基于深度学习的动态手势识别方法

摘要

针对人机交互中基于计算机视觉的多只人手的动态手势识别要求,提出一种基于深度卷积网络的高识别率动态手势识别方法.改进深度卷积网络模型中的相关参数和网络结构,并对多种改进的深度卷积网络手势识别模型做出研究,通过EgoHands手势数据集对深度卷积网络模型进行大量训练,得到一个可以识别多类别手势的高识别率动态手势识别模型.实验结果表明,上述方法可以在视频图像中识别出一张画面中4种不同手势各自区域位置及其所属类别,识别准确率达到85.9％,高于其它相关手势识别方法,识别速度较快,平均每秒可识别16.8帧图像.%To meet the dynamic and high precision gesture interaction requirement of human-computer interaction based on computer vision,this paper proposes a dynamic gesture recognition method based on deep learning.We modifyed the parameters and framework of deep learning model,researched the multiple modifying models,and then trained the model with the EgoHands dataset.The results of the experiment show that the method of deep learning model can recognize 4 different gestures' bounding boxes and each category simultaneously in videos,with the precision of 85.9％ and 16.8 frames per second,which is more precise and faster than other related gesture recognition method.