中文摘要：

基于稀疏运动矢量场,提出一种动态背景下的运动目标区域检测方法。根据运动矢量场特性分析进行全局运动参数估计和全局运动补偿,实现动态场景中的背景校正;利用最大树数据结构,基于运动矢量补偿误差分级表示视频帧中运动基本一致的连通区域,进行运动区域初始分类;根据运动目标在空间上的连通性和运动一致性的特点,选择区域相似性度量准则,进行区域合并和滤波,将具有相似运动的连通区域合并,实现运动目标区域检测。将检测出的运动目标区域作为运动矢量外点反过来又应用于全局运动参数估计过程中,全局运动估计和运动目标区域检测交替地进行,不仅加快了它们的计算速度,同时也提高了它们计算和检测的准确性。实验结果表明,本文算法能较好地补偿序列的全局运动,有效地检测出局部目标运动区域。

英文摘要：

A new method using the coarsely sampled motion vector fields to detect moving object regions from dynamic background is proposed. The proposed method estimates global motion parameters by analyzing properties of motion vectors, and compensates the global motion for removing background motion in motion vector fields. The motion vector compensated errors generated from global motion compensation are then utilized to create max-tree data structure representation for video frames. The connected regions that have almost the same movement in video frames are hierarchically represented,and the initial classifications of motion areas are completed. According to the spatial connectivity and motion consistency of the same motion object, the regional similarity measure criteria are selected and utilized for regions splitting and merging. The connected areas with similar movement are merged, and the moving object regions are then detected. The segmented moving object regions are then treated as outliers and rejected in the next round of global motion estimation. The alternation between global motion estimation and motion segmentation not only speeds up the computation, but also improves the precision of calculation and detectioru Experimental results demonstrate that the proposed approach can effectively compensate the background motion and detect moving object regions from dynamic background.