以投球語意單元為基底的棒球影片結構分析與階層式摘要

Abstract

在本篇論文中，我們提出一簡單的架構，僅使用少數的場景特徵如顏色及動量特徵有效率地分析MPEG-2棒球影片。我們的目標是偵測出大部分的重要事件並提供多階層式摘要。首先偵測投球場景，並將影片切割成一連串以投球場景為始的片段(投球語意單元，PSU)。接著利用主客場球衣顏色不同的特性，分析投手球衣顏色資訊找出換場事件(攻方、守方交替事件)，將影片半局為單位結構化。另外分析投球語意單元特徵，選出較長，較動態，要與球場上事件相關的內容當作重要事件。最後我們為每個選出的重要事件算分數以達成多層次摘要架構。在影片分析之後，我們能夠提供棒球影片的索引及多層次摘要內容，幫助使用者快速了解影片內容及增加影片的應用。實驗結果證實，我們方法是簡單有效的。

In this thesis, we propose a simple framework only using a few cinematic features, such as color features and motion features to analyze the MPEG-2 baseball video efficiently. Our objective is to detect most of the highlight events, and supply multi-level summaries. We first detect the pitching shots, and divide the baseball video into a sequence of segments starting with a pitching shot (Pitching Semantic Units, PSU). Then we use the property of different uniform colors between home team and visit team to find out the change events (an exchange of offensive and defensive). Moreover, we analyze the PSU information to take events that are longer, more active, and more relevant to in-filed content as the highlight events. Finally, we compute a score for each highlight event to implement the multi-level summarization framework. After the analysis process, we can provide the indices and the multi-level summaries of the baseball video to help user to comprehend the video content quickly and extend the applications of the video. Experimental results indicate that the proposed method is simple and effective.