Gesture recognition method, apparatus, device, and computer program product
CN117133016BActive Publication Date: 2026-08-28CHINA MOBILEHANGZHOUINFORMATION TECH CO LTD +1
Patent Information
- Application Number
- CN202311071810.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-23
- Publication Date
- 2026-08-28
- Estimated Expiration
- 2043-08-23
AI Technical Summary
Technical Problem
[0003]本申请实施例提供一种手势识别方法、装置、设备及计算机程序产品，用以解决现有手势识别方案存在手势识别不准确的技术问题
Benefits of technology
[0030]本申请实施例提供的手势识别方法、装置、设备及计算机程序产品，通过GNN(Graph Neural Networks，图神经网络)来进行手部关键点的建模，与CNN(ConvolutionalNeural Network，卷积神经网络)相比，GNN能够更好地对关键点之间的联系进行建模，从而提高准确率。无论是二维还是三维，手部关键点只有21个，因此，最终构成的是一个节点数量为21的图。小型图能够极大地减少计算量；为了进一步减少计算量，本申请使用ShufflenetV2(轻量级网络)进行编码，这种轻量级的网络结构可以快速将输入图像转换为具有多个特征的向量，并对二维手部关键点进行初步预测，实现准确的预测；由于GNN的邻接矩阵无法学习，直接使用GNN来进行预测，如果一个节点预测错误，就可能导致出现两个不同的子图。为了解决该问题，本申请应用Adaptive Graph U-Net(自适应图网络)来基于二维手部关键点预测三维手部关键点。Adatpvie Graph U-Net将邻接矩阵变为网络中可以学习的参数，使网络自己学习出邻接矩阵，令网络的预测结果更加鲁棒；本申请同时预测出手部的二维和三维坐标，可以用来满足不同应用的需求。
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN117133016B_ABST
Abstract
The application relates to the field of image processing, and provides a gesture recognition method, device, equipment and computer program product. The gesture recognition method comprises the following steps: performing feature extraction on a to-be-detected image through a lightweight network to obtain a feature vector; determining two-dimensional gesture key points based on the feature vector; and predicting three-dimensional gesture key points based on an adaptive graph convolution network and the two-dimensional gesture key points. The application models hand key points through a graph neural network, encodes using a lightweight network, and applies an adaptive graph network to predict three-dimensional hand key points based on two-dimensional hand key points, so that more accurate gesture recognition, faster calculation speed and the needs of different applications are achieved.
Need to check novelty before this filing date? Find Prior Art
Citation Information
Patent Citations
Feature extraction model training method and device, hand action recognition method and device and electronic equipment
CN114973424A