Research On ORB-SLAM Algorithm For Mobile Robot

With the continuous improvement of computer vision theory and the increasingly mature of AI technology,the Simultaneous Localization and Mapping(SLAM)of mobile robot has received widespread attention in recent years.The mobile robot system based on ORB(Oriented FAST and Rotated BRIEF)SLAM is one of the relatively perfect and easy-to-use systems,which has become a hotspot in the field of SLAM research.This paper mainly analyzes and researches the scene image preprocessing,visual odometer,loop closure detection and back-end optimization in the ORB-SLAM.Based on the application of mobile robots,we proposed an improved ORB-SLAM algorithm.The main work of this paper can be summarized as follows:(1)Aiming at the problem of background interference and invalid information in the scene image,we propose a pretreatment method of scene image.Based on the saliency algorithm of Aggregating Multi-level Convolutional Features(Amulet)to detect the region of interest,and optimizes it with depth information to eliminate a lot of background and invalid information in the image.Thus,it can reduce the computation of the system and improve the effect of the precision.(2)We improved and optimized the problem of the feature detection and feature matching of ORB algorithm in visual odometer.First,aiming at the problem that the ORB algorithm lacks scale invariance,a method for feature detection in a scale space with a large interval discretization is proposed to implement feature detection in scenes of image scale change.Secondly,aiming at the problem that the violent feature matching method has large errors and can't be applied to high-dimensional data,the multi-probe locality sensitive hashing(Multi-Probe LSH)algorithm is used to optimize the matching strategy,and combine the progressive sample consensus(PROSAC)algorithm to achieves more accurate matching.Then,the matching result is used as the initial value of iterative closest point(ICP)algorithm to improve the iterative efficiency,and prevent it falling into local extreme value because of initial value uncertainty,which results in the error of motion estimation.(3)Aiming at the erroneous closed loop of the closed-loop detection algorithm,we propose an improved closed loop detection algorithm based on the interest area of scene image.The algorithm combines the idea of stratified matching to enhance the similarity calculation,and improves the accuracy and recall of the closed loop.And we obtained more accurate motion trajectory and more practical 3D dense environment map through the pose graph and the Octomap.At the same time,we perform key frame extraction operations on image frames to remove redundant information and reduce the computation of loop closure detection.The experimental results show that the proposed algorithm has real-time performance and robustness,and can validly improve the positioning accuracy and operation efficiency of SLAM system.The 3D dense environment map can provide a good foundation for subsequent research and application.