【摘要】 手势识别技术是实现基于手势的人机交互的关键技术.但人手自由度高,结构复杂,同时由于手势自遮挡会引起信息丢失.本文在卷积网络框架下,提出了一种双视点下使用图像拼接技术的静态手势识手势识别方法.该方法使用3D模型手建立手势库,为解决双视点图像与神经网络的结合问题以及模型手建库带来的网络泛化能力不足,对手势两个视点的图像提出一种剪切拼接方法,在结合双视点信息的同时随机调整模型手的大小及位置来模拟真实情况,提高网络的泛化能力.实验结果表明,本文方法能对单视点下识别错误、存在严重自遮挡的手势进行准确的识别,同时还具有很好的泛化能力,能正确识别未加入训练集的真实手势图像.更多还原

【Abstract】 Gesture recognition technology is the key to gesture-based human-computer interaction. As the structure of the human hand is complicated,some important information will be missed when the self-occlusion appeared. Under the Convolutional Neural Network framework,this paper presents a new static gesture recognition method based on dual-views image stitching. This method first build a gesture dataset using a 3D hand model. In order to use dual-views gesture images in Convolutional Neural Network and enhance the generalization ability of the network,dual-views images were stitched and the size and position of hand model were adjusted randomly.The experimental results show that the proposed method can effectively recognize self-occlusion gestures which cannot be classified correctly using single-view methods. Furthermore,the generalization ability of this method is good enough to recognize real hand images which weren’t in the training dataset.更多还原