Scholarlink

<p>Traditional RGB-based hand gesture recognition is easily affected by problems such as light intensity and viewing angle. Bone-based hand gesture recognition is less affected by these problems and has become one of the mainstream methods. However, at present, the hand gesture recognition method, which is based on bone information, has many parameters, and the operation speed is slow. To solve these problems, a lightweight graph convolutional hand gesture recognition framework based on multistream fusion is proposed. First, the data fusing various types of hand joint, bone edge, joint motion and bone motion information are input into the spatial graph convolution module. Second, the spatial attention mechanism was added to the spatial graph convolution module to better extract the relationship between each joint. Finally, deep convolution and pointwise convolution are used in the temporal convolution module to reduce the number of parameters. Compared with the baseline network Semantic-Guided Network (SGN), the proposed network is improved by 2.3% under the 14-gesture setting evaluation, 1.9% under the 28-gesture setting evaluation, and the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\:0.12\times\:{10}^{6\:}\)</EquationSource> </InlineEquation> parameters are reduced under the SHREC 17 dataset, which verifies the effectiveness of the proposed network.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Lightweight Graph Convolutional Hand Gesture Recognition Algorithm Based on Multistream Fusion

  • 期刊论文

  • 发表日期：

    2025年11月24日

查看全文
  • Isack Bulugu

摘要

Traditional RGB-based hand gesture recognition is easily affected by problems such as light intensity and viewing angle. Bone-based hand gesture recognition is less affected by these problems and has become one of the mainstream methods. However, at present, the hand gesture recognition method, which is based on bone information, has many parameters, and the operation speed is slow. To solve these problems, a lightweight graph convolutional hand gesture recognition framework based on multistream fusion is proposed. First, the data fusing various types of hand joint, bone edge, joint motion and bone motion information are input into the spatial graph convolution module. Second, the spatial attention mechanism was added to the spatial graph convolution module to better extract the relationship between each joint. Finally, deep convolution and pointwise convolution are used in the temporal convolution module to reduce the number of parameters. Compared with the baseline network Semantic-Guided Network (SGN), the proposed network is improved by 2.3% under the 14-gesture setting evaluation, 1.9% under the 28-gesture setting evaluation, and the \(\:0.12\times\:{10}^{6\:}\) parameters are reduced under the SHREC 17 dataset, which verifies the effectiveness of the proposed network.