利用隱藏式馬可夫模型之棒球精彩事件短片偵測

Abstract

近年來棒球影像分析已有相當多的研究成果，但是對於精彩內容意涵事件偵測分類等影像處理技術的分析細膩度尚嫌不足。而這篇論文提出一個有效且有效率的棒球影片精彩內容意涵事件偵測分類系統，利用棒球影片中特定的場地規格及規律的場景轉換，使得此系統能夠辨識正在進行的精彩內容意涵事件。為了達到此一目標，本論文提出的系統概述如下:首先，我們可將一部棒球影片切割成精采短片，而每一個精彩短片開始於投手投球並結束於某些特定結束畫面，利用短片中每張畫面出現的物件特徵及順序，進而利用擷取出來的資訊用來發展以隱藏式馬可夫模型所建立的十二種精彩短片事件分類器。我們在此篇論文提出更精確的特徵擷取，使得棒球精采短片在分類上有著高準確性並且多樣性。此方法在運算上非常有效率，更重要的是，各種精采短片分類的實驗數據結果顯示相當良好的系統效能及準確性。

Despite a lot of research efforts in baseball video processing in recent years, little work has been done in analyzing the detailed semantic baseball event detection. This thesis presents an effective and efficient baseball event classification system for broadcast baseball videos. Utilizing the strictly-defined specifications of the baseball field and the regularity of shot transition, the system recognizes highlight in video clips and identifies what semantic baseball event of the baseball clips is currently proceeding. The semantic exploring system is proposed to achieve the objective. First, a video is segmented into several highlights starting with a PC (Pitcher and Catcher) shot and ending up with a close-up or some specific shots. Before every baseball event classifier is designed, several novel schemes including some specific features such as soil percentage and objects extraction such as first base are applied. The extracted midlevel cues are used to develop baseball event classifiers based on an HMM (Hidden Markov model). Due to specific features detection the proposed method not only improves the accuracy of the highlight classifier but also supports variety types of the baseball events. The proposed approach is very efficient. More importantly, the simulation results show that the classification of twelve significant baseball highlights is very promising.