视频中人体行为分析  

Analysis of Human Action in Video

作　　者：史东承[1] 冯占君 

机构地区：[1]长春工业大学计算机科学与工程学院,长春130012

出　　处：《吉林大学学报（信息科学版）》2014年第5期521-527,共7页Journal of Jilin University（Information Science Edition）

基　　金：国家自然科学基金资助项目(11071026)

摘　　要：为实现机器视觉代替人眼观察、认知世界以及减少背景和噪声对视频中人体特征提取的影响,以提高识别效果,在研究人体动作表征与识别的基础上,充分考虑局部和全局特征的优缺点,提出了基于局部时空兴趣点和全局累积边缘图像特征相结合的人体行为分析方法.首先,从视频序列中提取局部时空兴趣点和全局累积边缘图像特征;然后用加权字典向量法将两者有机地结合在一起;最后利用最近距离法进行人体行为分析和识别.该方法可有效获得人体时空特征、人体边缘轮廓、人的运动趋势和强烈程度.实验结果表明,该方法快速,相比其他算法识别率大致提高了2％～5％.In order to solve machine vision reduce the effects of background and noise on the video feature extraction,and to improve recognition results.We propose a human action analysis method by combining the local space-time interest points and global accumulated edge image feature based on the study of human action representation and the full consideration of the advantages and disadvantage of the local and global features.First,local space-time points of interest and global accumulated edge image feature are extracted from the video sequences.Then we use the weighted dictionary to combine them together organically,finally we use the minimums distance method for human action analysis and identification.This method can effectively obtain the spatial characteristics of human,human edge contour and the human movement trends and intensity,the experimental show that our method is faster and gains a higher recognition accuracy generally inceased by 2％ to 5％.

关 键 词：时空兴趣点 累积边缘图像 行为分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TG156[自动化与计算机技术—计算机科学与技术]