摘要
为完整有效地表征行为人体中的运动强度和运动趋势,提高人体行为识别算法准确率,利用光流特征算法对像素变化比率图进行改进,并采用多帧叠加和网格法进行特征的提取。该特征包含全局动态运动信息和局部运动位置信息,能够更好地描述行为。使用距离度量学习算法得到视频片段描述特征,使用支持向量机和多任务大边界最近邻的联合分类器进行分类,实现人体行为的识别。实验结果表明,该方法识别率相较之前算法能够提高约12%,与传统算法相比也有很大提高。
To effectively characterize the moving target exercise intensity and trends and improve the accuracy of human action recognition algorithms,optical flow algorithm was used to improve pixel change ratio map.Multi-frame overlays and grid method were used to do feature extraction and description.This feature contained global motion information and local position information,which described the behavior better.The video feature representations were obtained through the distance metric learning.Support vector machine classifier and multi-task large margin nearest neighbor were combined and scoring mechanism for joint classification was used to achieve action recognition.Experimental results show that the proposed scheme can obtain great increase in the accuracy by about 12% compared to the existing methods,and also it is superior to conventional algorithms.
出处
《计算机工程与设计》
北大核心
2016年第9期2515-2519,共5页
Computer Engineering and Design
基金
国家自然科学基金项目(61161006)
关键词
行为识别
光流特征
像素变化比率图
度量学习
支持向量机
action recognition
optical flow feature
pixel change ratio map
metric learning
support vector machine
