Research On Positioning And Mapping Method Of Robot Based On Stereo Vision

Simultaneous Localization and Mapping(SLAM)system refers to the realization of the robot's own positioning and mapping by obtaining environmental information through sensors.The visual SLAM system is a kind of SLAM system.It mainly uses the camera to obtain image information to complete positioning and mapping.The camera has a simple structure and rich image information,which has broad development prospects.There are still some shortcomings in the research of visual SLAM algorithms.So the paper improves and innovates the algorithm of feature extraction,matching and loop detection,processes the environment information obtained by the binocular camera,construct the corresponding sparse point map to complete the real-time positioning operation.In this research,the main contents are as follows: design of the binocular vision SLAM system framework: front-end,back-end,loop detection;in the front-end part,it is the image information collected by the binocular camera to achieve feature point extraction and matching;The back-end part mainly adopts BA graph optimization,adopts local optimization of pose and spatial points,and uses loop detection to realize global optimization,and establishes sparse point graph;the loop closure detection part uses an improved convolutional neural network method to correct the global map to ensure the accuracy of the map.The most efficient means to achieve the accuracy of feature extraction and matching in the front-end is mainly the ORB feature algorithm.Aiming at the deficiencies of the original ORB feature algorithm,such as edge detection performance and data point redundancy,the o FAST corner extraction algorithm in the ORB feature algorithm is improved,and the dynamic threshold is customized to improve the condition of feature point selection and promote the accuracy of matching;In feature matching process,the system uses brute force matching for preliminary matching,and uses Hamming distance and RANSAC algorithm to eliminate mismatch points,which improves the overall performance of the system.Usually the local position error is the main problem in the back end.This paper uses the BA diagram optimization method to jointly optimize the pose point and the space point to reduce the incidence of local error.In order to further improve the accuracy of the map,this paper improves the loop closure detection method based on the convolutional neural network,uses the VGG16 model to extract the image convolution features,and then uses Isomap to reduce the dimensionality of the image data.Finally,a formula is introduced to score the key frames to determine whether the system is looping.The experimental results in the EUROC-V203 data set and the actual environment show that the accuracy and recall rate of the loop detection algorithm are higher than the traditional bag-of-words model;at the same time,the established sparse point map also illustrates the applicability and stability of the system in realizing real-time positioning and mapping.