CN106127803A - Human body motion capture data behavior dividing method and system - Google Patents
Human body motion capture data behavior dividing method and system Download PDFInfo
- Publication number
- CN106127803A CN106127803A CN201610436925.XA CN201610436925A CN106127803A CN 106127803 A CN106127803 A CN 106127803A CN 201610436925 A CN201610436925 A CN 201610436925A CN 106127803 A CN106127803 A CN 106127803A
- Authority
- CN
- China
- Prior art keywords
- human body
- motion capture
- body motion
- data
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Abstract
The open a kind of human body motion capture data behavior dividing method of the present invention and system, the method includes: S1, using each frame data in human body motion capture data as each high dimensional data point, calculate the distance between each high dimensional data point and similarity, obtain distance matrix and similarity matrix;Human body motion capture data are processed by S2, local density and distance matrix according to each high dimensional data point, extract the behavior number in human body motion capture data and cluster centre;S3, according to similarity matrix, build undirected weighted graph;S4, human body motion capture data will be carried out behavior segmentation and be changed into undirected weighted graph is split, then be changed into spectral clustering, and use Spectral Clustering that human body motion capture data are clustered, obtain the behavior segmentation of human body motion capture data;S5, behavior segmentation is optimized, obtains the behavior cut-point of human body motion capture data.The present invention has preferable robustness, the suitability, effectiveness and accuracy.
Description
Technical field
The present invention relates to the process for human body motion capture data in computer animation.Have more particularly, to one
Imitate unsupervised human body motion capture data behavior dividing method and system.
Background technology
With computer hardware technique and the development of computer graphics techniques, computer animation becomes a kind of important
Digital Media form, the product that simultaneous computer animation combines as computer graphics and art, nowadays by extensively
General it is applied to numerous fields, such as film and television special effect making, three-dimensional large-sized game, virtual reality etc..
In recent years, the human body motion capture technology sustainable development of optically-based equipment, sense of reality human body movement capturing data
It is applied in computer animation.As it is shown in figure 1, human body motion capture data can preferably keep the verity of human motion
With naturalness, but the cost that exercise data catches but is abnormal high, and capture-process is extremely complex, therefore transports for human body
Dynamic reusing of data of seizure becomes necessary.Generally, human body motion capture Data Data amount is huge many, and behavior is enriched.Use people
Work method, it is desirable to extract the single behavior campaign number required for people from the human body motion capture data base that data volume is huge
According to, it is abnormal difficult and time-consuming.So, for human body motion capture data behavior split just become one indispensable
Processing procedure.Meanwhile, human body motion capture data behavior segmentation has the biggest researching value and meaning.
Lot of domestic and international scholar studies for human body motion capture data behavior dividing method.Muller et al.
(Muller et al.Efficient and robust annotation of motion capture data.In
The proceedings of the 2009 ACM SIGGRAPH. page number: 17-26) use the pass extracting human body motion capture data
Being feature, the method building relationship characteristic masterplate realizes the behavior segmentation of human body motion capture data.Lv et al. (Lv et
al.Recognition and segmentation of 3-d human action using hmm and multi-class
Adaboost.In Computer Vision-ECCV 2006. page number: 359-372) construct hidden Ma Erke by tranining database
Husband's model, uses many sorting techniques to realize the behavior segmentation of human body motion capture data.But, this people based on supervised learning
Body motor behavior dividing method, greatly can be limited by training dataset, can realize for the behavior comprised in data set
Preferably behavior segmentation, and training data concentrates the behavior lacked to arise that segmentation errors.Barbic et al. (Barbic et
al.Segmentation Motion Capture Data into Distinct Behaviors.Graphics
Interface, 2004. page numbers: 185-194) think that different behaviors can use different intrinsic dimensionalities to represent, use main
The method of component analysis carries out dimensionality reduction to human body motion capture data, then realizes human body by the analysis projection error on subspace
The behavior segmentation of movement capturing data.But this method, has discarded the information beyond main constituent during dimensionality reduction, point
Cut result and there is bigger error.On the basis of principal component analysis, Barbic et al. proposes row based on Probabilistic Principal Component Analysis
For dividing method, this method uses makes noise the information beyond main constituent, builds probabilistic model, by calculate geneva away from
From mode realize behavior segmentation.While the method realizes Data Dimensionality Reduction, process the information outside main constituent the most well, can
To obtain preferable behavior dividing method, but the deficiency that the method exists is to cannot be distinguished by human body motion capture data sequence
In, the identical behavior existed.
Accordingly, it is desirable to provide a kind of effective unsupervised human body motion capture data behavior dividing method and system.
Summary of the invention
It is an object of the invention to provide a kind of effective unsupervised human body motion capture data behavior dividing method and
System.
For reaching above-mentioned purpose, the present invention uses following technical proposals:
A kind of human body motion capture data behavior dividing method, comprises the steps:
S1, using each frame data in human body motion capture data as each high dimensional data point, calculate each high dimensional data point it
Between distance and similarity, obtain distance matrix and similarity matrix;
S2, according to the local density of each high dimensional data point and distance matrix, human body motion capture data are processed, extract people
The behavior number comprised in body movement capturing data and cluster centre;
S3, according to similarity matrix, build undirected weighted graph;
S4, human body motion capture data will be carried out behavior segmentation and be changed into undirected weighted graph is split, then by right
Undirected weighted graph carries out segmentation and is changed into spectral clustering, uses Spectral Clustering to cluster human body motion capture data, obtains
The behavior segmentation of human body motion capture data;
S5, behavior segmentation to human body motion capture data are optimized, and obtain the row of human body motion capture data
For cut-point.
Preferably, step S1 calculates the distance between each high dimensional data point and similarity, obtain distance matrix with similar
The detailed process of property matrix is:
Calculate the distance between each high dimensional data point, the distance between each high dimensional data point constitute distance matrix, each height
The computing formula of the distance between dimension strong point is as follows:
di,j=α d (pi,pj)+βd(vi,vj) i=1,2 ..., N;J=1,2 ..., N
Wherein, α and β is respectively weight, di,jFor representing high dimensional data point and the representative of the i-th frame human body motion capture data
Distance between the high dimensional data point of jth frame human body motion capture data, N is the frame number of human body movement capturing data;d(vi,vj)
wkIt is the speed difference between the i-th frame human body motion capture data and jth frame human body motion capture data, d (vi,vj)=| vi-vj|,
viIt it is the joint velocity in the i-th frame human body motion capture data;d(pi,pj)wkIt is the i-th frame human body motion capture data and jth frame
Posture between human body motion capture data is poor, following formula be calculated:
Wherein, | | pi,0-pj,0||2wkFor global translation distance;M is crucial joint number;wkFor kth key joint institute
The weight occupied;qi,kIt it is the quaternary number representation in k crucial joint of the i-th frame human body motion capture data;
Calculate the similarity between each high dimensional data point, the similarity between each high dimensional data point constitute similarity square
Battle array, the computing formula of the similarity between each high dimensional data point is as follows:
Wherein, wi,jFor representing the high dimensional data point of the i-th frame human body motion capture data and representing jth frame human motion and catch
Catch the similarity between the high dimensional data point of data;σiFor representing the phase of the high dimensional data point of the i-th frame human body motion capture data
Like parameter, following formula it is calculated:
Wherein, integer k ' the frame number that span is human body movement capturing data 1% to 10% between.
Preferably, the detailed process of step S2 is:
Calculate between local density and each high dimensional data point and the more highdensity high dimensional data point of each high dimensional data point
Relative distance, computing formula is as follows:
Wherein, ρiFor representing the local density of the high dimensional data point of the i-th frame human body motion capture data;IsFor each high dimension
The high dimensional data point set of strong point composition;dcFor clustering parameter, span is between 0.01 to 0.1;δiFor representing the i-th frame human body
The high dimensional data point of movement capturing data and the relative distance of more high density data point;
To between local density and each high dimensional data point and the more highdensity high dimensional data point of each high dimensional data point
Relative distance is normalized, and defines αi'=ρ 'i*δ'i;
The discrete derivative of the high dimensional data point that definition represents the i-th frame human body motion capture data is Δi=| α 'i-α'i-θ|, θ
For noise reduction parameters;
CalculateDiscrete derivative meansigma methodsAnd standard deviation εi, when meeting Δi> 3 εiTime think this α 'iValue meets
Condition, the high dimensional data o'clock of the i-th corresponding frame human body motion capture data is as a cluster centre, thus obtains human motion
Catch the cluster centre set { c included in data1,c2,c3,...,cK, number K of cluster centre is human motion and catches
Catch the behavior number comprised in data.
Preferably, the detailed process of step S3 is:
Using each frame in N frame human body motion capture data as a node of undirected weighted graph, obtain the node of figure
Set V={f1,f2,f3,...,fN, in figure, the limit weight between each node uses each internodal similarity to represent;
Calculating the degree of each node, by degree degree of the composition matrix of each node, the computing formula of the degree of each node is as follows:
Wherein, diFor undirected weighted graph represents the degree of the node of the i-th frame human body motion capture data.
Preferably, step S4 uses Spectral Clustering human body motion capture data are clustered, obtain human motion
The detailed process of the behavior segmentation catching data is as follows:
Use t arest neighbors, similarity matrix is changed into sparse matrix S;
Calculate Laplacian MatrixWherein, E is unit matrix, and D is degree matrix;
Calculating K the minimal characteristic vector of Laplacian Matrix L, building minimal characteristic vector matrix is U ∈ RN×K;
UseMode minimal characteristic vector matrix U is carried out
Normalized, obtains matrix
Use k-means algorithm cluster matrixN row become K class, the human body fortune by timing recovery, after being clustered
The dynamic behavior segmentation catching data.
Preferably, step S4 uses Spectral Clustering human body motion capture data are clustered, obtain human motion
The detailed process of the behavior segmentation catching data is as follows:
Randomly select l data point, build matrix A ∈ R respectivelyl×lWith B ∈ Rl×(N-l), A is for randomly selecting l high dimension
Similarity matrix between strong point, B is between l the high dimensional data point randomly selected and residue (N-l) individual high dimensional data point
Similarity matrix,;
Calculate matrix
Calculate matrix
Build matrixAnd it is carried out feature decomposition
CalculateK minimum characteristic vector for Laplacian Matrix L;
UseMode matrix U is normalized,
To matrix
Use k-means algorithm cluster matrixN row become K class, the human body fortune by timing recovery, after being clustered
The dynamic behavior segmentation catching data.
Preferably, the detailed process of step S5 is:
Energy function is used to be optimized the behavior segmentation of the human body motion capture data after cluster, described energy
Flow function is
Wherein, si'Represent be i-th in human body motion capture data ' individual behavior sequence, si'∈ S', S'={s1,s2,
s3,...,sK'It it is the behavior segmentation set of human body motion capture number;Oriental matrixWhen
Human body motion capture data f of the i-th frameiWhen belonging to class cIt is otherwiseBe i-th ' individual behavior sequence
Row and jth ' distance between individual behavior sequence,
Realize the Accurate Segmentation to human body motion capture data by the function defined based on dynamic programming principle, obtain people
The behavior cut-point of body movement capturing data, described function based on dynamic programming principle definition is:
J (S ")=min{J (S "-1)+min{F (S "-1, G) }
Wherein, S " by the data sequence fragment being made up of human body motion capture data, J (S') is data sequence fragment S "
Energy value.
A kind of human body motion capture data behavior segmenting system implementing said method, this system includes:
Distance matrix and similarity matrix build module, using each frame data in human body motion capture data as each higher-dimension
Data point, calculates the distance between each high dimensional data point and similarity, obtains distance matrix and similarity matrix;
Human body motion capture data behavior number extraction module, according to local density and the distance matrix pair of each high dimensional data point
Human body motion capture data process, and extract the behavior number and cluster centre comprised in human body motion capture data;
Undirected weighted graph builds module, according to similarity matrix, builds undirected weighted graph;
Human body motion capture data will be carried out behavior segmentation and be changed into undirected by human body motion capture data segmentation module
Weighted graph is split, more undirected weighted graph will carry out segmentation is changed into spectral clustering, uses Spectral Clustering to human motion
Catch data to cluster, obtain the behavior segmentation of human body motion capture data;
Optimize module, the behavior segmentation of human body motion capture data is optimized, obtains human body motion capture number
According to behavior cut-point.
Beneficial effects of the present invention is as follows:
Technical scheme of the present invention is possible not only to effectively realize human body motion capture data behavior segmentation, and permissible
The motion sequence fragment of the identical behavior in the human body motion capture data that distinguishing identifier is to be split.Technical scheme of the present invention
There is preferable robustness, for human body motion capture data sequence behavior segmentation on the suitability, effectiveness and accuracy all
There is certain superiority, it is possible to well meet actual demand.
Accompanying drawing explanation
Below in conjunction with the accompanying drawings the detailed description of the invention of the present invention is described in further detail.
Fig. 1 illustrates the schematic diagram that human body motion capture data behavior is split.
Fig. 2 illustrates the flow chart of the human body motion capture data behavior dividing method that embodiment 1 provides.
Fig. 3 illustrate two sections of human body motion capture data original in embodiment 1 and DTW distance solve after two sections of human bodies fortune
Dynamic seizure Data Comparison figure, wherein, Fig. 3-a illustrates that original two section human body motion capture datagram, Fig. 3-b illustrate DTW distance
Two sections of human body motion capture datagrams after solving.
Fig. 4 illustrate human body motion capture data behavior dividing method that embodiment 1 provides and existing methodical accuracy rate with
Recall rate comparison diagram.
Fig. 5 illustrates the schematic diagram of the human body motion capture data behavior segmenting system that embodiment 2 provides.
Detailed description of the invention
In order to be illustrated more clearly that the present invention, below in conjunction with preferred embodiments and drawings, the present invention is done further
Bright.Parts similar in accompanying drawing are indicated with identical reference.It will be appreciated by those skilled in the art that institute is concrete below
The content described is illustrative and be not restrictive, and should not limit the scope of the invention with this.
As in figure 2 it is shown, the human body motion capture data behavior dividing method that the present embodiment provides comprises the steps:
S1, using each frame data in human body motion capture data as each high dimensional data point, calculate each high dimensional data point it
Between distance and similarity, obtain distance matrix and similarity matrix;
Human body motion capture data are processed by S2, local density and distance matrix according to each high dimensional data point, carry
Take the behavior number and cluster centre comprised in human body motion capture data, in this, as the priori of Spectral Clustering;
S3, according to similarity matrix, build undirected weighted graph;
S4, human body motion capture data will be carried out behavior segmentation and be changed into and undirected weighted graph is split, i.e. people
Body movement capturing data behavior segmentation problem is changed into figure and cuts problem, more undirected weighted graph will carry out segmentation, and to be changed into spectrum poly-
Class, i.e. cuts the relation between problem and spectral clustering by analysis chart, and NP (non-deterministic polynomial) is difficult
The figure problem of cutting be changed into spectral clustering problem and solve, use Spectral Clustering that human body motion capture data are clustered,
Obtain the behavior segmentation of human body motion capture data;
S5, behavior segmentation to human body motion capture data are optimized, and obtain the row of human body motion capture data
For cut-point.
Wherein,
Step S1 is " using each frame in human body motion capture data as each high dimensional data point, each high dimensional data point composition height
Dimension data point set Is, calculate the distance between each high dimensional data point and similarity, obtain distance matrix and similarity matrix " tool
Body process is as follows:
The human body motion capture data that the present embodiment uses are CMU (Carnegie Mellon University Ka Neiji
Mei Long university) ASF/AMC file in data base.The human skeleton model of this document definition is 31 joints, has 62 certainly
By spending, having 1 in 31 joints is root joint, and root joint has 6 degree of freedom.Human body motion capture data N frame, the i-th frame altogether
Posture p of human body motion capture dataiFor removing other the articulate anglecs of rotation composition beyond root node, pi={ ai,1,
ai,2,ai,3…ai,56, i=1,2 ..., N, wherein, ai,j'Be the jth in the i-th frame human body motion capture data ' individual degree of freedom,
Represent an Euler angle, j'=1,2 ..., 56.Joint velocity v in i-th frame human body motion capture dataiBy adjacent pass
Euclidean distance between joint posture is tried to achieve, such as formula (1):
Represent the high dimensional data point of the i-th frame human body motion capture data and represent the height of jth frame human body motion capture data
Distance d between dimension strong pointi,jFormula (2) such as is used to calculate:
di,j=α d (pi,pj)+βd(vi,vj) i=1,2 ..., N;J=1,2 ..., N (2)
α and β is respectively weight, is disposed as 0.5 in the present embodiment, d (vi,vj) it is the i-th frame human body motion capture data
And the speed difference between jth frame human body motion capture data, d (vi,vj)=| vi-vj|, d (pi,pj) represent the i-th frame human motion
Catch the posture between data and jth frame human body motion capture data poor, posture difference d (pi,pj) it is calculated by formula (3):
Wherein, | | pi,0-pj,0||2Represent is global translation distance;For posture gap, it is chosen at Behavioral change
In play a decisive role m crucial joint (m=8 in the present embodiment);wkRepresent is the power occupied by kth key joint
Weight, in the present embodiment, the value of weight occupied by each joint is as shown in table 1, and the occupied weight in remaining joint is set to 0,
qi,kIt it is the quaternary number representation in k crucial joint of the i-th frame human body motion capture data.For N frame human body motion capture number
According to, after being calculated the distance between each high dimensional data point, the distance square of N*N can be made up of the distance between each high dimensional data point
Battle array Dist, for the element d in Distance matrix D isti,j, di,j=dj,i(i ≠ j) and di,j=0 (i=j).
Table 1: joint weighted value
Joint title | Weight wk |
Left and right hip joint | 1.0000 |
Left and right knee joint | 0.0901 |
Left and right shoulder joint | 0.7884 |
Left and right elbow joint | 0.0247 |
Represent the high dimensional data point of the i-th frame human body motion capture data and represent the height of jth frame human body motion capture data
Similarity w between dimension strong pointi,jFormula (4) such as is used to calculate:
In formula (4), represent the high dimensional data point of the i-th frame human body motion capture data and represent jth frame human body motion capture
Distance d between the high dimensional data point of datai,jIt is calculated by formula (2), represents the height of the i-th frame human body motion capture data
Similar parameter σ at dimension strong pointiNeighbouring adaptive scale mode is used to calculate, computational methods such as formula (5):
In formula (5), k' is the ε % of frame number contained by human body movement capturing data, i.e. k'=N ε % generally can choose taking of ε
Value scope is between 1 to 10, i.e. the value of k' is N% to 10N%, ε=1 in this enforcement, i.e. k'=N%.
After being calculated the similarity between each high dimensional data point, the similarity between each high dimensional data point constitute N*N
Similarity matrix W.
Step S2 " according to the local density of each high dimensional data point and distance matrix to human body motion capture data at
Reason, extracts the behavior number and cluster centre comprised in human body motion capture data, and the priori in this, as Spectral Clustering is known
Know " detailed process as follows:
The present embodiment uses a kind of unsupervised clustering method, and it can automatically extract clusters number, by experiment
Understanding for human body motion capture data, the clustering method that the present embodiment uses is extracted cluster number and is this human body fortune
The dynamic behavior number caught included in data.
Calculate between local density and each high dimensional data point and the more highdensity high dimensional data point of each high dimensional data point
Relative distance, method is as follows:
Wherein, ρiFor representing the local density of the high dimensional data point of the i-th frame human body motion capture data;IsFor each high dimension
The high dimensional data point set of strong point composition;dcFor clustering parameter, usual span is between 0.01 to 0.1, takes in the present embodiment
Value is 0.05;δiThe relative distance of high dimensional data point Yu more high density data point for representing the i-th frame human body motion capture data;
di,jFor representing the high dimensional data point of the i-th frame human body motion capture data and representing the high dimension of jth frame human body motion capture data
Distance between strong point.
The condition meeting the high dimensional data point becoming cluster centre is: needs have bigger local density ρ simultaneouslyiAnd
Relative distance δ between bigger and more highdensity high dimensional data pointi.Owing to density p and distance δ are skimble-scamble, first
It is normalized, defines αi'=ρ 'i*δ'i.Hopping behavior according to data seeks the α meeting conditioni' number, from
And try to achieve cluster numbers, i.e. behavior number, method is as follows:
The discrete derivative of the high dimensional data point that definition represents the i-th frame human body motion capture data is Δi=| α 'i-α'i-θ|,
Noise reduction parameters θ be disposed to reduce error noise, calculateDiscrete derivative meansigma methodsAnd standard deviation εi, when full
Foot Δi> 3 εiTime think this α 'iValue meets condition, and the high dimensional data o'clock of the i-th corresponding frame human body motion capture data is as one
Individual cluster centre, thus obtain behavior number K and cluster centre set { c included in human body motion capture data1,c2,
c3,...,cK}。
Step S3 " according to the similarity between movement capturing data frame, build undirected weighted graph and calculate the degree of each node "
Detailed process as follows:
Using each frame in N frame human body motion capture data as a node of undirected weighted graph, obtain the node of figure
Set V={f1,f2,f3,...,fN, in figure, the limit weight between each node uses each internodal similarity to represent, similarity
Gaussian kernel form is used to try to achieve as shown in formula (4).Calculate the degree of each node, computational methods such as formula (8):
In formula (8), diFor undirected weighted graph represents the degree of the node of the i-th frame human body motion capture data.
After being calculated the degree of each node, by degree degree of the composition matrix D of each node.
Step S4 " will carry out behavior segmentation to be changed into and split undirected weighted graph, i.e. to human body motion capture data
Human body motion capture data behavior segmentation problem is changed into figure and cuts problem, more undirected weighted graph will be carried out segmentation and be changed into spectrum
Cluster, i.e. cuts the relation between problem and spectral clustering by analysis chart, NP (non-deterministic polynomial)
The difficult figure problem of cutting is changed into spectral clustering problem and solves, and uses Spectral Clustering to gather human body motion capture data
Class, obtains the behavior segmentation of human body motion capture data " in " human body motion capture data will be carried out behavior segmentation turn
Become undirected weighted graph is split, i.e. human body motion capture data behavior segmentation problem be changed into figure and cut problem " tool
Body process is as follows:
For figure cuts problem, the similarity within subgraph is big, and the similarity between different subgraph is less, this analogy
In the behavior segmentation problem of human body motion capture data, for identical behavior, similarity is relatively big, and between different rows is
Similarity is less.Therefore human body motion capture data behavior segmentation problem can be changed into figure and cut problem.
The sequence of given human body motion capture data, the Frame that each behavior is comprised is constituted in undirected weighted graph
Subgraph F1,F2,...,FK, as a example by there are two subgraphs, following relation exists for node between two subgraphs:
Similarity between two subgraphs is:
For this human body motion capture data sequence, segmentation to be realized, it is only necessary to make by human body motion capture number
The figure segmentation Least-cost built according to sequence frame:
Thus can obtain, human body motion capture data behavior segmentation problem is converted into figure and cuts problem.
Step S4 " will carry out behavior segmentation to be changed into and split undirected weighted graph, i.e. to human body motion capture data
Human body motion capture data behavior segmentation problem is changed into figure and cuts problem, more undirected weighted graph will be carried out segmentation and be changed into spectrum
Cluster, i.e. cuts the relation between problem and spectral clustering by analysis chart, NP (non-deterministic polynomial)
The difficult figure problem of cutting is changed into spectral clustering problem and solves, and uses Spectral Clustering to gather human body motion capture data
Class, obtains the behavior segmentation of human body motion capture data " in " undirected weighted graph will carry out segmentation again, and to be changed into spectrum poly-
Class, i.e. cuts the relation between problem and spectral clustering by analysis chart, and NP (non-deterministic polynomial) is difficult
The figure problem of cutting be changed into spectral clustering problem and solve " detailed process as follows:
For figure cuts problem, two kinds of method for solving of Ratiocut and Ncut can be used, solve target and be respectively as follows:
As a example by the sequence of aforementioned given human body motion capture data, two classification problems to be realized, use as above
Two kinds of figures cut method for solving, solve target as follows:
That represent is subgraph FcThe weights sum on middle limit, fixed for human body motion capture data sequence frame
Justice instruction vector I=(I1,I2,I3,...,IN)'∈RN, for IiFor, definition
For undirected weighted graph, build its Laplacian Matrix L, L=D-W (D is degree matrix, and W is similarity matrix).
L=D-W=D-1/2(D-W)D-1/2=E-D-1/2WD-1/2
What E represented is unit matrix, by formula (15) can for the Laplacian Matrix of undirected weighted graph,
Wushu (14) substitutes into above formula and can obtain:
For the N frame human body motion capture data that the present embodiment is used, vol (V) is constant value, then target letter
Number is just equivalent to:
In sum, it is desirable to solution figure cuts the least cost function of problem and is equivalent to solve following problem:
Assuming LI=λ I, wherein λ and I is respectively eigenvalue and the characteristic vector of L, takes advantage of for LI=λ I the right and left simultaneously
With I', I'LI=λ I'I (I'I=vol (V)) can be obtained, then minimize I'LI problem and be the eigenvalue λ problem minimizing it.
So problem is converted to solve 2 eigenvalues of minimum of Laplacian Matrix L and characteristic vector problem.
Therefore, cut the relation between algorithm (Rationcut and Ncut) and spectral clustering by analyzing typical figure, send out
The target that solves the most between the two is of equal value, therefore the figure that NP is difficult can be cut problem and use spectral clustering to solve.
Step S4 " will carry out behavior segmentation to be changed into and split undirected weighted graph, i.e. to human body motion capture data
Human body motion capture data behavior segmentation problem is changed into figure and cuts problem, more undirected weighted graph will be carried out segmentation and be changed into spectrum
Cluster, i.e. cuts the relation between problem and spectral clustering by analysis chart, NP (non-deterministic polynomial)
The difficult figure problem of cutting is changed into spectral clustering problem and solves, and uses Spectral Clustering to gather human body motion capture data
Class, obtains the behavior segmentation of human body motion capture data " in " use Spectral Clustering human body motion capture data are entered
Row cluster, obtain the behavior segmentation of human body motion capture data " detailed process as follows:
The present embodiment is respectively adopted the algorithm of two kinds of spectral clusterings, processes N frame human body motion capture data.Two kinds of spectral clusterings
Algorithm is respectively Spectral Clustering and the Spectral Clustering of application Nystrom algorithm of the sparse matrix of t arest neighbors.
(1) human body motion capture data are clustered by the Spectral Clustering using the sparse matrix of t arest neighbors, obtain people
The detailed process of the behavior segmentation of body movement capturing data is as follows:
Input: N frame human body motion capture data sequence { f1,f2,f3,...,fN, each frame is seen as a high dimensional data
Point, K is the behavior number that step S2 is tried to achieve, { c1,c2,c3,...,cKIt it is the cluster centre set tried to achieve of step S2;
1) similarity matrix W ∈ R is built by formula (4)N×N;
2) using t arest neighbors, similarity matrix W is changed into sparse matrix S, and (generally, t may be defined as human body motion capture
The 2% to 10% of data frame number, in the present embodiment, t is defined as the 2% of human body motion capture data frame number, thinks in the present embodiment
Having limit between data point within the scope of this to be connected, between the data point within the scope of this, limit weight is not 0);
3) Laplacian Matrix L is calculated,Wherein, E is unit matrix, and D is degree matrix, and S is sparse
Matrix;
4) calculating K the minimal characteristic vector of Laplacian Matrix L, building minimal characteristic vector matrix is U ∈ RN×K, its
In to solve the method for its K minimal characteristic vector under conditions of known Laplacian Matrix L be mathematics side of the prior art
Method;
5) useMode minimal characteristic vector matrix U is entered
Row normalized, obtains matrix
6) k-means algorithm cluster matrix is usedN row become K class, by timing recovery, the human body after being clustered
The behavior segmentation of movement capturing data.
(2) human body motion capture data are clustered by the Spectral Clustering using application Nystrom algorithm, obtain human body
The behavior segmentation of movement capturing data " detailed process as follows:
Input: N frame human body motion capture data sequence { f1,f2,f3,...,fN, each frame is seen as a high dimensional data
Point, K is that step S2 tries to achieve behavior number, { c1,c2,c3,...,cKIt is the cluster centre set tried to achieve of step S2, set l as defeated
The 2% to 10% of the human body motion capture data frame number N entered, is specifically set as human body motion capture data frame number in the present embodiment
2%.
1) randomly select l data point, build matrix A ∈ R respectivelyl×lWith B ∈ Rl×(N-l), A is for randomly selecting l higher-dimension
Similarity matrix between data point, B is between l the high dimensional data point randomly selected and residue (N-l) individual high dimensional data point
Similarity matrix;
2) matrix is calculated
3) matrix is calculated
4) matrix is builtAnd it is carried out feature decompositionV is matrix
Characteristic vector, Σ be corresponding to characteristic vector V eigenvalue constitute diagonal matrix;
5) calculateK minimum characteristic vector for Laplacian Matrix L;
6) useMode matrix U is normalized place
Reason, obtain matrix
7) k-means algorithm cluster matrix is usedN row become K class, by timing recovery, after i.e. can being clustered
The behavior segmentation of human body motion capture data.
By obtaining above, by cluster mode, human motion is clustered into for N frame human body motion capture data and has caught
Catch several behavior segmentation.
The behavior segmentation of human body motion capture data " is optimized, obtains human body fortune more accurately by step S5
The dynamic behavior cut-point catching data " detailed process as follows:
By step S4, for the human body motion capture data of input, the behavior segmentation of available human body motion capture number
Set of segments S'={s1,s2,s3,...,sK', si'Represent be i-th in human body motion capture data ' individual behavior sequence.For
Reduction behavior segmentation error, is defined as follows energy function:
For human body motion capture data, oriental matrix G is defined asPeople when the i-th frame
Body movement capturing data fiWhen belonging to class cIt is otherwiseI.e. when human body motion capture data f of the i-th frameiFor cluster
During centerIt is otherwise
dist(si',sj') distance between ' individual behavior sequence and jth ' the individual behavior sequence that is i-th, use DTW mode to enter
Row calculates:
Fig. 3 illustrate original two section human body motion capture data and DTW distance solve after two sections of human body motion capture numbers
According to comparison diagram, Fig. 3-a, Fig. 3-b show between two short sequences align distance solve example.
For the cluster result after timing recovery, i.e. the behavior segmentation of the human body motion capture data after cluster,
The energy function be given using formula (21) is optimized process, and short sequence is grouped into the behavior sequence closest with it
In, the function based on dynamic programming principle definition be given eventually through formula (23) realizes the essence to human body motion capture data
Really segmentation:
J (S ")=min{J (S "-1)+min{F (S "-1, G) } (23)
In formula, " being defined as the data sequence fragment being made up of human body motion capture data, J (S') is defined as data to S
Sequence fragment S " energy value, for each frame human body motion capture data, will they be divided into it represented by
Among that behavior segment that behavior is consistent, therefore energy value needs to meet minimum.
The human body motion capture data behavior dividing method provided for the present embodiment, by great many of experiments, uses standard
Error rate, recall rate framework show the effectiveness of our method.Result is as shown in Figure 4.
Embodiment 2
As it is shown in figure 5, the human body motion capture data behavior segmenting system that the present embodiment provides, including:
Distance matrix and similarity matrix build module, using each frame data in human body motion capture data as each higher-dimension
Data point, calculates the distance between each high dimensional data point and similarity, obtains distance matrix and similarity matrix;
Human body motion capture data behavior number extraction module, according to local density and the distance matrix pair of each high dimensional data point
Human body motion capture data process, and extract the behavior number and cluster centre comprised in human body motion capture data, with this
Priori as Spectral Clustering;
Undirected weighted graph builds module, according to similarity matrix, builds undirected weighted graph;
Human body motion capture data will be carried out behavior segmentation and be changed into undirected by human body motion capture data segmentation module
Weighted graph is split, and i.e. human body motion capture data behavior segmentation problem is changed into figure and cuts problem, then will be to undirected weighting
Figure carries out segmentation and is changed into spectral clustering, i.e. cuts the relation between problem and spectral clustering by analysis chart, NP (non-
Deterministic polynomial) the difficult figure problem of cutting is changed into spectral clustering problem and solves, and uses Spectral Clustering
Human body motion capture data are clustered, obtains the behavior segmentation of human body motion capture data;
Optimize module, the behavior segmentation of human body motion capture data is optimized, obtains human body more accurately
The behavior cut-point of movement capturing data.
Obviously, the above embodiment of the present invention is only for clearly demonstrating example of the present invention, and is not right
The restriction of embodiments of the present invention, for those of ordinary skill in the field, the most also may be used
To make other changes in different forms, cannot all of embodiment be given exhaustive here, every belong to this
What bright technical scheme was extended out obviously changes or changes the row still in protection scope of the present invention.
Claims (8)
1. a human body motion capture data behavior dividing method, it is characterised in that the method comprises the steps:
S1, using each frame data in human body motion capture data as each high dimensional data point, calculate between each high dimensional data point
Distance and similarity, obtain distance matrix and similarity matrix;
Human body motion capture data are processed by S2, local density and distance matrix according to each high dimensional data point, extract people
The behavior number comprised in body movement capturing data and cluster centre;
S3, according to similarity matrix, build undirected weighted graph;
S4, human body motion capture data will be carried out behavior segmentation and be changed into undirected weighted graph is split, then will be to undirected
Weighted graph carries out segmentation and is changed into spectral clustering, uses Spectral Clustering to cluster human body motion capture data, obtains human body
The behavior segmentation of movement capturing data;
S5, behavior segmentation to human body motion capture data are optimized, and the behavior obtaining human body motion capture data divides
Cutpoint.
Method the most according to claim 1, it is characterised in that step S1 calculates the distance between each high dimensional data point and
Similarity, the detailed process obtaining distance matrix and similarity matrix is:
Calculate the distance between each high dimensional data point, the distance between each high dimensional data point constitute distance matrix, each high dimension
The computing formula of the distance between strong point is as follows:
di,j=α d (pi,pj)+βd(vi,vj) i=1,2 ..., N;J=1,2 ..., N
Wherein, α and β is respectively weight, di,jFor representing the high dimensional data point of the i-th frame human body motion capture data and representing jth frame
Distance between the high dimensional data point of human body motion capture data, N is the frame number of human body movement capturing data;d(vi,vj) it is i-th
Speed difference between frame human body motion capture data and jth frame human body motion capture data, d (vi,vj)=| vi-vj|, viIt is i-th
Joint velocity in frame human body motion capture data;d(pi,pj) it is the i-th frame human body motion capture data and jth frame human motion
Catch the posture between data poor, following formula be calculated:
Wherein, | | pi,0-pj,0||2For global translation distance;M is crucial joint number;wkOccupied by kth key joint
Weight;qi,kIt it is the quaternary number representation in k crucial joint of the i-th frame human body motion capture data;
Calculate the similarity between each high dimensional data point, the similarity between each high dimensional data point constitute similarity matrix, respectively
The computing formula of the similarity between high dimensional data point is as follows:
Wherein, wi,jFor representing the high dimensional data point of the i-th frame human body motion capture data and representing jth frame human body motion capture number
According to high dimensional data point between similarity;σiFor representing the similar ginseng of the high dimensional data point of the i-th frame human body motion capture data
Number, is calculated by following formula:
Wherein, integer k ' the frame number that span is human body movement capturing data 1% to 10% between.
Method the most according to claim 2, it is characterised in that the detailed process of step S2 is:
Calculate the phase between local density and each high dimensional data point with the more highdensity high dimensional data point of each high dimensional data point
Adjusting the distance, computing formula is as follows:
Wherein, ρiFor representing the local density of the high dimensional data point of the i-th frame human body motion capture data;IsFor each high dimensional data point
The high dimensional data point set of composition;dcFor clustering parameter, span is between 0.01 to 0.1;δiFor representing the i-th frame human motion
Catch the high dimensional data point of data and the relative distance of more high density data point;
Relative between local density and each high dimensional data point with the more highdensity high dimensional data point of each high dimensional data point
Distance is normalized, and defines αi'=ρi'*δi';
The discrete derivative of the high dimensional data point that definition represents the i-th frame human body motion capture data is Δi=| α 'i-α'i-θ|, θ is fall
Make an uproar parameter;
CalculateDiscrete derivative meansigma methodsAnd standard deviation εi, when meeting Δi> 3 εiTime think this α 'iValue meets bar
Part, the high dimensional data o'clock of the i-th corresponding frame human body motion capture data is as a cluster centre, thus obtains human motion and catch
Catch the cluster centre set { c included in data1,c2,c3,...,cK, number K of cluster centre is human body motion capture
The behavior number comprised in data.
Method the most according to claim 3, it is characterised in that the detailed process of step S3 is:
Using each frame in N frame human body motion capture data as a node of undirected weighted graph, obtain node set V of figure
={ f1,f2,f3,...,fN, in figure, the limit weight between each node uses each internodal similarity to represent;
Calculating the degree of each node, by degree degree of the composition matrix of each node, the computing formula of the degree of each node is as follows:
Wherein, diFor undirected weighted graph represents the degree of the node of the i-th frame human body motion capture data.
Method the most according to claim 4, it is characterised in that use Spectral Clustering to human body motion capture in step S4
Data cluster, and the detailed process of the behavior segmentation obtaining human body motion capture data is as follows:
Use t arest neighbors, similarity matrix is changed into sparse matrix S;
Calculate Laplacian MatrixWherein, E is unit matrix, and D is degree matrix;
Calculating K the minimal characteristic vector of Laplacian Matrix L, building minimal characteristic vector matrix is U ∈ RN×K;
UseMode minimal characteristic vector matrix U is normalized
Process, obtain matrix
Use k-means algorithm cluster matrixN row become K class, by timing recovery, the human motion after being clustered is caught
Catch the behavior segmentation of data.
Method the most according to claim 4, it is characterised in that use Spectral Clustering to human body motion capture in step S4
Data cluster, and the detailed process of the behavior segmentation obtaining human body motion capture data is as follows:
Randomly select l data point, build matrix A ∈ R respectivelyl×lWith B ∈ Rl×(N-l), A is for randomly selecting l high dimensional data point
Between similarity matrix, B is similar between l high dimensional data point randomly selecting and residue (N-l) individual high dimensional data point
Property matrix,;
Calculate matrix
Calculate matrix
Build matrixAnd it is carried out feature decomposition
CalculateK minimum characteristic vector for Laplacian Matrix L;
UseMode matrix U is normalized, obtain square
Battle array
Use k-means algorithm cluster matrixN row become K class, by timing recovery, the human motion after being clustered is caught
Catch the behavior segmentation of data.
7. according to the method described in claim 5 or 6, it is characterised in that the detailed process of step S5 is:
Energy function is used to be optimized the behavior segmentation of the human body motion capture data after cluster, described energy letter
Number is
Wherein, si'Represent be i-th in human body motion capture data ' individual behavior sequence, si'∈ S', S'={s1,s2,
s3,...,sK'It it is the behavior segmentation set of human body motion capture number;Oriental matrixWhen
Human body motion capture data f of the i-th frameiWhen belonging to class cIt is otherwisedist(si',sj') be i-th ' individual behavior sequence
Row and jth ' distance between individual behavior sequence,
Realize the Accurate Segmentation to human body motion capture data by the function defined based on dynamic programming principle, obtain human body fortune
The dynamic behavior cut-point catching data, described function based on dynamic programming principle definition is:
J (S ")=min{J (S "-1)+min{F (S "-1, G) }
Wherein, S " by the data sequence fragment being made up of human body motion capture data, J (S') is data sequence fragment S " energy
Value.
8. the human body motion capture data behavior segmenting system implementing method as claimed in claim 1, it is characterised in that should
System includes:
Distance matrix and similarity matrix build module, using each frame data in human body motion capture data as each high dimensional data
Point, calculates the distance between each high dimensional data point and similarity, obtains distance matrix and similarity matrix;
Human body motion capture data behavior number extraction module, local density and distance matrix according to each high dimensional data point are to human body
Movement capturing data processes, and extracts the behavior number and cluster centre comprised in human body motion capture data;
Undirected weighted graph builds module, according to similarity matrix, builds undirected weighted graph;
Human body motion capture data will be carried out behavior segmentation and be changed into undirected weighting by human body motion capture data segmentation module
Figure is split, more undirected weighted graph will carry out segmentation is changed into spectral clustering, uses Spectral Clustering to human body motion capture
Data cluster, and obtain the behavior segmentation of human body motion capture data;
Optimize module, the behavior segmentation of human body motion capture data is optimized, obtains human body motion capture data
Behavior cut-point.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610436925.XA CN106127803A (en) | 2016-06-17 | 2016-06-17 | Human body motion capture data behavior dividing method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610436925.XA CN106127803A (en) | 2016-06-17 | 2016-06-17 | Human body motion capture data behavior dividing method and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106127803A true CN106127803A (en) | 2016-11-16 |
Family
ID=57470733
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610436925.XA Pending CN106127803A (en) | 2016-06-17 | 2016-06-17 | Human body motion capture data behavior dividing method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106127803A (en) |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107169988A (en) * | 2017-05-12 | 2017-09-15 | 江苏大学 | A kind of extraction method of key frame based on COS distance hierarchical clustering |
CN107832713A (en) * | 2017-11-13 | 2018-03-23 | 南京邮电大学 | A kind of human posture recognition method based on OptiTrack |
CN111353543A (en) * | 2020-03-04 | 2020-06-30 | 镇江傲游网络科技有限公司 | Motion capture data similarity measurement method, device and system |
CN112055255A (en) * | 2020-09-15 | 2020-12-08 | 深圳创维-Rgb电子有限公司 | Shooting image quality optimization method and device, smart television and readable storage medium |
WO2021056750A1 (en) * | 2019-09-29 | 2021-04-01 | 北京市商汤科技开发有限公司 | Search method and device, and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20070103471A1 (en) * | 2005-10-28 | 2007-05-10 | Ming-Hsuan Yang | Discriminative motion modeling for human motion tracking |
CN103679757A (en) * | 2013-12-31 | 2014-03-26 | 北京交通大学 | Behavior segmentation method and system specific to human body movement data |
CN104964686A (en) * | 2015-05-15 | 2015-10-07 | 浙江大学 | Indoor positioning device and method based on motion capture and method |
CN105046720A (en) * | 2015-07-10 | 2015-11-11 | 北京交通大学 | Behavior segmentation method based on human body motion capture data character string representation |
-
2016
- 2016-06-17 CN CN201610436925.XA patent/CN106127803A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20070103471A1 (en) * | 2005-10-28 | 2007-05-10 | Ming-Hsuan Yang | Discriminative motion modeling for human motion tracking |
CN103679757A (en) * | 2013-12-31 | 2014-03-26 | 北京交通大学 | Behavior segmentation method and system specific to human body movement data |
CN104964686A (en) * | 2015-05-15 | 2015-10-07 | 浙江大学 | Indoor positioning device and method based on motion capture and method |
CN105046720A (en) * | 2015-07-10 | 2015-11-11 | 北京交通大学 | Behavior segmentation method based on human body motion capture data character string representation |
Non-Patent Citations (2)
Title |
---|
JIUN-YU KAO等: "GRAPH-BASED APPROACH FOR MOTION CAPTURE REPRESENTATION AND ANALYSIS", 《2014 IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING》 * |
XIAOHONG ZHAO等: "Sports Video Segmentation Using Spectral Clustering", 《JOURNAL OF MULTIMEDIA》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107169988A (en) * | 2017-05-12 | 2017-09-15 | 江苏大学 | A kind of extraction method of key frame based on COS distance hierarchical clustering |
CN107832713A (en) * | 2017-11-13 | 2018-03-23 | 南京邮电大学 | A kind of human posture recognition method based on OptiTrack |
CN107832713B (en) * | 2017-11-13 | 2021-11-16 | 南京邮电大学 | Human body posture recognition method based on OptiTrack |
WO2021056750A1 (en) * | 2019-09-29 | 2021-04-01 | 北京市商汤科技开发有限公司 | Search method and device, and storage medium |
CN111353543A (en) * | 2020-03-04 | 2020-06-30 | 镇江傲游网络科技有限公司 | Motion capture data similarity measurement method, device and system |
CN112055255A (en) * | 2020-09-15 | 2020-12-08 | 深圳创维-Rgb电子有限公司 | Shooting image quality optimization method and device, smart television and readable storage medium |
CN112055255B (en) * | 2020-09-15 | 2022-07-05 | 深圳创维-Rgb电子有限公司 | Shooting image quality optimization method and device, smart television and readable storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Chen et al. | Destruction and construction learning for fine-grained image recognition | |
Peng et al. | Correlation congruence for knowledge distillation | |
Babenko et al. | Aggregating deep convolutional features for image retrieval | |
CN106127803A (en) | Human body motion capture data behavior dividing method and system | |
Savva et al. | Shrec’17 track large-scale 3d shape retrieval from shapenet core55 | |
Gao et al. | View-based 3D object retrieval: challenges and approaches | |
Yan et al. | Beyond spatial pyramids: A new feature extraction framework with dense spatial sampling for image classification | |
Ming et al. | Simple triplet loss based on intra/inter-class metric learning for face verification | |
Oliveira et al. | Sparse spatial coding: A novel approach for efficient and accurate object recognition | |
CN109063649B (en) | Pedestrian re-identification method based on twin pedestrian alignment residual error network | |
CN103902964B (en) | A kind of face identification method | |
CN101866429A (en) | Training method of multi-moving object action identification and multi-moving object action identification method | |
CN104616316A (en) | Method for recognizing human behavior based on threshold matrix and characteristics-fused visual word | |
CN103605952A (en) | Human-behavior identification method based on Laplacian-regularization group sparse | |
CN102289685B (en) | Behavior identification method for rank-1 tensor projection based on canonical return | |
CN104881640A (en) | Method and device for acquiring vectors | |
Tümen et al. | Feature Extraction and Classifier Combination for Image-based Sketch Recognition. | |
Kong et al. | Qualitative and quantitative analysis of multi-pattern wafer bin maps | |
CN102915448A (en) | AdaBoost-based 3D (three-dimensional) model automatic classification method | |
Singh et al. | Leaf identification using feature extraction and neural network | |
Lee et al. | Filter pruning and re-initialization via latent space clustering | |
Khokher et al. | A super descriptor tensor decomposition for dynamic scene recognition | |
CN112560894A (en) | Improved 3D convolutional network hyperspectral remote sensing image classification method and device | |
Sun et al. | Weak supervised learning based abnormal behavior detection | |
Li et al. | A spectral clustering based filter-level pruning method for convolutional neural networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20161116 |
|
RJ01 | Rejection of invention patent application after publication |