Research On Indoor Localization And Mapping Algorithm Based On Multi-Feature

At present,visual Simultaneous Localization And Mapping(SLAM)technology can be seen in many indoor applications,such as automatic food delivery vehicles,AR and VR,and warehouse collaborative robots.However,indoor localization is often challenged by factors such as low texture,illumination,and structure repetition,which will make it difficult to perform feature matching normally in the process of camera ego-motion estimation,which will reduce the accuracy of localization and the accuracy of the reconstruction model.In addition,maps reconstructed by SLAM technology are often composed of sparse points,which is not conducive to subsequent high-level tasks,such as path planning and navigation,and human-machine interaction.Taking these problems as the starting point,this paper studies the improvement of the accuracy and robustness of indoor visual SLAM from the perspective of combining multiple features,reduces the algorithm’s dependence on a single feature,and builds dense maps to enrich environmental information.The specific work of this study is as follows:(1)In the indoor environment with artificial landmarks,this paper proposes a multi-feature combination monocular localization algorithm based on markers.The algorithm explores the structural information of markers,uses markers to extract planes in the environment,and combines point,landmark and plane features to estimate camera pose.The algorithm also uses the plane to establish the geometric relationship between the markers,so that geometric constraints can be added to the constraints,thereby effectively reducing the cumulative drift error of features and key frames.Based on the scale and encoding properties of the markers,the scale correction module is introduced to correct the map scale,and the loop detection and correction functions are added to improve the loop detection speed and reduce the closed-loop misjudgment problem.On the SPM dataset,compared with the UcoSLAM system using the marker feature,the root mean square error in the absolute trajectory is reduced by 0.7-3.6cm.On the UcoCeiling dataset,the system can track and map stably in large scenes.(2)In the indoor environment without artificial landmarks,this paper proposes an RGB-D localization and dense mapping algorithm based on the combination of point and plane.In the process of estimating the camera pose step by step with points and planes based on the assumption of Manhattan world,a data association algorithm of point feature assisted plane matching is proposed to improve the accuracy of plane association.And an optimization module is added to the system to reduce feature drift and make the system utilizes better environmental structure information.In the dense mapping part,the map draws plane information in real time,and introduces surfel elements to express the indoor structure.Compared with the sparse map,the information is richer,and the drawing process is real-time and efficient.On the ICL-NUIM synthetic dataset,the RMSE is reduced to the range of0.5-1.7cm.On the TAMU RGB-D real dataset,the tracking trajectories have less drift and the tracking process is more stable than the state-of-the-art framework.