摘要
目前基于视觉的动态手势识别问题仍是研究的难点,在大多数应用背景情况下很难提高手势识别率。传统的动态手势识别手段主要是利用智能传感设备以及单个或多个摄像头进行数据采集的视觉方法来实现,效率低,准确度差。近年来,随着深度神经网络技术的快速发展,利用网络自主学习的方法来提取手势姿态有关特征得到了广泛关注。本文针对传统动态手势识别准确率低的问题构建了Inception-CNN网络和LSTM网络融合的方法。在Cambridge-Gesture、VIVA以及Sheffield Kinect Gesture Dataset(SKIG)三个动态手势数据集上实验结果表明融合Inception-LSTM级联网络的识别率高,与现有的传统方法和当下流行的多种卷积神经网络方法相比,本文手势平均识别率和各个类别的手势识别率均高于现有方法,充分证明了本文方法的有效性和鲁棒性。
At present,the problem of dynamic gesture recognition based on visio is still a difficult point of research.It is difficult to improve the gesture recognition rate in most application backgrounds.The traditional dynamic gesture recognition method is mainly using the smart sensing device and the visual method of single or multiple cameras for data acquisition.It is low efficiency and poor accuracy.In recent years,the deep neural network technology rapids development.Using the advantages of network autonomous learning to extract gestures relevant features has received extensive attention.For the problem of traditional dynamic gesture recognition low accuracy,this paper builds an Inception-CNN network and LSTM network fusion method.The experimental results on three dynamic gesture da-tasets of Cambridge-Gesture,VIVA and Sheffield Kinect Gesture Dataset(SKIG)show that the Inception-LSTM cas-cade network has better adaptability.Comparing with existing traditional methods and a variety of popular convolu-tional neural network methods.The gesture average recognition rate and the gesture recognition rate in each category are higher than the existing methods,fully proving the validity and robustness of the proposed method.
作者
张国山
赵阳
ZHANG Guo-shan;ZHAO Yang(School of electrical automation and information engineering,Tianjin University,Tianjin 300072,China)
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2021年第4期373-381,共9页
Journal of Optoelectronics·Laser
基金
国家自然科学基金(61473202)资助项目。
