基于CNN与HOG特征融合的视觉手势识别  

Visual gesture recognition based on fusion of VGG16 and HOG features

作　　者：崔劲杰 韩晶[1] 李洁[1] 杨玉兵 任兵 CUI Jinjie;HAN Jing;LI Jie;YANG Yubing;REN Bing(North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学,太原030051

出　　处：《兵器装备工程学报》2024年第12期289-297,共9页Journal of Ordnance Equipment Engineering

基　　金：中央引导地方科技发展资金项目(YDZJSX2022A024,YDZJSX2023A026)。

摘　　要：手势的多样性和复杂性会对识别的可靠性和准确性带来较大影响,而基于视觉的手势识别通常采用单一的特征来分类,但是单一的特征提取到的特征信息有限。为了解决该问题,提出了基于卷积神经网络(VGG16)与梯度方向直方图(HOG)特征融合的手势识别方法,融合后的特征包括图像的深度纹理信息和局部区域梯度方向信息,以一对一方式构建组合式SVM分类器完成手势识别模型的训练和检验。实验结果表明:在公开的American Sign Language(ASL)数据集测试下,融合后的特征提取分类识别率达到了97.86%,较HOG特征分类提高了20.89%,较VGG16特征提取方式提高了19.35%;与网络DenseNet-18,ResNet-121对比,识别率高10%左右。通过实物小车实验,验证了算法的可靠性与实用性。The diversity and complexity of gestures can have a large impact on the reliability and accuracy of the recognition,while vision-based gesture recognition usually uses a single feature used for classification,but the feature information extracted from a single feature is limited.Therefore,the paper proposes a gesture recognition method based on the fusion of convolutional neural network(VGG16)and histogram of gradients(HOG)features,the fused features include the depth texture information of the image and the gradient direction information of the local region,and a combinatorial SVM classifier is constructed in a one-to-one manner to complete the training and testing of the gesture recognition model.The experimental results show that the recognition rate of the fused feature extraction classification reaches 97.86%under the test of the publicly available American Sign Language(ASL)dataset,which is 20.89%higher than the HOG feature classification,and 19.35%higher than the VGG16 feature extraction approach.Compared with other networks DenseNet-18,ResNet-121,the recognition rate is about 10%higher.Finally,the physical car experiment is carried out to verify the reliability and practicability of the algorithm.

关 键 词：视觉手势识别 特征融合 VGG16 HOG 支持向量机 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP23[自动化与计算机技术—计算机科学与技术]