基于改进卷积神经网络的人体姿态估计

Human Pose Estimation Algorithm Based on Improved Convolutional Neural Network

卷积神经网络是人体姿态估计中应用最成功的深度学习模型,但仍存在着一些诸如关节搜索空间过于巨大以及不同卷积核得到的抽象特征被平等对待等缺陷。为此,提出了一种基于改进卷积神经网络的人体姿态估计算法,利用先验分布减小关节搜索空间,改进卷积神经网络结构建立新的关节外观模型。改进的网络利用单个卷积核对应的全局和局部抽象特征计算关节的初始定位概率,通过对所有卷积核对应的关节初始定位概率进行线性组合来计算关节的最终定位概率,利用线性组合中不同的权值来体现不同抽象特征在定位关节时所起的不同作用。仿真实验表明,与现有基于卷积神经网络的人体姿态估计算法相比,所提出的算法具有更低的计算复杂度和更高的估计准确度。

The convolutional neural network is the most successful deep learning model for human pose estimation, but there are still some deficiencies such as the joint search space is too large and the abstract features obtained by different convolution kernels are treated equally. For overcoming these two deficiencies, a human pose estimation algorithm based on an improved convolutional neural network is proposed, in which the prior distribution is used to reduce the joint search space and improve the structure of traditional convolutional neural network to establish a new joint appearance model. The joint initial location probability is calculated by using the global and local abstract features corresponding to the convolution kernel, and the final location probability is obtained through linear combination of the initial location probabilities, in which the different weights indicate the different effects of different abstract features. The simulation results show that the proposed algorithm has lower computational complexity and higher estimation accuracy than the human pose estimation algorithms based on the traditional convolutional neural network.