籃球影片中的球員追蹤與戰術分析

Abstract

隨著電視轉播技術的發展，越來越多人觀賞籃球比賽，但是大多數的人對於籃球知識並不是非常了解。我們也許會因為球員投進壓哨三分球而尖叫，或為一個強力灌籃而興奮，但是不見得知道球員是如何擺脫防守者進行投籃。目前已經有一些籃球影片內容的研究，例如精采畫面擷取和記分板辨識，但是這些仍然無法幫助觀眾對於籃球有更深入的了解。所以我們希望能設計一個系統來提供觀眾一些比較深入的籃球知識，而不只是表面上的資訊。籃球比賽中，觀眾最有興趣的就是得分。但是得分背後的戰術是一門很深奧的學問，因為籃球是一項五個人的運動，不可能只靠一個球員去對抗另外一隊，也就是說單一球員很難靠他自己擊潰對方的防守並且進行得分。大部分的得分都是經由執行戰術而來的。所以我們的目標就是自動辨認出籃球比賽中執行的戰術，並且把這些收集來的資訊帶給觀眾，讓他們能更了解籃球這項運動；甚至可以提供教練和球員，作為他們訓練及了解敵隊的攻防策略之用。 籃球戰術種類眾多，很難用單一演算法一以概之，因此我們著重於大多數戰術中都會使用的「掩護」，加以偵測並且分類，藉此分析出戰術執行的模式。我們開發的系統執行步驟如下。在比賽一開始先收集整場比賽都不會變的資訊，包含球場地板顏色以及兩隊球衣顏色。首先我們計算攝影機的參數，並且產生一張表示球場範圍的遮罩。第二步我們在球場範圍內計算出現次數最多的顏色，也代表著地板顏色。接著利用背景相減法，我們可以從球場範圍減去地板得到前景物體。最後我們利用顏色資訊將前景分成兩群，分別代表著兩支球隊的球衣顏色。因為這些資訊在整場比賽中都不會改變，所以我們可以利用它們降低往後的計算量，並且提升系統效能。在比賽中，針對每一次球權先分辨哪一隊是進攻方，了解雙方球員的行為模式才能判斷執行的戰術。利用先前得到的資訊並追蹤雙方球員的軌跡。在一波進攻結束的時候，根據追蹤到的雙方球員軌跡來判斷執行的掩護。經由實驗結果，我們開發的系統對於掩護的偵測和分類準確度相當令人滿意，因此在戰術分析上也有著顯著的幫助。這些被辨識出來的戰術會存入資料庫，於是觀眾就可以查詢他們有興趣的戰術並且學習。

Thanks to the development of TV broadcasting technology, there are more and more people watching basketball games. Most of us, however, do not know the basketball sport very well. We may scream for a buzzer beater three-point shot or get excited about a slam dunk, but we do not exactly realize how a player gets rid of defenders and makes shots. There have been some researches on basketball video, such as highlight extraction and scoreboard recognition, but they still cannot help people further understand this sport. Therefore, we intend to design a system which provides audience with further knowledge of basketball instead of superficial information. In basketball games, people are most interested in scoring events. Nevertheless, scoring is not that simple as it looks. It can be an abstruse subject since basketball is a five-person sport and one player is not able to fight against the opponent team. That is, it is difficult for an individual player to break the defense and score by himself. Most shots are made through execution of tactics. Consequently, our goal is to automatically identify tactics executed in basketball games and bring audience the collected information so that they can learn more about the basketball sport. There is plenty of basketball tactics, and it is hard to model them by a single algorithm. Hence, we focus on “screen,” which is widely used in most basketball tactics. We detect and classify screens, and regard their patterns as certain tactics. Our proposed system performs with the following steps. First of all, we gather some consistent information at the beginning of the game, including the floor color and the jersey colors of the two teams. We first compute the camera calibration and generate a court mask indicating the court region. Second, we calculate the dominant color within the court region, which represents the floor color. Next, we obtain the foreground objects by subtracting the floor from the court region. This procedure is similar to a background subtraction mechanism. Finally, we divide the foreground region into two clusters with color information. Thus, the two clusters denote the jersey colors of the two teams respectively. Since this information is consistent through the entire game, we can utilize it to reduce computational cost and accelerate the computation in the following frames. During the game, we first distinguish which team is on offense in each possession since we have to learn the behaviors of offensive and defensive players respectively in order to identify tactics. Next, we extract players of the two teams with the previously obtained information and track them. At the end of a possession, we identify what screens are set by the trajectories of the players. Through our experiment, the accuracy of screen detection and classification is satisfactory, which significantly helps analysis of basketball tactics. The identified tactics are then inserted into a database from which audience can query tactics they are interested in.