基於場景與活動分析分辨運動影片類別

Abstract

數以萬計的影片被儲存在網路上，對於使用者來說，有效率地搜尋特定的影片變成一件困難的任務，因此影片分辨技術成為一項重要的研究議題。傳統的影片分辨方法是基於分析低層次視覺特徵。然而，低層次視覺特徵較難處理高度變化的影片。在這篇論文中，我們提出一個創新的運動類影片分辨方法藉由分析影片場景與活動。我們主要分析的影片活動為運動員之移動軌跡。運動員移動軌跡可代表運動員在運動時之活動行為，這可視為分類運動影片之線索。我們提出的方法可分成兩個步驟。第一步，我們辨別出包含有用的運動員移動軌跡之影片片段。第二步，我們整合運動員移動軌跡的特性與該影片片段場景的視覺特徵用於訓練運動影片分類器。我們實驗用的運動影片有五類，分別為棒球、籃球、網球、羽球與排球。根據我們的實驗顯示，比起單純使用低層次視覺特徵的方法，我們提出的方法在整體的分辨表現上是較令人滿意的。

As hundred thousands of videos are created and communicated on Internet, it is a difficult task for people to efficiently find a certain video. Thus, the video classification has become an important research issue. The conventional methods focus on the analysis of low-level visual features. However, the low-level visual features may encounter the difficulty in dealing with high variations of videos. In this paper, we propose a novel method for sport video classification by scene and activity analysis. We mainly focus on analyzing the activities: the moving trajectories of players in sport shots. The moving trajectories of players can represent the activity behaviors of players in game playing. It is a clue for sport video categorization. The proposed method contains two main steps. First, we recognize the video shots which contain the useful information of the moving trajectories of players. Second, we utilize the properties of the moving trajectories and the visual features of the scenes to train the classifier for sport video classification. We perform the experiments for sport video classification with five categories of sports: baseball, basketball, tennis, badminton and volleyball. The overall performances of our method are more satisfactory than the methods based on low-level features.