Visual Reconstruction And Relocalization Based On Stereo Vision Of Mobile Robotics

Scene reconstruction and relocalization based on mobile robots is an important basis for the subsequent path planning,navigation and tracking problems,and it is essential to construct an efficient,high-precision and robust global offline maps.Whether in industry or academia,the use of redundant observations provided by binocular cameras can greatly improve the reliability and accuracy of visual odometry.Although visual odometry has made a lot of progress compared to the past,in the two fields of visual offline mapping and visual relocation,the existing open source solutions in the industry have not taken full advantage of the advantages of multiple sensors,which has led to inconsistent scene composition.Accurate,not robust enough for relocation.Therefore,the vision-based navigation scheme is still far from practical.In this paper,based on binocular sensors,combined with two types of tasks,sparse 3D reconstruction and visual relocation in mobile robot scenes,the following work is done:First,in the case of VIO pose estimation,in order to eliminate the cumulative error of VIO,this paper proposes a multi-sensor-based fusion mapping algorithm.This algorithm uses the triangulation + RANSAC algorithm process proposed in this paper to get a better point cloud estimation,and thus a better point cloud estimation.In order to combine multi-sensor information,this paper uses VIO results and the direction of gravity acceleration to provide multi-sensor constraints,which effectively solves the problem of mapping failure under pure visual constraints.Second,in the absence of VIO pose estimation results,in order to solve the problem of pose drift in the incremental mapping link,this paper makes full use of binocular constraints to provide a complete mapping algori thm.This algorithm uses the initial map construction link provided in this article,and tightly couples the binocular constraints on the back-end constraints.The experimental results show that the algorithm effectively solves the problem of attitude drif t in incremental mapping and can provide more efficient and robust The result of the mapping.Third,in order to solve the problem of global positioning in known scenes,this paper uses the results of the previous mapping to propose a visual positioning sc heme based on binocular vision.This scheme makes full use of the expression of the Plucker line,and has a unified mathematical representation for the decentralized camera model,which can form an extension to the global positioning problem of the multi-eye camera.Experimental results show that the algorithm can make full use of the advantages of multiple cameras to perform global positioning in the scene where the map is built.Finally,the algorithm proposed in the paper tested multiple data sets,including self-collected data sets,KITTI data sets containing multiple types of scenes,and Eu Ro C data sets collected by indoor drones.The proposed algorithm has achieved good results and accuracy on these data sets.Compared with the current open source algorithms,the scheme proposed in this paper has stronger robustness and precision of the industry preface.In addition,in order to further analyze the intermediate results of the algorithm,this paper also made a user interface based on QT,which can analyze the intermediate data while satisfying the user's operation.