CN104778459A - Feature fusion method for actions of multiple athletes in football match video - Google Patents

Abstract

The invention discloses a feature fusion method for actions of multiple athletes in a football match video. The feature fusion method for the actions of the multiple athletes in the football match video comprises the following steps: a, performing a feature dimension reduction operation on extracted features on the basis of extracting features including athlete clothing colors, outlines, motion trails and football field lines; b, utilizing a multi-feature fusion technology to fuse the features; c, describing athlete actions by utilizing the fused features to perform action identification. The feature fusion method for the actions of the multiple athletes in the football match video can overcome the defects that the operation process is complex, the consumed time is long and the reliability is low in the prior art so as to achieve the advantages that the operation process is simple, the consumed time is short and the reliability is high.

Description

The Feature fusion of the behavior of the person of doing more physical exercises in a kind of section of football match video

Technical field

The present invention relates to technical field of video processing, particularly, relate to the Feature fusion of the behavior of the person of doing more physical exercises in a kind of section of football match video.

Background technology

Sportsman's behavior in section of football match video is a kind of team's behavior of planned, the high synergitic person of doing more physical exercises (intelligent body).The understanding of team's behavior and identification are one of the important research problems in computer vision research field, there is many-sided application perhaps, as video monitoring, object video summary, man-machine interaction, Sports Video Analysis, sportsman's supplemental training, the auxiliary penalty of match and video frequency searching are browsed, great economic worth and social value can be obtained.

In the video monitoring of football match, because single features is difficult to the feature of the behavior effectively describing the person of doing more physical exercises, so carry out with single features the unreliability that Activity recognition can cause result of calculation.

Realizing in process of the present invention, inventor finds at least to exist in prior art the defect such as the long and reliability of operating process complexity, spended time is low.

Summary of the invention

The object of the invention is to, for the problems referred to above, propose the Feature fusion of the behavior of the person of doing more physical exercises in a kind of section of football match video, with realize operating process simple, take a short time and advantage that reliability is high.

For achieving the above object, the technical solution used in the present invention is: the Feature fusion of the behavior of the person of doing more physical exercises in a kind of section of football match video, comprising:

A, extraction comprise on the basis of the feature of athletic clothing color, profile, movement locus and court line, to extract feature carry out Feature Dimension Reduction operation;

B, use multiple features fusion technology merge these features;

C, with merge after feature carry out Describing Motion person's behavior, carry out Activity recognition.

Further, described step a, specifically comprises:

KPCA algorithm is used to carry out Nonlinear Dimension Reduction to the feature extracted;

? in space, the training characteristics collection T of a given M element _x={ X ₁, X ₂..., X _m, the object of sub-space learning is at lower dimensional space in find an embedding data collection E _y={ Y ₁, Y ₂..., Y _m;

On H, principal component analysis (PCA) is applied to mapping (enum) data T _φ={ φ (X ₁), φ (X ₂) ..., φ (X _m); If k is a positive semidefinite kernel function, through type (1) definition two proper vectors with between nonlinear relationship:

$k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)=(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))---\left(1\right)$

The coefficient problem finding major component in H space can be summed up as the diagonalization of interior nuclear matrix κ:

$\mathrm{\γ\λ}\stackrel{\→}{e}=\mathrm{\κ}\stackrel{\→}{e}---\left(2\right)$

Wherein, ${\mathrm{\κ}}_{\mathrm{ij}}=k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j),\stackrel{\→}{e}={[e_1,e_2,...,e_{\mathrm{\γ}}]}^T;$

With representing main shaft, being mapped to a jth main shaft Z by newly putting X ^jbe expressed as:

$(Z^j\·\mathrm{\φ}\left(\stackrel{\→}{x}\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^j(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^{j}k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)---\left(3\right)$

Obtain after comprising the embedded space of first d major component, any one video v is mapped as an association track T of d dimensional feature space _o={ O ₁, O ₂..., O _t.

Further, described step b, specifically comprises:

Generation RBF neural is automatically adopted to merge various features;

The input layer of network is made up of N number of neuron, uses identical excitation function φ, and d is the distance utilizing training data to calculate; α ∈ [0,1] is the reduction parameter of corresponding neuron i, then have:

$\left\{\begin{array}{c}{s}_i={\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\\ \mathrm{\φ}\left(d_i\right)=\exp (-{\mathrm{\γ}}_i{\left(d_i\right)}^2)\end{array}\right.---\left(4\right)$

L1 layer is used for calculating the trust block m of i-th module be connected with front one deck i-th neuron _i;

$\left\{\begin{array}{c}{m}_i\left(\right\{w_q\left\}\right)={\mathrm{\α}}_i{u}_{q,i}\mathrm{\φ}\left(d_i\right)\\ m_i\left(\mathrm{\Ω}\right)=1-{\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\end{array}\right.---\left(5\right)$

Wherein, u _q,imember's degree of every category feature, q={1,2 ..., M};

L2 layer utilizes Dempster-Shafer rule of combination in monolithic, merge N number of different block function, and rule of combination is such as formula (6):

$m\left(A\right)=(m_1\⊕m_2\⊕...\⊕m_N)=\underset{B_1\∩B_2\∩...\∩B_N=A}{\mathrm{\Σ}}\underset{i=1}{\overset{N}{\mathrm{\Π}}}{m}_i\left(B_i\right)---\left(6\right)$

The excitation vector of definition i-th module is obtain by formula (7) recursive calculation;

$\left\{\begin{array}{c}{u}_1=m_1\\ u_{i,j}=u_{i-1,j}{m}_{i,j}+u_{i-1,j}{m}_{i,M+1}+u_{i-1,M+1}{m}_{i,j}\\ u_{i,M}=u_{i-1,M}{m}_{i,M+1}\end{array}\right.---\left(7\right)$

In block computation process, the excitation vector of consideration primitive character calculates and exports to reduce above-mentioned impact, calculates with (8):

$\left\{\begin{array}{c}{O}_j=\frac{{\mathrm{\Σ}}_{i=1}^N{u}_{i,j}}{{\mathrm{\Σ}}_{i=1}^N{\mathrm{\Σ}}_{j=1}^{M+1}{u}_{i,j}}\\ p_q=O_q+O_{M+1}\end{array}\right.---\left(8\right).$

Further, described step c, specifically comprises:

U, alpha, gamma is the important coefficient of each feature in Fusion Features process, and obtained by the study of network, the study of neural network, s and p is decided by the Gradient Descent of output error.

The Feature fusion of the behavior of the person of doing more physical exercises in the section of football match video of various embodiments of the present invention, owing to comprising: a, extraction comprise on the basis of the feature of athletic clothing color, profile, movement locus and court line, to extract feature carry out Feature Dimension Reduction operation; B, use multiple features fusion technology merge these features; C, with merge after feature carry out Describing Motion person's behavior, carry out Activity recognition; Thus the defect that operating process in prior art is complicated, spended time is long and reliability is low can be overcome, with realize operating process simple, take a short time and advantage that reliability is high.

Other features and advantages of the present invention will be set forth in the following description, and, partly become apparent from instructions, or understand by implementing the present invention.

Below by drawings and Examples, technical scheme of the present invention is described in further detail.

Accompanying drawing explanation

Accompanying drawing is used to provide a further understanding of the present invention, and forms a part for instructions, together with embodiments of the present invention for explaining the present invention, is not construed as limiting the invention.In the accompanying drawings:

Fig. 1 is the 3-D view deriving from the behavior mapping trajectories in subspace in the present invention at KPCA;

Fig. 2 is the RBF neural merging multiple features in the present invention;

Fig. 3 is the learning process of network in the present invention.

Embodiment

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are described, should be appreciated that preferred embodiment described herein is only for instruction and explanation of the present invention, is not intended to limit the present invention.

According to the embodiment of the present invention, as shown in Figure 1-Figure 3, the Feature fusion of the behavior of the person of doing more physical exercises in a kind of section of football match video is provided.

The Fusion Features of the behavior of the person of doing more physical exercises in section of football match video

Technical solution of the present invention is considered on the basis of extracting the features such as athletic clothing color, profile, movement locus and court line, first Feature Dimension Reduction operation is carried out to the feature extracted, then multiple features fusion technology is used to merge these features, and carry out Describing Motion person's behavior by the feature after merging, carry out Activity recognition.

Feature Dimension Reduction

In order to obtain compact description and effective calculating, technical solution of the present invention uses KPCA algorithm to carry out Nonlinear Dimension Reduction to the feature extracted.Mainly consider two aspects: 1. KPCA provides the nonlinear organization that a kind of effective sub-space learning method finds " action space ".2. KPCA can be applied to any new data point simply, and the Method of Nonlinear Dimensionality Reductions such as ISOMAP, LLE are still unclear to how describing new data point.

? in space, the training characteristics collection T of a given M element _x={ X ₁, X ₂..., X _m, the object of sub-space learning is at lower dimensional space in find an embedding data collection E _y={ Y ₁, Y ₂..., Y _m.For core principle component analysis method, each vector X _ifirst pass through by Nonlinear Mapping in Hilbert space H.Then, on H, principal component analysis (PCA) is applied to mapping (enum) data owing to employing " interior geo-nuclear tracin4 ", this mapping process can omit.If k is a positive semidefinite kernel function, through type (1) definition two proper vectors with between nonlinear relationship.

$k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)=(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))---\left(1\right)$

The coefficient problem finding major component in H space can be summed up as the diagonalization of interior nuclear matrix κ:

$\mathrm{\γ\λ}\stackrel{\→}{e}=\mathrm{\κ}\stackrel{\→}{e}---\left(2\right)$

Wherein, ${\mathrm{\κ}}_{\mathrm{ij}}=k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j),\stackrel{\→}{e}={[e_1,e_2,...,e_{\mathrm{\γ}}]}^T.$

With representing main shaft, being mapped to a jth main shaft Z by newly putting X ^jcan be expressed as:

$(Z^j\·\mathrm{\φ}\left(\stackrel{\→}{x}\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^j(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^{j}k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)---\left(3\right)$

Gaussian kernel function is used in experiment.

Obtain after comprising the embedded space of first d major component, any one video v can be mapped as an association track T of d dimensional feature space _o={ O ₁, O ₂..., O _t.Fig. 1 shows the behavior mapping trajectories (PTM) of document [14] data centralization, and wherein, the time sequencing mark of frame is unsharp.

The Fusion Features strategy of the person's of doing more physical exercises behavior

Consideration three problems are needed: 1. need to merge which characteristic information in the Fusion Features of the person's of doing more physical exercises behavior.Different characteristic informations to be merged according to different application scenarios choose reasonable; 2. on what level, Fusion Features is carried out.Fusion Features can be selected to carry out in low-level image feature level, intermediate key word level and high-level semantics rank; 3. the selection of convergence strategy.Comprise data normalization process, merge probability expression etc.Current multiple features fusion strategy has multiplicative fusion, Weighted Fusion and discrete Karhunen-Loeve (being called for short K-L) conversion to merge.Multiplicative merges the joint distribution adopting the method for feature weight quadrature to calculate multiple feature, can effectively improve motion target tracking precision, but may amplify noise; Weighted Fusion regulates each feature weight coefficient according to the confidence level of different characteristic, then utilizes weighted sum to calculate total feature weight, and weighted feature merges insensitive for noise, but can not improve the confidence level of fusion tracking ^[143].It is carry out self-adaptation fusion to multiple features that Karhunen-Loeve transformation is merged, and has advantages such as protecting entropy, protect that energy, decorrelation and energy are redistributed and concentrated, but calculates more complicated, lack learning process.The automatic generation RBF neural that technical solution of the present invention proposes as shown in Figure 2 merges various features.

The input layer of network is made up of N number of neuron, uses identical excitation function φ, and d is the distance utilizing training data to calculate.α ∈ [0,1] is the reduction parameter of corresponding neuron i.Then have:

$\left\{\begin{array}{c}{s}_i={\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\\ \mathrm{\φ}\left(d_i\right)=\exp (-{\mathrm{\γ}}_i{\left(d_i\right)}^2)\end{array}\right.---\left(4\right)$

L1 layer is used for calculating the trust block m of i-th module be connected with front one deck i-th neuron _i.

$\left\{\begin{array}{c}{m}_i\left(\right\{w_q\left\}\right)={\mathrm{\α}}_i{u}_{q,i}\mathrm{\φ}\left(d_i\right)\\ m_i\left(\mathrm{\Ω}\right)=1-{\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\end{array}\right.---\left(5\right)$

Wherein, u _q,imember's degree of every category feature, q={1,2 ..., M}.

L2 layer utilizes Dempster-Shafer rule of combination in monolithic, merge N number of different block function, and rule of combination is such as formula (6).

$m\left(A\right)=(m_1\⊕m_2\⊕...\⊕m_N)=\underset{B_1\∩B_2\∩...\∩B_N=A}{\mathrm{\Σ}}\underset{i=1}{\overset{N}{\mathrm{\Π}}}{m}_i\left(B_i\right)---\left(6\right)$

The excitation vector of definition i-th module is available formula (7) recursive calculation obtains.

$\left\{\begin{array}{c}{u}_1=m_1\\ u_{i,j}=u_{i-1,j}{m}_{i,j}+u_{i-1,j}{m}_{i,M+1}+u_{i-1,M+1}{m}_{i,j}\\ u_{i,M}=u_{i-1,M}{m}_{i,M+1}\end{array}\right.---\left(7\right)$

The output of the output layer of network is very responsive to the number of primitive character, there is less change and may cause fusion output characteristic that larger change occurs in number of features, therefore in block computation process, the excitation vector of consideration primitive character calculates output to reduce above-mentioned impact, calculates with (8).

$\left\{\begin{array}{c}{O}_j=\frac{{\mathrm{\Σ}}_{i=1}^N{u}_{i,j}}{{\mathrm{\Σ}}_{i=1}^N{\mathrm{\Σ}}_{j=1}^{M+1}{u}_{i,j}}\\ p_q=O_q+O_{M+1}\end{array}\right.---\left(8\right)$

Fusion parameters is estimated

U, alpha, gamma is the important coefficient of each feature in Fusion Features process, can be obtained by the study of network, and as shown in Figure 3, s and p is decided by the Gradient Descent of output error in the study of neural network.

This section introduces feature extraction and the fusion of the person's of the doing more physical exercises behavior in section of football match video.First describe and utilize background subtraction to carry out moving Object Segmentation, by moving target and court background separation from section of football match video image.Then utilize the clothing color of sportsman and judge, profile to follow the tracks of sportsman and judge, and utilize BSAS clustering algorithm to sportsman and the judge of classifying; The degree of association of the reference model of the comparison collection of candidate region and ball sample in court is utilized to follow the tracks of and detect ball; Utilize statistical learning method to the position of moving target T-ground of deriving; The extraction of feature has a significant impact last recognition result, and technical solution of the present invention is by extracting the contour feature of the clothing color histogram of sportsman and judge and color moment feature, sportsman and judge; Section of football match video image is converted into binary map, utilizes Hough transform to extract the coordinate parameters feature of court line, and obtain accurate rectilinear coordinates with gray scale matching; Utilize Kalman filter follow the tracks of moving target and predict movement locus, utilize track growth method to extract the low-level image features such as the track characteristic of moving target to alleviate the burden of high-rise recognizer; Before the feature of the behavior to the person of doing more physical exercises in the section of football match video extracted merges, KPCA algorithm is utilized to carry out Nonlinear Dimension Reduction to the feature extracted; Finally build and automatically generate RBF neural, utilize the study of Dempster-Shafer rule of combination and network to carry out Fusion Features; U in fusion parameters, alpha, gamma is obtained by the study of network, s and p is decided by the Gradient Descent of output error.

To sum up, on the basis of the Activity recognition present Research of the person of doing more physical exercises of technical solution of the present invention in comprehensive review section of football match video, proposition fuzzy inference system carries out the person's of the doing more physical exercises Activity recognition in section of football match video.The deficiency existed in research process for the current person's of doing more physical exercises Activity recognition launches research, and groundwork and the innovative point of paper are as follows:

(1) in video image preprocessing process, the Fractal Wavelet adaptive denoising algorithm based on multivariate statistical model and the section of football match video algorithm for image enhancement based on Orthogonal wavelet analysis and Pseudo Col ored Image is proposed: in video image denoising process, by in conjunction with multivariate statistical model and Fractal Wavelet denoising method, various relevant information can be estimated more accurately, select the image space of high-quality.In best subtree territory, nearly excellent father and son tree is found according to piecing distance together under the noise variance of appropriateness.Thus dope muting image Fractal Wavelet coding, reach the object optimizing denoising.Based on multivariate statistical model Fractal Wavelet adaptive denoising algorithm removal noise while, effectively can keep edge and the textural characteristics of image, retain the fine structure of image well, achieve good denoising effect.Owing to have employed prediction Wavelet-fractal coding, optimize algorithm structure, the processing speed of algorithm is than very fast.Section of football match video algorithm for image enhancement based on Orthogonal wavelet analysis and Pseudo Col ored Image both can overcome the defects such as the partially bright and contrast of the image after adopting the process of Orthogonal wavelet analysis section of football match video algorithm for image enhancement is poor, and the section of football match video algorithm for image enhancement that can overcome again Pseudo Col ored Image fully can not process the defect of some detailed information in image.

(2) propose to merge multiple features with automatic generation RBF network: in order to overcome illumination, block, the impact such as dimensional variation, meet the requirement of real-time identification, propose the movement locus feature of the clothing color moment characteristics with the sportsman extracted in the person's of the doing more physical exercises Activity recognition process automatically generated in RBF network integration section of football match video, sportsman and the contour feature of judge, the coordinate parameters feature of court line and moving target.Define a behavioral characteristics model, first extract the principal character of the person's of the doing more physical exercises behavior in relevant section of football match video, when these features have been not enough to Understanding and reasoning, system has progressively extracted the minutia of candidate.Simultaneously by adopting 3D local direction histogram feature, can effectively solve the diversity of blocking with attitudes vibration, making the person's of the doing more physical exercises Activity recognition in section of football match video have stronger robustness with understanding.

(3) propose space-time driving force model and the fuzzy inference system of the group behavior pattern in section of football match video: utilize limited by space-time track, change in time and space between the number of change and behavior the behavioural characteristic that forms of set be that group behavior in section of football match video carries out modeling.Using the dynamic process of the space-time driving force that the group behavior in section of football match video distributes as an area intensive, replace the generation of motion with simple discrete loci point set.By by F (t _n, x, y) Lie group's non-linearity manifold spatial transformation be f (t _n, x, y) the linear space of Lie algebraically carry out the calculated amount of simplified model greatly.Model takes full advantage of the low level feature such as position and speed obtained from movement locus, and the study of model is simple.The Fusion Features ability of model is stronger, compares superior performance with other model.Model has versatility and dirigibility in modeling complex behavior pattern.

Propose the fuzzy inference system of the person's of the doing more physical exercises Activity recognition in section of football match video, by abstract for behavior model be event model, set up pass, shoot, control ball, the inference rule of the event such as ball, goal, corner-kick, free kick, offside, no play, red and yellow card of dribbling, lose, system applies the person's of doing more physical exercises Activity recognition that these inference rules are carried out in section of football match video.

(4) devise dimension self-adaption local space time feature Harris detection operand and solve the illumination in complex background, multiple dimensioned and occlusion issue: according to globality and the level principle of visual analysis, spatial pyramid model is promoted and is applied in local space time's feature, devising dimension self-adaption selects local space time feature Harris to detect operand, this operand method is simple, computing velocity is fast, the illumination variation in complex background and Issues On Multi-scales can be solved, and occlusion issue can be solved to a certain extent.

(5) first transfer learning algorithm introduced the person's of the doing more physical exercises Activity recognition in section of football match video and solve various visual angles and occlusion issue in understanding: using for reference transfer learning at Images Classification, the successful experience of the area researches such as gesture identification, devise the local space time's code book prototype building algorithm based on transfer learning, can sharing feature between the code book that this algorithm makes different visual angles, the person's of the doing more physical exercises behavior in section of football match video is represented in compacter mode, various visual angles can be solved to a certain extent, the person's of doing more physical exercises Activity recognition in raising section of football match video and the robustness of understanding method.

(6) propose the tree construction hybrid classifer based on priori and artificial neural network, improve accuracy rate and the recognition speed of identification: propose a kind of tree construction hybrid classifer based on priori and artificial neural network.The decision-tree model of optimum is combined with neural network, priori is added the classification accuracy being used for improving sorter in the structure of sorter, maximally utilised the characteristic of Nonlinear Classification accurately of priori and the neural network obtained in learning process.Decision tree provides possibility for building decision tree forest on a data set.Utilize the independence of neural network and adaptivity solve single sorter be difficult to unceasing study and conform, illumination, sportsman's number change the person's of doing more physical exercises Activity recognition problem.

In a word, technical solution of the present invention is started with from key problems such as the constructing technologies of feature extraction, team's behavior representation, team's behavior modeling and sorter the Activity recognition of the person of doing more physical exercises studied section of football match video.First proposed the visual effect improving video image by the Fractal Wavelet adaptive denoising algorithm based on multivariate statistical model and the section of football match video algorithm for image enhancement based on Orthogonal wavelet analysis and Pseudo Col ored Image, then propose with automatically generating the multiple features that RBF network merges extraction, propose the space-time driving force model of the group behavior pattern in section of football match video, and carry out the group behavior in modeling section of football match video with it, have more versatility and dirigibility.Devise dimension self-adaption local space time feature Harris detection operand and solve the illumination in complex background, multiple dimensioned and occlusion issue, and occlusion issue can be solved to a certain extent.Propose the tree construction hybrid classifer based on priori and artificial neural network, improve accuracy rate and the recognition speed of identification, utilize the independence of neural network and adaptivity solve single sorter be difficult to unceasing study and conform, illumination, sportsman's number change the person's of doing more physical exercises Activity recognition problem.Finally first transfer learning algorithm introduced the person's of the doing more physical exercises Activity recognition in section of football match video for various visual angles and occlusion issue and in understanding, solve partial occlusion problem and various visual angles problem.The method that technical solution of the present invention proposes improves the recognition performance of the Activity recognition of the person of doing more physical exercises in section of football match video, is conducive to promoting that the Activity recognition technology of the person of doing more physical exercises in section of football match video constantly advances with further practical.

Last it is noted that the foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, although with reference to previous embodiment to invention has been detailed description, for a person skilled in the art, it still can be modified to the technical scheme described in foregoing embodiments, or carries out equivalent replacement to wherein portion of techniques feature.Within the spirit and principles in the present invention all, any amendment done, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (4)

1. a Feature fusion for the behavior of the person of doing more physical exercises in section of football match video, is characterized in that, comprising:

A, extraction comprise on the basis of the feature of athletic clothing color, profile, movement locus and court line, to extract feature carry out Feature Dimension Reduction operation;

B, use multiple features fusion technology merge these features;

C, with merge after feature carry out Describing Motion person's behavior, carry out Activity recognition.

2. the Feature fusion of the behavior of the person of doing more physical exercises in section of football match video according to claim 1, is characterized in that, described step a, specifically comprises:

KPCA algorithm is used to carry out Nonlinear Dimension Reduction to the feature extracted;

? in space, the training characteristics collection T of a given M element _x={ X ₁, X ₂..., X _m, the object of sub-space learning is at lower dimensional space in find an embedding data collection E _y={ Y ₁, Y ₂..., Y _m;

On H, principal component analysis (PCA) is applied to mapping (enum) data T _φ={ φ (X ₁), φ (X ₂) ..., φ (X _m); If k is a positive semidefinite kernel function, through type (1) definition two proper vectors with between nonlinear relationship:

$k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)=(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))---\left(1\right)$

The coefficient problem finding major component in H space can be summed up as the diagonalization of interior nuclear matrix κ:

$\mathrm{\γ\λ}\stackrel{\→}{e}=\mathrm{\κ}\stackrel{\→}{e}---\left(2\right)$

Wherein, ${\mathrm{\κ}}_{\mathrm{ij}}=k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j),\stackrel{\→}{e}={[e_1,e_2,...,e_{\mathrm{\γ}}]}^T;$

With representing main shaft, being mapped to a jth main shaft Z by newly putting X ^jbe expressed as:

$(Z^j\·\mathrm{\φ}\left(\stackrel{\→}{x}\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^j(\mathrm{\φ}\left({\stackrel{\→}{x}}_i\right)\·\mathrm{\φ}\left({\stackrel{\→}{x}}_j\right))=\underset{i=1}{\overset{\mathrm{\γ}}{\mathrm{\Σ}}}{e}_i^{j}k({\stackrel{\→}{x}}_i,{\stackrel{\→}{x}}_j)---\left(3\right)$

Obtain after comprising the embedded space of first d major component, any one video v is mapped as an association track T of d dimensional feature space _o={ O ₁, O ₂..., O _t.

3. the Feature fusion of the behavior of the person of doing more physical exercises in section of football match video according to claim 1 and 2, is characterized in that, described step b, specifically comprises:

Generation RBF neural is automatically adopted to merge various features;

The input layer of network is made up of N number of neuron, uses identical excitation function φ, and d is the distance utilizing training data to calculate; α ∈ [0,1] is the reduction parameter of corresponding neuron i, then have:

$\left\{\begin{array}{c}{s}_i={\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\\ \mathrm{\φ}\left(d_i\right)=\exp \left({-{\mathrm{\γ}}_i\left(d_i\right)}^2\right)\end{array}\right.---\left(4\right)$

L1 layer is used for calculating the trust block m of i-th module be connected with front one deck i-th neuron _i;

$\left\{\begin{array}{c}{m}_i\left(\right\{w_q\left\}\right)={\mathrm{\α}}_i{u}_{q,i}\mathrm{\φ}\left(d_i\right)\\ m_i\left(\mathrm{\Ω}\right)=1-{\mathrm{\α}}_i\mathrm{\φ}\left(d_i\right)\end{array}\right.---\left(5\right)$

Wherein, u _q,imember's degree of every category feature, q={1,2 ..., M};

L2 layer utilizes Dempster-Shafer rule of combination in monolithic, merge N number of different block function, and rule of combination is such as formula (6):

$m\left(A\right)=(m_1\⊕m_2\⊕...\⊕m_N)=\underset{B_1\∩B_2\∩...\∩B_N=A}{\mathrm{\Σ}}\underset{i=1}{\overset{N}{\mathrm{\Π}}}{m}_i\left(B_i\right)---\left(6\right)$

The excitation vector of definition i-th module is obtain by formula (7) recursive calculation;

$\left\{\begin{array}{c}{u}_1=m_1\\ u_{i,j}=u_{i-1,j}{m}_{i,j}+u_{i-1,j}{m}_{i,M+1}+u_{i-1,M+1}{m}_{i,j}\\ u_{i,M}=u_{i-1,M}{m}_{i,M+1}\end{array}\right.---\left(7\right)$

In block computation process, the excitation vector of consideration primitive character calculates and exports to reduce above-mentioned impact, calculates with (8):

$\left\{\begin{array}{c}{O}_j=\frac{{\mathrm{\Σ}}_{i=1}^N{u}_{i,j}}{{\mathrm{\Σ}}_{i=1}^N{\mathrm{\Σ}}_{j=1}^{M+1}{u}_{i,j}}\\ p_q=O_q+O_{M+1}\end{array}\right.---\left(8\right).$

4. the Feature fusion of the behavior of the person of doing more physical exercises in section of football match video according to claim 1 and 2, is characterized in that, described step c, specifically comprises:

U, alpha, gamma is the important coefficient of each feature in Fusion Features process, and obtained by the study of network, the study of neural network, s and p is decided by the Gradient Descent of output error.