籃球影片之場景偵測及其在戰術分析之應用

Abstract

運動影片之分析是近年來多媒體影像處理領域中一項重大議題，其中籃球影片由於場景與場地的複雜度較高，成為最富挑戰性的研究。目前已有相關論文利用事件偵測技術進而找尋比賽精華片段，然而對於專業籃球教練與球員，觀看籃球影片的目的則須提升至戰術分析。因此我們運用以GOP為基礎之場景變換偵測方法找出關鍵畫面，將影片切割成多個片段，並以球場主要顏色分布及片段長度作為場景分類的依據，分出近景、中景與遠景三纇片段。取出含有較多比賽資訊的遠景片段作進一步分析，利用顏色、形狀等資訊找出可能是籃球的區塊並追蹤球的軌跡，最後利用相機參數估算以及球軌跡之物理特性將二維軌跡對應到三維真實球場，並推論可能的出手點位置。

Sports video analysis has been a major issue of multimedia in recent years. For basketball videos, most researches only put emphasis on searching highlights of the game since the content of basketball video is too complicated. In order to look for more information of tactics from basketball videos, we propose a system that can automatically segment a basketball video into several clips by GOP-based scene change detection method. The length of each clip and the number of dominant color pixels of each frame could be used to classify shots into close-up view, medium view, and full-court view. We choose full-court view shots to do advanced analysis such as tracking the ball, and finding the transformation parameters from 3D real-world court to 2D image by camera calibration techniques. After that, we match the 2D ball trajectory to the corresponding coordinate in a real-world court and compute the statistics of shooting positions. Eventually we obtain information of the most possible shooting positions.