RGB-D Simultaneous Localization And Mapping In Complicated Environments

Visual Simultaneous Localization and Mapping(SLAM)mainly uses visual sen-sors for pose estimation and map construction,and is widely used in mobile robots,augmentation/virtual reality,smart phones,and autonomous driving,etc.The depth of the image pixels can be directly obtained by RGB-D cameras,which makes the calcula-tion of depth in visual SLAM simpler.Therefore,using RGB-D cameras can effectively improve the robustness and reliability of visual SLAM system.However,the existing RGB-D SLAM method still faces the challenge of being accurate and robust in com-plicated environments.On the one hand,it is easy to locate inaccurately for RGB-D SLAM method in complicated environments.For example,it is difficult for existing approaches to perform feature matching effectively in large rotating scenes,eliminate false association caused by dynamic objects in dynamic scenes,and extract effective image features in low-textured scenes,which results in inaccurate motion estimation On the other hand,the RGB-D SLAM method is prone to failure in complicated en-vironments.For example,it is difficult for existing approaches to robustly deal with scenes,such as with few overlapping regions between frames,with moving objects,or with low textures.Therefore,in these complicated environments it is easy to track lost for mobile robotsIn order to solve the above mentioned problems,this paper studies RGB-D SLAM in complicated environments,and proposes effective approaches to improve its accuracy and robustness.The main contributions of this paper are as follows·A RGB-D SLAM approach based on point feature matching using spatial con-sistency for large rotating scenes is proposed.Firstly,image feature points are extracted and feature matching is performed based on local descriptors.Then,point feature match-ing is performed using the spatial consistency of the point features,and the matching problem is transformed into the optimization problem to solve.The combination of the local descriptors and global characteristics of feature points can reduce the false match of point features in large rotating scenes.Furthermore,it can reduce the tracking error of system and avoid tracking lost,which achieves accurate and robust visual SLAM in large rotating environments.·A RGB-D SLAM approach based on point and line features for dynamic scenes is proposed.Both point and line features are extracted form RGB images and effective point and line feature matching is performed.Point features and line features are com-bined in SLAM process.In order to reduce the influence of dynamic objects on pose estimation,static and dynamic features are distinguished by static weights of point and line features.Finally,the remaining static point features and line features are jointly used to estimate camera motion.The proposed approach can effectively solve the problem of low accuracy and low robustness of the existing RGB-D SLAM method in dynamic scenes,and improve the accuracy and robustness of visual SLAM system in dynamic environments.·A RGB-D SLAM approach based on point and plane features for dynamic scenes and low-textured scenes is proposed.Point features are extracted from the RGB images and plane features are extracted from the point cloud data.Then,effective point and plane feature matching is performed.Point features and plane features are combined in SLAM process.By take advantage of the spatial relationship between point features and plane features,student t-distribution mixture model(TMM)is established to describe the point features.Then,Expectation-Maximization(EM)algorithm is used to solve the problem.The point features are divided into static and dynamic points.Finally,the camera pose is estimated based on static point and plane features.The proposed approach reduces the tracking error and tracking failure of RGB-D SLAM in dynamic environments and low-textured environments,and improves the accuracy and robustness of visual SLAM system.