A Research Of Mobile Robot SLAM System Based On Uniformly Feature Selection And Bidirectional-projection Matching

Simultaneous Localization and Mapping(SLAM)is a problem of locating sensors on an online constructed map.SLAM technology allows robots have such ability to construct maps and use the map to locate themselves in an unknown environment,independent of the external infrastructure.SLAM technology is a precondition for robot autonomous localization and navigation,as well as intelligent applications.It plays a crucial role in mobile robots,unmanned driving,virtual reality/enhanced reality and other related fields.Visual SLAM(vSLAM)mainly uses visual camera(s)as the only sensor(s)for estimating robot poses and mapping.Cameras are cheap,lightweight,low powered,and rich information.The common vSLAM framework mainly includes two parts: the front-end and the back-end.The main task of the front-end is to build the data association from frame to frame(or frame to map),including feature detection,feature matching and initial pose estimation.The back-end mainly includes loop detection and global optimization,which used for reducing the impact of accumulated error.There are some vSLAM systems improve the localization performance by using the highly robust visual features such as SIFT,SURF for feature extraction and matching,which would take up much computational complexity.However,we use the simple binary descriptors ORB(Oriented FAST and Rotated BRIEF)for feature detection and selection,and make them uniformly distributed over the whole image.Strict bidirectional verification used for feature matches,and is benifit for improving the algorithmic localization performance.This paper proposes a robust visual odometry framework based on uniformly feature selection and strict feature matching.Firstly,the stable feature selection is achieved by setting adaptive feature detection threshold and selecting limited number of features in each local region.Secondly,the precise correspondence matching is achieved by double verification based on the constant-speed motion model.Finally,the translation and rotation of camera are computed separately,and the five-point method is combined with RANSACbased outlier rejection scheme for rotation estimation.The experimental results show our proposed method can achieve the accurate localization with 15 Hz on the public KITTI dataset.