Improvement on Pose Estimation of an Object in the Robotic Vision

Abstract

In order to solve the problem of inaccurate pose estimation by the classical direct linear transformation(DLT) and efficient perspective-n-point(EPNP) methods, an improved method for pose estimation is proposed. The computation process of the DLT method is optimized for easy solving. The nonlinear optimization is introduced to improve its accuracy. A proper cost function is proposed based on the Levenberg-Marquardt(LM) algorithm for solving Jocabian matrix easily. The Li group and Li algebra are introduced for representing the tiny transformation of the pose matrix, which simplifies the solution of Jocabian matrix and iterative process of optimization. The experimental results show that the proposed method is much more accurate than the DLT and the EPNP methods, and is more accurate than the DLT + numerical value-based LM algorithm. The total mean image re-projection error is 0.269 0 pixel. The time consuming experiments indicate that the proposed method needs less time compared with the DLT + numeriacl value-based LM algorithm. Its total average time is 67.48 ms per frame. These prove that the proposed method has comprehensively good performance in precision and time cost, which has good practical value.

Keywords

• robot
• pose estimation
• direct linear transformation
• nonlinear optimization
• camera calibration