针对视频中的背景变化剧烈、目标尺度差异明显和视角时变性强的特点,提出一种鲁棒的针对multi⁃egocen⁃tric视频的多目标检测及匹配算法。首先,构建基于boosting方法的多目标检测模型对各视频序列中的显著目标进行粗检测,并提出一种基于局部相似度的区域优化算法对粗检测显著目标的轮廓进行优化,提高Egocentric视频中显著目标轮廓检测和定位的准确性。在显著目标检测基础上,对不同视角中的显著目标构建基于HOG特征的SVM分类器,实现多视角的多目标匹配。在Party Scene数据集上的实验验证了本文算法的有效性。%In this paper, a robust multi⁃object detection and matching algorithm for a multi⁃egocentric video is pro⁃posed by considering the characteristics of multi⁃egocentric videos, for example, sudden changes in background, and variable target scales and viewpoints. First, a multi⁃target detection model based on a boosting method is con⁃structed, to roughly detect any salient objects in the video frames. Then an optimization algorithm based on local similarity is proposed for optimizing the salient⁃object area and improving the accuracy of salient⁃object detection and localization. Finally, a SVM classifier based on HOG features is trained to realize multi⁃target matching in multi⁃egocentric videos. Experiments using Scene Party datasets show the effectiveness of the proposed method.
展开▼