摘要
针对人体行为识别提出一种基于深度学习的方法,使用CNN和LSTM以及MLP来构建的模型。用CNN提取视频的空间信息,LSTM提取视频的时间信息,使用MLP实现最后的分类,为提高训练速度,对视频剪辑进行稀疏下采样预处理。该模型在UCF-101数据集上达到了令人满意的效果,在与该领域中的同类算法比较中表现优异。
A deep learning-based method for the human action recognition was proposed,CNN,LSTM and MLP were used to construct the model.The CNN and LSTM were used to extract spatial and temporal information of videos respectively.MLP was used to achieve the final classification.The video clips were preprocessed by sparse down-sampling to improve the training speed.The model achieves satisfactory results on the UCF-101 data set and performs well in comparison with similar algorithms in the field.
作者
陈煜平
邱卫根
CHEN Yu-ping;QIU Wei-gen(School of Computers,Guangdong University of Technology,Guangzhou 510006,China)
出处
《计算机工程与设计》
北大核心
2019年第5期1445-1450,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61572142)
广东省科技计划基金项目(14ZK0180)
