基于注意力引导空域图卷积SRU的动态手势识别  

Dynamic gesture recognition based on attention-guided spatial graph convolutional SRU

作　　者：陈炫琦 佘青山[1] 张波涛[1] 马玉良[1] 张建海 CHEN Xuan-qi;SHE Qing-shan;ZHANG Bo-tao;MA Yu-liang;ZHANG Jian-hai(College of Automation,Hangzhou Dianzi University,Hangzhou 310018,China;Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学自动化学院,杭州310018 [2]浙江省脑机协同智能重点实验室,杭州310018

出　　处：《控制与决策》2023年第11期3083-3092,共10页Control and Decision

基　　金：国家自然科学基金项目(61871427);浙江省重点研发计划项目(2019C04018);浙江省自然科学基金重点项目(LZ22F010003);国家重点研发计划项目(2017YFE0116800)。

摘　　要：基于手部骨骼的动态手势识别是计算机视觉和人机交互领域的一个研究热点.手势涉及的关节在空间上分布更紧密,相关性更强.针对目前基于骨骼的动态手势识别存在空间特征复杂、识别计算速率缓慢等问题,提出一种注意力引导空域图卷积简单循环单元(ASGC-SRU)网络.首先,将空域图卷积嵌入至SRU的门结构中,使得具有高速并行计算能力的SRU能够对复杂手势的时域和空域信息进行建模;然后,引入一种指关节注意力引导模块,使得更重要的指关节具有更高的关注度;最后,引入一种注意力增强空域图丢弃(ASD)的正则化方法,缓解网络过拟合的弊端.为验证所提出方法的有效性,在公认的动态手势数据集SHREC’17和DHG 14/28上进行大量实验,实验结果表明,所提出方法取得了较高的识别准确率,同时保持优良的计算效率.Dynamic gesture recognition based on hand skeleton is a research hotspot in the field of computer vision and human-computer interaction.The joints involved in gestures are more closely distributed in space and have stronger correlation.Aiming at the problems of complex spatial features and slow recognition calculation speed in skeleton-based dynamic gesture recognition,an attention-guided spatial graph convolution simple recurrent units(ASGC-SRU)network is proposed.Firstly,the spatial graph convolution is embedded into the gate structure of the SRU,so that the SRU with high-speed parallel computing ability can model the temporal and spatial information of complex gestures.Then,an attention-guided module of finger joints is introduced to make more important finger joints have higher attention.Finally,an attention-enhanced spatial dropgraph(ASD)regularization is introduced to alleviate the effect of the over-fitting problem.To verify the effectiveness of the proposed method,a series of experiments are conducted on the recognized dynamic gesture datasets SHREC’17 and DHG 14/28.The results show that the proposed method has achieved high recognition accuracy and simultaneously maintains excellent calculation efficiency.

关 键 词：动态手势识别 图卷积 简单循环单元 注意力引导 过拟合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]