KR100963744B1 - A detecting method and a training method of event for soccer video - Google Patents

Abstract

The present invention is a method of learning and automatically detecting an event occurrence in accordance with the domestic broadcast situation by applying a hidden Markov model to a football video multimedia, the process of determining whether to learn the event pattern of the football video, and learning the event pattern Converting each scene of the soccer video into a symbol if the symbol is present; Determining an observation sequence from the transformed symbol; And learning the hidden Markov model by using the determined observation sequence. The present invention also provides a method of determining whether or not each event of a soccer video is detected; Converting each scene into a symbol when detecting the event; Determining the observation sequence from the transformed symbol and using the determined observation sequence, and detecting the combined probability from the hidden Markov model learning process result, so that the event detection rate is suitable for the soccer video of domestic video broadcasting. To improve and facilitate the analysis of the football game.

Description

Event learning and detection method of soccer video {A DETECTING METHOD AND A TRAINING METHOD OF EVENT FOR SOCCER VIDEO}

1 is a functional block diagram of an event detection apparatus of a general soccer video;

2 is a flow chart of the soccer video event learning method according to the present invention;

3 is a flowchart illustrating a method for detecting a soccer video event according to the present invention;

4 is a symbol classification state diagram according to the degree of expansion and contraction of the soccer video screen;

5 is a state diagram of symbol classification according to field positions of the present invention soccer video screen;

6 is a low level information state diagram by a camera photographing target of a soccer video;

7 is a diagram showing the analysis results of the 2002 FIFA World Cup and 2002 Busan Asian Games.

8 is a 252 symbol scene shape distribution diagram of the soccer video of the present invention;

9 is a pattern analysis diagram according to each sequence of the soccer video of the present invention;

10 is a diagram showing an example of a representative A, B, C pattern of the present invention soccer video,

11 is a distribution chart of each event occurrence pattern bar chart of the present invention soccer video;

12 is a state diagram of an event detection system of the present invention soccer video;

Fig. 13 is a diagram showing 84 symbolic analysis values of the present invention soccer video.

14 is a view showing 108 symbolic analysis values of a soccer video of the present invention;

Fig. 15 is a diagram showing the distribution of 84 symbols and 108 symbols of a soccer video.                 

FIG. 16 is a diagram illustrating a shooting event detection frequency analysis value of a 2002 FIFA World Cup soccer game and a 2002 Busan Asian game of the present invention soccer video.

Figure 17 is a contrast between the football video event detection method of the present invention and the event detection method suitable for foreign circumstances.

          ** Explanation of symbols on the main parts of the drawing **

10: video unit 20: video analysis unit

30: DL part 40: browsing tool part

50: detector

The present invention is to search for and output a specific event from the video multimedia, in particular, by applying the hidden Markov model (HIDDEN MARKOV MODEL) to the soccer video multimedia to learn the scene where the event occurs in accordance with the domestic broadcast situation and automatically search output An event learning and detection method of a soccer video.

Due to the development of computer related technology, rapid performance improvement, network speed, and mass storage device development, it is possible to use multimedia information of video that can be easily recognized and communicated quickly. Increasingly, the importance of record keeping is being confirmed.

The multimedia signal is efficiently processed and managed as a digital signal, and since the data capacity is very large, a large capacity storage device is required, and a DVL (data storage device) for managing such a large amount of multimedia data is required. DIGITAL VIDEO LIBRARY system technology is being developed, and VOD (VIDEO ON DEMAND) services are provided by the above-described DVL technology.

The multimedia contents (CONTENTS) information is recorded and stored in a mass storage device, and, if necessary, indexed to quickly, easily and accurately retrieve desired information from the mass storage device in which the multimedia information is recorded as described above. INDEXING) to store and manage. For indexing as described above, a browsing tool or a summary file should be created.

However, in order for a person to work on such indexing directly, there is a problem that requires a lot of labor and is expensive, but it is not objective and does not obtain accurate results.

In particular, it is necessary to develop a technology for automatically searching for a desired scene or shot by video analysis, low level analysis, intermediate analysis, and advanced analysis, and quickly detecting and outputting necessary video multimedia information.

Hereinafter, an event detection method of a soccer video according to the prior art will be described with reference to the accompanying drawings.

For the purpose of illustrating the prior art, Figure 1 is a functional block diagram of an event detection apparatus of a general soccer video.

Referring to FIG. 1, the multimedia signal of the soccer video output from the video unit 10 is applied to the video analysis unit 20, and is divided into scene information of low and intermediate levels, and advanced. Each event is classified by classifying each type of scene information of the level.

As described above, each piece of information analyzed and output as low, intermediate, and advanced levels of information from the video analyzing unit 20 is applied to the DL unit 30 and stored.

The browsing tool unit 40 connected to the DVL DIGITAL VIDEO LIBRARY unit 30 analyzes the stored event information and applies a control signal to the detection unit 50 to easily detect a desired content CONTENTS. The detection unit 50 detects and outputs a desired event from the information of the DL unit 30 according to a control signal applied from the browsing tool unit 40.

As described above, the video multimedia information analysis technology according to the related art is mainly, for example, low- and intermediate-level video information, such as remote video information, near video information, replay video information, video information of a substantially fixed background area, and the like. Will be automatically extracted.

For example, in the case of video multimedia information of a soccer sports game, a technique for searching and extracting a long distance shot of a stadium, a close shot, a replay scene, and a spectator scene with little movement becomes a low or intermediate level technology. .

Image information used for indexing by analyzing the video signal at a low level may include information such as color information, object motion, and edge histogram (EDGE HISTOGRAM).

By using the low level analysis information as described above, middle level video information such as enlargement and reduction of the screen, field position, camera shooting target, playback speed, and use of graphic information can be extracted.

In the analyzed video multimedia, for example, there is a domain (DOMAIN) divided into general interesting videos, recording videos, war videos, sports videos, and the like. In order to analyze each domain (DOMAIN), domain knowledge ( DOMAIN KNOWLEDGE), using the domain knowledge to model the corresponding content, search each event scene at an advanced level, and highlight important events by gradual scene change by the Wipe or Dissolve method. do.

The indexing method of analyzing the video signal is largely classified into a statistical method and a slice method. The statistical method uses a difference between the current video frame and the previous video frame, and detects a color histogram, a motion, or an edge of each frame, and shows a portion where the amount of change is prominent. Indexing to the boundary point of the

However, the statistical method as described above has a problem in that it is difficult to find a scene boundary for indexing from a frame of a scene having a high momentum while being performed at a certain place such as a sports video signal.

The slice method is a method for easily finding a scene boundary for indexing from a sports video signal, and dividing the image frame into slices, and using the difference in cohesion between the divided slices to determine a boundary point of the scene. Drew (MS DREW) suggests in VIDEO DISSOLVE AND WIPE DETECTION VIA SPATIO-TEMPORAL IMAGE OF CHROMATIC HISTOGRAM DIFFERENCES, PROC OF IEEE INTERNATIONAL CONFERENCE ON IMAGE PROCESSING, VOL 3, PP 929-932, 2000 ) Was proposed in DETECTION OF GRADUAL TRANSITIONS THROUGH TEMPORAL SLICE ANALYSIS, PROC OF IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR'99), PP 36-41, 1999.

Since the slice method is not focused on the change of the whole frame and is not sensitive to the partial change, it shows a good performance for analyzing a video signal such as soccer.

In particular, high-level video analysis in sports video multimedia such as soccer is called EVENT analysis, and examples of such events include goal scoring, shooting, corners, free kicks, penalty kicks, and fouls.

As the information used for indexing by detecting the event (EVENT) as described above, caption information (CLOSED CAPTION), blurring phenomenon, ball (BALL) and players and goal post (GOAL POST) tracking (TRACKING) There are four types of state transition diagrams.

The caption information is provided by T. KITAHASHI in September 2000, EXTRACTING ACTORS, ACTIONS AND EVENTS FROM SPORTS VIDEO-A FUNDAMENTAL SPPROACH TO STORY TRACKING, PROC OF IEEE INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION (CIPR'00), VOL As suggested by 4, BARCELONA, SPAIN, SEPT, 2000, it uses caption information, which is an additional medium for multimedia videos. For example, in the case of American football, it detects the word TOUCH DOWN and locates the event. .

The blurring phenomenon information is provided by shoe X. XU in 2002, CONTENT-BASED ANALYSIS AND INDEXING OF SPORTS VIDEO, PROC OF SPIE CONFERENCE ON STORAGE AND RETRIVAL FOR MEDIA DATABASE, SAN JOSE, USA, JAN 21-23, 2002 Proposed by, to detect the blurring that occurs because the video screen suddenly tries to follow the fast ball (BALL) when shooting occurs, it is to detect the shooting (SOOTING) and long pass (LONG PASS).

In the blurring phenomenon, when the edge EDGE is found using the MARR-HILDRETHS 'ZERO CROSSING algorithm, the amount of edges of the frame in which the blurring phenomenon occurs is significantly reduced.

In addition, tracking and using the ball, the players, and the goalpost as information detects various events such as a goal when a ball passes the goal, AF BOBICK in 1995 and 1995. Proposed by B. LIU and RJ QIAN in 2001, the OBJECT SEGMENTATION that separates the background from the players finds the movements of the players and traces them into a continuous mosaic of the background (MOSAIC). TRAJECTORY to track the movement of each player on the playing field.

The last state transition diagram is proposed by R. MIGLIORATI in SEMANTIC INDEXING OF MULTIMEDIA DOCUMENTS, IEEE MULTIMEDIA, VOL 9, NO 2, PP 44-51, APR-JUN, 2002 in 2002. An algorithm that detects goals, corners and free kicks using events that occur in a regular pattern.

In other words, if one team attacks the other team's goal before the goal is scored, the camera will panning or zooming in a fast motion and will shoot in the next order. To catch the scene where (BALL) enters the GOAL POST, when the CAMERA is not moving and the goal is scored, it shows the shot of the player who shot or shot the shot from a different angle of the camera. UP), a shot cut occurs.

The state transition diagram is used to find out that the goal occurs in the above sequence.

However, in order to detect the events of the prior art as described above, the first method of using caption information is different from football, in the current situation in Korea, when each broadcasting station broadcasts or broadcasts sports caption information. Since it does not provide, there is a fundamental problem that cannot be used.

The second method of using the blurring phenomenon is that the two events can be distinguished from each other because they use the characteristics of the screen when shooting and long pass are issued, but there is a problem that other events cannot be distinguished.                         

The third method of tracking the ball, the players and the goal, consumes too much load to create an IMAGE mosaic (MOSAIC) to find object separation and trajectory, There is a limiting problem of detecting only events that occur in the far scene.

In addition, the fact that the players overlap each other and that the ball is too small to be detected and obscured by the players makes it difficult to separate objects, while soccer video multimedia with a lot of camera movement can be used for moving backgrounds and foregrounds. There is a big problem that object separation is relatively difficult.

The fourth, state transition diagram method is an advanced level of information analysis method. It is assumed that a regular pattern (PATTERN) occurs when an event occurs by using a state transition graph. Detect and output.

However, it is a characteristic of sports video multimedia that various exceptions occur even in the pattern of one event EVENT, and from the soccer video in which the irregular pattern as described above occurs, using the prior art, in particular, Detecting each event by football video analysis by level, there is a problem that does not get a good result.

The present invention classifies the scene type from the football video multimedia, solves the exception by statistical learning and modeling in the temporal flow using the hidden markov model (HIDDEN MARKOV MODEL), and absorbs noise and deformation The purpose of the present invention is to automatically detect an event and provide an event learning and detection method of a soccer video suitable for the Korean broadcast situation.

The present invention devised to achieve the above object, Learning 1 process for determining whether to learn the event pattern of the soccer video; A learning 2 process of converting each scene into a symbol when the event is learned in the learning 1 process; A learning 3 process of determining an observation sequence from the symbols transformed in the learning 2 process; A learning 4 process of learning the hidden Markov model is performed using the observation sequence determined in the learning 3 process.

In addition, the present invention devised to achieve the above object, the detection 1 process of determining whether to detect each event of the soccer video; A second detection step of converting each scene into a symbol when the event is detected in the first detection step; A detection step 3 for determining an observation sequence from the symbols transformed in the detection step 2; The detection sequence is determined by using the observation sequence determined in the detection process 3 and detects the coupling probability from the hidden Markov model learning process.

In addition, the present invention devised to achieve the above object is characterized in that it detects the important highlight event of the video football game and transmits through wired and wireless television broadcasting and multimedia wireless communication equipment.

Hereinafter, a method of detecting an event of a soccer video according to the present invention will be described with reference to the accompanying drawings.

2 is a flowchart illustrating a soccer video event learning method according to the present invention.
As shown in FIG. 2, in the soccer video event learning method according to the present invention, a process of determining whether training of various event occurrence patterns (PATTERN) for distinguishing important shots (SHOT) of a soccer video is determined. In operation S110, when the event occurrence pattern is learned in step S110, a process of converting each scene into a symbol SYMBOL may be performed in operation S120. In the process (S120), the soccer video scene is enlarged and reduced to a value of 3, a field position of 7, a camera shooting target value of 2, a playback speed value of 2, and graphic information use value. Given the value of 2, it is analyzed by 252 scene-type symbols.
In addition, the soccer video event learning method according to the present invention performs a step (S130) of determining an observation sequence (OBSERVATION SEQUENCE) from the converted symbol (SYMBOL) in the step (S120), the current event scene in the observation sequence Seven symbols are determined, including the symbol, the previous three symbols, and the three subsequent symbols.
In addition, the soccer video event learning method according to the present invention performs the process of learning the hidden Markov model (HMM: HIDDEN MARKOV MODEL) using the observation sequence determined in the step (S130). That is, in step S130, the hidden Markov model HMM is trained by FORWARD PROCESS and BACKWARD PROCESS of BAUM to reduce the amount of calculation.
3 is a flowchart illustrating a method for detecting a soccer video event according to the present invention.

delete

delete

delete

delete

As shown in FIG. 3, the method for detecting a soccer video event according to the present invention includes determining whether to detect (DETECT) each event classifying an important scene of a soccer video (S210), and detecting the event. In operation S220, the scenes are converted into symbols SYMBOL. In the process (S220), the degree of zooming in of the soccer video scene is 3, the value of 7 as the field position, the value of 2 as the shooting target value of the camera, the value of 2 as the value of the playback speed, and whether the graphic information is used. We give 2 values each and analyze them as symbols of 252 scenes.

delete

Soccer video event detection method according to the invention, the step of determining the observation sequence from the transformed symbol (S230), and using the determined observation sequence, detecting the combined probability from the hidden Markov model (HMM) learning processing results The process (S240) is performed. In step S230, the observation sequence determines seven symbols including the current event scene symbol, the previous three symbols, and the three subsequent symbols.

delete

Hereinafter, an event learning and detection method of a soccer video according to the present invention having the above configuration will be described in detail with reference to the accompanying drawings.

In general, a computer (PC) is a model that converts the input signal to facilitate processing, it is effective to use the deterministic model (DETERMINISTIC MODEL), the noise (ISE) and the distortion (DISTORTION) using the specific signals already known or secured There are two types of statistical models to solve.

There is a Markov model as a statistical model, but since only processing for an observable signal is possible, there are many limitations in the application.

The above problem is solved by the Hidden Markov Model (HMM), which is a dual STOCHASTIC model, and allows the signal pattern to express the distortion and similarity between the patterns.

The Hidden Markov Model (HMM) provides an appropriate solution with its own modeling and learning capabilities, even when it does not have accurate knowledge of problem solving, and eliminates noise, loss and distortion. There is this.

In the soccer video, since the information pattern of the intermediate level is generated according to the time flow before and after the event, the hidden Markov model is used to probabilistically model the change over time.

Therefore, as shown in FIG. 1, when the user wants to learn various event patterns PATTERN from the soccer video already captured and secured by the video or the like as described above (S110), the controller (not shown) is respectively. The event scene is converted into 252 symbols (SYMBOL) (S120).

For example, each event generated from the soccer video is set as a symbol. For example, three screens, such as GLOBAL / LONG SHOT, MEDIUM / WOOM-IN SHOT, and CLOSE-UP SHOT, may be used. The value is set to 3 because it is divided into symbols, and the location of the field is, for example, a penalty area of each goal post, each of the left and right corners, and the midfield area (MID-FIELD). It is divided into 7 symbols by AREA) and gives a value of 7.

As described above, the actual state of symbol division based on the degree of enlargement and reduction of the screen is shown in FIG. 4, and the actual state of symbol division by field position is shown in FIG. 5.

In addition, a value of 3 is given to classify the symbol into three symbols of a field, a player, or an audience, and the playback speed is a motion vector or a gradual scene. It is judged by the editing effect of GRADUAL SHOT TRANSITION, and it is given a value of 2 to divide it into two symbols of the normal speed playback and the replay screen played at a slow speed. The value of 2 is given to separate the use of into two symbols.

It is for automatically determining what is being photographed by the camera, and the intensity distribution of edges, the segment length, the distribution of the direction, and the color tone of the hue. The frequency distribution and the like are used, and the presence or absence of graphic information is used by the edge information and the high frequency distribution.

FIG. 6 is a graph illustrating edge histogram extraction state of low level information and shooting state and graphic information for distinguishing camera shooting targets, and FIG. 7 is a value obtained by analyzing 10 games of the 2002 FIFA World Cup and the 2002 Busan Asian Games in the form of 252 symbols. 8 illustrates a scene shape distribution diagram using 252 symbol values of a soccer video.

Events that occur in soccer videos include GOAL, SHOOTING, CORNER KICK, FREE KICK, PENALTY KICK, and Foul. An example of an event detected from ten soccer video games is shown.

(Table 1)

event goal Shooting corner kick free kick Penalty kick Foul Detection value / 10 game 20/10 215/10 108/10 215/10 2/10 328/10 Average 2/1 22/1 11/1 22/1 0-2 / 1 33/1

According to Table 1, when the shooting event occurs 215 times in 10 randomly selected soccer games, and when the shooting event occurs, for example, the previous three scenes (SHOT), the scene where the event occurs and after the event occurs Three scenes are determined as one observation sequence (S130).
FIG. 9 illustrates a pattern analysis diagram according to each sequence of a soccer video, and FIG. 10 illustrates an example of representative A, B, and C patterns of a soccer video.

As shown in Figs. 9 and 10, the 215 sequences are represented by an A pattern, a B pattern, a C pattern, and a large number of other patterns, with the A, B, and C patterns accounting for more than 80% of all events. do.

In FIG. 10, the A pattern that appears when the shooting event occurs, the game is interrupted and resumed in the corner area before the event occurs, followed by a close-up of the shooter after shooting. The replay scene is a continuous pattern. The replay scene also includes a scene in which the ball flies toward the goal post, a scene in which the player kicks the ball in front of the goal post.

In FIG. 10, the B pattern is similar to the A pattern, but a shooting event occurs after a corner corner, and closes the player shooting after the event, followed by a re-extension.

The replay shows the players fighting in front of the goalpost before the corner is taken, then replays the shooting scenes following the corner, shows the player, followed by the replay, and shows the progress of the game over a distance. Shows the player kicking the ball at close range.

In FIG. 10, the C pattern has a pattern before an event occurs, and the goal area is photographed at a long distance, followed by a short distance scene, and then a shooting is generated at a long distance scene. After shooting the game, the shooting player is CLOSE-UP, followed by a REPLAY scene, and the game is played after the replay scene. Show you from a distance

Through the above analysis, A, B, C pattern is progressed by various attack methods before the shooting occurs, and shows that a certain scene occurs to some extent after the event occurs.
FIG. 11 is a distribution chart of each event generation pattern bar chart of a soccer video, and shows distribution patterns of shooting events.
In FIG. 11, the pattern in which the shooting occurs in a soccer game has a high specific gravity of the A, B, and C patterns, and the small number of the remaining large and minor patterns is very small.

delete

That is, when the shooting event occurs, a total of four patterns exist in the sequence before and after the shooting event, with the main three patterns of the A, B, and C patterns and the rest as one pattern, but there are exceptions in these patterns. It can be seen that a lot of phenomenon occurs.

However, it can be seen that there are not many cases where a shooting event does not occur in a sequence similar to the pattern of the shooting event. As a method of learning and modeling a pattern in a temporal flow, the distortion and similarity of the pattern can be expressed. A hidden Markov model having a feature capable of fast processing while absorbing noise or distortion is used (S140).

When the event learning of the soccer video is completed as described above, as shown in FIG. 3, it is determined whether a soccer video event is detected (S210). When the event is to be detected in the determination, the controller converts each scene into a symbol SYMBOL in the same manner as in the event learning process (S220), and determines the observation sequence in the same way as in the event learning process (S220). S230).

When the observation sequence is determined as described above (S230), the result of the learning process by the hidden Markov model is input, and the coupling probability is detected as follows.

(Mathematical formula)

P (O | λ) = P (O ₁ , O ₂ , .., O _T | λ)

λ = {A, B, π}

A = {aij}, aij = P {qj (t + 1) | qi (t)}: State transition probability distribution

B = {bj (k), bj (k) = P {vk (t) | qj (t)}: Observation symbol probability distribution in state J

π = (πi}, πi = P {qi (1)}: Initial state probability distribution

As described above, each process of learning and detecting an event from a soccer video can be more easily understood with reference to FIG. 12.
12 is a diagram illustrating a configuration of an event detection system of a soccer video.

12 shows that the learning process and the detection process are processed according to their respective order, and finally, the event is detected in the detection process using the results learned in the learning process.
FIG. 13 is a diagram showing an analysis value of 84 symbol types of a soccer video in the present invention, FIG. 14 is a view showing an analysis value of 108 symbol types of a soccer video in the present invention, and FIG. 15 is a football video in the present invention. 84 shows the distribution state of the 108 symbols and 108 symbols.
As described above, each scene is defined as 252 symbols (SYMBOL) from the soccer video, but in reality, when the shooting target of the camera is an audience or a bench, the shot (SHOT) forms are formed regardless of the position information on the field. If the symbol is defined and the position information on the field is defined as one symbol even in a close-up of the athlete, the entire symbol can be defined as 84 symbols, and the definition of the symbol is shown in FIG.

delete

In addition, it is difficult to know the position information of the player in the close-up of the player, but when the player is the goalkeeper, it is possible to find out the goal area, and even if not the goalkeeper, it is possible to find the position of the currently close-up player in the front and rear scenes.

In this manner, as illustrated in FIG. 14, the symbols of 252 scenes may be defined as the symbols of 108 scenes.

As described above, in the case of the 252 symbol, the utilization ratio is low, and in the case of the symbols 84 and 108, the utilization ratio of the symbol is high and the overall distribution is not changed. have.

That is, the case where 84 symbols or 108 symbols are used rather than 252 symbols (SYMBOL) is used, but the processing pattern is simplified because the event occurrence pattern is simple, but it can be seen that there is no significant change in the event pattern.

FIG. 16 shows the number of shooting events detected as a result of experimenting with the 2002 FIFA World Cup Soccer and the 2002 Busan Asian Game using the present invention.

Conventional event detection method uses a state transition diagram, which does not reflect the characteristics of the exception occurring in a sporting event, and shows a low detection rate, but in the present invention, using a hidden Markov model, it is suitable for domestic broadcast situation, Statistical analysis of exceptions solves sequence recognition problems and improves detection rates.
17 is a contrast diagram of an event detection method suitable for a foreign country football video event detection method of the present invention.

FIG. 17 is a diagram illustrating a comparison between an existing event detection method suitable for a foreign situation and an event detection method suitable for a domestic situation, and shows that the present invention is the most suitable method for a domestic broadcasting situation.

As described above, the present invention can be applied to various multimedia communication devices including a portable terminal capable of video communication in addition to wired / wireless TV broadcasting.

In particular, when the present invention is applied to a multimedia portable terminal that charges a communication fee in proportion to the amount of data transmission, the main highlight event is detected and transmitted instead of the entire soccer game content. It is possible to transmit a video signal to easily understand the contents.

Therefore, when using a communication device such as a multimedia portable terminal, since only the highlight event portion is detected and transmitted, the amount of data to be transmitted is reduced and the communication cost is paid less, thus improving the quality of service and video processing technology of the wireless communication device. And improve the usability and stability of the system.

As described above, the present invention has the industrial use effect of improving the event detection rate from the soccer video by the domestic video broadcasting situation using the hidden Markov model.

In addition, since each event is detected and provided at a high rate from the soccer video, there is a convenient effect in use to facilitate the analysis of the soccer game.

In addition, it detects only the important highlight events of the football game and enjoys in a short time, saving time, and has a convenient effect that can be applied to the multimedia communication equipment and transmitted.

In addition, in the multimedia wireless communication method that charges a communication fee in proportion to the amount of transmission data, there is a convenient effect that can easily grasp the entire content of the football game with a minimum cost and a short time.

Claims (10)

Determining whether or not the event pattern of the football video learning, In the case of learning the event pattern, the process of converting each scene of the football video into a scene-shaped symbol corresponding to the degree of expansion and reduction of the football video scene, the field position, the shooting target of the camera, the playback speed, and the use of graphic information; ; Determining an observation sequence from the transformed symbol; And learning to conceal an event generation pattern using the determined observation sequence. delete The method of claim 1, wherein the determining process, And a current event scene symbol, three previous symbols, and three subsequent symbols as the observation sequence. The method of claim 1, wherein the learning process comprises: A method of learning an event of a football video, characterized by reducing the amount of computation by learning the hidden Markov model by forward processing and backward processing of Baum. Determining whether each event of the soccer video is detected; Converting each scene into a corresponding scene-type symbol by reflecting the extent to which the moving scene is reduced, the field position, the photographing target of the camera, the playback speed, and the use of graphic information when detecting the event; Determining an observation sequence from the converted symbol; And using the determined observation sequence, detecting a joint probability from a hidden Markov model learning process result. delete The method of claim 5, wherein the determination process, The event detection method of a soccer video, comprising the current event scene symbol, the previous three symbols and the next three symbols in the observation sequence. The method of claim 5, wherein the detecting process, Detecting the combined probability from the hidden Markov model learning process result. delete delete

Images