Research On Visual Simultaneous Localization And Mapping Algorithm Based On ORB Features

With the popularization of intelligent robots in daily life and the development of autonomous driving technology,Simultaneous Localization and Mapping(SLAM)has received extensive attention.How mobile robots self-locate and construct maps in an unknown environment is a problem that SLAM needs to solve.At present,SLAM is one of the important topics in the research of indoor mobile robots and outdoor driverless cars.SLAM is divided into Lidar SLAM and Visual SLAM according to the type of sensor.Lidar is expensive and not suitable for popular use in daily life.Cameras are relatively cheap and with high scene recognition ability,and can obtain more texture information than the laser radar can do,so it is widely applied to indoor positioning robots.In this paper,visual SLAM is studied,the realization principle of SLAM system framework is introduced,and the characteristics of filter optimization and nonlinear optimization are discussed and analyzed.A visual SLAM system based on nonlinear optimization of ORB features is designed.The main contents of this article are as follows:Firstly,because traditional Random Sample Consensus(RANSAC)algorithm is vulnerable to noise,the accuracy of deleting mismatches is not high enough.An improved RANSAC algorithm,LO~*-RANSAC(LO~*for short),is proposed.First,the inner points generated by the conventional RANSAC are iteratively filtered to further narrow the selection range of the inner points.Secondly,the estimated model is optimized by minimizing the error with Bundle Adjustment(BA).The published data set is used in the experiment and the results are compared with some popular SLAM system.The experimental results show that the improved RANSAC algorithm can improve the positioning accuracy of SLAM compared with some popular SLAM systems.Secondly,RGB-D camera can obtain the depth of the image,and the acquired depth has noise.In order to reduce influence of depth noise on the positioning accuracy of SLAM,an adaptive graph optimization method is proposed.That is,different optimization methods are adopted for different situations.For optimization of estimated camera motion between frames,only poses are considered for local map optimization and global map optimization,poses and observation points are simultaneously optimized.A 3D-2D re-projection error model is established to selectively optimize poses and observation points by minimizing re-projection errors.Experimental results show that the proposed method improves positioning accuracy compared with traditional method of optimizing poses.Thirdly,traditional loop closure detection uses a binary dictionary to represent the weight of a word.Loop closure is determined according to the degree of similarity of images,which results in low accuracy of loop closure detection and SLAM positioning.In this regard,this paper proposes an improved loop closure detection algorithm.The weight of the word is calculated using the Frequency-Inverse Document Frequency(TF-IDF),and the similarity of the image is calculated using the Kullback-Leibler(KL)divergence.In addition,in view of the problem that the absolute similarity mechanism misdetects loop closure in similar scenes,this paper uses the relative similarity mechanism,that is,uses the relative similarity between images to judge the loop closure.Experimental results show that the improved loop closure detection algorithm improves the positioning accuracy.