基于深度神经网络和投影树的高效率动作识别算法  

HIGH EFFICIENT ACTION RECOGNITION ALGORITHM BASED ON DEEP NEURAL NETWORK AND PROJECTION TREE

作　　者：郭洪涛[1] 龙娟娟[2] Guo Hongtao;Long Juanjuan(School of Information Technology,Luoyang Normal University,Luoyang 471934,Henan,China;School of Digital Media,Jiangnan University,Wuxi 214122,Jiangsu,China)

机构地区：[1]洛阳师范学院信息技术学院,河南洛阳471934 [2]江南大学数字媒体学院,江苏无锡214122

出　　处：《计算机应用与软件》2020年第4期273-279,289,共8页Computer Applications and Software

基　　金：江南大学卓越课程建设项目阶段性成果(ZYKC201813);江苏高校哲学社会科学基金指导项目(2016SJD760046)。

摘　　要：为了提高视频中动作识别的准确率和速度,提出一种基于深度神经网络和投影树的高效率动作识别算法。采用三维Harris角点检测时空域中发生显著变化的局部结构,划分动作识别的主要区域和次要区域;设计两种Siamese神经网络以及相应的损失函数,考虑连续帧间的局部一致性,学习视频的主要区域特征;为兴趣点的特征建立投影树,提高查询的匹配速度。基于公开数据集的仿真实验结果表明,该算法实现了较好的无监督学习效果,并且具有较高的效率。In order to improve the accuracy and speed of action recognition in videos,this paper proposes a high efficient action recognition algorithm based on deep neural network and projection tree.We adopted three dimensional Harris corner to detect the significant local structure variations in both space domain and time domain,and distinguished the major regions and minor regions for action recognition.Then,two types of Siamese neural networks and two respective loss functions were designed.We considered the local consistency between consecutive frames and learned the main regional features of the video.We constructed a projection tree for features of interested points to improve the speed of query matching.Simulation experimental results based on the public datasets indicate that the proposed algorithm realizes a good unsupervised learning performance,and it has a higher efficiency.

关 键 词：三维角点检测 人工神经网络 无监督学习 投影树 投票机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]