基于多级稠密块沙漏网络的人体姿态估计

Multistage dense block hourglass network for human pose estimation

为解决基于卷积神经网络进行人体姿态估计只使用最后一层特征、忽略卷积层之间的联系从而造成信息丢失的问题,提出多级稠密块沙漏网络。多级稠密块沙漏网络在基于堆多尺度捕获并整合人体关节点的堆叠沙漏网络基础上,引进稠密网络。稠密网络具有每层互相连接这一特性,可以极大减少堆叠沙漏网络中的信息丢失,提高识别精度。实验证明:该网络结构在两个国际公认标准人体姿态估计数据集FLIC和MPII上的评价结果都优于目前广泛应用的堆叠沙漏网络。

In order to solve the problem of information loss caused by using only the last feature layer and ignoring the connection between convolution layers in human pose estimation based on convolution neural network,a multilevel dense block hourglass network is proposed. Dense network has the characteristic of interconnecting each layer,which can greatly reduce the information loss in stacked hourglass network and improve the recognition precision. Experiments show that the evaluation results on FLIC and MPII are prior to stacked hourglass networks which are widely used at present.