Abstract: During the research process of gesture recognition, it is difficult for the manual selection features to adapt to the variability of gestures. In view of this, this paper proposes a new method of gesture recognition, which combines color models and convolution neural network. In terms of the collected gesture image in different circumstances, firstly, it uses Gaussian skin color model to segment the gesture area. Then it takes advantage of convolution neural network to build a gesture recognition model which combines the process of the extraction with that of classification of gesture feature and simulates visual transduction and cognition, effectively avoiding the subjectivity and limitation of the manual features selection. Gesture recognition model regards the gray information in gesture area as input, and simultaneously takes advantage of the weight sharing and pooling techniques and so on to decrease the number of network weight value and lower the complexity of the model. The results of the experiment show that via the method of Convolution Neural Network（CNN）, the feature learning can be realized effectively, and the average rate of gesture recognition can reach 95% under different data sets. By comparing with the traditional methods, it shows that the method in this paper owns higher recognition rate and real-time feature.

Key words: gesture recognition, Gaussian skin color model, deep learning, convolution neural network

摘要： 在手势识别研究过程中，人工选取特征难以适应手势的多变性。提出了一种结合肤色模型和卷积神经网络的手势识别方法，对采集的不同背景下的手势图像，首先用肤色高斯模型分割出手势区域，然后采用卷积神经网络建立手势的识别模型，该模型融合了手势特征提取和分类过程，模拟视觉传导和认知，有效避免了人工特征提取的主观性和局限性。识别模型以手势区域的灰度信息为输入，同时利用权值共享和池化等技术减少网络权值个数，降低了模型的复杂度。实验结果表明，卷积神经网络（CNN）方法能够有效进行特征学习，在不同数据集下对手势的平均识别率都达到95%以上，与传统方法进行对比实验，表明该方法具有较高的识别率和实时性。

关键词: 手势识别, 高斯肤色模型, 深度学习, 卷积神经网络