Design And Implementation Of Stereo Vision SLAM With Fused Point And Line Features

Vision SLAM is a technology that uses cameras as sensors for simultaneous localization and map building.It plays a key role in fields such as unmanned vehicles,autonomous mobile robots and virtual reality.Point feature-based visual SLAM uses point features in images for localization,which has good performance in real-time and localization accuracy and has high practical value.However,in weak texture environments,the insufficient number and uneven distribution of point features lead to the poor reliability of such SLAM algorithms.Considering that weak-texture environments are usually in man-made buildings with richer line features,SLAM using straight lines for localization has gradually become a research hotspot in recent years.In order to improve the stability of SLAM in weak texture environment,this paper adds line segment features to the traditional point feature-based SLAM;in order to improve the speed degradation problem brought by additional line segment features,the traditional line segment feature processing algorithm is improved and a semi-direct line segment tracking algorithm is proposed,and the main research contents are as follows.(1)A new binocular visual odometry based on point features and line segment features is designed.The spatial positions of point features and line segment features are calculated in each frame by binocular alignment of the features;then the camera pose is estimated by minimizing the reprojection error of point features and line segment features based on feature matching between frames.In order to reduce the impact caused by the line segment tracking error,the reprojection error of the line segment is expressed as the distance between the two endpoints of the line segment in the current frame and the corresponding projected line segment.(2)The tracking algorithm of line segment features is improved.Different processing methods are implemented for key frames and non-key frames: the traditional method is used to extract and match line segment features in key frames,while the semi-direct line segment tracking algorithm is implemented for known line segments in non-key frames,and outlier segment rejection and key frame determination are performed.Since the semi-direct line segment tracking algorithm avoids line segment detection and line segment matching,it can improve the overall operation speed of the visual SLAM.(3)The back-end structure of the visual SLAM is designed.In the local map building part,a sparse feature map is used to reduce the memory occupation;data association of map landmarks is performed by co-view to build a local map;BA optimization of the local map is executed after inserting key frames to further reduce the trajectory error.In the loopback detection part,Bo W-based loopback detection algorithm is used;when loopback closure occurs,the loopback error is corrected using graph optimization method.(4)The improved SLAM algorithm in this paper is tested under KITTI binocular dataset and Eu Ro C binocular dataset respectively,and several sets of comparison experiments are conducted under realistic scenes using ZED binocular camera and Vicon motion capture system.The experimental results show that this algorithm not only has better performance in trajectory accuracy and environment adaptability,but also has significantly improved operation speed.