Loop Closure Detection Algorithm Research In Visual SLAM Based On Stereo Camera

Simultaneous Localization and Mapping(SLAM)is a key technology for robots or other intelligent carriers to achieve their own position and orientation as well as understand the perception of the surrounding environment.Nowadays,SLAM has quite high research value for the rapid development of intelligent mobile robots and the increasing demand for navigation and positioning.Loop Closure Detection,as an important module in the visual SLAM system,corrects the global positioning trajectory by recognizing a scene that the robot or other vehicle has reached already.Loop closure detection can effectively eliminate the accumulated error of the visual navigation system,thereby significantly improving the positioning accuracy of the system and constructing a globally consistent map.It is a very practical research direction in the SLAM framework.Based on the current mainstream bag-words loop closure detection method,this paper studies how to detect the loop and how to optimize the global pose trajectory after detecting the closing loop combined with the overall structure of visual SLAM.The main contents and results are:1)Two kinds of bag of words(BOW)model algorithms based on ORB and SURF features were studied and the effect of loop detection as well as the time used were analyzed through experiments.This validates the efficiency of the ORB bag of words model in loop closure detection.2)Aiming at the problem that the traditional BOW method does not consider the spatial structure of the feature words to cause false positive detection,an improved BOW algorithm based on the depth information of co-visibility feature word is proposed.By adding the spatial depth information constraint in the original BOW algorithm,it effectively avoids the occurrence of perceptual aliasing in the loop detection process.Experiments based on datasets and self-measured data were performed to verify that the proposed improved method effectively improved the accuracy of loop closure detection.3)To undertake the research on the loop detection in the first half of this paper,two global optimization methods based on bundle adjustment and pose graph optimization are theoretically deduced to complete modeling and analysis.And through the simulation experiment based on the actual situation,the two algorithms and their respective correction effects on the camera carrier pose and the landmark coordinates are analyzed.4)Combined with the front-end visual odometry and point cloud construction in the SLAM system,multiple sets of datasets and self-measured data experiments were performed to examine the effects of proposed algorithm in positioning and building.The results show that the algorithm used in this paper has an intuitive improvement effect on the map constructed by the SLAM system,and also significantly improves the global tracking accuracy.