CN105956543A - Multiple athletes behavior detection method based on scale adaptation local spatiotemporal features - Google Patents

Abstract

The invention discloses a multiple athletes behavior detection method based on scale adaptation local spatiotemporal features, and the method comprises the following steps: S1, a normalized Laplace operand is adopted to estimate a local scale and establish a scale adaptation local spatiotemporal feature detection algorithm, and a Harris spatiotemporal interest point detection operand and the Laplace operand are combined to infer a Harris-Laplace spatiotemporal interest point detection operand; S2, the normalized Laplace operand is generalized to a spatiotemporal domain; S3, local spatiotemporal feature descriptors are generalized to three-dimensional soccer match video images; then before a K-means clustering algorithm is adopted to generate a spatiotemporal codebook, each spatiotemporal interest point is normalized to ensure that the each spatiotemporal interest point is invariable under scaling and shifting.

Description

A kind of person's of doing more physical exercises behavioral value method based on dimension self-adaption local space time feature

Technical field

The invention belongs to Activity recognition field, be specifically related to a kind of special based on dimension self-adaption local space time The person's of the doing more physical exercises behavioral value method levied.

Background technology

The quality of the person's of the doing more physical exercises behavior representation in section of football match video directly affects the accurate of Activity recognition Rate.During the person's of doing more physical exercises behavior analysis in football match and identification, the feature chosen is the most, Description to behavior is the most abundant.But choose too much feature and can cause the redundancy mistake between data Greatly, the dimension of characteristic vector is too high, and the essential laws of data distribution is not easy to be found, and trains behavior Identify that the data volume needed for model is excessive, algorithm computationally intensive, it is unfavorable for the real-time of Activity recognition Process.Therefore, what the person's of the doing more physical exercises Activity recognition in section of football match video was studied key challenge is how The feature effectively describing extraction can characterize behavior well.

Behavior representation is conducted extensive research by past researcher, wherein has than more typical research: BOBICK and DAVIS utilizes background subtraction to derive the expression of time template.The letter of behavior method for expressing Single, but the most affected by noise.WANG uses adaptive multi-thresholding to select Optical-flow Feature as behavior Represent and identify the team's behavior in section of football match video.But study limitation is in processing three class team row For, and easily by noise jamming.Existing employing seasonal effect in time series cause effect relation is described as behavior, And utilize spatial behavior path matching to identify behavior of men, but it is difficult to be generalized to the row of the person of doing more physical exercises For identifying, with four complete dimensional objects-real-time, interactive tensor, team's behavioral pattern is described, Reduce kernel by learning and optimizing tensor product, make it agglomerate to a diacritic space-time interaction square In Zhen.

The expression of target recognition based on model and probability Plan recognition constitutes four main assumptions: 1. in intelligence When can be engaged in team appointments between body, single intelligent body target is the natural former of the time-space relationship specified Subrepresentation unit；2. during the Activity recognition of the multiple agent of highly structural, the time of behavior The high-level description of structure uses less low order time-space relationship collection and logic to limit and just be enough to express intelligence Relation between energy body；3. the multiple source that Bayesian network is uncertain visually-perceptible feature provides A kind of suitable syncretizing mechanism；When 4. can merge uncertain with the Bayesian network automatically generated State information and calculating object trajectory data set are the probabilities of the behavior of a specific multiple agent.By Easily cause the biasing problem of labelling in the restriction assumed, and be difficult to.An energy is introduced bright for this Dynamic team member really encodes and adapts to the plan of new multiple agent of Plan recognition form Representation.The local time's dependency extracted from the plan representation method of multiple agent can be repaiied significantly Cut the hypothesis collection of potential effective team plan.Pruning process is relatively time-consuming, it is impossible to meet real-time football The requirement of match Activity recognition.

Hereafter during unsupervised time division, it is proposed that a kind of Gauss based on Dynamic Time Series The Activity recognition method of mixed model coupling behavior.The method suppose behavioral data be one can be abundant The how unidirectional label described, make use of space-time characteristic by assuming that the priority orders of behavior identifies many People's behavior.But in the character subset of team's Activity recognition, there will be multiple simultaneous events.Special In the case of, the behavior that the priority that is previously set is high is probably unessential behavior, so have impact on many The accuracy of athlete's Activity recognition.LAPTEV is succinct expression of video data to propose space-time interest Point, and inquired into the advantage of the behavior utilizing space-time interest points to describe people.

Use for reference the thought of LAPTEV, need a kind of new detection method to ensure that its zooming and panning not Degeneration.

Summary of the invention

For the deficiencies in the prior art, it is an object of the invention to provide a kind of based on dimension self-adaption local time The person's of doing more physical exercises behavioral value method of empty feature has been capable of before cluster generates code book, to each Space-time interest points has all carried out normalization, to ensure its zooming and panning invariance.

A kind of person's of doing more physical exercises behavioral value method based on dimension self-adaption local space time feature, described side Method comprises the following steps:

S1, uses normalized Laplace operand to estimate local scale, sets up dimension self-adaption

Local space time's feature detection algorithm, by Harris space-time interest points detection operand and Laplace

Operand combines and derives Harris-Laplace space-time interest points detection operand；

S2, is generalized to normalized Laplace operand in time-space domain；

S3, is generalized to local space time's Feature Descriptor in three-dimensional section of football match video image.

Preferably, described S1 mesoscale adaptive local space-time characteristic detection algorithm is specially

$H={\mathrm{\λ}}_1^3(\mathrm{\α}\mathrm{\β}-k(1+\mathrm{\α}+\mathrm{\β}\left)\right)---\left(1\right)$

Wherein, the positive local maximum of H corresponds to λ₁、λ₂And λ₃High level point, λ₁＜ λ₂＜ λ₃, Ratio, α=λ₂/λ₁, β=λ₃/λ₁.From requiring that H >=0 can obtain k≤α β/(1+ alpha+beta)³。

Preferably, local space time's Feature Descriptor is generalized to three-dimensional section of football match video figure by described S3 Computational methods in Xiang include: histogram calculation, the calculating of direction quantization and average gradient calculation.

Preferably, described histogram calculation and direction quantify calculating particularly as follows:

For a given cube c=(x_c,y_c,t_c,l_c,w_c,h_c), it is classified as S × S × S sub-block b_i。 For each sub-block, utilize formula (4) to calculate its average gradient, then willIt is quantified as q_bi.For Each region c, the average gradient vector after being quantified by summation calculates its rectangular histogram h_c。

$h_c=\underset{i=1}{\overset{S^3}{\mathrm{\Σ}}}{q}_{bi}---\left(3\right)$

For a given positive n face body, its center is exactly the initial point of three-dimensional system of coordinate, 3D gradient vector Being quantified by the translation of coordinate points relative to the direction of initial point, each point in matrix p is polyhedron The center p of each₁,p₂,…,p_n, p=(p₁,p₂,…,p_n)^T, wherein, p_i=(x_i,y_i,t_i)^T, then in 20 Heart point is:

Wherein, golden section

?On projection can pass through formula (5) and calculate；

${\hat{q}}_b={\[{\hat{q}}_{b1},{\hat{q}}_{b2},...,{\hat{q}}_{bn}\]}^T=\frac{p{\stackrel{\‾}{g}}_b}{\left|\right|{\stackrel{\‾}{g}}_b|{|}_2}---\left(5\right),$

WillThresholding is processed asThen formula (6) is utilized to carry out quantification treatment；

$q_b=\frac{\left|\right|g_b|{|}_2{\hat{q}}_{bth}}{\left|\right|q_{bth}|{|}_2}---\left(6\right).$

Preferably, described average gradient calculate particularly as follows:

By section of football match video sequence v, (x, y, t) along x, the partial derivative in y, t direction is divided into and being designated asThen Formula (7) can be used to calculate

$\left\{\begin{array}{c}{\mathrm{iv}}_{\∂x}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂x}(x^{\′},y^{\′},t^{\′})\\ {\mathrm{iv}}_{\∂y}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂y}(x^{\′},y^{\′},t^{\′})\\ {\mathrm{iv}}_{\∂t}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂t}(x^{\′},y^{\′},t^{\′})\end{array}\right.---\left(7\right)$

For a 3D cube b=(x, y, t, l, w, h), wherein, (x, y, t)^TAnd l, w, h represent cube respectively The position of body and length and width and height；

Formula (8) can be used quickly to calculate its average gradient vector'sPoint Amount；

$\begin{array}{c}{\stackrel{\‾}{g}}_{b\∂x}=\[{\mathrm{iv}}_{\∂x}(x+w,y+h,t+l)-{\mathrm{iv}}_{\∂x}(x,y+h,t+l)\\ -{\mathrm{iv}}_{\∂x}(x+w,y,t+l)+{\mathrm{iv}}_{\∂x}(x,y,t+l)\]\\ -\[{\mathrm{iv}}_{\∂x}(x+w,y+h,t)-{\mathrm{iv}}_{\∂x}(x,y+h,t)\\ -{\mathrm{iv}}_{\∂x}(x+w,y,t)+{\mathrm{iv}}_{\∂x}(x,y,t)\]\end{array}---\left(8\right)$

In like manner, can quickly calculate

Technical scheme has the advantages that

The present invention provides a kind of person's of doing more physical exercises behavioral value based on dimension self-adaption local space time feature Method, utilizes space-time interest points to carry out behavior knowledge to the person's of the doing more physical exercises behavior representing in section of football match video Not.The person's of doing more physical exercises behavior in section of football match video sequence is regarded as the space-time interest in three dimensions The set of point.Use rectangular histogram quantification technique that space-time interest points set is quantified as the Nogata that dimension is fixing Figure (instant null word), uses K-means clustering algorithm to generate space-time code book.At cluster generated code Before Ben, each space-time interest points is carried out normalization, to ensure its zooming and panning invariance. Test result indicate that, method in this paper can greatly reduce the amount of calculation of algorithm, with based on yardstick The person's of doing more physical exercises behavior representation in the section of football match video of adaptive local space-time characteristic is done more physical exercises The accuracy rate of identification can be significantly improved during member's Activity recognition.

Accompanying drawing explanation

Below by drawings and Examples, technical scheme is described in further detail.

Fig. 1 a is the present invention person of doing more physical exercises behavioral value side based on dimension self-adaption local space time feature Whole description of method；

Fig. 1 b is the present invention person of doing more physical exercises behavioral value side based on dimension self-adaption local space time feature The histogram calculation of method；

Fig. 1 c is the present invention person of doing more physical exercises behavioral value side based on dimension self-adaption local space time feature The direction gradient of method quantifies；

Fig. 1 d is the present invention person of doing more physical exercises behavioral value side based on dimension self-adaption local space time feature The average gradient of method calculates.

Detailed description of the invention

In order to have a clear understanding of technical scheme, its detailed knot be will be set forth in the description that follows Structure.Obviously, concrete execution the deficiency of the embodiment of the present invention is limited to those skilled in the art and is familiar with Specific details.The preferred embodiments of the present invention are described in detail as follows, except these described in detail are implemented Exception, it is also possible to there is other embodiments.

With embodiment, the present invention is described in further details below in conjunction with the accompanying drawings.

Use for reference LINDEBERG herein about the adaptively selected method of local scale feature in space, will Harris detection operand is generalized in section of football match video time-space domain, it is proposed that dimension self-adaption local Space-time characteristic detection algorithm (1), can be simultaneously in time and space at defined in local time spatial domain one Obtain the difference operation number of maximum in dimension, use normalized Laplace operand to estimate local Yardstick.Formula (1) is utilized to be tied mutually with Laplace operand by Harris space-time interest points detection operand Harris-Laplace space-time interest points detection operand is derived in conjunction.

$H={\mathrm{\λ}}_1^3(\mathrm{\α}\mathrm{\β}-k(1+\mathrm{\α}+\mathrm{\β}\left)\right)---\left(1\right)$

Wherein, the positive local maximum of H corresponds to λ₁、λ₂And λ₃High level point, λ₁＜ λ₂＜ λ₃, Ratio, α=λ₂/λ₁, β=λ₃/λ₁.From requiring that H >=0 can obtain k≤α β/(1+ alpha+beta)³.Assuming that α=β=1, The maximum value possible of k is 1/27.When k value is sufficiently large, point corresponding to the positive local maximum of H along time Between and direction in space image value change acutely.Working as the maximum of α and β in spatial domain especially is 23 Time, k ≈ 0.005.Therefore, football ratio can be detected by the positive local space time's maximum in detection H Space-time interest points in match sequence of video images v.

Formula (2) is utilized to be generalized in time-space domain by normalized Laplace operand.

${\▿}_{norm}^{2}r=r_x^2+r_y^2+r_t^2={\mathrm{\σ}}_r^2{\mathrm{\τ}}_r^{1/2}(r_x^2+r_y^2)+{\mathrm{\σ}}_r{\mathrm{\τ}}_r^{3/2}{r}_t^2---\left(2\right)$

Dimension self-adaption space-time after formula 1 describes popularization in detail detects specifically performing of operand Journey.

In section of football match video, the expression of regional area is an open problem.Utilization orientation ladder herein Degree rectangular histogram (Histogram Oriented Gradient, HoG) describes son by section of football match video It is considered as " Space Time " cuboid, and HoG is described son is generalized in three-dimensional section of football match video image.

For local space time's area-of-interest, local description represents this district by a characteristic vector Territory.Image or whole section of football match video sequence are represented by one group under different yardsticks and position Set of eigenvectors.In order to be able to effectively utilize local feature vectors to carry out in section of football match video The identification of the person's of doing more physical exercises behavior, objectively requires that these Feature Descriptors have stronger differentiation energy Power, simultaneously the most by illumination, the impact of the interference factors such as slight deformation.HoG is described son be generalized to The computational methods of 3D are as it is shown in figure 1, crucial calculating theing be directed to has: average gradient calculates, side The calculating of vectorization, histogram calculation.These three described in detail below calculates, in conjunction with Fig. 1 a to Fig. 1 d. Histogram calculation: the rectangular histogram of direction gradient needs to calculate in the set of a gradient vector. Circular is as follows: for a given cube c=(x_c,y_c,t_c,l_c,w_c,h_c), it is classified as S × S × S sub-block b_i.For each sub-block, utilize formula (4) to calculate its average gradient, then will It is quantified as q_bi.For each region c, it is straight that the average gradient vector after being quantified by summation calculates it Side figure h_c。

$h_c=\underset{i=1}{\overset{S^3}{\mathrm{\Σ}}}{q}_{bi}---\left(3\right)$

Direction gradient quantifies: in static 2D image, the histograms of oriented gradients of a n-bin is permissible Seeing as and surrounded by the regular polygon on n limit, each limit of regular polygon be equivalent in rectangular histogram One bin.In the 3 d space, regular polygon is extended to regular polygon (such as Fig. 1 (c)), often The limit number of the regular polygon seen has 4,6,8,12,20.Therefore, the polyhedron limit number used herein It is 20.For a given positive n face body, it is assumed that its center is exactly the initial point of three-dimensional system of coordinate, 3D Gradient vector is quantified by the translation (i.e. multiplication of matrices) of coordinate points relative to the direction of initial point. The center p that each point is each of polyhedron in matrix p₁,p₂,…,p_n, p=(p₁,p₂,…,p_n)^T, its In, p_i=(x_i,y_i,t_i)^T.Then 20 central points are:

Wherein, golden section?On projection can pass through formula (5) and calculate.

${\hat{q}}_b={\[{\hat{q}}_{b1},{\hat{q}}_{b2},...,{\hat{q}}_{bn}\]}^T=\frac{p{\stackrel{\‾}{g}}_b}{\left|\right|{\stackrel{\‾}{g}}_b|{|}_2}---\left(5\right)$

It follows that willThresholding is processed asThen formula (6) is utilized to carry out quantification treatment.

$q_b=\frac{\left|\right|g_b|{|}_2{\hat{q}}_{bth}}{\left|\right|q_{bth}|{|}_2}---\left(6\right)$

The calculating of average gradient: in order to calculate 3D histograms of oriented gradients it may first have to the most effective Calculate the gradient vector (such as Fig. 1 (d)) of image.Due to must take into different space scales and Time scale, an effective solution uses space-time pyramid exactly, precalculates each different Direction gradient under yardstick.Although but this method reduces time complexity, substantially increases Space complexity.Calendar year 2001, Viola etc. proposes integrogram concept, and using integrogram as Harr The intermediate representation method of feature calculation is applied to face detection system, substantially increases computational efficiency.Ginseng The method examining the proposition such as Viola, by section of football match video sequence v (x, y, t) along x, the partial derivative in y, t direction It is divided into and being designated asFormula (7) then can be used to calculate

$\left\{\begin{array}{c}{\mathrm{iv}}_{\∂x}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂x}(x^{\′},y^{\′},t^{\′})\\ {\mathrm{iv}}_{\∂y}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂y}(x^{\′},y^{\′},t^{\′})\\ {\mathrm{iv}}_{\∂t}(x,y,t)=\underset{x^{\′}\≤x,y^{\′}\≤y,t^{\′}\≤t}{\mathrm{\Σ}}{v}_{\∂t}(x^{\′},y^{\′},t^{\′})\end{array}\right.---\left(7\right)$

For a 3D cube b=(x, y, t, l, w, h), wherein, (x, y, t)^TAnd l, w, h represent cube respectively The position of body and length and width and height.

Formula (8) can be used quickly to calculate its average gradient vector'sPoint Amount.

$\begin{array}{c}{\stackrel{\‾}{g}}_{b\∂x}=\[{\mathrm{iv}}_{\∂x}(x+w,y+h,t+l)-{\mathrm{iv}}_{\∂x}(x,y+h,t+l)\\ -{\mathrm{iv}}_{\∂x}(x+w,y,t+l)+{\mathrm{iv}}_{\∂x}(x,y,t+l)\]\\ -\[{\mathrm{iv}}_{\∂x}(x+w,y+h,t)-{\mathrm{iv}}_{\∂x}(x,y+h,t)\\ -{\mathrm{iv}}_{\∂x}(x+w,y,t)+{\mathrm{iv}}_{\∂x}(x,y,t)\]\end{array}---\left(8\right)$

In like manner, can quickly calculateSo, by using for reference the thought of integrogram, Ke Yi great The big space-time interest points improving the person's of doing more physical exercises behavior represents

Space-time interest points refers to the point that temporally and spatially Strength Changes is bigger.Space-time interest points table Show that method is the new frame sequence low-level image feature method for expressing of a kind of comparison.If section of football match video space-time diagram As sequence isBy v with there is different spaces varianceAnd time varianceEach to Anisotropic Gaussian core convolution constructs linear-scale space representation

$R(\·;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)=g(.;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)*v(\·)---\left(9\right)$

The separable gaussian kernel of space-time is defined by formula (10):

$g(x,y,t;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)=\frac{\exp (-(x^2+y^2)/2{\mathrm{\σ}}_r^2-t^2/2{\mathrm{\τ}}_r^2)}{\sqrt{{\left(2\mathrm{\π}\right)}^3{\mathrm{\σ}}_r^4{\mathrm{\τ}}_r^2}}---\left(10\right)$

Time domain and spatial domain are used separately single scale parameter it is critical that, because event time Between and spatial component be in general independent.And, detect event with space-time interest point operators and depend on In the time observed and space scale, accordingly, it would be desirable to separately process corresponding scale parameterWith

Consider with 3 × 3 space-time second-order moments matrixes being made up of single order room and time yardstick, and with height This weighting function is averaging.

$g(.;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)\mathrm{\μ}=g(.;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)*\left(\begin{array}{ccc}{r_x}^2& r_x{r}_y& r_x{r}_t\\ r_x{r}_y& {r_y}^2& r_y{r}_t\\ r_x{r}_t& r_y{r}_t& {r_t}^2\end{array}\right)---\left(11\right)$

UtilizeWithBy the scale parameter in formula (9)WithIt is fused to local scale parameterWithIn, and define first derivative and be: the calculating speed of gradient in 3d space

$\left\{\begin{array}{c}{r}_x(x,y,t;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)=\∂x(g*v)\\ r_y(x,y,t;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)=\∂y(g*v)\\ r_t(x,y,t;{\mathrm{\σ}}_r^2,{\mathrm{\τ}}_r^2)=\∂t(g*v)\end{array}\right.---\left(12\right)$

In order to detect space-time interest points, search in μ in video v and there is marked feature value λ₁、λ₂And λ₃District Territory.In the distinct methods of region of search, determinant and tracking expanded definition in conjunction with μ are for spatial domain Harris Angle function be time-space domain Harris Angle function:

H=λ₁λ₂λ₃-k(λ₁+λ₂+λ₃)³,(λ₁≤λ₂≤λ₃) (13)

The present invention provides a kind of person of doing more physical exercises behavioral value side based on dimension self-adaption local space time feature Method, utilizes space-time interest points to carry out behavior knowledge to the person's of the doing more physical exercises behavior representing in section of football match video Not.The person's of doing more physical exercises behavior in section of football match video sequence is regarded as the space-time interest in three dimensions The set of point.Use rectangular histogram quantification technique that space-time interest points set is quantified as the Nogata that dimension is fixing Figure (instant null word), uses K-means clustering algorithm to generate space-time code book.At cluster generated code Before Ben, each space-time interest points is carried out normalization, to ensure its zooming and panning invariance. Test result indicate that, method in this paper can greatly reduce the amount of calculation of algorithm, with based on yardstick The person's of doing more physical exercises behavior representation in the section of football match video of adaptive local space-time characteristic is done more physical exercises The accuracy rate of identification can be significantly improved during member's Activity recognition.

Finally should be noted that: above example is only in order to illustrate technical scheme rather than right It limits, although the present invention being described in detail with reference to above-described embodiment, and art general The detailed description of the invention of the present invention still can be modified or equivalent by logical technical staff, this A bit without departing from any amendment or the equivalent of spirit and scope of the invention, the power all awaited the reply in application Within the scope of profit is claimed.

Claims (5)

1. the person's of doing more physical exercises behavioral value method based on dimension self-adaption local space time feature, its It is characterised by, said method comprising the steps of:

S1, uses normalized Laplace operand to estimate local scale, sets up dimension self-adaption Local space time's feature detection algorithm, by Harris space-time interest points detection operand and Laplace Operand combines and derives Harris-Laplace space-time interest points detection operand；

S2, is generalized to normalized Laplace operand in time-space domain；

S3, is generalized to local space time's Feature Descriptor in three-dimensional section of football match video image.

The person of doing more physical exercises based on dimension self-adaption local space time feature the most according to claim 1 Behavioral value method, it is characterised in that the detection of described S1 mesoscale adaptive local space-time characteristic is calculated Method is specially

$H={\mathrm{\λ}}_1^3(\mathrm{\α}\mathrm{\β}-k(1+\mathrm{\α}+\mathrm{\β}\left)\right)---\left(1\right)$

Wherein, the positive local maximum of H corresponds to λ₁、λ₂And λ₃High level point, λ₁<λ₂<λ₃, Ratio, α=λ₂/λ₁, β=λ₃/λ₁.From requiring that H >=0 can obtain k≤α β/(1+ alpha+beta)³。

The person of doing more physical exercises based on dimension self-adaption local space time feature the most according to claim 1 Behavioral value method, it is characterised in that local space time's Feature Descriptor is generalized to three-dimensional by described S3 Computational methods in section of football match video image include: the calculating peace that histogram calculation, direction quantify All gradient calculation.

The person of doing more physical exercises based on dimension self-adaption local space time feature the most according to claim 3 Behavioral value method, it is characterised in that calculating that described histogram calculation and direction quantify particularly as follows:

For a given cube c=(x_c,y_c,t_c,l_c,w_c,h_c), it is classified as S × S × S sub-block b_i； For each sub-block, utilize formula (4) to calculate its average gradient, then willIt is quantified as q_bi.For Each region c, the average gradient vector after being quantified by summation calculates its rectangular histogram h_c；

$h_c=\underset{i=1}{\overset{S^3}{\&Sigma;}}{q}_{bi}---\left(3\right)$

For a given positive n face body, its center is exactly the initial point of three-dimensional system of coordinate, 3D gradient vector Being quantified by the translation of coordinate points relative to the direction of initial point, each point in matrix p is polyhedron The center p of each₁,p₂,…,p_n, p=(p₁,p₂,…,p_n)^T, wherein, p_i=(x_i,y_i,t_i)^T, then in 20 Heart point is:

Wherein, golden section

?On projection can pass through formula (5) and calculate；

${\hat{q}}_b={\&lsqb;{\hat{q}}_{b1},{\hat{q}}_{b2},...,{\hat{q}}_{bn}\&rsqb;}^T=\frac{p{\stackrel{\&OverBar;}{g}}_b}{\left|\right|{\stackrel{\&OverBar;}{g}}_b|{|}_2}---\left(5\right),$

WillThresholding is processed asThen formula (6) is utilized to carry out quantification treatment；

$q_b=\frac{{\left|\right|g_b\left|\right|}_2{\hat{q}}_{\mathrm{bth}}}{{\left|\right|q_{\mathrm{bth}}\left|\right|}_2}---\left(6\right).$

The person of doing more physical exercises based on dimension self-adaption local space time feature the most according to claim 4 Behavioral value method, it is characterised in that described average gradient calculate particularly as follows:

By section of football match video sequence v, (x, y, t) along x, the partial derivative in y, t direction is divided into and being designated asThen Formula (7) can be used to calculate

$\left\{\begin{array}{c}{\mathrm{iv}}_{\&part;x}(x,y,t)=\underset{x^{\&prime;}\&le;x,y^{\&prime;}\&le;y,t^{\&prime;}\&le;t}{\mathrm{\&Sigma;}}{v}_{\&part;x}(x^{\&prime;},y^{\&prime;},t^{\&prime;})\\ {\mathrm{iv}}_{\&part;y}(x,y,t)=\underset{x^{\&prime;}\&le;x,y^{\&prime;}\&le;y,t^{\&prime;}\&le;t}{\mathrm{\&Sigma;}}{v}_{\&part;y}(x^{\&prime;},y^{\&prime;},t^{\&prime;})\\ {\mathrm{iv}}_{\&part;t}(x,y,t)=\underset{x^{\&prime;}\&le;x,y^{\&prime;}\&le;y,t^{\&prime;}\&le;t}{\mathrm{\&Sigma;}}{v}_{\&part;t}(x^{\&prime;},y^{\&prime;},t^{\&prime;})\end{array}\right.---\left(7\right)$

For a 3D cube b=(x, y, t, l, w, h), wherein, (x, y, t)^TAnd l, w, h represent cube respectively The position of body and length and width and height；

Formula (8) can be used quickly to calculate its average gradient vector'sPoint Amount；

$\begin{array}{c}{\stackrel{\&OverBar;}{g}}_{b\&part;x}=\&lsqb;{\mathrm{iv}}_{\&part;x}(x+w,y+h,t+l)-{\mathrm{iv}}_{\&part;x}\left(x,y+h,t+l\right)\\ -{\mathrm{iv}}_{\&part;x}\left(x+w,y,t+l\right)+{\mathrm{iv}}_{\&part;x}\left(x,y,t+l\right)\&rsqb;\\ -\&lsqb;{\mathrm{iv}}_{\&part;x}\left(x+w,y+h,t\right)-{\mathrm{iv}}_{\&part;x}\left(x,y+h,t\right)\\ -{\mathrm{iv}}_{\&part;x}\left(x+w,y,t\right)+{\mathrm{iv}}_{\&part;x}\left(x,y,t\right)\&rsqb;\end{array}---\left(8\right)$

In like manner, can quickly calculate