Research On RGBD SLAM Based On Point And Line Features Fusion

Simultaneous Localization and Mapping(SLAM)is a hotspot in the field of computer vision.It is widely used in driverless vehicle,robots and AR.Visual SLAM is highly dependent on environmental information.Insufficient features or unstable tracking between images may result in system location failure.Due to the rich line features extracted in the artificial structured environment,more scene information can be provided for the visual SLAM,and the map constructed by the point line feature fusion SLAM can restore a more realistic scene structure during the mapping process.In this paper,a point line features fusion SLAM based on RGBD sensor is proposed.At the same time,point features and line features are extracted in the front-end image.In the process of front-end pose and back-end optimization,joint optimization of point and line features is used to improve the positioning accuracy and robustness of visual SLAM in low texture environment.The common methods of camera pose estimation in visual SLAM include feature method and direct method.Since the matching process of feature extraction and descriptor calculation is time-consuming,this paper attempts to use word bag model and direct method to reduce the time cost of line feature in tracking process.This paper proposes a semi-direct method using point and line features in the front-end,and compares two front-end pose estimation methods based on feature description and semidirect method feature matching.At the same time,the use of point and line features for nonlinear optimization,especially in the local bundle adjustment method and the global bundle adjustment optimization method,leads to more objects and larger error than single feature.This paper presents the analytical solution in the optimization process and compares it with the numerical solution for optimization time.In this paper,line features are not only used in frontend pose estimation,but also in back-end mapping threads and closed-loop detection.In the back-end optimization process,the keyframes and 3D point and line features transmitted from the front end need to be further filtered and optimized.Common optimization methods include local bundle adjustment method,global bundle adjustment method and pose map optimization.In this paper,the bundle adjustment method and the pose optimization method are verified in the corresponding datasets.The results show that the optimization method has obvious improvement on the back-end optimization.At the same time,the SLAM algorithm is validated for different features of point features,line features and point line features.The experimental results show that the point line fusion SLAM has obvious advantages over other single features.In the back-end composition of visual SLAM,this paper proposes a composition method based on point and line features,which focuses on comparing the difference between point line features and simple point features for 3D reconstruction.The mapping test is performed in both the dataset and the actual environment.It shows that the surface reconstruction using the point and line features is more suitable for the actual scene structure.In the composition process,as the RGB-D sensor is used in this paper,other composition methods are also tried,such as dense mapping,semi-dense mapping,and Octomap construction.Finally,the point line features fusion visual SLAM was tested in the open standard dataset and compared with other visual SLAMs.The experimental results show that the proposed algorithm has better positioning accuracy.Finally,this paper summarizes the research work carried out and looks forward to the follow-up research direction.