一种改进的基于3D-BN-GRU网络的行为识别算法

An Improved Behavior Recognition Algorithm Based on 3D-BN-GRU Network

行为识别是计算机视觉研究一大热点,为了改善其计算量大、识别率低的问题,提出了一种基于三维卷积神经网络(3Dimension Convolutionnal Neural Network,3D-CNN)与门控循环单元网络(Gated Recurrent Unit,GRU)相融合的行为识别算法。该算法采用keras框架,首先对3D-CNN结构进行优化,采用把大的卷积核用若干个小的串联起来的Block结构;然后在每层卷积层后采用批量归一化处理,并添加Dropout层以提高网络泛化能力;最后与GRU网络融合,使用Softmax进行分类得出结果。实验结果表明,所设计的融合网络有较高的识别率,达到94. 5%。

Behavior recognition is a hot topic in the research about computer vision. To improve the problem of large computation but low recognition rate,a behavior recognition algorithm based on the fusion of 3D convolutional neural network(3D-CNN) and gated recurrent unit(GRU) is proposed. Keras framework is adopted by this algorithm. Firstly,the 3D-CNN structure is optimized. Then,several small convolution kernels are used to connect the large convolution kernels in series to form a Block structure. Next,batch normalization is used to process after each convolution layer. At the same time,Dropout layer is added to improve the generalization ability of the network. Finally,it is fused with GRU network and classified by Softmax to obtain the results. The experiment shows that the fusion network designed in this paper has a high recognition rate of 94.5%.