基于神经网络空时特征提取的动态手势识别

针对现有的动态手势识别3D卷积方法计算参数量大和对2D卷积长时间序列的空时特征难以提取的问题，提出一种基于2D卷积神经网络和长短期记忆网络相结合的提取时空域特征的动态手势识别方法。首先基于2D卷积神经网络提取空域特征，再通过长短期记忆网络进行序列图像时序上的相互关联提取时间维度上的信息。为验证算法的有效性，使用自采集的7种动态手势动作和IsoGD公开数据集对本文所提算法进行验证。实验结果表明，在线增强算法下实验在自采集的动态手势集上的识别率达到87.14%。在IsoGD公开数据集上的识别率达到57.89%，相对于现有的其他方法有所提升。

Dynamic Gesture Recognition Based on Space-time Feature Extraction of Neural Network

Aiming at the existing 3D convolution method of dynamic gesture recognition with large number of computational parameters, it is difficult to extract 2D convolution of long-time-series images in terms of time dimension. In this paper, a gesture recognition method based on the combination of 2D convolutional neural network and long and short term memory network is proposed. Firstly, spatial features are extracted based on 2D convolutional neural network, and then features in time dimension are extracted by interrelation of sequential images through long and short term memory network. In order to verify the validity of the algorithm in this paper, using the acquisition of 7 kinds of dynamic hand gestures and IsoGD public data sets to verify this proposed algorithm, the experimental results show that under using online enhancement algorithm the paper in the collection on a set of dynamic hand gestures recognition rate reached 87.14%, IsoGD public data sets on recognition rate of 57.89%, compared with the existing method, the recognition rate is improved.