摘要：

传统二维卷积神经网络因遗漏时间维度信息导致不能识别微动手势。为此,提出一种基于视频流的微动手势识别方法。对输入视频流进行简单预处理,利用改进型多维卷积神经网络提取手势的时空特征,融合多传感器信息并通过支持向量机实现微动手势识别。实验结果表明 ,该方法对手势的背景和光照都具有较好的鲁棒性,且针对各类动态手势数据集能达到87%以上的识别准确率。

关键词: 计算机视觉, 手势识别, 二维卷积神经网络, 多维卷积神经网络, 支持向量机, 鲁棒性

Abstract:

For the traditional Two Dimensional Convolutional Neural Network(2D-CNN),the time dimension information is lost,and thus the dynamic gesture cannot be recognized.This paper proposes a novel dynamic hand gesture recognition method based on video streams,which can effectively improve the overall performance of hand gesture recognition.The input data is simply preprocessed.The spatio temporal feature extraction operation is performed by using improved Multiple Dimensional Convolutional Neural Network(MD-CNN).A multi-sensor fusion method is provided and the dynamic gesture recognition is realized by using Support Vector Machine(SVM).Experimental results show that the proposed method performs well in robustness with respect to the gesture background and illumination.Furthermore,the method achieves the high recognition accuracy beyond 87% for every kind of dynamic gesture dataset.

Key words: computer vision, hand gesture recognition, Two Dimensional Convolutional Neural Network(2D-CNN), Multiple Dimensional Convolutional Neural Network(MD-CNN), Support Vector Machine(SVM), robustness

中图分类号: 

• TP393