CN104102910A - Sports video tactical behavior recognition method based on space-time local mode - Google Patents

Abstract

The invention relates to a sports video tactical behavior recognition method based on a space-time local mode, and belongs to the technical field of video content understanding. A local space-time mode is used for realizing the tactical video behavior recognition of teamwork match in sports videos. The sports video tactical behavior recognition method is different from the existing method using track features for tactical behavior recognition. For video frame sequence images, a measure of improving a space-time local regression kernel to be used as a feature detector for detecting a movement significance region is provided; and the movement significance region is directly used as a feature word for building a visual feature word bag model and is used for tactical behavior recognition. The defect of influence on the multi-target track extraction due to factors such as complicated backgrounds in the multi-target track extraction process is overcome; the local space-time mode features and the space-time distribution of the local space-time mode features are used as tactical expression; the complexity of the recognition method is reduced; and meanwhile, the practicability of the method is improved. By aiming at large data volume of the videos, the detecting efficiency of a detecting operator is effectively improved.

Description

Sports video tactics behavior recognition methods based on space-time local mode

Technical field

The present invention relates to video content and understand technical field, be particularly related to the tactics behavior identification for sports video, espespecially a kind of sports video tactics behavior recognition methods based on space-time local mode.

Background technology

Follow the develop rapidly of video technique and the lifting of consumption demand, multimedia messages analysis, understanding and retrieval correlation technique based on semantic have wide development of demand space.As one of main multimedia video content, sports video is because it has huge commercial value, amusement function and huge audient colony, therefore be subject to extensive concern for detection, semantic understanding and the label technology excellent events such as shooting, shooting in sports video, and obtained a large amount of achievements in research.

In general, in sports video, the detection of excellent event and mark only can meet the amusement requirement of General Visitors, further go deep into understanding and excavating Tactical Mode and the strategy of game in match for professional spectators, sportsman and professional and technical personnel by video, thereby training plan and the tactics strategy of formulating targetedly one's own side, meaning is very micro-.Current sports speciality personnel are in order effectively to obtain tactical analysis result during the games, the modes that adopt manual analysis mark more, expend so a large amount of manpower and materials, be subject to again the restriction of personnel self professional knowledge simultaneously, therefore propose a kind of tactics identification and analysis that utilizes video its data automatically to realize to race in video both sides and have great importance for the development that meets sports video related industry future.

Summary of the invention

The object of the present invention is to provide a kind of sports video tactics behavior recognition methods based on space-time local mode, solved the problems referred to above that prior art exists.The present invention uses local space time's pattern to realize the tactics video behavior identification of team collaboration's match in sports video.Be different from the current method for distinguishing of knowing using track characteristic as tactics behavior.To sequence of frames of video image, propose to improve space-time local regression core and detect motion salient region as feature detection, be different from the current method of mainly utilizing light stream appraisal procedure to detect motion salient region.Directly use motion salient region to build visual signature word bag model as feature word, identify for tactics behavior.The present invention is different from current sport event recognition methods and utilizes the method for track as the statement of tactics behavior pattern feature, for sports video case, propose a kind of space-time local regression that improves and check motion salient region in video monitoring regional and detect and location, by motion salient region with participate in competing that higher both sides sportsman poised for battle is corresponding for liveness.The motion salient region obtaining for detection builds possesses the visual signature word band model that spatiotemporal mode distribution represents, realize the expression to specific tactics behavior in video, realize automatic analysis and the identification to tactics behavior in video in conjunction with clustering method, recognition result feeds back to user with brief description form.

Above-mentioned purpose of the present invention is achieved through the following technical solutions:

Sports video tactics behavior recognition methods based on space-time local mode, comprises the following steps:

S1: sports video input, choose excellent sport event video-frequency band, comprise goal event, attacking and defending event, as input

S2: Video segmentation, input sports video is implemented to equal time length and cut apart, obtain some isometric sub-video sections;

S3: tactics information extraction, for extracting the feature that represents tactics behavior, step is as follows: first, propose to improve the conspicuousness of two field picture pixel in space-time local regression core judgement sub-video section, build the motion salient region in sub-video section; Then, application court cut zone label identifies salient region, corresponding label salient region frequency of occurrence in statistics sub-video section, construction feature histogram; Finally, according to sequential sub-series video-frequency band feature histogram as tactics behavioural characteristic;

S4: Classification and Identification is carried out in the tactics behavior that S3 is extracted;

S5: feed back to the tactics behavior in user video event with brief description.

Described tactics information extraction, for extracting the feature that represents tactics behavior, specifically:

S31: identification is detected in region, place, by ball park line in video frame images, comprise goal line, sideline, offside line, forbidden zone camber line, middle astragal, detect and intersection point location, realize the mapping that scene is image from video frame images to true coordinate, realize the division in region, court and label identification is carried out in zoning in conjunction with place line simultaneously;

S32: video motion salient region detects, application enhancements space-time local regression is examined existing pixel motion conspicuousness judgement, builds motion Saliency maps and detects the motion salient region in video; The sub-detection mode of this feature detection is that the weighted linear of orthogonal space plane detected value separately merges as this aerial image vegetarian refreshments conspicuousness testing result value;

S33: salient region Feature Words bag model builds, in sub-video section, utilize region, the court identification label that S31 realizes to identify as the motion salient region correspondence of S32 extraction, utilize salient region frequency of occurrence to build the feature histogram in sub-video section, according to the feature histogram building in sub-video sequential sub-series video-frequency band, realize the character representation of sports video case tactics behavior.

The described sports video tactics behavior recognition methods based on space-time local mode is not only suitable for the tactics behavior identification in goal video, is also applicable to other tactics behavior pattern recognition in team collaboration's athletics event video.

In the present invention, court Region Segmentation adopts the corresponding true projected image of corresponding sub-video section the first two field picture, and this sub-video section subsequent frame image is all mapped on this true projected image.The local feature of tactics behavior is as the motion salient region in method and similar characteristics description.Feature detection is for improving space-time local regression core, for judging video frame images pixel conspicuousness.The sub-detection mode of this feature detection is that the weighted linear of orthogonal space plane detected value separately merges as this aerial image vegetarian refreshments conspicuousness testing result value.Use the motion salient region of sub-video section detection as feature word construction feature histogram, the normalization histogram of series connection is as the feature describing mode of tactics behavior chronologically.

Beneficial effect of the present invention is: technical solution of the present invention design philosophy is dependent on, in team collaboration match, the execution of tactics behavior is general higher with the higher team member's degree of correlation of liveness, and the higher sportsman of liveness is generally in the direct or indirect implementation that participates in tactics behavior; Otherwise the contribution degree that the sportsman that liveness is lower carries out for tactics is relatively little.Therefore, in the sports video of shooting with video-corder at fixed cameras, utilize the improvement space-time local regression core proposing as feature detection operator, build the motion salient region in video, in order to the higher sportsman of corresponding liveness, by statistics different time sections in the distribution of motion salient region in court, with realize tactics behavior represent and for identification.Innovative point of the present invention is, the feature of current description sports video tactics behavior is main mainly with sportsman's track, the impact that the factors such as the present invention overcomes in multi-target track leaching process due to complex background, blocks, object matching are extracted for multi-target track, utilize local space time's pattern feature and spatial and temporal distributions thereof to represent as tactics, reduce recognition methods complicacy, improve the practicality of method simultaneously.Meanwhile, space-time local mode feature detection operator---the space-time local regression core proposing in invention, for video big data quantity, effectively improves the detection efficiency that detects operator.

Brief description of the drawings

Accompanying drawing described herein is used to provide a further understanding of the present invention, forms the application's a part, and illustrative example of the present invention and explanation thereof are used for explaining the present invention, do not form inappropriate limitation of the present invention.

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is sports ground Region Segmentation mark schematic diagram of the present invention;

Fig. 3 is space-time local mode feature detection operator schematic diagram of the present invention;

Fig. 4 is that S6 space-time local mode operator of the present invention detects effect schematic diagram.

In figure: V1, sideline, court; V2, offside line; V3, vertical direction forbidden zone line and extension line H1 thereof, sideline, court; H2, horizontal direction forbidden zone line and extension line thereof; R1, middle astragal; S6: space-time local mode feature detection, a kind of space-time local regression core that improves, by concentric three normal surface XY, XT, uses 2-D local regression core conspicuousness testing result to merge as central value separately on YT; S7: the video frame images sequence that video-frequency band transforms.

Embodiment

Further illustrate detailed content of the present invention and embodiment thereof below in conjunction with accompanying drawing.

Referring to shown in Fig. 1 to Fig. 4, the sports video tactics behavior recognition methods based on space-time local mode of the present invention, comprises the steps:

S1: sports video input.Be divided into long distance, middle distance and low coverage video for the capture video of sports tournament, and long distance video not only occupies most of ratio of capture video, and this video can embody the information competition field from overall angle.Therefore the long distance capture video section (goal event, attacking and defending event) of, choosing excellent sport event is converted into video frame images as input.

F＝{f ₁,f ₂,…,f _i,…,f _N}

Wherein F represents the video-frequency band that the specific sport time is corresponding, f _i, i=1,2,3 ... N represents i frame video image, and N represents that input video section changes into the frame number of video frame images.

S2: Video segmentation, input sports video is implemented to equal time length and cut apart, obtain some isometric sub-video sections, if wherein the time span of sub-video section is shorter, more careful to follow-up tactics behavior identification statement.The sub-video section that this step obtains provides the elementary cell of tactics information extraction for S3.

F＝{F ₁,F ₂,F ₃,…,F _M}

F _j={ f _j1, f _j2..., f _jq..., f _jmand jm=pm, j ≠ p, j, p=1,2 ... M, q=1,2 ..., m

Wherein F _j, j=1,2 ... M represents j sub-video-frequency band after Video segmentation, f _jqrepresent j the q two field picture in sub-video-frequency band, M is cut apart sub-video hop count amount.

S3: tactics information extraction, for extracting the feature that represents tactics behavior, this step can be subdivided into three and carry out sub-step, in turn for identification is detected in region, place, motion salient region detects and the statement of tactics behavioural characteristic.

S31: identification is detected in region, place.

Sportsman in match, event all occur in the regional extent of court, therefore for the detection and the processing of identification for follow-up S32 and S33 in ball park region, comprise that tactics behavior is identified in interior numerous Video Events analyses particularly necessary with processing.First be white feature for the numerous tag lines in inside, court (goal line, sideline, forbidden zone camber line and middle astragal), and turf is pruned and is produced the offside line that visual color forms through different direction, use the line in digital image processing techniques to detect above-mentioned tag line and extension line thereof, the orientation and segmentation of feasible region with circle detection algorithm.Then utilize the line and the line intersection point calculation camera calibration matrix that detect, thereby realize the conversion that is tied to real space coordinate system from the video image coordinate of haul distance collected by camera, demarcate for S32 motion salient region provides real space coordinate system.Divide and more carefully will represent for space-time local mode space distribution more specifically, higher for Tactical Mode discrimination for region, court, identify more accurate.

Shown in Figure 2, divide and be identified as example with the region of left side half-court, use line and circle detection algorithm in digital image processing techniques to determine the tag line in region, court and extension line thereof in video image, calculate the intersecting point coordinate of vertical direction tag line and horizontal direction marking line, calculate the calibration matrix of video image coordinate system and real space coordinate system by camera calibration method, video image is mapped in real space coordinate system, and mapping mode adopts sub-video F _jtwo field picture f _jq, q=2,3 ..., m is mapped to this sub-video the first frame f _j1on corresponding real space coordinate system figure, and realize the division to this region, true picture court.Identify for the cut zone utilization numeral after dividing, as shown in Figure 2.Utilize successively { 1,2,3,4} and { 13,14,15,16} difference corresponding the right road and left side road, { 5,6,7,8} and { 9,10,11, the 12} corresponding right road of difference and left road.Region result after mark positions and identifies for the motion salient region that S32 is detected.

S32: video motion salient region detects.

Application enhancements space-time local regression core detects operator as motion significant characteristics, builds motion Saliency maps, and detection and location campaign salient region, for S33 uses.Specifically be implemented as follows:

S321: discuss in detail emphatically and improve space-time local regression core feature detection, as shown in Figure 3, the present invention puts basic thought, the time empty sequence forming for effective detection video frame images, by XY representation space dimension, with XT, YT represents time dimension, utilize 2-D local regression core locating with one heart XY, XT, YT is the significance probability value of computing center's point on coordinate surface separately, thereby then merges calculated value separately and obtain the conspicuousness value at this of place event, thereby the space-time local motion significant characteristics of realizing on sequence of frames of video detects.Specific implementation step is as follows:

1. determine the conspicuousness discriminant function of video image vegetarian refreshments.Concrete function expression is as follows:

${\mathrm{Pix}}_{\mathrm{Val}}\left(i\right)=\left\{\begin{array}{cc}1& {\mathrm{Pix}}_{\mathrm{Loc}}\left(i\right):P({\mathrm{Pix}}_{\mathrm{Val}}\left(i\right)=1|F)\≥\mathrm{th}\\ 0& \mathrm{other}\end{array}\right.$

Wherein Pix _loc(i)=(x, y, t _j), i=1,2 ... I, j=1,2 ... m represents video frame images pixel coordinate, Pix _val(i) i pixel is conspicuousness point assignment 1, otherwise assignment 0.

Judgment basis is by P (Pix _val(i)=1|F)>=th determine, calculate under feature F, Pix _val(i) whether=1 probable value is greater than a certain predefined thresholding th.It is as follows that probable value is calculated expression:

P(Pix _Val(i)＝1|F)＝P(Pix _Val(i)＝1|Fusion(F _XY,F _XT,F _YT))

＝P(Pix _Val(i)＝1|(a ₁F _XY+a ₂max(F _XT,F _YT)))

＝a ₁P(Pix _Val(i)＝1|F _XY)+a ₂max(P(Pix _Val(i)＝1|F _XT),

P(Pix _Val(i)＝1|F _YT)))

＝a ₁max(P(Pix _Val(i)＝1|F _R),P(Pix _Val(i)＝1|F _G),

P(Pix _Val(i)＝1|F _R))+a ₂max(P(Pix _Val(i)＝1|F _G),

P(Pix _Val(i)＝1|F _B)))

s.tF＝[F ₁,F ₂,…,F _W]

F _XY＝max(F _R,F _G,F _B),F _XT＝F _YT＝F _gray,

a ₁+a ₂＝1

Wherein Fusion (F _xY, F _xT, F _yT) represent with Pix _loc(i)=(x, y, t _j) be three mutually orthogonal XY of the heart, XT, the upper Fusion Features function of YT, F is illustrated in the eigenvectors matrix extracting in neighborhood window, and other F marks represent same meaning, further limit its usable range after adding subscript, and W represents neighborhood window size.F _xY, F _xT, F _yTbe illustrated in the feature of extracting separately on three normal surfaces, F _xYfusion Features value representation on the corresponding R/G/B tri-look channels of video frame images on XY coordinate surface for feature, a ₁, a ₂for weights.

In the present invention on XY coordinate surface video frame images feature extraction to adopt Feature fusion on R/G/B tri-look channels be effectively to distinguish in foreground target sportsman in background court, and at XT, on YT coordinate surface, only extracting frame of video along time shaft T gray scale channel characteristics, is to consider that gray-value variation can effectively represent that pixel describes the variation of regional movement conspicuousness on time shaft T

2. solving condition probability general formula

In the present invention, concentrate on for the solving of conditional probability for asking for of the conspicuousness discriminant function of above-mentioned proposition, solution procedure brief description is as follows:

(1) Bayes' theorem framework solving condition probability

$S=P({\mathrm{Pix}}_{\mathrm{Val}}\left(i\right)=1|F)=\frac{p({\mathrm{Pix}}_{\mathrm{Val}}\left(i\right)=1)p\left(F\right|{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right)=1)}{p\left(F\right)},$

Initial p (Pix _val(i)=1), the prior distribution of p (F) is for being uniformly distributed constant,

S=P (Pix _val(i)=1|F) ∝ p (F|Pix _val(i)=1), be further reduced to

S=P (Pix _val(i)=1|F)=β p (F|Pix _val(i)=1), for constant.

(2) p (F|Pix _val(i)=1) estimate

1) feature F calculates and represents

With at XT coordinate surface to Pix _val(i) estimate to describe, other each color spaces and coordinate surface method of estimation are similar.Conditional probability is estimated to adopt norm of nonparametric kernel density method of estimation, and kernel function is used 2-D local regression core,

$\begin{array}{c}K({\mathrm{Pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))\\ =\frac{\sqrt{\det \left(C_l\right)}}{h^2}\exp \left(\frac{{({\mathrm{Pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))}^T{C}_l({\mathrm{Pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))}{-2h^2}\right)\end{array}$

Wherein Pix _val(l) be with Pix on XT coordinate surface _val(i) pixel in center neighborhood window, C _lfor pixel value covariance matrix in neighborhood, h suppresses the smoothing parameter of noise.

Core value after normalization is as feature F, and normalization expression formula is as follows:

$F_i:W_i({\mathrm{pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))=\frac{K({\mathrm{Pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))}{\underset{l}{\mathrm{\Σ}}K({\mathrm{Pix}}_{\mathrm{Val}}\left(l\right)-{\mathrm{Pix}}_{\mathrm{Val}}\left(i\right))}$

2) p (F|Pix _val(i)=1) estimate

Calculation expression is as follows: $\hat{p}\left(F\right|{\mathrm{pix}}_{\mathrm{Val}}\left(i\right)=1)=\frac{G_i({\stackrel{\‾}{F}}_i-{\stackrel{\‾}{F}}_j)}{\underset{j}{\mathrm{\Σ}}{G}_i({\stackrel{\‾}{F}}_i-{\stackrel{\‾}{F}}_j)}$

Wherein $G_i({\stackrel{\‾}{F}}_i-{\stackrel{\‾}{F}}_j)=\exp \left(\frac{-1+\mathrm{\ρ}({\stackrel{\‾}{F}}_i,{\stackrel{\‾}{F}}_j)}{{\mathrm{\σ}}^2}\right),$ for normalization proper vector,

$\mathrm{\ρ}({\stackrel{\‾}{F}}_i,{\stackrel{\‾}{F}}_j)=\mathrm{trace}\left(\frac{{F_i}^T{F}_j}{{\left|\right|F_i\left|\right|}_F{\left|\right|F_j\left|\right|}_F}\right)\∈[-\mathrm{1,1}],$ || || _ffor Schmidt's norm.

S322. build Saliency maps, detect motion salient region.

The conspicuousness point set being distributed in video-frequency band of being determined by S321 forms Saliency maps, thereby realizes the detection of salient region in video.Motion salient region detects schematic diagram and illustrates result as accompanying drawing 4.

S33: salient region Feature Words bag model builds, in sub-video section, utilize region, the court identification label that S31 realizes to identify as the motion salient region correspondence of S32 extraction, build the feature histogram in sub-video section, according to the feature histogram building in sub-video sequential sub-series video-frequency band, realize the character representation of sports video case tactics behavior.Concrete execution step is as follows:

1, motion salient region is demarcated.Demarcation mode utilize this motion salient region be distributed in court cut zone label under true coordinate system 1,2 ..., 16} marks it, i.e. S _j(i)=l, l ∈ 1,2 ..., 16}, wherein S _j(i) represent sub-video section F _jon i motion salient region, and this areal distribution is on cut zone l, so it is carried out to label with label l;

2, step 1 is detected to calibrated motion salient region and carry out color and vein classification, for distinguishing the both sides sportsman of host and guest team on court, a certain sportsman on a corresponding court of salient region;

3, the motion salient region to the sportsman of host and guest team after the mark obtaining in step 1 and 2, statistics sub-video section F _jsalient region number separately in different cut zone, and calculate a certain region significance region frequency of occurrence in this sub-video section build sub-video section F _jon salient region Feature Words bag histogram H _j;

4, press histogram H in sub-video section sequential series connection step 3 _j, obtain the character representation { H of tactics behavior ₁, H ₂... H _j... H _m.

Return to tactics behavioural characteristic { H ₁, H ₂... H _j... H _m.

S4: Classification and Identification is carried out in the tactics behavior that S3 is extracted

For given tactics behavioural characteristic, classifying identification method is divided into two steps: coarseness identification and fine granularity identification.

Concrete execution step is as follows:

1. coarseness identification.

(1) attack pattern description.

Taking football attack and score the goal event as analyze sample, other team collaboration's class sport event analytical approachs are similar.Take into account that the Tactical Mode of breaking through door is divided into four large classes, that is: mid-way attacking, flank attack, coordinated attack and place kick attack (the direct attack of mid-front field corner-kick) substantially.

Local space time's pattern that mid-way attacking is corresponding presents { 5,6,7,8} and { on 9,10,11,12} region, entirety is definitely dominant in spatial and temporal distributions.

Local space time's pattern corresponding to flank attack presents { 1,2,3,4} and { on 13,14,15,16} region, entirety is definitely dominant in spatial and temporal distributions.

Local space time's pattern corresponding to coordinated attack is different from mid-way attacking and flank attack in spatial and temporal distributions in above-mentioned 4 regions.

Local space time's pattern corresponding to place kick attack presents in spatial and temporal distributions that { on 7,8,9,10} region, entirety is definitely dominant.

(2) Classification and Identification.

The tactics behavioural characteristic of utilizing S1 to S3 step to extract for given initial typical training video sample, be characterized by histogram as corresponding label location motion salient region frequency of occurrence is higher, and forming the tactics behavior classification under coarseness by clustering algorithm, similarity measurement standard can be used current histogram measuring similarity standard (Euclidean distance tolerance, the side's of card matching measurement, Manhattan tolerance etc.).Then, utilize professional to carry out manual review calibration, and be finally that above-mentioned initial training video sample pastes note class label, use successively number designation { 1,2,3,4} represents above-mentioned four class tactics behaviors, 1: mid-way attacking, 2: flank attack, 3: coordinated attack, 4: place kick attack.

2. fine granularity identification

Adopt the method in coarseness identification (2) to carry out separately further clustering to the training sample that distributes class label, manually mark for division result, be respectively A: breakaway attack, B: short pass infiltration.

S5: feed back to the tactics behavior in user video event with brief description

Return string and word annotation,

1A: Road breakaway attack; 1B: Road short pass infiltration

2A: wing breakaway attack; 2B: wing short pass infiltration

3A: collaborative breakaway attack; 3B: collaborative short pass infiltration

4: place kick attack

The present invention realizes for the method for goal tactics behavior and is equally applicable to other tactics behavior identifications in football match, attack, defence, and the attacking and defending of regional area coordinates.The inventive method is equally applicable to the tactics behavior identification of team collaboration's class sport event.

The foregoing is only preferred embodiment of the present invention, be not limited to the present invention, for a person skilled in the art, the present invention can have various modifications and variations.All any amendments made for the present invention, be equal to replacement, improvement etc., within protection scope of the present invention all should be included in.

Claims (3)

1. the sports video tactics behavior recognition methods based on space-time local mode, is characterized in that: comprise the following steps:

S1: sports video input, choose excellent sport event video-frequency band, comprise goal event, attacking and defending event, as input

S2: Video segmentation, input sports video is implemented to equal time length and cut apart, obtain some isometric sub-video sections;

S3: tactics information extraction, for extracting the feature that represents tactics behavior, step is as follows: first, propose to improve the conspicuousness of two field picture pixel in space-time local regression core judgement sub-video section, build the motion salient region in sub-video section; Then, application court cut zone label identifies salient region, corresponding label salient region frequency of occurrence in statistics sub-video section, construction feature histogram; Finally, according to sequential sub-series video-frequency band feature histogram as tactics behavioural characteristic;

S4: Classification and Identification is carried out in the tactics behavior that S3 is extracted;

S5: feed back to the tactics behavior in user video event with brief description.

2. the sports video tactics behavior recognition methods based on space-time local mode according to claim 1, is characterized in that: described tactics information extraction, for extracting the feature that represents tactics behavior, specifically:

S31: identification is detected in region, place, by ball park line in video frame images, comprise goal line, sideline, offside line, forbidden zone camber line, middle astragal, detect and intersection point location, realize the mapping that scene is image from video frame images to true coordinate, realize the division in region, court and label identification is carried out in zoning in conjunction with place line simultaneously;

S32: video motion salient region detects, application enhancements space-time local regression is examined existing pixel motion conspicuousness judgement, builds motion Saliency maps and detects the motion salient region in video; The sub-detection mode of this feature detection is that the weighted linear of orthogonal space plane detected value separately merges as this aerial image vegetarian refreshments conspicuousness testing result value;

S33: salient region Feature Words bag model builds, in sub-video section, utilize region, the court identification label that S31 realizes to identify as the motion salient region correspondence of S32 extraction, utilize salient region frequency of occurrence to build the feature histogram in sub-video section, according to the feature histogram building in sub-video sequential sub-series video-frequency band, realize the character representation of sports video case tactics behavior.

3. the sports video tactics behavior recognition methods based on space-time local mode according to claim 1 and 2, it is characterized in that: the described sports video tactics behavior recognition methods based on space-time local mode is not only suitable for the tactics behavior identification in goal video, is also applicable to other tactics behavior pattern recognition in team collaboration's athletics event video.