【作者】 王伟；

【导师】 付兴武；

【作者基本信息】 辽宁工程技术大学 ， 软件工程（专业学位）， 2017， 硕士

【摘要】 根据图像中二维人体关节点坐标重建三维人体姿势时常会遇到一个非凸优化问题,给人体姿势的求解带来很大的困难。三维人体姿势估计的凸松弛方法虽然解决了该问题,但是存在算法的收敛速度较慢、姿势估计准确度不高的不足。针对这些不足,提出一种基于遗传优化的自适应凸松弛三维人体姿势估计改进算法。首先根据自适应控制的思想,利用误差评价函数对原凸松弛方法中控制算法迭代步数的变量更新方式进行自适应处理,然后再利用遗传优化算法为该变量的初始值进行寻优,最后利用寻优结果对原凸松弛方法中影响姿势估计准确度的闭式解表达式进行改进。实验结果表明,提出的三维人体姿势估计改进算法对单幅人体图像的处理速度更快,姿势估计准确度更高,更有利于实际应用。更多还原

【Abstract】 When reconstructing 3D human pose based on 2D human joint point coordinates in images,a nonconvex optimization problem is often encountered.It brings great difficulties to the solution of human pose.Although the convex relaxation approach for 3D human pose estimation has been solved this problem,its convergence speed is slow and accuracy of human pose estimation is not high.To overcome these shortcomings,this article proposes an adaptive convex relaxation approach based on genetic optimization for 3D human pose estimation.Firstly,adaptive processing is performed on the variable which controls iterative step in the original convex relaxation approach by using error evaluation function according to the idea of adaptive control.Then it uses the Genetic Algorithm to optimize initial value of the variable.Finally it improves the expression of closed-form solution in the original convex relaxation approach by using the result of optimization.The experimental results show that the improved 3D human pose estimation algorithm has faster speed and higher accuracy of human pose estimation results when processing single human image,which could be more beneficial to practical applications.更多还原