3D Vision Based Simultaneous Localization And Mapping For Indoor Environments

Simultaneous Localization and Mapping(SLAM)is the technology that constructing a scene map with the observation data obtained by the robots' carried sensors,and simultaneously estimates the position and motion trajectory of the robots in the constructed map.Vision based SLAM(V-SLAM)refers to the SLAM whose observation sensor is an imaging device,and V-SLAM is the simplest and cheapest SLAM solution.With the development and application of the low-cost 3D vision sensors(RGB-D camera),3D vision based SLAM has become a new research focus in the field of both robotics and 3D computer vision.This thesis focuses on the SLAM technology that fusion the visual features and depth data for the real-world application requirements.The main contributions of this thesis are as follows.(1)The methods to estimate the relative pose of different frames can be divided into two categories: direct method and feature-based method.It's well know that the direct method easily induces local extremum and the feature-based method is low realtime performance,this thesis proposed an RGB-D SLAM that combines sparse direct method and feature point based method.In the SLAM front end,a fast visual odometery is constructed by jointly using the sparse depth and photometric residual direct method,and the feature-based optimization is used in the SLAM backend.The real-time performance of SLAM is improved without affecting its accuracy.(2)For the problem that the performance of pose estimation is worse in the textureless scence,the reason is that the number of extracted feature points and the high gradient pixels in the captured image is low.This thesis proposed a point-plane based SLAM method,which extracts planar features to construct stable constraints for the pose estimation in texture-less scenes,and it constructs new plane landmarks in the SLAM backend.In addition,as the indoor scene's spatial structure satisfies the ”Manhattan world” assumption,a extened point-plane SLAM based on the ”Manhattan” constraint has been proposed,which further improves the pose estimation accuracy.(3)As the mismatches is high in dynamic scenes,this thesis proposed a novel SLAM method based on convolutional neural network and multi-view geometry,which can detect the moving objects and achieve their pixelwise segmentation,and the static feature points and ground plane are used for the pose estimation and map management,and the regions of moving objects in keyframes are repaired to facilitate the reuse of constructed maps.In addition,this thesis proposed a multiple objects' tracking method that combines spatial constraints and KCF algorithm,which can provide the 3D trajectory of the robot in realtime.(4)To reconstructe the map of large-scale indoor scenes rapidly,this thesis proposed a system based on multi-robot cooperative,including three parts: the robot system constructs a multi-feature map based on the three research contents above,the map information transmission between the robot and the server is realized by the UDP protocol,the server system registers the sub-maps to obtain a globally consistent large indoor scene's map.