Research On Loop Closure Detection Based On Point And Line Feature For Visual SLAM

As an important part of visual SLAM(Simultaneous Localization And Mapping),loop closure detection is related to the accuracy of trajectory estimation and map construction.Loop closure detection can provide data association between current frames and historical frames,and help the robot to relocate after the tracking algorithm fails.It can improve the accuracy and stability of the whole SLAM system.At present,most of the methods of location and map construction are based on point features,while line features exist widely in various environments,and are roboust with change of illumination and viewpoint,which is more conducive to data association.Therefore,in this paper line features are used to the detect loop,with proposing a closed-loop detection algorithm based on point and line features.The main research work and achievements of this paper are as follows:1.Summarize the research background and significance of this topic,and then analyze the research status and key difficulties of visual SLAM and closed-loop detection algorithms.2.Elaborate the related principles of each part of visual SLAM framework,including sensors,visual odometry,loop closure detection,optimization and mapping,which provides theoretical basis for the following chapters.3.Study the detection and description algorithms of point and line features,and compare three different point-detecting algorithms:SURF,SIFT and ORB.The ORB algorithm with rotation invariance and fast extraction speed is selected to extract and describe points.While LBD binary descriptor is used to describe line segments detected by LSD,which improving computational efficiency.4.A hybrid dictionary tree combining point and line features is constructed and applied in the closed-loop detection part of SLAM to adapt to a wider environment.5.Considering the independence between visual words and the lack of spatial information and geometric constraints in BoW,we introduce the depth information of visual words.The depth information of feature points obtained from the front end of SLAM system is applied to the geometric verification,which reduces the complexity of the algorithm and improves the accuracy of loop closure detection.6.Through off-line construction of visual dictionary on a large number of data sets,and contrasting experiments with different algorithms in different environments,the results show that the proposed method is effective and has higher accuracy in detecting closed-loop.