融合多层卷积特征的双视点手势识别技术研究

Dual-views Hand Gesture Recognition Using Fusion Features of Multi-convolution Layers

在人机交互技术领域,基于视觉的手部交互技术凭借其良好的舒适性和自然性被广泛研究和应用.手势识别是手势交互技术的核心内容之一.本文提出一种基于深度学习网络的识别方法,构建双视点网络框架,采用支持向量机对各视点下提取的特征进行分类识别,降低手势自遮挡的影响,提高识别精度;同时对各视点卷积网络,根据训练样本卷积特征的累计贡献率实现不同深度层的卷积特征的融合,补充深层网络丢失的浅层特征信息,增强特征鲁棒性.实验结果表明,较传统方法本文方法能有效提高手势识别准确率,同时基于预训练的学习方法能有效提高手势识别的时间效率.

In the field of human-computer interaction,the technology of vision-based hand interaction is widely researched and applied due to its comfortability and naturalness. The hand gesture recognition is one of the core technology about hand gesture interaction.This paper proposes a recognition method based on deep learning to construct the network frame in dual views,which adopts Support Vector Machine( SVM) classifier to recognize the extracted features from each view,to enhance the recognition accuracy by reducing effects of self-occlusion. Meanwhile,for each network in each view,aiming to enhance robustness of the extracted features,this method concatenates the features from multiple convolution layers,according to their accumulative contribution rate,for supplementing the lost information from the shallowlayers. Experiments shows that the method proposed in this paper can effectively improve the accuracy for hand gesture recognition,and improve the efficiency of recognition due to its pre-trained method.