CN101127866A - A method for detecting wonderful section of football match video - Google Patents

Abstract

The utility model provides a method to detect wonderful clips of the soccer video, aiming at the incapacity of the prior method in judging the type of the wonderful clips in soccer sport competitions. The method of the utility model comprises the following steps: first a compressed video of the soccer competitions in the data base is decoded via a decoder; then video shot boundary is detected; feature extraction is conducted to other shots in the shot types, and the soccer video shots are classified based on the corresponding feature extraction; then video shot clips boundary is detected according to the result of the shot classification; observation vector of each clip is extracted, and the type of the clip is judged according to the observation vector extracted from the clip; the confirmed wonderful clips are classified; finally the classified wonderful clips are excerpted and stored in the data base.

Description

A kind of method that detects wonderful section of football match video

Technical field

The present invention relates to the method that detects wonderful in the sports tournament video to a kind of, particularly a kind of method that detects wonderful in the section of football match video.

Background technology

Football is a kind of motion that people like, the football video program is one of glad program of appreciating of people.The HD digital TV occurs, the foundation of high-speed video transmission network, for the propagation of football video program provides important means.Every day, different channel all had the interior at one time broadcast of a large amount of football videos.For the football fan, it can miss any football match program.Therefore football match is provided the summary and the navigation scheme of wonderful, can save a large amount of valuable time of people, and have no lack of understanding again the football game situation.Just need carry out detection method in order to reach this purpose to the section of football match video content based on wonderful.

The method that detects wonderful in a kind of sports tournament video is disclosed among the Chinese patent ZL02156973.8.The inventor uses the movable information of video camera in this patent, judges whether to take place at a slow speed playback camera lens and carries out that the sports tournament wonderful judges, and think that replaying is from same video camera.Yet specific to football match, the foul in football video, shooting, place kick and goal all need to provide corresponding replay camera lens.This method that detects according to the wonderful of whether replaying just can not be classified to wonderful, thereby can not judge the type that fragment is affiliated.Because different user is to the degree of concern difference of football wonderful, this method can't provide browsing of specific wonderful in a kind of needed football match to the user.

Summary of the invention

The present invention has wonderful classification and the judgement that sports tournament video highlight fragment detection method is not suitable for the football match camera lens now in order to solve, and has proposed a kind of method that detects wonderful in the section of football match video.

For reaching above purpose, the present invention adopts following technical scheme to be achieved:

A kind of method that detects wonderful section of football match video is characterized in that, comprises the steps:

Step 1 is decoded by decoder to the compressed video of football match in the database earlier;

Step 2 is carried out video shot boundary and is detected, and is used for detecting the border of video camera lens, and is divided into replay sign, other camera lens, replay camera lens three parts according to the type of camera lens;

Step 3 is carried out feature extraction to other camera lens in the lens type, is used for compression domain and pixel domain are carried out feature selecting and signature analysis, extracts feature and comprises: main color characteristic, camera motion feature, textural characteristics and captions feature; During main color characteristic extracted, wherein each frame is calculated two features: the green area in the main color connected region accounted for the ratio G of whole image _RRatio G with the shared whole image of motion subject area in the main color connected region _PAdopt in the camera motion feature based on global motion vector field GMVF and carry out the judgement of camera motion pattern; In texture feature extraction, the texture strength information of extracting in each frame is Tx;

Step 4 comes the section of football match video camera lens is classified according to main color, camera motion, texture and captions feature that step 3 is extracted;

Step 5 is carried out the video lens segment boundaries according to the section of football match video shot classification result of step 4 and is detected, to determine the border between the fragment in the football video; And extract measurement vector in each fragment;

Step 6 is carried out clip types and judge that the measurement vector that is extracted is judged the type under this fragment, according to the sorting technique based on hidden Markov model HMM clip types from a fragment.This fragment is divided into a kind of in following 5 types, is respectively: shooting fragment, foul fragment, goal fragment, place kick fragment and common fragment; Wherein, selecting shooting fragment, goal fragment, place kick fragment is wonderful;

Step 7, the wonderful of determining is divided, be about to detected wonderful in the step 6 and be divided in two teams in the football match according to the direction of primary motion of video camera, with obtain two teams on the wonderful number contrast and the goal type judged;

Step 8 is taken passages sorted wonderful and store in the database.

In the such scheme, in the described step 3, in extracting the camera motion feature, described employing is carried out the concrete grammar that the camera motion pattern judges based on global motion vector field GMVF and is, according to the correlated characteristic of describing the global motion of video camera in formula (3)-formula (6):

${\mathrm{\θ}}_{\mathrm{var}}=\frac{1}{N\×M}\underset{\mathrm{xi}=1}{\overset{N}{\mathrm{\Σ}}}\underset{\mathrm{yi}=1}{\overset{M}{\mathrm{\Σ}}}{|{\mathrm{\θ}}_{(\mathrm{xi},\mathrm{yi})}-\stackrel{\‾}{\mathrm{\θ}}|}^2---\left(3\right)$

$\left\{\begin{array}{c}\mathrm{GMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVxi}\\ \mathrm{GMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVyi}\end{array}\right.---\left(4\right)$

$\left\{\begin{array}{c}\mathrm{AGMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVxi}\right|\\ \mathrm{AGMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVyi}\right|\end{array}\right.---\left(5\right)$

$\left\{\begin{array}{c}\mathrm{GMVDx}=\mathrm{var}\left(\right|\mathrm{GMVxi}\left|\right)\\ \mathrm{GMVDy}=\mathrm{var}\left(\right|\mathrm{GMVyi}\left|\right)\end{array}\right.---\left(6\right)$

Wherein, θ _{(xi, yi)},

And θ _VarRepresent motion vector (GMVxi, GMVyi) pairing angle value, the angle variance of motion vector in pairing evaluation angle of all motion vectors and the entire image in the global motion vector field.(GMVx, GMVy), (AGMVx is AGMVy) with (GMVDx GMVDy) represents the mean value of average motion vector, motion vector magnitude and the variance vector of corresponding motion vector magnitude respectively.

In the described step 4, the video lens classification step has been used two graders: earlier camera lens is carried out rough sort by the rough sort device, and then by the disaggregated classification device full shot in the rough sort camera lens is carried out disaggregated classification; Wherein, the camera lens rough sort is color and vein characteristic vector＜G that step 3 is extracted _R, G _P, Tx〉and be divided into close-up shot, grandstand camera lens, full shot and medium camera lens; The camera lens disaggregated classification is the vector theta of being extracted in the formula (6) according to formula (3) _Var, GMVx, GMVy, AGMVx, AGMVy, the global motion characteristic vector of one 7 dimension that GMVDx and GMVDy constituted is divided into Zoom lens, tracking lens and still frame with the full shot in the camera lens rough sort once more.

In the described step 5, the method that described segment boundaries detects is to judge the full shot of fragment and the border of non-full shot, and the original position of a fragment and final position all are the switching position of non-full shot to full shot.

In step 6, described sorting technique based on hidden Markov model HMM clip types is, with a fragment according to its corresponding observation vector O=O ₁... O _n, O _N+1, O _N+2, O _N+3HMM model parameter λ with corresponding shooting, foul, goal, place kick and common fragment _i=(A _i, B _i, π _i) judge that this fragment belongs to above-mentioned 5 types any, the method for judgement is to calculate measurement vector in the fragment at 5 above-mentioned HMM model parameter λ _i=(A _i, B _i, π _i) probability P (O| λ _i) and judge wherein the pairing classification S of maximum according to formula (7):

$S=\arg \underset{i}{\max }P\left(O\right|{\mathrm{\λ}}_i)---\left(7\right)$

This fragment is divided in pairing that type of maximum then.

In step 7, described goal type comprises four classes: free kick is scored, and corner-kick is scored, penalty stroke goal, common goal.

The present invention's advantage compared to existing technology is: according to main color, texture, camera motion and captions feature the section of football match video camera lens is classified, the measurement vector that carries out on the basis of camera lens component in segment boundaries detection and the fragment extracts; Carry out clip types on this basis and judge, use method to be divided into shooting fragment, foul fragment, goal fragment, place kick fragment and common fragment a fragment based on hidden Markov model; And to the wonderful in the section of football match video: shooting fragment, goal fragment, place kick fragment are carried out wonderful and are divided, score the wonderful in the football match is divided in two teams in the match and institute's goal type is divided into free kick according to its goal mode, corner-kick is scored, penalty stroke goal, common goal.We not only can detect a wonderful in the section of football match video and can also determine the team under the wonderful and the type of goal like this.

Description of drawings

Fig. 1 is the flow chart of steps of the inventive method.

The idiographic flow of Fig. 2 in the step 4 among Fig. 1 video lens being classified.

Fig. 3 carries out football video segment boundaries location schematic diagram for step 5 among Fig. 1.

Measurement vector extracts schematic diagram to Fig. 4 in the football video fragment for step 5 among Fig. 1 is carried out.

Fig. 5 is for browsing main window to wonderful section of football match video in the step 8 among Fig. 1, wherein W0 is for browsing football video File Open button, W1 is a wonderful content-browsing window, and W2 is the anchor window of wonderful content, and W3 and W4 carry out the button that browsing content is selected.

Fig. 6 browses window for the section of football match video splendid contents that click W3 button among Fig. 5 enters.

Fig. 7 is for clicking the section of football match video goal content-browsing window that the W4 button enters among Fig. 5.

Embodiment

The present invention is described in further detail below in conjunction with drawings and Examples.

Fig. 1 has provided the step structured flowchart that among the present invention the wonderful in the section of football match video is detected.Comprise step 1, earlier the compressed video of football match in the database is decoded by decoder; Execution in step 2 is carried out the video shot boundary detection then, is used for detecting the border of video camera lens, and is divided into replay sign, other camera lens, replay camera lens three parts according to the type of camera lens; Execution in step 3 is carried out feature extraction to other camera lens in the lens type, is used for compression domain and pixel domain are carried out feature selecting and signature analysis, mainly extracts main color characteristic, camera motion feature, textural characteristics and captions feature; Execution in step 4 again, come the section of football match video camera lens is classified according to main color, camera motion, texture and captions feature that step 3 is extracted; Then execution in step 5 detects the football video highlight segment boundaries, to determine wonderful and non-wonderful and the border between them in the football video; And extract measurement vector in each fragment; Execution in step 6 then, carrying out clip types judges, the measurement vector that is extracted from a fragment is judged a type under the fragment, be divided into a kind of in following 5 types according to this fragment of big young pathbreaker of likelihood score, these 5 types correspond to respectively: shooting fragment, foul fragment, goal fragment, place kick fragment and common fragment; Execution in step 7 then, and the wonderful of determining is divided, and are about to detected wonderful in the step 6: shooting, goal and place kick fragment are divided in two teams in the football match, to obtain two contrasts of team on the wonderful number; Last execution in step 8 is taken passages sorted wonderful and store in the database.

In step 2, carrying out shot boundary detects, can adopt document H.Pan, B.Li, and M.Sezan, " Automatic detection of replay segments in broadcast sports programsby detecting of logos in scene transitions; " in Proc.IEEE ICASSP, Orlando, FL, May 2002, vol.4, method described in the pp.3385-3388: adopt method to detect the replay camera lens based on hidden Markov model, and the detection of the frame around replay camera lens replay sign image, just section of football match video can being divided into replays indicates and replay accordingly camera lens and other camera lens.

In step 3, when other camera lenses are carried out feature extraction, select for use and extract in the main color characteristic, we calculate two features to each frame wherein: main color (green in the picture) zone accounts for the ratio G of whole image _RRatio G with the shared whole image of motion subject area in the main color connected region _P, concrete computational methods are known in the art.

In extracting the camera motion feature, carry out overall motion estimation and extract corresponding feature to utilize its disaggregated classification in carry out step 4.The method of estimation of global motion can adopt X.Qian, and G.Liu, " Global motion estimation from randomly selected motion vector groupsand GM/LM based applications; " method described in the Signal, Image and Video Processing is estimated global motion parameter m=[m ₀, m ₁..., m ₅].After we estimated the global motion parameter, for the influence that the MVF that eliminates in the compression domain judges for the camera motion model, we adopted global motion vector field (GMVF) to carry out the judgement of camera motion pattern here.Among the GMVF at coordinate (x _i, y _i) (GMVxi, computational methods GMVyi) as shown in Equation (1) to locate corresponding global motion vector

$\left\{\begin{array}{c}{\mathrm{GMVx}}_i=x_i^{\′}-x_i\\ {\mathrm{GMVy}}_i=y_i^{\′}-y_i\end{array}\right.---\left(1\right)$

(x wherein _i, y _i) and (x _i', y _i') represent respectively in present frame and the reference frame through global motion pairing position

$\left\{\begin{array}{c}{x}_i^{\′}=m_0{x}_i+m_1{y}_i+m_2\\ y_i^{\′}=m_3{x}_i+m_4{y}_i+m_5\end{array}\right.---\left(2\right)$

We adopt correlated characteristic in formula (3)-formula (6) to describe the global motion pattern of video camera on the basis of GMVF:

${\mathrm{\θ}}_{\mathrm{var}}=\frac{1}{N\×M}\underset{\mathrm{xi}=1}{\overset{N}{\mathrm{\Σ}}}\underset{\mathrm{yi}=1}{\overset{M}{\mathrm{\Σ}}}{|{\mathrm{\θ}}_{(\mathrm{xi},\mathrm{yi})}-\stackrel{\‾}{\mathrm{\θ}}|}^2---\left(3\right)$

$\left\{\begin{array}{c}\mathrm{GMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVxi}\\ \mathrm{GMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVyi}\end{array}\right.---\left(4\right)$

$\left\{\begin{array}{c}\mathrm{AGMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVxi}\right|\\ \mathrm{AGMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVyi}\right|\end{array}\right.---\left(5\right)$

$\left\{\begin{array}{c}\mathrm{GMVDx}=\mathrm{var}\left(\right|\mathrm{GMVxi}\left|\right)\\ \mathrm{GMVDy}=\mathrm{var}\left(\right|\mathrm{GMVyi}\left|\right)\end{array}\right.---\left(6\right)$

Wherein, θ _{(xi, yi)},

And θ _VarRepresent motion vector (GMVxi, GMVyi) pairing angle value, the angle variance of motion vector in pairing evaluation angle of all motion vectors and the entire image in the global motion vector field.(GMVx, GMVy), (AGMVx is AGMVy) with (GMVDx GMVDy) represents the mean value of average motion vector, motion vector magnitude and the variance vector of corresponding motion vector magnitude respectively.

In texture feature extraction, we obtain the texture strength information Tx in each frame in other camera lens 12, are known about the computational methods of texture information in this area.

Be characterised in that in the captions feature extraction and judge having or not and the appearance frame and the disappearance frame of such captions of the short-term captions that manually add in the section of football match video.The short-term captions of the manual interpolation here refer in football match foul, goal etc. given captions in the rear videos take place.Such captions detect and the definite of frame and disappearance frame occur and can adopt X.Qian, G.Liu, H.Wang and R.Su, " Text Detection; Localization and Tracking in Compressed Videos ", method described in the Signal Processing:Image Communication: use the texture that the AC coefficient is represented in each piece of the I frame in the compressed video to carry out the detection and the location of captions, and adopt caption area DC picture differences value in every frame to judge the initial sum abort frame of each captions.

Fig. 2 carries out the substep of video lens classification step 4 for the feature according to extracting among Fig. 1, its characteristics are to have used two graders: earlier to the camera lens 4-1 that goes out to classify, and then by the disaggregated classification device full shot in the rough sort camera lens is carried out disaggregated classification 4-2 by the rough sort device.Wherein, the camera lens rough sort is according to the color and vein characteristic vector＜G that is extracted among Fig. 1 with other camera lens _R, G _P, Tx〉and be divided into the close-up shot shown in Fig. 2, grandstand camera lens, full shot and medium camera lens; Camera lens disaggregated classification 4-2 be according to formula (3) extracted in the formula (6) by θ _Var, GMVx, GMVy, AGMVx, AGMVy, the global motion characteristic vector of one 7 dimension that GMVDx and GMVDy constituted is divided into Zoom lens, tracking lens and still frame with the full shot in the camera lens rough sort once more.On this basis, in conjunction with the amplitude AGMVx and the direction character of camera motion

Once more top division result is segmented.For example we can be divided into enlarging lens and dwindle camera lens for Zoom lens; Be divided into general still frame and be still in the goal camera lens for still frame; Can be divided into tracking lens and tracking lens at a slow speed fast for tracking lens according to amplitude.Can also be again can be divided into from left to right quick tracking, right quick tracking, the left-to-right tracking at a slow speed and right tracking at a slow speed to a left side to a left side in conjunction with the direction of motion of video camera.

The rough sort device among Fig. 2 and the feature of disaggregated classification device all are to adopt SVMs (SVM), concrete being applied in this area is known (C.J.C.Burges, " A tutorial on supportvector machines for pattern recognition " Data.Mining and KnowledgeDiscovery.No.2, vol.2,1998, pp.121-167.).The rough sort device is in some close-up shots, grandstand camera lens, full shot and the medium camera lens of manually choosing from other camera lens in other camera lens and extraction＜G _R, G _P, Tx〉and to carry out model parameter study.The disaggregated classification device is some Zoom lens, tracking lens and the still frame of manually choosing from full shot in other camera lens, and extracts by θ _Var, GMVx, GMVy, AGMVx, AGMVy, the global motion characteristic vector of one 7 dimension that GMVDx and GMVDy constituted is to carry out model parameter study.

Carrying out above-mentioned step 4 by thick shot classification to essence finally can be divided into whole video sequence: grandstand, close-up shot, medium camera lens, replay and indicate, the replay camera lens amplifies, and dwindles, quick tracking from left to right, quick tracking from right to left, tracking at a slow speed from left to right, tracking at a slow speed from right to left, general static and be still in 13 kinds such as goal.

Carry out segment boundaries during the segment boundaries that shows Fig. 3 in the step 5 detects and determine method.Its objective is for the camera lens of dividing in the video lens classification step 4 and be referred in the different fragments.The method that segment boundaries detects is to judge the border of full shot and non-full shot.The original position of a fragment and final position all are the switching position of non-full shot to full shot.

Fig. 4 shows the schematic diagram that how to extract measurement vector in the football video fragment in through a fragment in the Boundary Detection.It is characterized in that we extract the pairing measurement vector of this fragment according to (overall) two aspects on time last (temporal) and the general overview for each fragment.Suppose that a fragment is made of n camera lens, a kind of corresponding in above-mentioned 13 kinds of each camera lens wherein, then measurement vector is: O=O ₁..., O _n, O _N+1..., O _N+kDimension be n+k.N measured value also is the temporal part corresponding to the camera lens of this fragment before among the observation sequence O, and each measured value wherein is a kind of in the above-mentioned classification.Back k measured value O _N+1..., O _N+kIt is general overview to this fragment.Can merge the feature of variform here, k=3 among the present invention.O wherein _N+1For judging whether the global portions in current fragment and the next fragment short-term captions occur, the output of this measured value is to judge according to the detection that captions extract to draw, wherein O _N+2Whether expression exists long pause, if in the fragment number of non-full shot greater than 5 we think to exist for a long time in this fragment and pause, otherwise think and do not exist.O wherein _N+3Whether expression exists interested playback in the fragment, if replay is arranged then think, otherwise thinks and is not.

In segment boundaries detection and fragment, on the basis of measurement vector extraction step 5, carry out clip types and judge 6,, carry out the type identification of fragment and could well the causal relation in the fragment be taken into account dividing on the segment boundaries basis.Here we are divided into 5 classifications that the public is familiar with for the section of football match video fragment: shooting fragment (shoot), foul fragment (foul), goal fragment (goal), place kick fragment (placed kick) and common fragment (normal kick).Wherein place kick fragment comprises three types of corner-kick fragment (corner kick), free kick fragment (free kick) and penalty kick fragments (penalty).A football video sequence can complete the showing by the fragment of top 5 classifications.This mode can reflect defined 5 types fragment and can not produce the situation of omission.

The judgement of above-mentioned clip types is the sorting technique that adopts based on hidden Markov model (HMM) clip types.Specifically be according to its corresponding observation vector O=O with a fragment ₁... O _n, O _N+1, O _N+2, O _N+3HMM model parameter λ with corresponding shooting, foul, goal, place kick and common fragment _i=(A _i, B _i, π _i) (wherein i=1,2,3,4,5 corresponding expressions do not shoot, break the rules, goal, place kick and common fragment) judge that this fragment belongs to above-mentioned 5 types any.The method of judging is to calculate measurement vector in the fragment at 5 above-mentioned HMM model parameter λ _i=(A _i, B _i, π _i) probability P (O| λ _i) and judge wherein the pairing classification S of maximum according to formula (7):

$S=\arg \underset{i}{\max }P\left(O\right|{\mathrm{\λ}}_i)---\left(7\right)$

This fragment is divided in pairing that type of maximum then.Such as, we calculate P (the O| λ of a fragment from formula (7) ₂) be maximum, also promptly: $\arg \underset{i}{\max }P\left(O\right|{\mathrm{\λ}}_i)=2,$ We judge that the type of current fragment is foul so.

Wherein the parameter of HMM model needs to learn (L.R.Rabiner before classification, " Atutorial on hidden markov models and selected applications in speechrecognition; " Proceedings of the IEEE, vol.77, no.2, pp.257-285,1989.).In the HMM parameter learning, the manual common set of segments that is used to train, the shooting set of segments that is used to train, the foul set of segments that is used to train chosen from training video, the goal set of segments that is used to train, and the place kick set of segments that is used to train.Type and execution fragment measurement vector under these training fragments of mark extract.We adopt the ergodic topological structure of 6 states in the HMM parameter learning.

In carrying out above-mentioned clip types judgement 6, wherein, selecting shooting fragment, goal fragment, place kick fragment is wonderful; If a given fragment is confirmed as place kick, we can be divided into free kick, corner-kick and penalty kick with place kick again according to the direction character on sportsman's distribution situation and border in this fragment.As adopt J.Assfalg, M.Bertini, C.Colombo, A.D.Bimbo, and W.Nunziati, " Semantic annotation of soccer videos:automatic highlightidentification; " Computer Vision and Image Understanding, vol.6, No.4, pp.285-305, the described method among the Aug.2003.

After carrying out clip types determining step 6, for a given section of football match video, we are divided into a series of fragment sequences with clear and definite type with it.On this basis, carry out wonderful partiting step 7, it is characterized in that realizing wonderful is divided in two teams in the football match and to the goal type and judge.The method that team divides under the wonderful is to divide according to the direction of primary motion of video camera.For example, the direction of primary motion of wonderful is from left to right, then this wonderful is divided into the Na Zhi team that belongs to the left side; Otherwise this wonderful is divided into the Na Zhi team that belongs to the right.The judgement of goal type is to determine that according to fragment ordinal relation in time for example the fragment of a goal fragment front is a corner-kick, then this is carried out fragment and is divided into the goal (corner-kick goal) that is caused by corner-kick.Like this, carry out wonderful partiting step 7 and the goal type can be divided into following four classes: the goal (free kick goal) that causes by free kick, the goal that corner-kick causes (corner-kick goal), the goal that penalty kick causes (penalty stroke goal), and common shooting goal (common goal).

After carrying out wonderful type partiting step 7, then can carry out wonderful and generate step 8, it is characterized in that wonderful stored in the video database and can provide corresponding wonderful browsing.

Fig. 5 has provided the result of browse figure of the football match wonderful that obtains by the inventive method to Fig. 7.Can browse wonderful in the football match of being taken passages according to following three aspects:

(1) based on the video tour and the summary scheme of all shootings in the whole video, the wonderfuls of scoring;

(2) based on the video tour and the summary scheme of the wonderful of team;

(3) score based on all corner-kicks, free kick is scored, penalty stroke goal, the video tour of wonderfuls such as common goal and summary scheme etc.

Fig. 5 shows that the section of football match video splendid contents browses main window.Wherein W0 selects the browsing video file button, W5 closes the button of browsing window, and W1 is the video pictures broadcast window that shows splendid contents, W2 be splendid contents data list such as position appears in video, W3 shows excellent interior window, and W4 shows excellent window of scoring.Be that we can select the football video that will browse when pressing W0.After selected literary composition institute wanted browser document, we can click picture or the click W4 that W3 enters as shown in Figure 6 and enter as shown in Figure 7 picture.In Fig. 6, shown the general overview correction data of two team's splendid contents in the match among the window W33, and alternative team (A of team or the B of team or both) and corresponding splendid contents option in window W34, have been provided, after clicking navigation button, can in window W31, play wonderful, and in window W32, provide in the fragment time in the whole video appearance.In Fig. 7, shown in the match specifically the score general overview correction data of type of two teams among the window W43, and alternative team (A of team or the B of team or both) and corresponding goal content options in window W44, have been provided, after clicking navigation button, can in window W41, play wonderful, and in window W42, provide in the selected goal fragment time in the whole video appearance.

Claims (6)

1. a method that detects wonderful section of football match video is characterized in that, comprises the steps:

Step 1 is decoded by decoder to the compressed video of football match in the database earlier;

Step 2 is carried out video shot boundary and is detected, and is used for detecting the border of video camera lens, and is divided into replay sign, other camera lens, replay camera lens three parts according to the type of camera lens;

Step 3 is carried out feature extraction to other camera lens in the lens type, is used for compression domain and pixel domain are carried out feature selecting and signature analysis, extracts feature and comprises: main color characteristic, camera motion feature, textural characteristics and captions feature; During main color characteristic extracted, wherein each frame is calculated two features: the green area in the main color connected region accounted for the ratio G of whole image _RRatio G with the shared whole image of motion subject area in the main color connected region _PAdopt in the camera motion feature based on global motion vector field GMVF and carry out the judgement of camera motion pattern; In texture feature extraction, the texture strength information of extracting in each frame is Tx;

Step 4 comes the section of football match video camera lens is classified according to main color, camera motion, texture and captions feature that step 3 is extracted;

Step 5 is carried out the video lens segment boundaries according to the section of football match video shot classification result of step 4 and is detected, to determine the border between the fragment in the football video; And extract measurement vector in each fragment;

Step 6 is carried out clip types and judge that the measurement vector that is extracted is judged the type under this fragment, according to the sorting technique based on hidden Markov model HMM clip types from a fragment.This fragment is divided into a kind of in following 5 types, is respectively: shooting fragment, foul fragment, goal fragment, place kick fragment and common fragment; Wherein, selecting shooting fragment, goal fragment, place kick fragment is wonderful;

Step 7, the wonderful of determining is divided, be about to detected wonderful in the step 6 and be divided in two teams in the football match according to the direction of primary motion of video camera, with obtain two teams on the wonderful number contrast and the goal type judged;

Step 8 is taken passages the wonderful after dividing and store in the database.

2. the method for detection wonderful section of football match video according to claim 1, it is characterized in that, in the described step 3, in extracting the camera motion feature, described employing is carried out the concrete grammar that the camera motion pattern judges based on global motion vector field GMVF, according to the correlated characteristic of describing the global motion of video camera in formula (3)-formula (6):

${\mathrm{\θ}}_{\mathrm{var}}=\frac{1}{N\×M}\underset{\mathrm{xi}=1}{\overset{M}{\mathrm{\Σ}}}\underset{\mathrm{yi}=1}{\overset{M}{\mathrm{\Σ}}}{|{\mathrm{\θ}}_{(\mathrm{xi},\mathrm{yi})}-\stackrel{\‾}{\mathrm{\θ}}|}^2---\left(3\right)$

$\left\{\begin{array}{c}\mathrm{GMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVxi}\\ \mathrm{GMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\mathrm{GMVyi}\end{array}\right.---\left(4\right)$

$\left\{\begin{array}{c}\mathrm{AGMVx}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVxi}\right|\\ \mathrm{AGMVy}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\left|\mathrm{GMVyi}\right|\end{array}\right.---\left(5\right)$

$\left\{\begin{array}{c}\mathrm{GMVDx}=\mathrm{var}\left|\mathrm{GMVxi}\right|\\ \mathrm{GMVDy}=\mathrm{var}\left|\mathrm{GMVyi}\right|\end{array}\right.---\left(6\right)$

Wherein, θ _{(xi, yi)}

And θ _VarRepresent motion vector (GMVxi; GMVyi) pairing angle value, the angle variance of motion vector in pairing evaluation angle of all motion vectors and the entire image in the global motion vector field, (GMVx, GMVy), (AGMVx, AGMVy) and (GMVDx GMVDy) represents the mean value of average motion vector, motion vector magnitude and the variance vector of corresponding motion vector magnitude respectively.

3. the method for detection wonderful section of football match video according to claim 1, it is characterized in that, in the described step 4, the video lens classification step has been used two graders: earlier camera lens is carried out rough sort by the rough sort device, and then by the disaggregated classification device full shot in the rough sort camera lens is carried out disaggregated classification; Wherein, the camera lens rough sort is the color and vein characteristic vector＜G that step 3 is extracted _R, G _P, Tx〉and be divided into close-up shot, grandstand camera lens, full shot and medium camera lens; The camera lens disaggregated classification is the vector theta of being extracted in the formula (6) according to formula (3) _Var, GMVx, GMVy, AGMVx, AGMVy, the global motion characteristic vector of one 7 dimension that GMVDx and GMVDy constituted is divided into Zoom lens, tracking lens and still frame with the full shot in the camera lens rough sort once more.

4. the method for detection wonderful section of football match video according to claim 1, it is characterized in that, in the described step 5, the method that described segment boundaries detects is to judge the full shot of fragment and the border of non-full shot, and the original position of a fragment and final position all are the switching position of non-full shot to full shot.

5. the method for detection wonderful section of football match video according to claim 1 is characterized in that, in step 6, described sorting technique based on hidden Markov model HMM clip types is, with a fragment according to its corresponding observation vector O=O ₁... O _n, O _N+1, O _N+2, O _N+3HMM model parameter λ with corresponding shooting, foul, goal, place kick and common fragment _i=(A _i, B _i, π _i) judge that this fragment belongs to above-mentioned 5 types any, the method for judgement is to calculate measurement vector in the fragment at 5 above-mentioned HMM model parameter λ _i=(A _i, B _i, π _i) probability P (O| λ _i) and judge wherein the pairing classification S of maximum according to formula (7):

$S=\arg \underset{i}{\max }P\left(O\right|{\mathrm{\λ}}_i)---\left(7\right)$

This fragment is divided in pairing that type of maximum then.

6. the method for detection wonderful section of football match video according to claim 1 is characterized in that, in step 7, described goal type comprises four classes: free kick is scored, and corner-kick is scored, penalty stroke goal, common goal.

Images