一种鲁棒的Multi-Egocentric视频中的多目标检测及匹配算法  

A robust multi-object detection and matching algorithm for multi-egocentric videos

作　　者：李龙[1] 尹辉[1,2] 许宏丽[1] 欧伟奇 LI Long Yin Hui XU Hongli OU Weiqi(Department of Computer Science and Technology, Beijing Jiaotong University, Beijing 100044, China Beijing Key Lab of Transportation Data Analysis and Mining, Beijing Jiaotong University, Beijing 100044, China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]北京交通大学交通数据分析与挖掘北京市重点实验室,北京100044

出　　处：《智能系统学报》2016年第5期619-626,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61472029;61473031)

摘　　要：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-egocentric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.

关 键 词：Multi-Egocentric视频 多目标检测 多目标匹配 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]