Research On Visual SLAM Based On Points And Line Features

With the development of robots and intelligent sensing technology,visual SLAM is playing an increasingly important role in the service robots industry.At present,the method based on points and lines is generally used to improve the algorithm’s representation ability and robustness in complex environments,especially in low-textured structured environments.In visual SLAM system based on points and lines,a good weight distribution scheme can greatly improve the accuracy of describing environmental information.This paper proposes the concept of globality and balance according to the degree of attention of the human visual attention mechanism to specific targets.And a visual SLAM based on points and lines is implemented on the basis of weight adaptive model for uneven features distribution environments.At the same time,we improve the stability of point features and line features in the extraction and matching process through algorithm optimization.The superiority of the proposed algorithm in pose estimation and environmental adaptability were verified by experiments.The main contents are as follows:First,the research of extraction and description algorithm based on point features.Aiming at the problem that SURF algorithm takes too long in the extraction and description process,a point feature optimization scheme based on SURF-BRIEF is proposed.First,the DOH detector in the SURF algorithm is used to detect the point information,and the corner descriptor is described using the BRIEF descriptor on the basis of determining the main direction,and then the SURF-BRIEF point feature extraction and description algorithm is implemented.Experimental verification shows that the algorithm can reduce the running cost while ensuring the stability of point feature extraction and description.The average running time was reduced by 43.21%.Secondly,the research of accurate initialization based on line features.Aiming at the problem that the end points cannot be accurately matched during the initialization of line features,which affects the accuracy of line feature initialization,an alignment initialization based on the SAD algorithm is proposed.First,the LSD-LBD algorithm is used to extract and describe line features.During the matching process,the SAD algorithm is used to accurately detect the endpoints of the line features based on the alignment initialization.The local pixel difference value obtained by the SAD algorithm is used to determine the true correspondence between the endpoints of the left and right images.The experimental comparison indirectly verifies the effectiveness of the SAD algorithm to improve the accuracy of line feature initialization.Thirdly,a weight adaptive model based on uneven distributed features is established,and a point-line fusion visual SLAM system based on the weight adaptive model is established.Firstly,we simulate the phenomenon that the human visual attention mechanism focuses more attention on specific targets,and propose the concept of focusing on globality and balance in the feature description process.Secondly,the method of region division and region growth is used to establish an adaptive weight model.During the calculation of the reprojection error of the line features,the reprojection errors of the two endpoints are calculated separately for improving the fusion of the points and line features.Then,combined with the weight adaptive model,a visual SLAM system based on scenes with uneven distributed features is established.Finally,an experiment research of the proposed visual SLAM system based on points and line features is carried.Firstly,the influence of the setting of the region division coefficient n on the weight is explored.Then,based on the selection of a reasonable region partitioning coefficient n,Kitti and Euroc datasets is used as processing objects.The pose estimation accuracy of this algorithm and ORB-VO and PL-SVO algorithms are compared and analyzed.The accuracy and robustness of the proposed algorithm in pose estimation are demonstrated.Finally,an online test is performed on the algorithm to verify that the proposed algorithm can meet the real-time requirements.