Research On Localization And Semantic Mapping Algorithm Of Mobile Robot In Outdoor Large Scale Scene

Simultaneous localization and mapping is a basis for navigation and environmental perception for robots.In order to better understand the environment,the semantic perception of the scene around the robot is also crucial.This thesis mainly focus on the localization of the robots using stereo vision,visual-joint-RTK localization and dense semantic reconstruction of 3D scenes in outdoor large-scale scenes.Accurate and reliable pose estimation is the premise of mapping of the robot.For the localization problem,a stereo visual odometry based on LK optical flow method is firstly built.The odometry is mainly composed of online tracking of adjacent frames and optimization of back-end sliding window.In order to increase the computational efficiency to make the front-end positioning have higher real-time performance,the online tracking part adopts a semi-direct method.By using the LK optical flow to track keypoints,the calculation speed of the data correlation part is increased.In the optimization part of the back end,the sliding window optimization method is used to optimize the poses of a fixed number of key frames and the observed map points.Since the visual odometry is a relative positioning method,the positioning result is reliable in a short time,but a large cumulative error often occurs under long-time localization.In outdoor scenes,GPS is also a sensor that mobile robots often use to locate.Compared to the results of visual odometry,GPS has the advantage of accurate global positioning.In this paper,a fusion algorithm based on pose-graph is proposed to realize the joint positioning of vision and GPS.This paper proposes a visual-joint-GPS localization algorithm based on pose-graph to fuse visual odometry and high-precision real time kinematic positioning.By merging the GPS positioning information,the problem of trajectory drift caused by the error accumulation of the visual odometry under long-term localization is solved.At present,most SLAM systems for outdoor scenes often focus on online real-time localization.The constructed maps are often sparse maps composed of feature points or semi-dense maps.This is not enough for the complete modeling of the 3D environment..In order to make the robot better model the environment,the mapping algorithm at the back end of this paper has densely reconstructed the 3D scene through the dense restoration of binocular parallax.In the process of reconstruction,in order to save storage space and improve the quality of construction,this paper adopts the construction method based on the truncated distance symbol function(TSDF).In addition,in order to make the robot better perceive the semantic information in the surrounding scene,this paper integrates the semantic segmentation network deeplab-v3+ into the mapping algorithm to complete the semantic mapping of the scene.Finally,in order to verify the effectiveness of the above algorithms,the experiments were carried out on the KITTI datasets and the dataset collected by test car under the actual road scene.From the positioning accuracy and real-time of the visual odometer,the improvement of the positioning accuracy and the reconstruction effect of the map were analyzed after the integration of GPS.