Construction Of Mobile Robot RGB-D Point Cloud Map Based On Graph Optimization

With the continuous development of artificial intelligence and machine vision,in order to allow robots to complete tasks more autonomously,higher requirements are placed on simultaneous positioning and map construction of robots,simultaneous positioning and map construction are hot topics in robot research.In recent years,there has been an upsurge in the construction of color depth maps.However,due to the influence of front-end visual sensors and the external environment,there are still problems that need to be solved for simultaneous positioning and map building front-ends.A good visual front-end is the basis for constructing a globally consistent map.This paper focuses on the research of point cloud map construction algorithms based on color depth images for mobile robots.Some improvements have been made on the basis of traditional simultaneous positioning and map construction framework.The main research contents are as follows:Firstly,the background and significance of the research are summarized,It has been analyzed the actuality of the development and research about mobile robots at home and abroad,the simultaneous positioning of mobile robots and the problem of map construction are analyzed at the meanwhile.the movement model of the mobile robot is modeled including the pinhole camera model and the double camera model,as well as the color depth camera model,additionally analyzes the camera's movement model,which is the visual theory basis of the map construction algorithm.Secondly,analyze the visual front end feature extraction and matching classical algorithms comprehensively,because the slow speed of the classic algorithm in feature extraction and matching can't meet the real-time requirements of the back-end map construction.I come up with the image feature extraction and matching algorithm based on the(SURF)feature detection operator and ORB the feature descriptor.At the same time,it removes the corners of different regions by comparing the Hamming distances of the key points,and then achieves correct matching through the random sampling consistency algorithm.Through experimental simulation,the improved algorithm is proved to improve the correctness of the match and verify the effectiveness of the proposed algorithm.Thirdly,for the construction of RGB-D point cloud map,accurate registration and space saving are needed.An improved point cloud splicing algorithm is proposed.The improved algorithm completes rough matching by building a visual lexicon to improve the initial search time of(ICP)algorithm.Through the combination of PNP and ICP algorithm to establish the least squares model to complete the optimization match,during the point cloud registration stage,the(BNB)global optimization algorithm is combined with the traditional ICP point cloud registration algorithm to solve the problem that the ICP excessively depends on the initial value generation.The problem of local optimization rather than global optimization has achieved good results through experiments,reduced computational time,and achieved accurate matching.Finally,in the process of constructing an RGB-D point cloud map,the visual effects of the map are affected due to the influence of the visual sensor accuracy and the environmental noise,and in order to remove extra noise points and achieve better visual effects,a voxel filter is added.The device performs down sampling and a good point cloud map is obtained through experiments.