈ [1, K] is wherein any class frame sequence, the number of the frame that m comprises for such, then representative frame V _r(V _r∈ (V _k1, V _k2..., V _km)) meet as the wherein frame in frame sequence in such

$d_{\mathrm{sum}}=\min \left(\underset{t=k1}{\overset{\mathrm{km}}{\mathrm{\Σ}}}\left(\right||V_r-(R_{\mathrm{rt}}*V_t+T_{\mathrm{rt}})||+\mathrm{\λ}|t_r-t_t\left|\right)\right)$

5) step 2 is repeated) ~ 4), until meet convergence end condition.

Wherein restraining end condition is: reach the largest frames of representative frame and other frames in maximum iteration time times or section apart from being less than threshold value threshold.

6) final cluster result and representative frame V thereof is preserved _r.

Finally, obtain all kinds of interior frame index continuous print cluster result, start carry out data processing by class and complete coding transmission respectively, individual transmission is carried out to all kinds of representative frame simultaneously.

Claims (2)

1. the frame clustering method of Grid-oriented animation progressive transmission, is characterized in that the method comprises the steps:

Mesh animation sequence A (M is imported at server end ₁, M ₂..., M _f), wherein M _l, l ∈ [1, F] represents the static network lattice model that every frame is corresponding, and its number of vertices is N, and F represents the frame number of this mesh animation;

Step 1, mates each frame model; Utilize ICP algorithm to find transformation relation between frame grid model coordinates matrix, obtain rotation matrix R and the translation matrix T of frame and interframe; Specific as follows:

If V _i, V _jrepresent the coordinates matrix of the i-th frame and jth frame static network lattice model in mesh animation sequence respectively, R _ij, t _ijrepresent respectively from V _itransform to V _jtime corresponding rotation matrix and translation vector, i.e. R _ijand t _ijmeet the error function value E (R in following formula _ij,t _ij) minimum;

$E(R_{\mathrm{ij}},t_{\mathrm{ij}})=\underset{R,t}{\arg \min }\left|\right|V_j-(V_i\·R_{\mathrm{ij}}+t_{\mathrm{ij}})\left|\right|$

Secondly, utilize class kmeans clustering algorithm that all frame mesh coordinate data are carried out cluster, it is specific as follows:

1) random selecting K frame is as the initial representative frame of K cluster, is kept in array Sframes by its index;

2) select the i-th frame data as present frame, calculate the frame pitch d of present frame to each representative frame j _ij, the i-th frame is classified as the C when obtaining minimum frame distance _jin; Wherein i=1,2 ...., N, j=Sframes (1) ..., Sframes (K);

3) step 2 is repeated) F time; Now all frames are classified all, and this time cluster terminates;

4) for each cluster, the representative frame that such is new is found out;

Wherein representative frame is defined as: set and perform cluster result that frame clustering algorithm obtains as A (C ₁, C ₂..., C _k), wherein C _k=(V _k1, V _k2..., V _km), k ∈ [1, K] is wherein any class frame sequence, the number of the frame that m comprises for such, then representative frame V _rmeet as the wherein frame in frame sequence in such

$d_{\mathrm{sum}}=\min \left(\underset{t=k1}{\overset{\mathrm{km}}{\mathrm{\Σ}}}\left(\right||V_r-(R_{\mathrm{rt}}*V_t+T_{\mathrm{rt}})||+\mathrm{\λ}|t_r-t_t\left|\right)\right)$

5) step 2 is repeated) ~ 4), until meet convergence end condition;

Wherein restraining end condition is: reach the largest frames of representative frame and other frames in maximum iteration time times or section apart from being less than threshold value threshold;

6) final cluster result and representative frame V thereof is preserved _r;

Finally, obtain all kinds of interior frame index continuous print cluster result, start carry out data processing by class and complete coding transmission respectively, individual transmission is carried out to all kinds of representative frame simultaneously.

2. the frame clustering method of Grid-oriented animation progressive transmission according to claim 1, is characterized in that: when definition frame distance, not only replaces Euclidean distance with grid residual values, also will consider key frame time difference; With the frame pitch d of the i-th frame to jth frame _ijfor example, its final expression formula is:

d _ij＝||V _j-(V _i·R _ij+T _ij)||+λ|t _i-t _j|

Wherein R _ijwith T _ijmatrix represents that the jth frame transform that utilizes ICP algorithm to calculate is to the rotation matrix of the i-th frame and translation matrix respectively, and t _iwith t _irepresent the i-th frame and the time coordinate of jth frame in unit grids animation time respectively;

When λ=0, show that frame pitch does not consider key frame time difference; Now, due to the ambiguity of mesh animation change, therefore after execution clustering algorithm, all kinds of comprised frame sequence not necessarily keeps continuous, and when λ value increases gradually, two interframe the time interval, proportion increased to a certain value time, all kinds of comprised frame frame index value starts continuously, namely to carry out contiguous segmentation to original mesh animation sequence; But when λ value is excessive, frame cluster will be determined primarily of time difference; In order to ensure that cluster is accurate, getting and making frame index start continuous print critical value as final λ value.