多策略驱动的动态手势识别算法
A multi- strategy driven algorithm for dynamic gesture recognition
投稿时间:2020-06-04  修订日期:2020-06-04
DOI:
中文关键词: 动态手势识别  多模输入  卷积神经网络  自适应融合策略
英文关键词: dynamic gesture recognition  multi-mode input  Convolutional Neural Networks  adaptive fusion strategy
基金项目:国家自然科学基金资助项目(61701548),中央高校基本科研业务费专项资金资助项目(CZT20003),企业委托项目资助(HZY18010)
作者单位E-mail
项俊 中南民族大学 电子信息工程学院 junxiang@scuec.edu.cn 
王超 中南民族大学 电子信息工程学院  
沙洁 中南民族大学 电子信息工程学院  
麻建 中南民族大学 电子信息工程学院  
摘要点击次数: 313
全文下载次数: 0
中文摘要:
      在动态手势识别方法中,现有多模输入策略易忽略手势类别局部信息与全局信息的差异性,同时现有决策级融合对各通道数据预测结果采用均值操作,不能充分利用在识别中起关键性作用的特征信息。本文提出了一种多线索自适应驱动算法:采用多线索输入策略融合各模态数据的局部与全局信息,构建2D与3D卷积神经网络级联框架提取手势特征,设计自适应融合策略为各通道预测结果赋予不同的权重。在SKIG数据集和isoGD公开数据集上的实验结果证明了该算法的有效性,取得了与当前主流算法相当的动态手势识别效果。
英文摘要:
      In dynamic gesture recognition, existing multi-mode input strategy ignores the difference between local and global information about gesture category. Furthermore, current decision-level fusion schemes perform averaging over each channel’s prediction, which is unable to emphasize the important features in recognition. To address the above issues, a multi-strategy driven algorithm is proposed. A multi-cue input scheme is adopted to fuse local and global information in each input mode. 2 Dimension Convolutional Neural Networks (2DCNNs) and 3DCNNs are then cascaded for feature extraction. Finally, an adaptive fusion strategy is designed to assign different weights to each channel’s prediction. Experiments conducted on both SKIG and isoGD datasets demonstrate the validity of the proposed method, which achieves competitive performance when compared with several state-of-the-art dynamic gesture recognition methods.
View Fulltext   查看/发表评论  下载PDF阅读器
关闭