基于时域循环增强图卷积网络的手势识别  

Gesture Recognition Based on Time-Domain Cyclic Enhanced Graph Convolutional Networks

作　　者：高广荣 李云峰[1] GAO Guang-rong;LI Yun-feng(School of Mechatronics Engineering,Henan University of Science and Technology,Luoyang 471003,China)

机构地区：[1]河南科技大学机电工程学院,洛阳471003

出　　处：《印刷与数字媒体技术研究》2025年第2期201-212,共12页Printing and Digital Media Technology Study

基　　金：河南省重大科技专项(No.221100220100)。

摘　　要：针对现有手势识别方法中存在的网络模型参数量大、识别过程中忽略了帧间差异、无法充分捕获视频流关键帧细节特征等问题,本研究提出了一种基于时域循环增强图卷积网络(Time-Domain Cyclic Enhanced GraphConvolutionalNetworks, TCE-GCN)的手势识别方法。首先,采用VGG-19网络的前10层提取图像特征,结合轻量型OpenPose提取手势骨骼的特征关键点坐标,充分关注手势的全局图像特征和局部骨骼信息。其次,引入时域循环增强模块,用于双流自适应图卷积网络中捕捉手势的演变,提高对帧与帧之间手势差异的理解。最后,引入坐标注意力模块,利用坐标信息进一步捕获手势的位置信息,提高模型的泛化能力。该算法在Jester手势数据集上准确率为95.9%,在SHREC’17数据集上识别准确率为96.3%,相对于其他基线算法有更高的手势识别精度。Aiming at the problem that the existing gesture recognition methods have a large number of network model parameters,ignoring the inter-frame differences in the recognition process,and failing to capture the detailed features of the key frames of the video stream,a gesture recognition method based on Time-Domain Cyclic Enhanced Graph Convolutional Networks(TCE-GCN)was proposed in this study.Firstly,the first ten layers of VGG-19 network were used to extract image features,and the coordinates of the feature keypoints of the gesture bones were extracted by combining with the lightweight OpenPose,which paid full attention to the global image features of the gesture and the local bone information.Secondly,a time-domain cyclic enhancement module was introduced for dual-stream adaptive graph convolutional networks to capture the evolution of gestures and improve the understanding of the differences in gestures between frames.Finally,a coordinate attention module was introduced to further capture the positional information of the gesture using coordinate information to improve the generalization ability of the model.The algorithm has an accuracy of 95.9%on the Jester gesture dataset and 96.3%recognition accuracy on the SHREC’17 dataset,which provides higher gesture recognition accuracy relative to other baseline algorithms.

关 键 词：手势识别 轻量型OpenPose 时域循环增强 坐标注意力 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]