Research On Feature Extraction And Mapping Technology Based On ORB-SLAM2

With the continuous development of electronic industry,sensor technology,communication technology and artificial intelligence,mobile robot technology is gradually popularized in various fields,which significantly liberates productivity and improves people's quality of life.With the continuous development of various sensors,SLAM(Simultaneous Localization and Mapping)technology has gradually become one of the hot topics in the field of robotics.ORB-SLAM2,as one of the best SLAM systems,has the advantages of small amount of calculation and fast running speed,but it also has the problems of light invariance,uniformity and poor human-computer interaction of sparse feature point map.Therefore,this paper optimizes the feature point extraction and map construction of ORB-SLAM2 system.In the aspect of feature points extraction,as an important part of SLAM system,the illumination invariance,scale invariance and uniformity of feature points determine the authenticity of the constructed map to a certain extent.Therefore,this paper takes ORB-SLAM2 algorithm front-end visual odometer image feature point extraction and quadtree uniform distribution algorithm as the research object.Aiming at the fixed threshold of ORB algorithm and the problem of false extraction and false matching in the environment of changing illumination,the adaptive threshold is used to optimize the ORB algorithm;Aiming at the problem that quadtree is prone to over splitting when dealing with feature blocks,the method of optimizing the splitting depth and extraction region of quadtree is adopted to improve the timeliness and uniformity of the algorithm.The experimental results show that the illumination invariance and feature point uniformity of the improved ORB algorithm are significantly improved.In the aspect of map construction,map is an important bridge between SLAM system and upper application and researchers.Due to the sparse feature points constructed by ORB-SLAM2,the map retains fewer feature points.Although it ensures real-time performance,the human-computer interaction of the map is poor.Therefore,this paper takes map construction as the research object,Aiming at the problem of low readability of sparse feature point map constructed by ORB-SLAM2 system,the scheme of constructing dense point cloud map in real time is adopted;In view of the noise in the point cloud due to the sensor error in the process of building the point cloud map and the poor effect of single filtering algorithm in the filtering process,the comprehensive filtering method is used to eliminate the outliers.The experimental results on the TUM data set show that the proposed algorithm can construct dense point cloud map,and the integrated filtering algorithm can improve the timeliness of the algorithm and reduce the memory space of the point cloud map.