摘要
视频理解是当前计算机视觉领域中的研究热点.提高监控视频的计算效率和人体姿态的识别精度仍然是挑战,本文提出一种识别视频关键帧中人体姿态的方法.首先通过计算视频中帧间的覆盖率和失真率,提取关键帧;然后抽取关键帧中人体姿态轮廓的多种特征,建立多特征融合的姿态描述算子;在自采集和公用数据上构建标准姿态的特征库,用于训练基于支持向量机的多类分类器,以实现人体姿态的识别.实验表明,本文方法实现了11种人体运动姿态的识别,在识别效率和精度上具有令人满意的结果.
Machine understanding of video is a hot study area in computer vision. It's still keeping a challenge to improve the compu- tation efficiency for videos and recognization accuracy for human actions in surveillance. In this paper, an algorithm is proposed to recognize human body postures in key frames from a video. Firstly, key frames are extracted by computing coverage rate and distortion rate among frames in a video. Then, we construct a posture descriptor with multiple features fusion. The multiple features are extracted from posture silhouette in the key frame. Multi-categories supported vector machine classifier is trained with a standard feature library to recognize the different human postures. The feature library is constructed on the public available database and a part of selfcollected datum. The results demonstrate that 11 different human postures can be recognized using the proposed method with satisfying efficiency and accuracy.
出处
《小型微型计算机系统》
CSCD
北大核心
2014年第1期167-171,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61070114)资助
浙江省可视媒体智能处理技术研究重点实验室项目(2011E10003)资助
关键词
关键帧
姿态识别
覆盖率
失真率
多特征融合
支持向量机
key frame
posture recognition
coverage rate
distortion rate
multiple features fusion
supported vector machine