Research On Robustness Method Of Rigid Body Attitude Estimation Based On Manifold Structure

Pose estimation is widely used in many fields,such as human-computer interaction,it is need to obtain human pose information to understand human instruction;in the augmented reality technology,pose estimation is considered as one of the most basic problems;for unmanned aerial vehicle,it is also need to obtain the pose information of the aircraft to do some relevant manipulation.This paper mainly studies the rigid body pose estimation method based on manifold structure,analyzes the robustness of this method,and then proposes an improved pose estimation scheme.In this thesis,the robustness of rigid body pose estimation method based on manifold structure is discussed firstly.Because of the influence of noise and illumination,the manifold structure of the pose images will change.Therefore,a coefficient of difference between manifold structures is proposed to evaluate the manifolds of the pose images under different conditions.We consider that the larger difference of manifolds between the prior pose images and the pose images to be measured,the larger error of pose estimation.So we use this coefficient to evaluate influence of noise and illumination on the pose estimation accuracy.Experiments show that the factors of noise and illumination have great influence on pose estimation,and the robustness of this method is poor.Aiming at the poor robustness of existing method,this paper proposes an improved robust pose estimation method based on manifold structure.In order to eliminate the influence of noise and illumination,this paper applies image segmentation technique to image segmentation and preprocessing.Then,reduce the dimension of the segmented image data by manifold learning algorithm;and establish the regression relation model between the pose images and the manifold coordinates by support vector regression.For an arbitrary image of the given object,after image segmentation,the manifold coordinates are predicted by the regression model;then the pose angle can be estimated.At last,the pose images of the three-dimensional model are used as experimental data,the experiment is carried out under the conditions of Gaussian noise,salt and pepper noise,different illumination direction and light intensity.Experiments show that the proposed scheme can significantly reduce the difference of manifold between the priori images and the images to be measured.Various factors have less influence on the proposed method,which can achieve high accuracy.The scheme proposed in this paper can eliminate the interference of various factors to a great extent.The experiment proves the validity and feasibility of the method.The method has strong robustness.