摘要
视频中的人体运动分析是计算机视觉领域的重要课题,同时也是近年来备受关注的前沿研究方向之一.在明确实际视频中存在的若干种难点,如人体遮挡、视频模糊、拍摄视角变化等基础上,从经典的人体运动特征提取、特征选择以及特征融合3个方面,对基于视频序列的人体运动描述方法和研究现状进行了概述,归纳出人体运动描述算法的研究难点,并分析了人体运动分析的技术发展趋势.指出了利用不同特征间存在的互补性质探求高性能特征选择和特征融合机制是人体运动描述技术发展的必然趋势,从处理简单实验场景视频向挑战高难度实际场景视频的转化是运动视频分析未来发展的方向.
Recently analysis of human actions in videos has become an important issue in the field of computer vi- sion. Much attention has been paid to this frontier research. In this paper, we first explicitly defines several existing difficulties in real-world videos, such as body occlusion, video fuzzy, shooting angle change and then conducts a survey based on the popular methods and present situation research studies on human action representation. Next, we focus attention on three aspects of feature extraction, feature selection and feature fusion, and then summarize the research difficulties in algorithms of action description, and analyze the technical development trend of human action analysis. It was pointed out that the inevitable trend of human action representation technology is to explore high-performance feature selection and feature merging mechanism by making use of the complementary mechanism among different features, and the development trend of motion video analysis in the future is to change from pro- cessing simple experimental scene videos to the challenge of real-world scene videos with high difficulties.
出处
《智能系统学报》
CSCD
北大核心
2013年第3期189-198,共10页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金资助项目(60875050
60675025)
国家"863"计划资助项目(2006AA04Z247)
深圳市科学和技术创新委员会资助项目(JC201005280682A
JCYJ20120614152234873
CXC201104210010A)
关键词
视频序列
人体运动描述
特征提取
特征选择
特征融合
video sequences
human action representation
feature extraction
feature selection
feature fusion
