With the rapid development of multimedia capturing devices, it becomes easier for people to record their lives. A large amount of multimedia data, such as images and videos are produced and uploaded to the internet. Therefore, analyzing and understanding automatically the complex and compound multimedia data becomes an important issue. Sports video is one of the most popular multimedia data since sports games hold a lot of audiences worldwide. In recent years, sports videos analysis is a flourishing research area. There has been an explosive growth of researches focusing on analyzing sports videos due to the potential commercial benefits and entertainment demands. From a sports-watcher point of view, only some portions in a sports video are worth viewing. These video segments of interest are the semantic events which have certain high-level semantics, such as homeruns in baseball games and goals in soccer games. Various efforts are made for event detection, content understanding, and sports information retrieval, in order to provide the viewers with automatic annotation and enriched visual presentation. Therefore, algorithms have been developed for shot classification, highlight extraction and semantic annotation based on the fusion of audiovisual features and the game-specific rules. In this thesis, we focus on audio and visual features integration and algorithm development for sports video content analysis and understanding. Sports information retrieval, tactics analysis, enriched visual presentation can provide the audience and professionals a further insight into the games.
Generally, sports videos can be roughly classified into two categories [11]: real videos and broadcast videos. Real videos are captured by a single or several fixed cameras. In real videos, the camera is stationary and the background is almost static.
2
Broadcast videos are obtained from television programs while broadcasting the games.
In broadcast videos, the camera motions are inevitable, as shown in Fig. 1-1. The background is usually changing since the cameras always focus on the ball or the moving players. More and more researches are devoted to broadcast videos such as soccer, baseball, and basketball because of the potential commercial benefits and massive audiences. Unlike the sport mentioned above, volleyball games are not so popular since there are only a few channels broadcasting the volleyball games. On the contrary, there are plenty of videos recorded by coaches, players, or even the audiences with a fixed digital camera since the volleyball courts are relatively small so that the scenes can be seen clearly by only one camera, as shown in Fig. 1-2.
Fig. 1-1. Examples of broadcast videos.
3
Fig. 1-2. Videos recorded by coaches, players, or even the audiences.
By recording the volleyball matches, the audiences can enjoy watching the matches they are interested in, and the coaches and professional players can obtain the statistical information about opponent teams‟ and themselves. For the audience, they usually pay their attention to the exciting events, such as splendid spikings and wonderful digs, since it may take a long time to watch the entire matches. It is a trend to design automatic systems for content-based video retrieval and semantic analysis in order to display such selective events and shots. From the coaches‟ and the professional players‟ points of view, analyzing the tactic patterns executed by the opponent teams can help them work out corresponding strategies in the training process or even in the matches. However, it is time-consuming and labor-intensive to manually recognize the tactic patterns and collect the statistical data from a large amount of sport videos. Consequently, establishing a system which can automatically provide tactic patterns recognition and semantic event extraction is required.
Volleyball is an Olympic team sport in which two teams of six players are separated
4
by a net. Each team tries to score points by grounding a ball on the other team's court under organized rules [15]. In volleyball matches, spiking is the most direct and effective way to get points and it is always the most exciting part in the games. On the other hand, blocking can prevent opponents from getting scores by spiking, and even get points with successful blocking, which means grounding the spiking ball back on the opponents‟ court. The blocking patterns can be classified into four types based on the number of people included: none, single, double, and triple blockings, as shown in Fig. 1-3. The more people participate in blocking, the higher probabilities of successful blocking would make. In addition, marvelous spiking such as delayed spiking and alternate position spiking always results from brilliant set which leads to none or only single blocking, as examples are shown in Fig. 1-4. Delayed spiking utilizes close time and positions between the jumps of two spikers to confuse the blockers. And alternate position spiking avoids too many blockers by making a wide set. Both two tactics need a cloak, which means a player pretends to spike. In other words, the type of set plays an important role in tactics analysis. For this reason, classifying the blocking patterns can help the coaches analyze the successful tactics used by the setters. Therefore, our goal is to propose a system which not only provides specific spiking events detection but also classifies the blocking patterns for the spiking events.
5
(a) (b)
(c) (d)
Fig. 1-3. Four types of blocking patterns: (a) none (b) single (c) double (d) triple
(a) (b)
(c) (d)
Fig. 1-4. Two important tactics: (a)(b) alternate position spiking (c)(d) delayed spiking
6
Although increasing research effort of sports video processing concentrates on ball tracking and trajectory-based tactics analysis, the majority of existing work focuses on tennis and soccer video. Little work was done for volleyball video because it is much complex to track the ball and players in volleyball video due to the high density of players on the court and the frequent ball-player overlaps. In this thesis, we propose a scheme for detecting spiking events and simultaneously classifying blocking patterns by analyzing the moving pixels. analyze the moving pixels above the net and detect the spiking events. After detecting the spiking events, we further analyze the blocking patterns when the spiking occurs.
The proposed scheme enhances the accuracy of the subsequent analysis by preprocessing the video clips. A moving pixels analysis based algorithm is designed for easily computing and efficiently analyzing the blocking patterns.
The remaining of this thesis is organized as follows. In Chapter 2, we introduce the related work about sports videos analysis. Chapter 3 describes the methods we proposed. The experimental results are presented and discussed in Chapter 4. Finally, we make conclusions and discuss future work in Chapter 5.
7