基于多网络融合技术的人体姿态特征识别研究  

On the recognition of human postures with multi-network technology

作　　者：郝敏钗 卢嘉怡 HAO Min-chai;LU Jia-yi(School of Intelligent Manufacturing,Hebei Vocational University of Industry and Technology,Shijiazhuang,Hebei 050091,China;School of Chinese,Hainan Normal University,Haikou,Hainan 571158,China)

机构地区：[1]河北工业职业技术大学智能制造学院,河北石家庄050091 [2]海南师范大学文学院,海南海口571158

出　　处：《石家庄职业技术学院学报》2024年第6期7-14,共8页Journal of Shijiazhuang College of Applied Technology

基　　金：河北省社科基金项目(HB21TQ002)。

摘　　要：为了克服传统人体姿态特征识别方法在准确识别姿态特征方面的局限性和易导致识别精度低的问题,提出了一种基于多网络融合技术的人体姿态特征识别算法.该算法首先通过改进的VGG-19网络获取人体姿态信息,同时加入残差模块以细化关节位置和编码获取到的初步特征;其次,利用解码网络执行反卷积采样操作,以获取更大的细粒度分辨率;再次,设计了图位姿细化模块,提取并分析动作特征关键点之间的关系,以解决人体被部分遮挡时姿态特征识别准确率低的问题.在公开数据集和自建数据集上的实验结果表明,该算法在准确性和有效性方面表现较好,准确率分别达到了99%和98%.To overcome the limitations and inefficiency of traditional methods in identifying human postures,a recognition algorithm is proposed based on multi-network technology,which initially acquires relevant information through an improved VGG19 network,incorporating residual modules to refine joint positions and encode the initially obtained features.A decoding network is,then,utilized to perform deconvolution sampling operations,achieving a higher level of fine-grained resolution.Furthermore,a posture refining module is designed to extract and analyze the relationships between key action feature points,addressing the issue of low accuracy in posture feature recognition when parts are occluded.Public and self-constructed datasets are experimented,and results demonstrate that this algorithm exhibits commendable performance in terms of accuracy and effectiveness,achieving accuracy rates of 99%and 98%,respectively.

关 键 词：人体姿态 特征 识别 网络 卷积神经网络 残差模块 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]