3D Indoor Environment Map Building Based On Monocular Vision

In recent years，the3D mapping based on visual system has served as an excellent researchdomain for mobile robots. In order to work in an unknown indoor environment, mobile robots mustdetermine the environmental representation of the surrounding space by carrying sensors. Thispaper proposes a3D map building algorithm based on improved visual odometry method andtree-based network optimization, which contains three modules: visual feature model approach,incremental3D map building depending on the6D visual odometry, loop closure detection methodand map optimization.Firstly, this paper introduces the visual system related technologies before creating3D maps,including the imaging model and the principle of the camera; in visual feature processing module,visual features of each RGB image are extracted stably by SURF; image and depth information ofindoor environments is acquired and fused by Kinect sensor, and then loading point cloud library(PCL) to make3D map displayed.Secondly, for the problem of3D map data association, this paper proposes a method based onimproved visual odometry algorithm. Three-dimensional model of visual information is associatedwith six degrees of freedom of visual odometry to represent the transformation matrix. Using liealgebra coordinates of rigid body motion and maximizing photoconsistency, the registration resultsare evaluated by linearization of the energy equation. This algorithm can improve the success rateof data fusion so that the3D map can be built continuously and effectively in real time.Finally, for the problem of calibration of the color image and depth information, this paper usesthe distortion correction equation to correct the corresponding relationship between the RGB imageand depth map. Aiming at the problem of loop closure detection in3D map, this paper proposes amethod based on tree-based network optimization. A method that combines BoVW modeling andBayesian filtering algorithm is applied to track the historical detection which ensured a success rateand continuity. When a loop closure is detected, tree-based network optimizer is used to optimizethe map with a new constraint so that a global3D map of the closed indoor environment is builtfinally.