Research On Indoor Visual SLAM Algorithm Based On Geometric Features

Simultaneous Localization And Mapping(SLAM)enables mobile robots to perceive the surrounding environment and perform simultaneous localization and mapping.It has been widely used in smart homes,autonomous driving,and augmented reality in recent years.Among them,visual SLAM based on camera sensors has received extensive attention,due to its advantage of low cost,strong ease of use,and rich information.Although SLAM achieves great success recently,it exhibits several major flaws.First,traditional visual SLAM methods usually use point features for camera pose estimation,which can easily lead to missing features in indoor scenes with missing textures,thereby affecting the localization accuracy.Secondly,the method based on point features is prone to cause wrong data association in dynamic scenes,along with less matching information of static objects,it performs poorly in localization and tracking.Finally,for the global relocation with camera,most current visual SLAM research does not consider the preservation and reuse of maps,and the information of geometric features is not fully utilized in the process of optimization.This thesis focuses on the geometric features in visual SLAM,and is committed to further improving the localization accuracy of visual SLAM.The main of this thesis is as follows:(1)We propose a visual SLAM method based on geometric features to improve the accuracy and real-time performance of the SLAM algorithm in a low-texture environment.By introducing the EDline line feature extraction algorithm and the hierarchical clustering plane feature extraction algorithm,the accuracy and operating efficiency of the feature extraction module can be improved.Furthermore,since the motion changes are small during the matching process of adjacent frames,in this thesis we peopose to replace the descriptor with the geometric properties of the line in the process of line feature matching to improve the operation efficiency.At the same time,the centroid and covariance matrix of the plane are used in the process of plane feature matching in order to improve the matching accuracy.In addition,to improve the localization accuracy,the virtual right-eye line feature constraint and the supposed plane constraint are introduced under the cost function based on the point-line-plane reprojection error.The experimental comparison of multiple image sequences shows that the method proposed in this thesis has better localization accuracy and operating efficiency.(2)We propose a dynamic SLAM method based on the line feature optimization to solve the problem of low localization accuracy of traditional feature point-based methods in dynamic scenes.Specifically,this method uses a semantic segmentation model to cull potential dynamic objects.Aiming at the problem of high computational complexity and slow inference time of the semantic segmentation model,this thesis proposes a lightweight semantic segmentation network based on Deeplabv3+ and introduces a multi-view geometric algorithm to further extract the location of unlabeled dynamic objects.From there,given the problem that the remaining static object feature matching information is less,this thesis regards the line feature as one of the features of the static object and introduces the line feature static weight algorithm to eliminate the dynamic line,thereby improving the feature-based method in the localization accuracy and robustness in dynamic scenes.(3)We propose a camera relocation module for fusing plane features.This module embraces plane feature information during the map building process.Further,this thesis completes the function of map saving and loading for a relocation,and uses the reprojection error of feature points and feature planes for optimization,which delivers a more accurate relocation pose.