基于深度学习与运动信息的动作识别算法  

Action recognition algorithm based on deep learning and motion information

作　　者：吴志攀[1] 郑中韦 WU Zhi-pan1 , ZHENG Zhong-wei2(1. School of hforrnation Science and Technology, Huizhou University, Huizhou 516007, China; 2. School of Computer, Guangdong University of Technology, Guangzhou 510006, Chin)

机构地区：[1]惠州学院信息科学技术学院,广东惠州516007 [2]广东工业大学计算机学院,广东广州510006

出　　处：《计算机工程与设计》2018年第8期2668-2674,共7页Computer Engineering and Design

基　　金：惠州市科技计划基金项目(2014-01)

摘　　要：为提高视频中人体小幅度动作识别的准确率以及对大规模数据集的计算效率,提出一种基于双层核极限学习机与深度学习技术的动作识别算法。在双层核极限学习机的第一层,采用线性核极限学习机学习密集轨迹特征与深度学习特征;在第二层,将密集轨迹特征与深度学习特征进行融合。在深度学习特征中,将深度视频特征与视频RGB三色特征融合作为深度学习的特征。基于大规模真实数据集与小幅度手势数据集进行仿真实验,实验结果表明,该算法对大规模数据集与小幅度的手势动作具有较高的识别准确率。To improve the accuracy of small amplitude action recognition and the computational efficiency of big scale dataset of videos,an action recognition algorithm based on the extreme learning machine and motion information was proposed.In the first layer of double layer kernel extreme learning machine,linear kernel extreme learning machine was adopted to learn dense trajectories feature and deep learning feature.In the second layer,the dense trajectories feature and deep learning feature were fused.As for the deep learning feature,the deep video feature and the RGB color feature were fused as the deep learning feature.Experimental results based on the big scale real dataset and small amplitude gesture dataset show that,the proposed algorithm shows good recognition accuracy to big scale datasets and small amplitude gestures.

关 键 词：极限学习机 密集轨迹 深度学习 卷积神经网络 运动信息 人体动作识别 大规模数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]