Research On Key Technologies Of Monocular SLAM Based On Graph Optimization

With the development of robot technology,SLAM(Simultaneous Localization And Mapping)has become one of the key technologies for robot to realize autonomation.At the same time,with the application of unmanned driving,SLAM has drawn more and more attention from researchers.As a mainstream SLAM system,visual SLAM has some advantages over traditional laser-based SLAM in sensor's price and map building.However,in the current visual SLAM systems,there is problem of computational inefficiency,which affects the speed and real-time performance of the system.In the feature-based visual SLAM system,the main calculation consist in the features extraction and matching,building global data association,graph optimization and loop detection.In this paper,the main feature extraction algorithms were studied,and the time consumption of the feature extraction process was compared,in which we found that the ORB(Oriented Fast and Rotated BRIEF)algorithm was obviously better than the SIFT(Scale Invariant Feature Transform)and SURF(Speeded Up Robust Features)algorithm in the extraction speed.At the same time,because of the linear matching has low efficiency,this paper applied hierarchical clustering tree and LSH(Locality-Sentitive Hashing)algorithm to the matching process of ORB feature,and it had been proved by experiment that they were better than linear method in speed.This paper also compared the speed of the hierarchical clustering tree and LSH angorithm,and it was found that LSH algorithm was better in index building.Finally,this paper proposed a front-end frameword based on RANSAC(Random Sample Consensus),which can help to eliminate the error of feature matching.In this paper,we also proposed a method of loop detection based on the overall appearance of the image.First,the histogram of each image was extracted,and the image with similar appearance was selected by the fast nearest neighbor algorithm to reduce the searching scope of the loop detection.Then,based on the feature matching and location consistency check,the best loops were selected.Finally,a graph optimization model based on similarity transformation was constructed to closing the loop in the back end.Finally,a SLAM frameword based on graph optimization was designed,which included three parts:the front-end of visual odometry,the local BA(Bundle Adjustment)and the loop closing.The local BA optimized the poses of camera and the positions of points at the same time,when the pose optimization in loop closing part only optimized all poses of camera.