| Among various indoor positioning technologies,visual SLAM has the advantages of strong autonomy,low cost,and rich perception information.At present,there have been many researches on visual localization based on single features.However,in the indoor weak texture environment,limited by the high dependence on single features,complementary information between different types of features cannot be effectively utilized,and the adaptability and antiinterference capability after feature fusion are insufficient.Therefore,the fusion of information between features to achieve high-precision and reliable localization in weak texture environment is an urgent problem in indoor visual localization.In order to solve the above problems,this paper proposes a visual localization method combining point and plane features.This method effectively uses the structured plane and point feature information in the indoor environment,estimates the RGB-D camera pose based on the matching relationship of the point and plane features,and constructs a comprehensive map containing point and plane features at the same time,thus improving the localization accuracy and stability in indoor weak texture environment.The specific research contents and contributions are as follows:(1)The performance of joint bilateral filtering applied to the preprocessing of RGB-D camera perceptual information is verified.The joint bilateral filtering algorithm guided by color images is introduced to process RGB-D depth images,and compared with other three commonly used filtering algorithms in denoising experiments.The results show that the PSNR value of the joint bilateral filtering algorithm is the largest,which is about 38.960.It has the best denoising and repairing effect on depth images,thus effectively reducing the noise of candidate plane regions.(2)The effective method of expressing environmental information by point-plane features is discussed.Owing to that indoor weak texture scene has few point features,ORB point features with comprehensive speed advantage and matching accuracy are applied to carry out violent matching,and misregistration point pairs are eliminated by combining RANSAC algorithm.For structured point cloud generated by depth image,plane feature segmentation is realized based on AHC algorithm,and the extracted plane is uniquely represented by normalized Hessian vector.The plane matching conditions based on geometric constraints are set,and the plane mismatch pairs are eliminated by using consistent matching factors.(3)An indoor localization algorithm PP-RGB-D SLAM based on point-plane feature fusion is proposed.A factor graph optimization model is established by combining point-plane features and plane parallel vertical structural constraints,and Jacobian analytic form of pointplane features is derived.The keyframe selection method based on temporal and spatial rules and image similarity ensures the difference of keyframes and the richness and traceability of image information.Using g2 o optimization library,factor map optimization is completed,map optimization update management and loop closure including local map and global map are realized,and globally consistent sparse point-plane comprehensive map is generated.(4)Experimental verification and analysis of three indoor datasets and two sets of measured data were carried out.Compared with the existing typical algorithms,the experiments verified the wide applicability,higher positioning accuracy and stability of the proposed algorithm.Compared with ORB-SLAM2,the average percentage reduction of ATE RMSE of PP-RGB-D SLAM is about 33.75% on TUM RGB-D test data sequence and 57.12% on ICLNUIM dataset.The percentage reduction of loop error on TAMU RGB-D dataset is about44.76%.When ORB-SLAM2 is tracking lost,the algorithm in this paper can still keep stable trajectory tracking,pose calculation and mapping.For measured data,the algorithm in this paper has reliable tracking and pose estimation,and effectively constructs a sparse point-plane comprehensive map. |
PDF Full Text Request