摘要
目前,针对深度学习的人体行为识别研究,往往采用视频中的全局信息对人体行为进行分析.然而,局部信息缺失造成的特征提取不完备,同样会导致识别精度急剧下降.由此,提出了基于多流深度学习的人体行为识别方法,将人体局部信息与全局信息相结合,通过局部不同特征的精确识别,使人体行为识别更加准确.实验表明,与现有深度学习方法相比,提出的方法在数据集UCF101和HMDB51上识别精度分别平均提高了约4.0%和6.2%.
At present,in the research of human behavior recognition based on depth learning,the global information in video is often used to analyze human behavior.However,the incomplete feature extraction caused by the lack of local information will also lead to a sharp decline in recognition accuracy.Therefore,this paper proposes a method of human behavior recognition based on multi-stream depth learning,which combines local information with global information,and makes human behavior recognition more accurate by accurately recognizing different local features.Experiments show that compared with the existing deep learning methods,the accuracy of the proposed method on datasets UCF101 and HMDB51 has increased by an average of about 4.0%and 6.2% respectively.
作者
韩雪平
吴甜甜
Han Xue-ping;Wu Tian-tian(Department of Information Engineering,Henan Polytechnic College,Zhengzhou,450018,China;Faculty of Information Techndogy,Beijing University of Technology,Beijing 100124,China)
出处
《数学的实践与认识》
北大核心
2019年第24期133-139,共7页
Mathematics in Practice and Theory
基金
国家科技支撑计划(2014BAH09F00)
河南省重点研发与推广专项支持项目(182102210085)
河南省政府决策招标课题(2018B341)
河南职业技术学院重点科研基金项目(2016-HZK-07)
关键词
深度学习
人体行为识别
局部特征
全局特征
时空信息
deep learning
human activity recognition
local feature
global featur
e spatiotemporal information
