基于单应性与相似度矩阵的运动分割-《计算机技术与发展》

文章信息/Info

Author(s):: WU Xiao-lun; YANG Min; School of Automation & School of Artificial Intelligence,Nanjing University of Posts and Telecommunications, Nanjing 210023,China

Keywords:: motion segmentation; homography; affinity matrix; hypothesis estimation; ordered residual kernel

摘要:: 运动分割的关键问题是根据在运动场景中提取的视觉特征,对视频序列中不同的运动物体进行聚类划分。在对物体运动进行运动分割时,通常会遇到运动物体被遮挡的问题,而传统的运动分割方法如子空间聚类等都没能很好地解决这一问题。而且对于 Hopkins155 这个数据集来说,视频的场景基本上都比较小,而且相机大多固定,所以要对这种视频进行运动分割,应该有最适合的矩阵模型去对视频的运动进行拟合。针对这些问题,该文提出了一种基于相似度矩阵和单应性矩阵的运动分割方法,用假设估计的方法与单应性矩阵模型去拟合运动场景,有序残差核计算出相似度,再利用特征轨迹之间的相似度构造相似度矩阵,最后对其进行累加以解决遮挡问题。实验结果表明,此方法在对 Hopkins155 数据集进行运动分割的实验中,取得了不错的效果。

Abstract:: The key problem of motion segmentation is to cluster and divide different moving objects in the video sequ-ence according to the visual features extracted in the motion scene. When performing motion segmentation on object motion,the problem of occlusion of moving objects is usually encountered. Traditional motion segmentation methods such as subspace clustering have not been able to solve this problem well. And for the Hopkins155 data set, video scenes are basically relatively small, and the cameras are mostly fixed. Therefore, the most suitable matrix model should be used to fit the video motion to segment such video. In response to these problems,we propose a motion segmentation method based on the affinity matrix and the homography matrix. The hypothetical estimation method and the homography matrix model are used to fit the motion scene. The ordered residual kernel calculates the similarity and the similarity between feature trajectories constructs a affinity matrix. Finally,it is accumulated to solve the occlusion problem. The experiment shows that the proposed method has achieved ideal results in the experiment of motion segmentation on the Hopkins155 data set.