An Interactive Soccer Video Generation System using Scene Analysis of Multi-Camera Image

在本研究中，目的是实现一次互动直播。以足球视频为例，自动生成来自任选分配的视角方向的序列图像。当目标是在足球等大场地上进行的运动时，很难用一台摄像机捕捉到整个场景。无论使用多少摄像机，专家只选择一个摄像机进行广播。因此，需要满足观众需求的自动相机选择系统。这被称为交互式视频生成系统。研究结果如下：为了构建一个多摄像机的图像采集环境，在足球场周围安装了多个摄像机，并对采集到的图像进行了分析。转播的摄像机将根据分析的结果进行控制，利用多摄像机的图像序列对球员和球进行检测和跟踪，在一帧中检测一个球，采用模板匹配技术，并使用背景差图像来检测球员。在此基础上，提出了帧间跟踪算法，并进行了应用。提出了一种在同一摄像机拍摄的图像中，当一名球员被其他球员遮挡时，由于球员被遮挡而导致的错误纠正算法。遮挡不应该存在于其他摄像机的图像中，也不应该是对球员和球位置的估计，通过跟踪球和球员的特征值来确定最佳视角方向信息。通过对这些特征值向量的分析，给出了最佳视角方向。为了识别踢球类型的球员，检测了球周围的位置和速度信息，利用位置和速度信息来区分足球比赛，如传球、铲球等

In this research, to realize an interactive broadcasting is aimed. Soccer video is concerned as example, and sequential images from optionally assigned view direction, are generated automatically. When object is sports which is played in large fields like soccer, it is very much difficult to capture by one camera for whole scene. However many cameras are used, only one selected camera is selected by expert and broadcasted. So automatic camera selection systems yield to viewers demand are desired. This is called an interactive video generation systems. The results are as follows,To construct an environments of image capture from multi-camera Many cameras are fixed around the soccer field and those images are analyzed. Camera for broadcasting will be controlled by the results of analysis.Detection and tracking of players and ball using image sequence from multi-camera To detect a ball in one frame, template matching technique are applied, and to detect players, background difference images are used. After that, inter frame tracking algorithms are proposed and applied. The algorithms are well worked and useful to make loci of players.To develop an algorithm for error correction caused by occluded player When a player is occluded by other player in image from one camera. An occlusion should not exist in images from the other camera, or be estimated players and ball position.To decide the best view direction Information from tracking of ball and player are concerned to feature value. The analysis of those feature value vector, gives the best view directionTo recognize play type Players around the ball are detected, position and velocity information are used for distinguish soccer play such as pass, shot tackle and so on