Information Processing Of Visual Perception Uncertainty And Its Research In Mobile Robot Localization

One of the most important tasks of mobile robot to complete the task safely is to acquire the information of working environment through the sensor measurement,and then extract meaningful information from these measurements,which is the elementary aspect for mobile robot to establish the environment model and determine its localization.The traditional observation model of mobile robot is based on the ideal assumption,for example,in vision-based perception methods,it is usually assumed that environmental noise and measurement noise are isotropic,independent and identically distributed Gaussian noise,and can not fully reflect the influence of multi-source uncertainty such as environmental changes and sensor noise during the perception process.In order to ensure the accuracy of mobile robot's perception of the environment under the interference of multi-source uncertainty and improve the adaptability of mobile robot to the perception of uncertain information,it is of great significance to study the fusion method of multi-source uncertainty for improving the ability for mobile robots to perceive.Vision-based perception methods are mainly divided into direct method and feature-based method.Point feature is widely used in feature-based method because of its simple extraction process and fast matching process.Considering that the multi-source uncertainty affects the imaging mode,and then affects the feature point extraction and matching,this thesis takes the mobile robot visual perception system as the research object,and studies how to integrate the multi-source uncertainty in the visual perception process into the feature-point-based method,so as to improve the mobile robot's adaptability to uncertainty.The main contents of the thesis are as follows:1.Vision perception system modeling considering the depth uncertainty and direction uncertainty of feature points.Camera noise affects the depth estimation of feature points,and the depth uncertainty model of feature points is established by Gaussian Mixture Model in the neighborhood of feature points;environmental noise affects the imaging mode and the detection of feature points,and the direction uncertainty model of feature points is established by the affine matrix in the thesis.At the same time,the depth uncertainty model is integrated into the direction uncertainty model.Finally,a multi-source uncertainty model is established for the feature points in the thesis.2.The evidential fusion method of feature point matching considering multi-source similarity index.The descriptor of the traditional feature point ignores the other information of the neighborhood of feature points.Under the influence of multi-source uncertainty,they lack the multi-source evaluation index of feature point similarity.According to the advantages of evidence theory in dealing with multi-source information,this thesis establishes a multi-source evaluation index to measure the similarity of feature points,and then adopts evidence combination rules to fuse the multi-source evaluation indexes to obtain the basic probability assignment(BPA)function between matching point pairs.Lastly,the matching point pairs are selected according to the fused BPA.Aiming at the limitation of traditional evidence conflict factor to express evidence conflict,this thesis proposes an improved conflict degree measure,which is used in the process of selecting matching point pairs.Given the initial value of motion,the virtual feature points are as the center for searching.3.Construction of objective function for mobile robot pose estimation considering multi-source uncertainty.On the basis of the matching point pairs obtained by evidential fusion method,combined with the established the multi-source uncertainty model of the feature point,the multi-source uncertainty is intergrated into the objective function of the pose estimation,so that the original objective function could fully consider the influence of multiple uncertainty by the calculation of rotation matrix and translation vector,as well as can adapt to varying degrees of mixed uncertainty.The thesis establishes a multi-source mixed uncertainty model of feature points,simulates the pose estimation process under multi-source uncertainty,and uses the relative error of the rotation matrix and translation vector to measure the accuracy of the pose estimation,and verifies the feasibility of the peoposed method.