DYNAMIC SCENE ANALYSIS OF GROUP GAME

（一）.群体行为的定性分析从群体关系的时间转换角度对足球比赛中的群体行为进行了分析。通过直方图反投影将游戏场景中的玩家群体描述为分布。通过尺度空间滤波对分布图进行平滑，并计算出静态的群层次结构。我们将场景中的群体行为标注为群体层级的时间变迁。（二）、运动员动作的定性分析通过检测X-Y-T空间中的奇异点，从图像序列中识别出运动员的动作。对播放器图像序列进行二值化处理，并利用时间连续性细化算法提取三维骨架结构。基于骨架中的交叉点和分支来注释动作的特征。通过图像间的时间相关性检测周期性参数的变化，分割出周期性动作。（三）、开发了基于Active-NetPlayr的游戏场景中的Playr提取方法。在该算法中，主动网络应用于直方图反投影图像，以找到多色实体的边界。调整直方图反投影以描述制服的颜色。本文还提出了一种参数自动调整的方法来提高收敛速度，并提出了一种精确提取形状的网格化算法。

(1). Qualitative analysis of group behaviorGroup behavior of soccer game is analyzed from temporal transition of group relation. Playr groups in the game scene are described as the distribution of by Histogram Backprojection. The distribution maps are smoothed by scale space filtering and a static structure of group hierarchy is computed. We annotated the group behavior in the scene as the temporal transition of group hierarchy.(2). Qualtative analysis of playr's actionThe action of a playr is identified from a sequence of image by examining singular points in X-Y-T space. Image sequence of a playr is binarized and the 3-D skeletal structure is detedted by temporal-continuity-preserving thinning algorithm. The characteristics of the action is annotated based on crossing points and branches in the skeleton. Periodical actions are segmented from the change of the periodicity parameter detected by temporal correlation between images.(3). Playr extraction by Active-NetPlayr extraction method from game scene has been developed. In the algorithm, the Active-Net is applied to Histogram Backprojection images to find the boundary of multi-colored entities. Histogram Backprojection is tuned up to descriminate the colors of uniforms. We have also proposed an automatic parameter adjustment method to emprove the speed to converge, and a net-tering algorthm for exact shape extraction.