Performance Evaluation Of Feature Extractor Algorithms In Visual SLAM

Mobile robotic SLAM(Simultaneous Localisation and Map Building)is defined to build nn incremental map of the environment by the mobile robot through its pose estimation and perception of the environment,meanwhile realize the robot's localization using the map.With the development of computer technology,the increasing capacity and speed of data processing,as well as the unique advantages of visual sensor,achieving the problem of robotic SLAM by visual sensor if rising.As there are often rotation,scaling,illumination,blur,viewpoint changes in images captured by robot in visual SLAM,it if important to choose an efficient feature extraction algorithm for the visual SLAM.A comprehensive comparative research has been conducted between the classic algorithms SIFT、SURF used in the visual SLAM and the new feature extraction algorithms FAST9、BRIEF、ORB proposed recently.In order to make the performance comparison conveniently,FAST9 detector is combined with the descriptors SURF and BRIEP separately to generate two new algorithms FAST9+SURF and FAST9+BRIEF.Then the performance comparative research has been done among these five algorithms.Standard evaluation images and images captured by robot are chosen to simulate the all kinds of transformations in the visual SLAM.The rate of correct matches、repeat and run-time curve are used to compare the performance of these five algorithms.Experimental results show that ORB algorithm outperforms other four algorithms over the run time and matching under rotation,scaling,illumination,blur changes,but the performance of above five algorithms is decreased in viewpoint transformation.In this paper,a new affine-invariant algorithm AORB is proposed after we combine the idea of simulating view changes in ASIFT and the rapidity of ORB.The experimental results show that the new algorithm not only retains the affine-invariance of ASIFT but also inherits the rapidity of ORB.